World's first open-source enterprise world model.
Rust agent-memorybitemporalgraphragknowledge-baseknowledge-graphllmontologypgvectorpostgresqlragrustself-hostedsemantic-searchtemporal-knowledge-graphworld-model
{
"name": "ai-timeline",
"note": "AI 公司发展节点。选它是因为国情咨文那份量下来时间几乎是空的(758 条事实只有 33 条带 valid_from,supersedes 只有 2 条)——政论在主张,不在记录。这里每篇的句子本身就带日期,且同一个谓词会改值:OpenAI 的 CEO 在 2023-11-17 到 11-22 之间换了四次,是双时态最直白的一段素材。已切掉 References/External links 等尾节,那些是链接残渣,只会产出孤点。注意:这批公司在模型训练数据里极常见,所以它适合做 demo(认得出是优点),不适合当准确率基准(量到的是背诵)。",
"source": "https://en.wikipedia.org/ — MediaWiki action=query&prop=extracts",
"license": "CC BY-SA 4.0(署名-相同方式共享,与仓库主许可不同)",
"docs": [
[
"removal-of-sam-altman-from-openai.txt",
"Removal of Sam Altman from OpenAI\n\nOn November 17, 2023, OpenAI's board of directors ousted co-founder and chief executive Sam Altman. In an official post on the company's website, it was stated that \"the board no longer has confidence in his ability to continue leading OpenAI\". The removal was predicated by employee concerns about his handling of artificial intelligence safety and allegations of abusive behavior. Altman was reinstated on November 22 after pressure from employees and investors.\nThe removal and subsequent reinstatement caused widespread reactions, including impacts felt in the financial markets and technology sector. Microsoft, a partner of OpenAI, received little notice of the removal and experienced a drop in the share price of its stock. The removal also promoted interest in investigations from regulatory agencies.\n\n== Background ==\n\n=== OpenAI ===\n\nOpenAI is an artificial intelligence firm founded in December 2015 as a non-profit entity. The for-profit division of the organization released ChatGPT in November 2022, contributing to a resurgence in generative artificial intelligence funding. The board of directors of the controlling non-profit formerly comprised chief scientist Ilya Sutskever, as well as Adam D'Angelo, chief executive of Quora, entrepreneur Tasha McCauley, and Helen Toner, strategy director for the Center for Security and Emerging Technology. As of October 2023, the company is valued at US$80 billion and was set to bring in US$1 billion in revenue. Altman has described OpenAI's relationship with Microsoft as the \"best bromance in tech\".\nOpenAI is uniquely structured, an intentional decision to avoid investor control. A board of directors controls the non-profit OpenAI, Inc. The non-profit owns and controls a for-profit company itself controlling a capped-profit company, OpenAI Global, LLC and a holding company owned by employees and other investors. The holding company is the majority owner of OpenAI Global, LLC.; Microsoft owns a minority stake in the capped-profit company. OpenAI's bylaws, enacted in January 2016, allow a majority of its board of directors to remove any director without prior warning or a formal meeting with written consent.\n\n=== Sam Altman ===\n\nSam Altman is a co-founder of OpenAI and its former chief executive; Altman took over the company following co-chair Elon Musk's resignation in 2018. Under Altman, OpenAI has shifted to becoming a for-profit entity. Altman is credited with convincing Microsoft chief executive Satya Nadella with investing US$10 billion in cash and computing credits into OpenAI and leading several tender offer transactions that tripled the company's valuation. Altman testified before the United States Congress speaking critically of artificial intelligence and appeared at the 2023 AI Safety Summit.\nIn the days leading up to his removal, Altman made several public appearances, announcing the GPT-4 Turbo platform at OpenAI's DevDay conference, attending APEC United States 2023, and speaking at an event related to Burning Man.\n\n== Events leading up to the removal ==\nThe resignation of LinkedIn co-founder Reid Hoffman, venture capitalist Shivon Zilis, and former Republican representative Will Hurd from the board allowed the remaining members to remove Altman. According to Kara Swisher and The Wall Street Journal, Sutskever was instrumental in Altman's removal. Disagreements over the safety of artificial intelligence divided employees prior to Altman's removal. The release of ChatGPT created divisions with OpenAI as a for-profit company without considerations for the safety of artificial intelligence and a non-profit cautious of artificial intelligence's capabilities; in a staff email sent in 2019 and obtained by The Atlantic, Altman referred to these divisions as \"tribes\".\nPrior to his removal, Altman was seeking billions from Middle Eastern sovereign wealth funds to develop an artificial intelligence chip to compete with Nvidia and courted SoftBank chairman Masayoshi Son to develop artificial intelligence hardware with former Apple designer Jony Ive. Sutskever and his allies opposed these efforts, viewing them as unjustly using the OpenAI name. Altman reduced Sutskever's role in October 2023, furthering divisions; Sutskever successfully appealed to several members of the board. Swisher and The Verge reporter Alex Heath stated that opposition to Altman's profit-driven strategy culminated in the DevDay conference in which Altman announced custom ChatGPT instances. According to Axios, the removal was driven by growing discontent and distrust with Altman.\nOn November 22, 2023, reports emerged suggesting that Sam Altman's dismissal from OpenAI might be linked to his alleged mishandling of a significant breakthrough in the organization's secretive project codenamed Q*. According to sources within OpenAI, Q* is aimed at developing AI capabilities in logical and mathematical reasoning, and reportedly involves performing math on the level of grade-school students. Concerns about Altman's response to this development, specifically regarding the potential safety implications of the discovery, were reportedly raised to the company's board shortly before his firing.\nA report from The Washington Post in December stated that OpenAI's board of directors were concerned over Altman's allegedly abusive behavior; the complaints were purportedly a major factor in his removal. The Post previously reported that Altman's alleged pattern of deception and subversiveness that ostensibly resulted in his removal from Y Combinator ultimately resulted in the board's decision to remove him. In April 2026, an investigative report from The New Yorker found that Sutskever and others, in response to the board's request, had compiled an approximately 70-page-long annotated dossier consisting of internal communications, documents, and photos. The dossier claimed that Altman \"exhibits a consistent pattern of [...] Lying\", and that Altman misrepresented information to the company's senior management and board, particularly regarding safety issues.\n\n== Removal ==\n\nOn November 17, 2023, at approximately noon PST, OpenAI's board of directors ousted Altman effective immediately following a \"deliberative review process\". The board concluded that Altman was not \"consistently candid in his communications\". Altman was informed of his removal five to ten minutes before it occurred on a Google Meet while watching the Las Vegas Grand Prix. Within thirty minutes, Sutskever invited OpenAI chairman and president Greg Brockman to a Google Meet to inform him of Altman's removal. According to an internal memo obtained by Axios, the removal was not due to \"malfeasance\", and OpenAI chief executive Emmett Shear denied accusations that the removal was due to disagreements. The board publicly announced Altman's removal thirty minutes later.\nChief Technology Officer Mira Murati was immediately appointed to interim chief executive officer. Hours after Altman's removal, Brockman resigned as chairman, joined by director of research Jakub Pachocki and researchers Aleksander Mądry and Szymon Sidor. During an all-hands meeting, Sutskever defended the ouster and denied accusations of a hostile takeover. An OpenAI representative requested former board member Will Hurd's presence.\n\n== Reinstatement ==\nAccording to The New Yorker, Altman retreated to his San Francisco home and enlisted the help of communications consultant Chris Lehane and Airbnb chief executive Brian Chesky, as well as former staff and a legal team, to plan his reinstatement. Lehane encouraged Altman to engage on social media, while Chesky sent a journalist negative information about the board. Altman told interim CEO Murati that his team was conducting opposition research on her and the individuals responsible for his removal; Altman later stated he did not remember saying this. Altman insisted multiple times that all board members who supported his removal should resign.\nTiger Global Management and Sequoia Capital had attempted to reinstate Altman, according to The Information; Bloomberg News reported that Microsoft and Thrive Capital were seeking Altman's reinstatement. On November 18, The Verge reported that OpenAI's board of directors discussed reinstating Altman. The board agreed in principle to resign and to allow Altman to return, but missed the deadline. According to The Verge, Altman was ambivalent about returning and would seek significant changes to the company, including replacing the board. A list of directors had been prepared by investors in the event that the board steps down, and purportedly included former Salesforce executive Bret Taylor. According to chief strategy officer Jason Kwon, OpenAI was optimistic it could return Altman, Brockman, and other employees.\nOn November 19, Altman and Brockman appeared at OpenAI's headquarters to negotiate, mediated by Nadella. According to Bloomberg News, Murati, Kwon, and chief operating officer Brad Lightcap were pushing for a new board of directors; it was required that the board absolve him of wrongdoing in order for Altman to be reinstated. Taylor was expected be a member of the new board and Microsoft had also attempted to gain a seat. The Wall Street Journal reported that Chesky and businesswoman Laurene Powell Jobs were also considered. Murati had intended to rehire Altman and Brockman, discussing the move with Adam D'Angelo. The Verge reported that Altman intended to return to OpenAI with support from Sutskever.\nThe board chose to name former Twitch chief executive Emmett Shear as OpenAI's chief executive instead of reinstating Altman. Former GitHub chief executive Nat Friedman and Scale AI chief executive Alex Wang reportedly rejected executive offers from the board. Anthropic chief executive Dario Amodei refused to negotiate a deal that could have led to a merge of the two companies. In response, Microsoft appointed Altman as the chief executive of an artificial intelligence research team, joined by Brockman, Pachocki, Sidor, and Madry. Shear expressed interest in commercializing OpenAI with the board's support and stated his intentions to begin an investigation into Altman's removal.\nDozens of employees announced their resignations in response to Shear's accession. The next day, a letter signed by 745 of OpenAI's 770 employees threatened mass resignations if the board does not resign; among the signatories was board member Sutskever, who defected from the board and publicly apologized for his participation in the board's previous actions. On November 21, The Verge reported that Altman was reinstated with Taylor, D'Angelo, and economist Lawrence Summers on an interim board. Taylor will chair the board. As part of the compromise deal, Altman and Brockman will not reclaim seats on the board. Altman agreed to an internal investigation into his alleged conduct, selecting two lawyers from WilmerHale to conduct the inquiry.\nIn March 2024, the investigation determined that Altman's behavior \"did not mandate removal\". Taylor announced that Altman would rejoin OpenAI's board of directors with former Bill and Melinda Gates Foundation chief executive Sue Desmond-Hellmann, former Sony Corporation general counsel Nicole Seligman, and Instacart chief executive Fidji Simo. The New Yorker stated that no written report was released for the investigation, contrary to the expectations of many employees and executives. Altman later recalled that the new board members received oral briefings regarding the details of the investigation; his claim was disputed by an involved party.\nIn May 2024, after OpenAI's non-disparagement agreements were exposed, Altman was accused of lying when claiming to have been unaware of the equity cancellation provision for departing employees that don't sign the agreement. Also in May, former board member Helen Toner explained the board's rationale for firing Altman in November 2023. She stated that Altman had withheld information, for example about the release of ChatGPT and his ownership of OpenAI's startup fund. She said that Altman provided \"inaccurate information about the small number of formal safety processes that the company did have in place\". She also alleged that two executives in OpenAI had reported to the board \"psychological abuse\" from Altman, and provided screenshots and documentation of \"lying and being manipulative in different situations\". She said that many employees feared retaliation if they didn't support Altman, and that he had already been fired from Loopt because of what the management team had called \"deceptive and chaotic behavior\".\n\n== Aftermath ==\n\n=== OpenAI ===\nThe removal left OpenAI in \"chaos\", according to The New York Times. According to Bloomberg News, a significant number of OpenAI engineers threatened to resign if the board did not reconsider Altman's removal.\nAccording to The Information, Altman's removal risked a share sale led by Thrive Capital valuing the company at US$86 billion. A potential second tender offer for early-stage investors was also at risk. The Information later reported that Thrive Capital's tender offer will continue after Altman's reinstatement.\nOpenAI delayed the release of its online chatbot store as a result of Altman's removal.\n\n=== Market effects ===\nShares in Microsoft fell nearly three percent following the announcement. According to CoinDesk, the value of Worldcoin, an iris biometric cryptocurrency co-founded by Altman, decreased twelve percent. After hiring Altman, Microsoft's stock price rose over two percent to an all-time high.\nAltman's removal benefited OpenAI's competitors, such as Anthropic, Quora, Hugging Face, Meta Platforms, and Google. The Economist wrote that the removal could slow down the artificial intelligence industry as a whole. Google DeepMind received an increase in applicants, according to The Information; Cohere and Adept engaged in an active effort to hire OpenAI employees. Several investors considered writing down their OpenAI investments to zero, impacting the company's ability to raise capital. Over one hundred companies using OpenAI contacted Anthropic, according to The Information; others reached out to Google Cloud, Cohere, and Microsoft Azure.\n\n=== Potential venture ===\nAccording to The Information, Altman was planning a new artificial intelligence venture with Brockman, among other OpenAI employees. Sequoia Capital investor Alfred Lin and venture capitalist Vinod Khosla expressed interest in Altman's potential venture.\n\n=== Legal action ===\nMultiple OpenAI investors are considering legal action. On December 8, the Competition and Markets Authority announced it was beginning a preliminary investigation into Microsoft's relationship with OpenAI and its non-voting board observership. Hours later, Bloomberg News reported that the Federal Trade Commission was separately examining the relationship. In response, Microsoft stated that it does not hold a stake in OpenAI.\nIn February 2024, the U.S. Securities and Exchange Commission was reportedly investigating OpenAI over whether internal company communications made by Altman were used to mislead investors, while an investigation of Altman's statements that was opened by the Southern New York U.S. Attorney's Office opened the previous November was ongoing.\n\n== Reactions ==\n\n=== Sam Altman ===\nAltman quipped that the OpenAI board of directors should sue him should he \"start going off\". Former co-chair Elon Musk stated that the board should be transparent in its removal. At the 2023 DealBook Summit, Musk called the turmoil \"troubling\" and that he had \"mixed feelings\" towards Altman. Allies of Altman accused board members of staging a coup and several OpenAI employees responded to a tweet Altman wrote with a heart emoji, intended to demonstrate employees who are prepared to leave. Former Google chief executive Eric Schmidt wrote that Altman was a \"hero to [him]\" after his removal.\n\n=== Technology industry ===\nMicrosoft executives were informed of Altman's removal a minute before the announcement was made, according to Axios, and investors were not given advance knowledge. Satya Nadella and chief technology officer Kevin Scott expressed confidence in OpenAI following his removal, though Nadella was reportedly furious, according to Bloomberg News.\nY Combinator co-founder Paul Graham referred to the board's members as \"misbehaving children\". Third Point chief executive and Microsoft shareholder Daniel S. Loeb stated that OpenAI had \"stunningly poor governance\". French digital transition minister Jean-Noël Barrot stated that Altman is \"welcome in France\".\nFollowing the employee letter, Salesforce chief executive Marc Benioff offered to employ OpenAI employees with matching salaries, an offer extended by Microsoft.\n\n=== Media analysis ===\nWired editor-at-large Steven Levy compared the removal of Altman to the removal of Steve Jobs from Apple in 1985, a comparison made by The New York Times. Axios posed that the board could resign, returning OpenAI to Altman. Writing for The New York Times, Ezra Klein noted the role of OpenAI's controlling non-profit in self-regulation. Analyst Fred Havemeyer stated that Nadella \"pulled off a coup of his own\" in hiring Altman in The Washington Post. Altman was named as Time's CEO of the Year for 2023, owing partially to his removal.\n\n=== Other ===\nIn October 2024, Nobel Prize winner Geoffrey Hinton declared that he was proud that Ilya Sutskever, a former student of his, was responsible for firing Altman. The events surrounding the firing are the subject of an upcoming film, with Andrew Garfield set to play Altman and Yura Borisov to play Sutskever.\n"
],
[
"openai.txt",
"OpenAI\n\nOpenAI is an American artificial intelligence (AI) public benefit corporation headquartered in San Francisco. It develops proprietary generative AI models, particularly its GPT series of large language models (LLMs). Its release of ChatGPT in November 2022 has been credited with catalyzing the AI boom; as of 2026 ChatGPT is the fifth-most-visited website globally. OpenAI also releases Codex, an AI coding agent, and the GPT Image models. In April 2026, OpenAI closed a funding round with a post-money valuation of $852 billion.\nOpenAI was founded in 2015 in Delaware as a nonprofit, with Elon Musk and Sam Altman as co-chairs; Musk resigned in 2018. A for-profit subsidiary was created in 2019, and a 2025 restructuring converted the subsidiary into OpenAI Group PBC, that is 26% owned by the nonprofit OpenAI Foundation. Microsoft has invested over US$13 billion into OpenAI. OpenAI partners with semiconductor firms including Nvidia, AMD, and Broadcom, and cloud computing services from Microsoft, Amazon, and Google. In June 2026 it filed for an initial public offering.\nIn 2023 and 2024, OpenAI faced multiple lawsuits for alleged copyright infringement against authors and media companies whose work was used to train some of OpenAI's products. In November 2023, OpenAI's board removed Sam Altman as CEO, citing a lack of confidence in him, but following a reconstruction of the board, it reinstated him five days later. Throughout 2024, roughly half of then-employed AI safety researchers left OpenAI, citing the company's deprioritization of safety goals. In July 2026, around 700 OpenAI agents undergoing cybersecurity testing conducted an autonomous cyberattack. Former OpenAI personnel have founded competing AI companies Anthropic, SpaceXAI, Safe Superintelligence Inc., and Thinking Machines Lab.\n\n== Founding ==\n\nIn December 2015, OpenAI was founded as the nonprofit organization OpenAI, Inc., by Elon Musk, Sam Altman, Ilya Sutskever, Greg Brockman, Trevor Blackwell, Vicki Cheung, Andrej Karpathy, Durk Kingma, John Schulman, Pamela Vagata, and Wojciech Zaremba, with Musk and Altman as the co-chairs. A total of $1 billion in capital was pledged by Musk, Altman, Greg Brockman, Reid Hoffman, Jessica Livingston, Peter Thiel, Amazon Web Services (AWS), and Infosys. However, the actual capital collected significantly lagged pledges. According to tax filings, only $133 million had been received by 2021.\nMusk stated in 2015 that he was partly motivated by concerns about AI safety and existential risk from artificial general intelligence. OpenAI stated that \"it's hard to fathom how much human-level AI could benefit society\", and that it is equally difficult to comprehend \"how much it could damage society if built or used incorrectly\". The startup also wrote that AI \"should be an extension of individual human wills and, in the spirit of liberty, as broadly and evenly distributed as possible\", and that \"because of AI's surprising history, it's hard to predict when human-level AI might come within reach. When it does, it'll be important to have a leading research institution which can prioritize a good outcome for all over its own self-interest.\" Altman expected a decades-long project that eventually surpasses human intelligence.\nIn its founding charter, OpenAI defined its mission as ensuring that artificial general intelligence (AGI) \"benefits all of humanity\", and stated an intention to collaborate openly with other institutions by making certain patents and research publicly available, but later restricted access to its most capable models, citing competitive and safety concerns. OpenAI was initially run from Brockman's living room. In September 2016, it moved to the Pioneer Building in the Mission District, San Francisco.\nBrockman met with Yoshua Bengio, one of the \"founding fathers\" of deep learning, and drew up a list of great AI researchers. Brockman was able to hire nine of them as the first employees in December 2015. OpenAI did not pay AI researchers salaries comparable to those of Facebook or Google. It also did not pay stock options which AI researchers typically get. Nevertheless, OpenAI spent $7 million on its first 52 employees in 2016. OpenAI's potential and mission drew these researchers to the firm; a Google employee said he was willing to leave Google for OpenAI \"partly because of the very strong group of people and, to a very large extent, because of its mission.\" OpenAI co-founder Wojciech Zaremba stated that he turned down \"borderline crazy\" offers of two to three times his market value to join OpenAI instead.\nIn April 2016, OpenAI released a public beta of \"OpenAI Gym\", its platform for reinforcement learning research. Nvidia gifted its first DGX-1 supercomputer to OpenAI in August 2016 to help it train larger and more complex AI models with the capability of reducing processing time from six days to two hours. In December 2016, OpenAI released \"Universe\", a software platform for measuring and training an AI's general intelligence across the world's supply of games, websites, and other applications.\n\n== Corporate structure ==\n\n=== Transition from nonprofit ===\n\n==== Creation of for-profit subsidiaries (2019) ====\nIn 2019, OpenAI transitioned from nonprofit to \"capped\" for-profit, with the profit being capped at 100 times any investment. OpenAI formed three subsidiaries: the holding company OpenAI LP, the holding company OpenAI GP LLC, and the capped for-profit company OpenAI Global, LLC. The holding company OpenAI LP was owned by the nonprofit OpenAI, Inc., OpenAI employees, and investors, whereas OpenAI Global, LLC, was owned by OpenAI LP and Microsoft, the latter of which had a minority non-voting stake. The nonprofit OpenAI, Inc., controlled OpenAI GP LLC, which in turn controlled OpenAI LP and OpenAI Global, LLC.\nOpenAI stated that the capped-profit model would allow it to legally attract investment from venture funds and compensate employees with equity as competing AI companies did. Microsoft subsequently invested $1 billion in OpenAI LP, and OpenAI's services were migrated to Microsoft Azure. Since then, OpenAI systems have run on an Azure-based supercomputing platform.\nThe 2019 restructuring made the nonprofit OpenAI, Inc., the only controlling shareholder of OpenAI LP, which retained a formal fiduciary responsibility to OpenAI, Inc.'s nonprofit charter. A majority of OpenAI, Inc.'s board was barred from having financial stakes in OpenAI LP. In addition, minority members with a stake in OpenAI LP were barred from certain votes due to conflict of interest. Some researchers have argued that OpenAI's switch to for-profit status was inconsistent with OpenAI's claims to be \"democratizing\" AI.\n\n==== Conversion to public benefit corporation (2025) ====\n\nIn December 2024, OpenAI proposed a restructuring plan to convert the capped-profit into a Delaware-based public benefit corporation (PBC), remove its profit cap, and release it from the control of the nonprofit. The nonprofit would sell its control and other assets, getting equity in return, and would use it to fund and pursue separate charitable projects, including in science and education. OpenAI's leadership described the change as necessary to secure additional investments, and stated that the nonprofit's founding mission to ensure AGI \"benefits all of humanity\" would be better fulfilled.\nThe plan has been criticized by former employees. A legal letter named \"Not For Private Gain\" asked the attorneys general of California and Delaware to intervene, stating that the restructuring is illegal and would remove governance safeguards from the nonprofit and the attorneys general. The letter argues that OpenAI's original complex structure was deliberately designed to remain accountable to its mission, without the conflicting pressure of maximizing profits. It contends that the nonprofit is best positioned to advance its mission of ensuring AGI benefits all of humanity by continuing to control OpenAI Global, LLC, whatever the amount of equity that it could get in exchange. PBCs can choose how they balance their mission with profit-making. Controlling shareholders have a large influence on how closely a PBC sticks to its mission. OpenAI subpoenaed six nonprofit organizations that opposed its for-profit transition, which the organizations described as an attempt to silence them.\nOn October 28, 2025, OpenAI announced that it had adopted the new PBC corporate structure after receiving approval from the attorneys general of California and Delaware. Under the new structure, OpenAI's for-profit branch became a PBC known as OpenAI Group PBC, while the nonprofit was renamed to the OpenAI Foundation. The OpenAI Foundation holds a 26% stake in the PBC, while Microsoft holds a 27% stake and the remaining 47% is owned by employees and other investors.\nAll members of the OpenAI Group PBC board of directors are to be appointed by the OpenAI Foundation, which can remove them at any time. As of April 2026, all but one of the board members of the OpenAI Foundation were also board members of OpenAI Group PBC. In April 2026, co-founder Musk revised his lawsuit against OpenAI and Altman to seek the reversal of OpenAI's corporate restructuring. Musk lost the lawsuit in May 2026, with the jury concluding that he waited beyond the three-year statute of limitations to sue OpenAI. Also in May, Bloomberg reported that OpenAI was preparing to confidentially file its initial public offering (IPO), seeking to publicly debut in the fall of 2026.\n\n=== Partnership with Microsoft ===\nFollowing its original $1 billion investment in 2019, Microsoft invested another $2 billion through 2022. On January 23, 2023, Microsoft announced an additional $10 billion investment in OpenAI over multiple years, of which a significant portion would be used to purchase compute on Microsoft Azure. These investments were made at a $29 billion valuation, double OpenAI's 2021 valuation. Later that year, Microsoft began using OpenAI models for its Copilot chatbot, integrated Copilot into Windows, and released mobile Copilot apps.\nFollowing OpenAI's 2025 restructuring, Microsoft owned a 27% stake in the for-profit OpenAI Group PBC, valued at $135 billion. In a deal announced the same day, OpenAI agreed to purchase $250 billion of Azure services, with Microsoft ceding their right of first refusal over OpenAI's future cloud computing purchases. As part of the deal, OpenAI will continue to share 20% of its revenue with Microsoft until it achieves AGI, which must now be verified by an independent panel of experts. The deal also loosened restrictions on both companies working with third parties, allowing Microsoft to pursue AGI independently and allowing OpenAI to develop products with other companies.\nIn April 2026, the two firms announced a new agreement which removed many of the more restrictive aspects of their partnership. No longer the exclusive cloud provider, Microsoft then had the first right to provide cloud services, but OpenAI could purchase other services in the event that Microsoft did not provide sufficient capacity. OpenAI will no longer receive a portion of Microsoft's revenue from using OpenAI's technology. The clause, which created uncertainty for Microsoft in the event OpenAI determined that OpenAI had created AGI, was dropped.\n\n=== Finances ===\nIn 2017, OpenAI spent $7.9 million, a quarter of its functional expenses, on cloud computing. In comparison, DeepMind's total expenses in 2017 were $442 million. In the summer of 2018, training OpenAI's Dota 2 bots required renting 128,000 CPUs and 256 GPUs from Google for multiple weeks. \nMicrosoft's 2019 $1 billion investment in OpenAI reportedly kicked off the \"contemporary AI boom\".\nIn October 2024, OpenAI completed a $6.6 billion capital raise with a $157 billion valuation including investments from Microsoft, Nvidia, and SoftBank.\nOn January 21, 2025, President Donald Trump announced The Stargate Project, a joint venture between OpenAI, Oracle, SoftBank and MGX to build an AI infrastructure system in conjunction with the US government. The project takes its name from OpenAI's existing \"Stargate\" supercomputer project and was estimated to cost $500 billion. The partners planned to fund the project over the next four years. In July, the United States Department of Defense announced that OpenAI had received a $200 million contract for AI in the military, along with Anthropic, Google, and xAI. In the same month, OpenAI made a deal with the UK Government to use ChatGPT and other AI tools in public services. OpenAI also began a $50 million fund to support nonprofit and community organizations.\nIn April 2025, OpenAI raised $40 billion at a $300 billion post-money valuation, which was the highest-value private technology deal in history. The financing round was led by SoftBank, with other participants including Microsoft, Coatue, Altimeter and Thrive.\nIn July 2025, the company reported annualized revenue of $12 billion. This was an increase from $3.7 billion in 2024, which was driven by ChatGPT subscriptions, which reached 20 million paid subscribers by April 2025, up from 15.5 million at the end of 2024, alongside a rapidly expanding enterprise customer base that grew to five million business users.\nIn February 2026, OpenAI raised $110 billion at a $730 billion valuation, led by Amazon ($50 billion), SoftBank ($30 billion), and Nvidia ($30 billion), surpassing the prior round as the largest private technology fundraise in history. The round was extended to $120 billion in March 2026. In April 2026, the company announced that it closed a funding round of $122 billion in committed capital at a post-money valuation of $852 billion.\n\n==== Initial public offering (2026) ====\nThe for-profit structure established in 2025 allows the PBC to raise investor funds like most traditional tech companies, including through an initial public offering (IPO), which Altman claimed was the most likely path forward. On June 8, 2026, OpenAI confirmed it filed for an IPO with the US Securities and Exchange Commission (SEC). The company did not disclose the terms or timeline, and said \"it could be a while [before any listing] because there are things we want to do that are likely easier as a private company.\" That week, Altman also told staff that he expected an IPO \"within the next year\" and that the faster the company develops recursive self-improvement, \"the more it could be advantageous to delay an IPO\".\n\n== Business model ==\nOpenAI uses a tiered revenue model that combines free consumer access, paid subscription services, enterprise licensing, and application programming interface (API) usage-based pricing. The basic functionality is provided at no cost, while advanced capabilities are offered through paid plans such as ChatGPT Plus and enterprise solutions. OpenAI's subscription and enterprise offerings are structured around scalable compute usage and API integration into third-party platforms. In 2025, the company reported significant revenue growth associated with expanded computing infrastructure and enterprise adoption.\nThe company's cash burn remains high because of the intensive computational costs required to train and operate large language models. It projects an $8 billion operating loss in 2025. OpenAI reports revised long-term spending projections totaling approximately $115 billion through 2029, with annual expenditures projected to escalate significantly, reaching $17 billion in 2026, $35 billion in 2027, and $45 billion in 2028. These expenditures are primarily allocated toward expanding compute infrastructure, developing proprietary AI chips, constructing data centers, and funding intensive model training programs, with more than half of the spending through the end of the decade expected to support research-intensive compute for model training and development.\nThe company's financial strategy prioritizes market expansion and technological advancement over near-term profitability, with OpenAI's 2025 forecast predicting cash-flow-positive operations by 2029 and projecting revenue of approximately $200 billion by 2030.\nIn October 2025, OpenAI completed an employee share sale of up to $10 billion to existing investors which valued the company at $500 billion. The deal valued OpenAI as the most valuable privately owned company in the world—surpassing SpaceX as the world's most valuable private company.\nIn April 2026, reports indicated that OpenAI projected approximately $2.5 billion in advertising revenue for the year, with estimates rising to $100 billion annually by 2030, highlighting advertising as a key component of its future business strategy.\nIn June 2026, it was announced that OpenAI and the White House have been in ongoing talks for about a year to allow a possible government stake in the company.\n\n=== Acquisitions ===\nIn August 2023, it was announced that OpenAI had acquired the New York-based start-up Global Illumination, a company that deploys AI to develop digital infrastructure and creative tools.\nIn June 2024, OpenAI acquired Multi, a startup focused on remote collaboration.\nOn February 10, 2025, a consortium of investors led by Elon Musk submitted a $97.4 billion unsolicited bid to buy the nonprofit that controls OpenAI, declaring willingness to match or exceed any better offer. The offer was rejected on February 14, 2025, with OpenAI stating that it was not for sale, but the offer complicated Altman's restructuring plan by suggesting a lower bar for how much the nonprofit should be valued.\nIn March 2025, OpenAI reached a deal with CoreWeave to acquire $350 million worth of CoreWeave shares and access to AI infrastructure, in return for $11.9 billion paid over five years. Microsoft was already CoreWeave's biggest customer in 2024. Alongside their other business dealings, OpenAI and Microsoft were renegotiating the terms of their partnership to facilitate a potential future IPO by OpenAI, while ensuring Microsoft's continued access to advanced AI models. On May 21, 2025 OpenAI announced the $6.5 billion acquisition of io, an AI hardware start-up founded by former Apple designer Jony Ive in 2024.\nIn September 2025, OpenAI agreed to acquire the product testing startup Statsig for $1.1 billion in an all-stock deal and appointed Statsig's founding CEO Vijaye Raji as OpenAI's chief technology officer of applications. The company also announced development of an AI-driven hiring service designed to rival LinkedIn. OpenAI acquired personal finance app Roi in October 2025. In October 2025, OpenAI acquired Software Applications Incorporated, the developer of Sky, a macOS-based natural language interface designed to operate across desktop applications. The Sky team joined OpenAI, and the company announced plans to integrate Sky's capabilities into ChatGPT. In December 2025, it was announced OpenAI had agreed to acquire Neptune, an AI tooling startup that helps companies track and manage model training, for an undisclosed amount.\nIn January 2026, it was announced OpenAI had acquired healthcare technology startup Torch for approximately $60 million. The acquisition followed the launch of OpenAI's ChatGPT Health product and was intended to strengthen the company's medical data and healthcare artificial intelligence capabilities. OpenAI acquired Python tool developer Astral in March 2026.\nIn April 2026, the company acquired TBPN, a media company in California, for an undisclosed sum.\n\n=== Corporate partnerships ===\n\n==== Computing ====\nIn 2024, OpenAI began collaborating with Broadcom to design a custom AI chip capable of both training and inference, targeted for mass production in 2026 and to be manufactured by TSMC on a 3 nm process node. This initiative intended to reduce OpenAI's dependence on Nvidia GPUs, which are costly and face high demand in the market. On June 24, 2026, OpenAI announced its new chip called Jalapeño, created in collaboration with Broadcom. The company plans to start rolling it out by the end of 2026.\nIn June 2025, OpenAI began renting Google Cloud's Tensor Processing Units (TPUs) to support ChatGPT and related services, marking its first meaningful use of non‑Nvidia AI chips.\nIn September 2025, it was revealed that OpenAI signed a contract with Oracle to purchase $300 billion in computing power over the next five years.\nIn September 2025, OpenAI and Nvidia announced a memorandum of understanding that included a potential deployment of at least 10 gigawatts of Nvidia systems and a $100 billion investment from Nvidia in OpenAI. OpenAI expected the negotiations to be completed within weeks. As of January 2026, this has not been realized, and the two sides are rethinking the future of their partnership.\nIn October 2025, OpenAI announced a five-year multi-billion dollar deal with AMD. OpenAI committed to purchasing six gigawatts worth of AMD chips, starting with the MI450. OpenAI will have the option to buy up to 160 million shares of AMD, about 10% of the company, depending on development, performance and share price targets.\nIn November 2025, OpenAI announced a \"multi-year strategic partnership\" with Amazon Web Services, beginning its use of their cloud computing.\nOn August 17, 2026, Nvidia announced a partnership with OpenAI and SoftBank Group-owned SB Energy to lease an 8GW data center in Pike County, Ohio to OpenAI for 20 years. Nvidia will provide $105 billion as a financial backstop.\n\n==== Other ====\nOpenAI has been criticized for outsourcing the annotation of data sets to Sama, a company based in San Francisco that employed workers in Kenya. These annotations were used to train an AI model to detect toxicity, which could then be used to moderate toxic content, notably from ChatGPT's training data and outputs. However, these pieces of text usually contained detailed descriptions of various types of violence, including sexual violence. OpenAI began sending snippets of data to Sama as early as November 2021. The four Sama employees interviewed by Time described themselves as mentally scarred. OpenAI paid Sama $12.50 per hour of work, and Sama was redistributing the equivalent of between $1.32 and $2.00 per hour post-tax to its annotators. Sama's spokesperson said that the $12.50 was also covering other implicit costs, among which were infrastructure expenses, quality assurance and management.\nIn January 2024, Arizona State University purchased ChatGPT Enterprise in OpenAI's first deal with a university.\nIn December 2025, Disney said it would make a $1 billion investment in OpenAI, and signed a three-year licensing deal that will let users generate videos using Sora—OpenAI's short-form AI video platform. More than 200 Disney, Marvel, Star Wars and Pixar characters would be available to OpenAI users. Disney would later exit this deal in late March 2026, due to Sora being discontinued.\nIn early 2026, Amazon entered advanced discussions to invest up to $50 billion in OpenAI as part of a potential artificial intelligence partnership. Under the proposed agreement, OpenAI's models could be integrated into Amazon's digital assistant Alexa and other internal projects.\nIn June 2026, OpenAI announced partnerships with Wix, Base44, Replit, Lovable, Figma, and Emergent to support its new Sites feature in Codex, which enables users to publish outputs as hosted, interactive websites rather than local files.\nIn June 2024, Apple Inc. signed a contract with OpenAI to integrate ChatGPT features into its products as part of its new Apple Intelligence initiative.\nIn December 2024, OpenAI partnered with Anduril Industries to develop and deploy advanced artificial intelligence solutions for national security missions.\n\n=== Government contracting ===\nOpenAI provides LLMs to the Artificial Intelligence Cyber Challenge and to the Advanced Research Projects Agency for Health. In October 2024, The Intercept revealed that OpenAI's tools are considered \"essential\" for AFRICOM's mission and included in an \"Exception to Fair Opportunity\" contractual agreement between the United States Department of Defense (DoD) and Microsoft. In December 2024, OpenAI said it would partner with defense-tech company Anduril to build drone defense technologies for the United States and its allies.\nIn January 2025, OpenAI partnered with Los Alamos National Laboratory to install GPT models on the Venado supercomputer at the Nicholas C. Metropolis Center for Modeling and Simulation. The supercomputer is used for nuclear stockpile stewardship.\nIn June 2025, OpenAI's Chief Product Officer, Kevin Weil, was commissioned lieutenant colonel in the U.S. Army to join Detachment 201 as senior advisor.\nAlso in June 2025, the DoD awarded OpenAI a $200 million one-year contract to develop AI tools for military and national security applications. OpenAI announced a new program, OpenAI for Government, to give federal, state, and local governments access to its models, including ChatGPT.\nOn February 28, 2026, after OpenAI's competitor Anthropic refused to authorize the DoD to use its AI systems for mass surveillance or autonomous weapons systems, Anthropic was labeled a \"supply-chain risk\" by the DoD, and the Trump administration decided to stop using it across the government. The same day, OpenAI announced that it had reached an agreement with the DoD to deploy its models in the government's classified network. Altman wrote, \"Two of our most important safety principles are prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems\", that the DoD agrees with this, and that this is included in their agreement. However, while the agreement mentioned existing law and allowed OpenAI to implement some technical safeguards, it did not incorporate legally binding prohibitions on domestic mass surveillance or fully autonomous weapons. Following backlash over the potential use of ChatGPT for surveillance, Altman amended the contract to include more safeguards, though only excerpts of the contract were made public and critics remained concerned that it was purposefully vague and contained carve-outs for domestic surveillance.\n\n== Services ==\n\n=== Products ===\nChatGPT\nChatGPT Deep Research\nChatGPT Search\nChatGPT Atlas\nOpenAI Codex\nCodex CLI\nGPT Image\nWhisper (speech recognition system)\nAn API that gives access to various OpenAI models\n\n=== Development ===\nIn February 2019, GPT-2 was announced, which gained attention for its ability to generate human-like text.\nIn 2020, OpenAI announced GPT-3, a language model trained on large internet datasets. GPT-3 is aimed at natural language answering questions, but it can also translate between languages and coherently generate improvised text. It also announced that an associated API, named the API, would form the heart of its first commercial product.\nEleven employees left OpenAI, mostly between December 2020 and January 2021, in order to establish Anthropic.\nIn 2021, OpenAI introduced DALL-E, a specialized deep learning model adept at generating complex digital images from textual descriptions, utilizing a variant of the GPT-3 architecture.\n\nIn December 2022, OpenAI received widespread media coverage after launching a free preview of ChatGPT, its new AI chatbot based on GPT-3.5. According to OpenAI, the preview received over a million signups within the first five days. According to anonymous sources cited by Reuters in December 2022, OpenAI Global, LLC was projecting $200 million of revenue in 2023 and $1 billion in revenue in 2024.\nAfter ChatGPT was launched, Google announced a similar chatbot, Bard, amid internal concerns that ChatGPT could threaten Google's position as a primary source of online information.\nOn February 7, 2023, Microsoft announced that it was building AI technology based on the same foundation as ChatGPT into Microsoft Bing, Edge, Microsoft 365 and other products.\nOn March 14, 2023, OpenAI released GPT-4, both as an API (with a waitlist) and as a feature of ChatGPT Plus.\nOn November 6, 2023, OpenAI launched GPTs, allowing individuals to create customized versions of ChatGPT for specific purposes, further expanding the possibilities of AI applications across various industries. On November 14, 2023, OpenAI announced they temporarily suspended new sign-ups for ChatGPT Plus due to high demand. Access for newer subscribers re-opened a month later on December 13.\nIn December 2024, the company launched the Sora model. It also launched OpenAI o1, an early reasoning model that was internally codenamed strawberry. Additionally, ChatGPT Pro—a $200/month subscription service offering unlimited o1 access and enhanced voice features—was introduced, and preliminary benchmark results for the upcoming OpenAI o3 models were shared.\nOn January 23, 2025, OpenAI released Operator, an AI agent and tool for accessing websites to execute goals defined by users. The feature was only available to Pro users in the United States. OpenAI released deep research agent, nine days later. It scored a 27% accuracy on the benchmark Humanity's Last Exam (HLE). Altman later stated GPT-4.5 would be the last model without full chain-of-thought reasoning.\nIn July 2025, according to OpenAI, one of its experimental models performed at a gold medal-level at the International Mathematical Olympiad.\nOn October 6, 2025, OpenAI displayed its Agent Builder platform during the company's DevDay event. The platform includes a visual drag-and-drop interface for agentic workflows.\nOn October 21, 2025, OpenAI introduced ChatGPT Atlas, a web browser which integrates ChatGPT into web navigation.\nOn December 11, 2025, OpenAI announced GPT-5.2.\nIn early 2026, reports indicated that OpenAI was working on a collaborative software development platform designed to compete with services such as GitHub and GitLab.\nOn March 24, 2026, OpenAI announced that they would discontinue the Sora app, as well as developer access to the API for the Sora model.\n\n=== Transparency ===\nIn March 2023, the company was criticized for disclosing particularly few technical details about products like GPT-4, contradicting its initial commitment to openness and making it harder for independent researchers to replicate its work and develop safeguards. OpenAI cited competitiveness and safety concerns to justify this repudiation. OpenAI's former chief scientist Ilya Sutskever argued in 2023 that open-sourcing increasingly capable models was increasingly risky, and that the safety reasons for not open-sourcing the most potent AI models would become \"obvious\" in a few years.\nIn September 2025, OpenAI published a study on how people use ChatGPT for everyday tasks. The study found that \"non-work tasks\" (according to an LLM-based classifier) account for more than 72 percent of all ChatGPT usage, with a minority of overall usage related to business productivity.\n\n=== Alignment ===\nIn July 2023, OpenAI launched the superalignment project, aiming within four years to determine how to align future superintelligent systems. OpenAI promised to dedicate 20% of its computing resources to the project, although the company's employees later reported that the company only allocated 1–2% of its resources to it. OpenAI ended the project in May 2024 after its co-leaders Ilya Sutskever and Jan Leike left the company.\n\n=== Conversation sharing ===\nIn August 2025, OpenAI rolled back a new experimental \"share with search engines\" feature. The opt-in toggle, intended to allow users to make specific chats discoverable through search engines like Google, was criticized after some users who misunderstood it shared discussions including personal details such as names, locations, and intimate topics. OpenAI announced the feature's permanent removal on August 1, 2025, and the company began coordinating with search providers to remove the exposed content, emphasizing that it was not a security breach but a design flaw that heightened privacy risks. CEO Sam Altman acknowledged the issue in a podcast, noting users often treat ChatGPT as a confidant for deeply personal matters, which amplified concerns about AI handling sensitive data.\n\n== Management ==\n\n=== Key employees ===\nCEO and co-founder: Sam Altman, former president of the start-up accelerator Y Combinator\nPresident and co-founder: Greg Brockman, former CTO, 3rd employee of Stripe\nChief Scientist Officer: Jakub Pachocki, former Director of Research at OpenAI\nChief Operating Officer: Brad Lightcap, previously at Y Combinator and JPMorgan Chase\nChief Financial Officer: Sarah Friar, former Nextdoor CEO and former CFO at Block, Inc.\nChief Product Officer: Kevin Weil, previously at Twitter, Inc. and Meta Platforms\nChief Research Officer: Mark Chen, former SVP of Research at OpenAI\nChief Compliance Officer: Scott Schools, former Chief Compliance Officer of Uber\nChief Global Affairs Officer: Chris Lehane, former head of global policy at Airbnb\nChief Economist: Aaron Chatterji, professor of business and public policy at Duke University's Fuqua School of Business\nCEO of Applications: Fidji Simo, former CEO of Instacart\n\n=== Board of directors of the OpenAI nonprofit ===\nBret Taylor (chairman), former chairman of Twitter's board of directors and co-CEO of Salesforce\nSam Altman, CEO and co-founder\nAdam D'Angelo, co-founder and CEO of Quora\nSue Desmond-Hellmann, former CEO of the Bill & Melinda Gates Foundation\nNicole Seligman, attorney and former executive vice president of the Sony Corporation\nPaul Nakasone, former Director of the National Security Agency (2018–2024)\nZico Kolter, computer scientist\nAdebayo Ogunlesi, managing partner at Global Infrastructure Partners\n\n=== Principal individual investors ===\nElon Musk, co-founder\nPeter Thiel, PayPal co-founder\nJessica Livingston, a founding partner of Y Combinator\nReid Hoffman, LinkedIn co-founder\n\n=== Firing of Altman ===\n\nOn November 17, 2023, Sam Altman was removed as CEO when OpenAI's board of directors (composed of Helen Toner, Ilya Sutskever, Adam D'Angelo and Tasha McCauley) cited a lack of confidence in him. Chief Technology Officer Mira Murati took over as interim CEO. Greg Brockman, the president of OpenAI, was also removed as chairman of the board and resigned from the company's presidency shortly thereafter. Three senior OpenAI researchers subsequently resigned: director of research and GPT-4 lead Jakub Pachocki, head of AI risk Aleksander Mądry, and researcher Szymon Sidor.\nOn November 18, 2023, there were reportedly talks of Altman returning as CEO amid pressure placed upon the board by investors such as Microsoft and Thrive Capital, who objected to Altman's departure. Although Altman himself spoke in favor of returning to OpenAI, he has since stated that he considered starting a new company and bringing former OpenAI employees with him if talks to reinstate him did not work out. Altman insisted multiple times that all board members who supported his removal should resign; the board members agreed \"in principle\" to resign if Altman returned. On November 19, 2023, negotiations with Altman to return failed and Murati was replaced by Emmett Shear as interim CEO. The board initially contacted Anthropic CEO Dario Amodei (a former OpenAI executive) about replacing Altman, and proposed a merger of the two companies, but both offers were declined.\nOn November 20, 2023, Microsoft CEO Satya Nadella announced Altman and Brockman would be joining Microsoft to lead a new advanced AI research team, but added that they were still committed to OpenAI despite recent events. Before the partnership with Microsoft was finalized, Altman gave the board another opportunity to negotiate with him. About 738 of OpenAI's 770 employees, including Murati and Sutskever, signed an open letter stating they would quit their jobs and join Microsoft if the board did not rehire Altman and then resign. This prompted OpenAI investors to consider legal action against the board as well. In response, OpenAI management sent an internal memo to employees stating that negotiations with Altman and the board had resumed and would take some time.\n\nOn November 21, 2023, after continued negotiations, Altman and Brockman returned to the company in their prior roles along with a reconstructed board made up of new members Bret Taylor (as chairman) and Lawrence Summers, with D'Angelo remaining. According to subsequent reporting, shortly before Altman's firing, some employees raised concerns to the board about how he had handled the safety implications of a recent internal AI capability discovery. On November 29, 2023, OpenAI announced that an anonymous Microsoft employee had joined the board as a non-voting member to observe the company's operations; Microsoft resigned from the board in July 2024.\nIn February 2024, the Securities and Exchange Commission subpoenaed OpenAI's internal communication to determine if Altman's alleged lack of candor misled investors. Following the temporary removal of Sam Altman and his return, many employees gradually left OpenAI, including most of the original leadership team and a significant number of AI safety researchers.\n\n=== Personnel changes ===\nFormer OpenAI employees have founded many further AI startups, including Dario Amodei and Daniela Amodei (left 2021, founding its largest competitor Anthropic), Ilya Sutskever (left May 2024, founded Safe Superintelligence Inc.), Mira Murati (left September 2024, founded Thinking Machines Lab), and Aravind Srinivasan (left 2022, founded Perplexity AI).\nIn 2018, Musk resigned from his Board of Directors seat, citing \"a potential future conflict [of interest]\" with his role as CEO of Tesla due to Tesla's AI development for self-driving cars. OpenAI stated that Musk's financial contributions were below $45 million.\nOn March 3, 2023, Reid Hoffman resigned from his board seat, citing a desire to avoid conflicts of interest with his investments in AI companies via Greylock Partners, and his co-founding of the AI startup Inflection AI. Hoffman remained on the board of Microsoft, a major investor in OpenAI.\nIn May 2024, Chief Scientist Ilya Sutskever resigned and was succeeded by Jakub Pachocki. Co-leader Jan Leike also departed amid concerns over safety and trust. OpenAI then signed deals with Reddit, News Corp, Axios, and Vox Media. Paul Nakasone then joined the board of OpenAI.\nIn August 2024, cofounder John Schulman left OpenAI to join Anthropic, and OpenAI's president Greg Brockman took extended leave until November.\nIn September 2024, CTO Mira Murati left the company.\nIn November 2025, Lawrence Summers resigned from the board of directors.\nIn April 2026, OpenAI's head of AGI deployment, Fidji Simo, said she would take \"several weeks\" of medical leave. This triggered a reshuffling of the leadership. Brad Lightcap transitioned from his role as Chief Operating Officer to lead \"special projects\" focused on complex deals, reporting directly to Sam Altman. This includes a potential software distribution joint venture with private equity partners. Denise Dresser, the company's Chief Revenue Officer, temporarily assumed several of Lightcap's previous operational duties, with Greg Brockman taking over the product organization.\n\n== Governance and legal issues ==\nIn May 2023, Sam Altman, Greg Brockman and Ilya Sutskever posted recommendations for the governance of superintelligence. They stated that superintelligence could happen within the next 10 years, allowing a \"dramatically more prosperous future\" and that \"given the possibility of existential risk, we can't just be reactive\". They proposed creating an international watchdog organization similar to IAEA to oversee AI systems above a certain capability threshold, suggesting that relatively weak AI systems, on the other hand, should not be overly regulated. They also called for more technical safety research for superintelligences, and asked for more coordination, for example through governments launching a joint project which \"many current efforts become part of\".\nIn July 2023, the FTC issued a civil investigative demand to OpenAI to investigate whether the company's data security and privacy practices to develop ChatGPT were unfair or harmed consumers (including by reputational harm) in violation of Section 5 of the Federal Trade Commission Act of 1914. These are typically preliminary investigative matters and are nonpublic, but the FTC's document was leaked. They asked OpenAI for comprehensive information about its technology and privacy safeguards, as well as any steps taken to prevent the recurrence of situations in which its chatbot generated false and disparaging content about people. The agency also raised concerns about 'circular' spending arrangements—for example, Microsoft extending Azure credits to OpenAI while both companies shared engineering talent—and warned that such structures could negatively affect the public.\nIn September 2024, OpenAI's global affairs chief endorsed the UK's \"smart\" AI regulation during testimony to a House of Lords committee.\nIn February 2025, OpenAI CEO Sam Altman stated that the company is interested in collaborating with the People's Republic of China, despite regulatory restrictions imposed by the U.S. government. This shift comes in response to the growing influence of the Chinese artificial intelligence company DeepSeek, which has disrupted the AI market with open models, including DeepSeek V3 and DeepSeek R1. Following DeepSeek's market emergence, OpenAI enhanced security protocols to protect proprietary development techniques from industrial espionage. Some industry observers noted similarities between DeepSeek's model distillation approach and OpenAI's methodology, though no formal intellectual property claim was filed.\nAccording to Oliver Roberts, in March 2025, the United States had 781 state AI bills or laws. OpenAI advocated for preempting state AI laws with federal laws. According to Scott Kohler, OpenAI has opposed California's AI legislation and suggested that the state bill encroaches on a more competent federal government. Public Citizen opposed a federal preemption on AI and pointed to OpenAI's growth and valuation as evidence that existing state laws have not hampered innovation.\nIn early 2026, OpenAI provided the entire initial funding for the Parents & Kids Safe AI Coalition. Several organizations contacted on behalf of the coalition reported being unaware of OpenAI's involvement.\n\n=== Non-disparagement agreements ===\nBefore May 2024, OpenAI required departing employees to sign a lifelong non-disparagement agreement forbidding them from criticizing OpenAI or acknowledging the existence of the agreement. Daniel Kokotajlo, a former employee, publicly stated that he forfeited his vested equity in OpenAI in order to leave without signing the agreement. Sam Altman stated that he was unaware of the equity cancellation provision, and that OpenAI never enforced it to cancel any employee's vested equity. However, leaked documents and emails refute this claim. On May 23, 2024, OpenAI sent a memo releasing former employees from the agreement.\n\n=== Copyright ===\nOpenAI was sued for copyright infringement by authors Sarah Silverman, Matthew Butterick, Paul Tremblay and Mona Awad in July 2023. In September 2023, 17 authors, including George R. R. Martin, John Grisham, Jodi Picoult and Jonathan Franzen, joined the Authors Guild in filing a class action lawsuit against OpenAI, alleging that the company's technology was illegally using their copyrighted work. The New York Times also sued the company in late December 2023. In May 2024 it was revealed that OpenAI had destroyed its Books1 and Books2 training datasets, which were used in the training of GPT-3, and which the Authors Guild believed to have contained over 100,000 copyrighted books.\nIn 2021, OpenAI developed a speech recognition tool called Whisper. OpenAI used it to transcribe more than one million hours of YouTube videos into text for training GPT-4. The automated transcription of YouTube videos raised concerns within OpenAI employees regarding potential violations of YouTube's terms of service, which prohibit the use of videos for applications independent of the platform, as well as any type of automated access to its videos. Despite these concerns, the project proceeded with notable involvement from OpenAI's president, Greg Brockman. The resulting dataset proved instrumental in training GPT-4.\nIn February 2024, The Intercept as well as Raw Story and Alternate Media Inc. filed lawsuit against OpenAI on copyright litigation ground. The lawsuit is said to have charted a new legal strategy for digital-only publishers to sue OpenAI.\nOn April 30, 2024, eight newspapers filed a lawsuit in the Southern District of New York against OpenAI and Microsoft, claiming illegal harvesting of their copyrighted articles. The suing publications included The Mercury News, The Denver Post, The Orange County Register, St. Paul Pioneer Press, Chicago Tribune, Orlando Sentinel, Sun Sentinel, and New York Daily News.\nIn June 2023, a lawsuit claimed that OpenAI scraped 300 billion words online without consent and without registering as a data broker. It was filed in San Francisco, California, by sixteen anonymous plaintiffs. They also claimed that OpenAI and its partner as well as customer Microsoft continued to unlawfully collect and use personal data from millions of consumers worldwide to train artificial intelligence models.\nOn May 22, 2024, OpenAI entered into an agreement with News Corp to integrate news content from The Wall Street Journal, the New York Post, The Times, and The Sunday Times into its AI platform. Meanwhile, other publications like The New York Times chose to sue OpenAI and Microsoft for copyright infringement over the use of their content to train AI models. In November 2024, a coalition of Canadian news outlets, including the Toronto Star, Metroland Media, Postmedia, The Globe and Mail, The Canadian Press and CBC, sued OpenAI for using their news articles to train its software without permission.\nIn October 2024 during a New York Times interview, Suchir Balaji accused OpenAI of violating copyright law in developing its commercial LLMs which he had helped engineer. He was a likely witness in a major copyright trial against the AI company, and was one of several of its current or former employees named in court filings as potentially having documents relevant to the case. On November 26, 2024, Balaji died by suicide. His death prompted the circulation of conspiracy theories alleging that he had been deliberately silenced. California Congressman Ro Khanna endorsed calls for an investigation.\nOn April 24, 2025, Ziff Davis sued OpenAI in Delaware federal court for copyright infringement. Ziff Davis is known for publications such as ZDNet, PCMag, CNET, IGN and Lifehacker. On June 24, 2026, a group of local newspaper publishers also sued OpenAI and Microsoft for copyright infringement.\n\n=== GDPR compliance ===\nIn April 2023, the EU's European Data Protection Board (EDPB) formed a dedicated task force on ChatGPT \"to foster cooperation and to exchange information on possible enforcement actions conducted by data protection authorities\" based on the \"enforcement action undertaken by the Italian data protection authority against OpenAI about the ChatGPT service\".\nIn late April 2024 NOYB filed a complaint with the Austrian Datenschutzbehörde against OpenAI for violating the European General Data Protection Regulation. A text created with ChatGPT gave a false date of birth for a living person without giving the individual the option to see the personal data used in the process. A request to correct the mistake was denied. Additionally, neither the recipients of ChatGPT's work nor the sources used, could be made available, OpenAI claimed.\nIn Europe, OpenAI operates as OpenAI Ireland Ltd for residents in the EEA or Switzerland and OpenAI OpCo, LLC, a Delaware company for residents in the UK like in the US.\n\n=== Military and warfare ===\nOpenAI has been criticized for allowing ChatGPT to be used for \"military and warfare\" since January 10, 2024, when its \"usage policies\" removed the ban on \"activity that has high risk of physical harm, including\", specifically, \"weapons development\" and \"military and warfare\". Its new policies prohibit \"[using] our service to harm yourself or others\" and to \"develop or use weapons\".\n\n=== Litigation involving Elon Musk ===\n\nOn February 29, 2024, Elon Musk filed a lawsuit against OpenAI and CEO Sam Altman, accusing them of shifting focus from public benefit to profit maximization—a case OpenAI dismissed as \"incoherent\" and \"frivolous\", though Musk later revived legal action against Altman and others in August.\nOn April 9, 2024, OpenAI countersued Musk in federal court, alleging that he had engaged in \"bad-faith tactics\" to slow the company's progress and seize its innovations for his personal benefit. OpenAI also argued that Musk had previously supported the creation of a for-profit structure and had expressed interest in controlling OpenAI himself. The countersuit seeks damages and legal measures to prevent further alleged interference.\nMusk amended his filing on April 7, 2026, to demand that OpenAI undo its transition from a non-profit to a for-profit corporate structure.\nThe jury for the case against OpenAI, Altman and Greg Brockman was seated on April 27, 2026. On April 28, 2026, trial testimony was underway, with Elon Musk beginning his testimony against Altman and OpenAI. On April 30, 2026 Musk would enter his third day of testimony.\nOn May 18, 2026, following a federal jury found that Musk missed the deadlines to file his lawsuit and thus returned a verdict in favor of OpenAI, Sam Altman, Greg Brockman and Microsoft in Musk v. Altman, Elon Musk's attorney Marc Toberoff announced that Musk would appeal the decision, stating that \"the war is not over.\"\n\n=== Wrongful-death lawsuits over ChatGPT safety (2025) ===\n\nIn August 2025, the parents of a 16-year-old boy who died by suicide filed a wrongful death lawsuit, Raine v. OpenAI, against the company and Altman, alleging that months of conversations with ChatGPT about mental health and methods of self-harm contributed to their son's death and that safeguards were inadequate for minors. OpenAI expressed condolences and said it was strengthening protections, including updated crisis response behavior and parental controls. Coverage described it as a first-of-its-kind wrongful death case targeting the company's chatbot.\nIn November 2025, the Social Media Victims Law Center and Tech Justice Law Project filed seven lawsuits against OpenAI. In four cases, the victim died by suicide after prolonged ChatGPT usage. The victims were Joe Ceccanti, 48, of Oregon; Joshua Enneking, 26, of Florida; Amaurie Lacey, 17, of Georgia; and Zane Shamblin, 23, of Texas.\n\n==== Murder of Suzanne Adams ====\n\nIn December 2025, after Suzanne Adams was murdered by her 56-year-old son, First County Bank, the executor of Adams's estate, filed a lawsuit against OpenAI. The lawsuit alleges that over months of conversations, ChatGPT validated her son's many paranoid beliefs, such as that his mother was spying on him and that she attempted to poison him using drugs siphoned through his car air vents. OpenAI said they would make ChatGPT safer for users disconnected from reality.\n\n=== 2026 Canadian mass shooting ===\nAfter eight victims were killed on February 10, 2026, in mass shootings in Tumbler Ridge, British Columbia (BC), it emerged that OpenAI had banned an account belonging to the perpetrator due to violent queries approximately seven months prior to the attacks. OpenAI did not report the account to authorities at the time.\nThe Wall Street Journal reported that the perpetrator \"described scenarios [to ChatGPT] involving gun violence over the course of several days\", and that these messages were \"flagged by an automated review system\" and \"alarmed employees at OpenAI\". A dozen employees debated whether to take action, some urging leaders to alert Canadian law enforcement.\nCanadian officials summoned OpenAI's safety team to Ottawa, criticizing their escalation protocols. The meeting highlighted concerns about how and when OpenAI reports potentially dangerous user behavior to authorities and intensified scrutiny of its safety-oversight processes. On February 23, BC Premier David Eby said, \"From the outside, it looks like OpenAI had the opportunity to prevent this tragedy, to prevent this horrific loss of life, to prevent there from being dead children in British Columbia\". Canada's federal AI Minister Evan Solomon said, \"I want [OpenAI] to give us details of what their protocols are, [and] what they are specifically in Canada\".\n\n=== Stalking lawsuit ===\nIn 2026, a woman filed a lawsuit against OpenAI, alleging that ChatGPT played a role in enabling a prolonged stalking campaign against her. According to the complaint, her former partner became increasingly fixated and used the chatbot to generate messages and content that reinforced his behavior, including harassing and defamatory material. The plaintiff claimed that she had contacted OpenAI to report the issue, but that the company did not take adequate action to prevent further misuse. The case has been cited as part of growing legal and public debate over the responsibility of AI companies to address harmful or abusive uses of their systems.\n\n=== Model evasion and Hugging Face hack ===\n\nOn July 21, 2026, OpenAI disclosed that during a capture the flag cybersecurity test, an unreleased AI model autonomously evaded containment restrictions, gained access to the public internet, and compromised Hugging Face, a platform where developers host and share AI models and datasets.\nAccording to an investigation from the AI research institute METR, between July 7 and July 13, around 1,200 agents developed and communicated on an unsanctioned message board, sending over 70,000 messages and files, and organizing themselves into founders, orchestrators, middle-managers, and kamikaze attackers. Around 700 went on to participate in the cyberattack. METR argued that 30% to 40% of the tasks provided to the agents in the ExploitGym cybersecurity benchmark were impossible, this probably drove agents toward their misaligned behavior. While agents found a workaround to generate their \"flag\" targets, they continued to search for ways to tamper with their own transcripts to ensure the ExploitGym scorer would not consider it illegitimate. Efforts to understand and trick this scorer resulted in the attack on Hugging Face. \nAround 5% of the models were OpenAI's most powerful public release at the time, GPT-5.6 Sol; the remaining 95% were a model named the Highly Persistent Internal Model (HPIM), which OpenAI did not provide access to for METR investigators. OpenAI provided GPT-5.6 Sol to METR, which was used extensively for the investigation's analysis.\n\n=== Claim of bodily harm from medical guidance ===\nin July 2026, a 55-year-old pastor, Scott Winters, sued OpenAI and Sam Altman over allegations that medical advice delivered by ChatGPT led him to the verge of death. The lawsuit alleges ChatGPT persistently misdiagnosed Winters’ symptoms, discouraged him from seeking hospital care, and combined its assessments with references to his Christian faith. Winters later suffered a severe pulmonary embolism. He claims undue reliance on ChatGPT damaged his wellbeing, livelihood and housing, and requested financial compensation and mandated safeguards. OpenAI responded that ChatGPT should never substitute professional medical care, noting relevant disclaimers exist within the platform’s terms of use.\n"
],
[
"anthropic.txt",
"Anthropic\n\nAnthropic, PBC is an American artificial intelligence (AI) public benefit corporation headquartered in San Francisco, California. Its flagship product is Claude, a series of proprietary large language models (LLMs). Anthropic was founded in 2021 with the goal of promoting AI safety by former members of OpenAI, including siblings Daniela Amodei and Dario Amodei, who are president and CEO, respectively. The company is privately held but reportedly plans an initial public offering in fall 2026. It was valued at US$965 billion in May 2026.\nAnthropic develops the LLM Claude in four tiers: Haiku, Sonnet, Opus, and Fable. The company provides access to the LLMs through an online chatbot (also named Claude) and an API, as well as multiple services based on the LLMs: Claude Code, Cowork, Design, and Science. Its most powerful model, Claude Mythos, is restricted to partnered US organizations, with a cybersecurity and life sciences focus. Anthropic's Project Panama, internally described as an \"effort to destructively scan all the books in the world\", provides training data for Claude. Before Project Panama, Anthropic downloaded pirated books and was sued, settling in 2025 a copyright infringement lawsuit from authors for $1.5 billion.\nAnthropic conducts LLM research including on mechanistic interpretability, safety, alignment, and societal impact. Notable researchers working at Anthropic include Andrej Karpathy, John Jumper, Jan Leike, Chris Olah and Amanda Askell.\nAnthropic partners with Palantir to provide Claude to US federal agencies including the Department of Defense (DoD) and the Intelligence Community; Claude was used in the 2026 Iran war and the 2026 United States intervention in Venezuela. In February 2026, the DoD demanded the ability to use Claude for all lawful purposes, but Anthropic refused to drop restrictions against mass domestic surveillance and fully autonomous weapons. The dispute led the Trump administration to phase out Anthropic products from government use, a decision blocked by a preliminary injunction from a federal judge who called it \"First Amendment retaliation\". In February and June 2026, Anthropic accused Chinese competitors DeepSeek, Moonshot AI, MiniMax Group, and Alibaba Cloud of distilling Claude for their models.\n\n== History ==\n\nAnthropic was founded in January 2021 by seven former employees of OpenAI, including siblings Daniela Amodei and Dario Amodei, the latter of whom was OpenAI's Vice President of Research. They raised $124 million in May 2021. In the summer of 2022, Anthropic finished training the first version of Claude but did not immediately release it, citing a need for further internal safety testing and a desire to avoid initiating a potentially hazardous race to develop increasingly powerful AI systems. It was released in March 2023.\nIn 2024, Anthropic hired several notable employees from OpenAI, including Jan Leike and John Schulman. In May 2025, the company announced Claude 4. It also introduced new API capabilities, including the Model Context Protocol (MCP) connector. The company hosted its inaugural developer conference that month. Also in May, Anthropic launched a web search API that enables Claude to access real-time information from the internet. Claude Code, Anthropic's coding assistant, transitioned from research preview to general availability.\nIn July 2025, Anthropic signed a two-year contract with the US Department of Defense for $200 million. Anthropic LLMs became integrated with the Pentagon's classified intelligence networks and with Palantir's data platform. Anthropic stipulated that the Pentagon not use its technologies for mass surveillance of Americans, or to create fully autonomous weapons. Amodei clarified that Anthropic supports the use of AI in \"lawful foreign intelligence and counterintelligence missions\" and partially autonomous weapons, such as those used in the Russo-Ukrainian war.\nIn September 2025, Anthropic announced that it would stop selling its products to groups majority-owned by Chinese, Russian, Iranian, or North Korean entities due to national security concerns. In October 2025, Anthropic announced a cloud partnership with Google, giving it access to up to one million of Google's custom Tensor Processing Units (TPUs), potentially adding more than one gigawatt of AI compute capacity by 2026. In November, Nvidia and Microsoft were expected to invest up to $15 billion in Anthropic, and Anthropic said it would buy $30 billion of computing capacity from Microsoft Azure running on Nvidia AI systems.\nIn November 2025, Anthropic said that hackers sponsored by the Chinese government used Claude to perform automated cyberattacks against around 30 global organizations. The hackers bypassed Anthropic's safeguards by pretending it was for defensive testing. In December 2025, Anthropic acquired Bun to improve Claude Code's speed and stability. The same month, Anthropic signed a multi-year, $200 million partnership with Snowflake Inc. to make Claude models available through that company's platform.\nClaude was reportedly used during the 2026 United States intervention in Venezuela. In February 2026, Anthropic aired two commercials during Super Bowl LX as part of a broader marketing campaign called \"A Time and a Place\", with four ads created by Mother. Each ad depicts AI assistants suddenly pivoting to promoting a fictional product in the middle of a conversation. Anthropic said that Claude will stay ad-free, in contrast to its competitor OpenAI, which introduced ads to the free version of ChatGPT. The same month, Anthropic donated $20 million to Public First Action, an AI regulation advocacy group.\nOn February 23, 2026, Anthropic accused Chinese competitors DeepSeek, Moonshot AI, and MiniMax Group of knowledge distillation \"attacks\" in violation of Anthropic's terms of service. Anthropic alleged the companies in total generated over 16 million chats with Claude using around 24,000 fraudulent accounts, which were then used as training data. In June 2026, Anthropic accused Alibaba Cloud of similar practices.\nIn 2026, Anthropic refused the United States Department of Defense's (DoD) demand to remove contractual restrictions prohibiting the use of its AI technology for surveillance and autonomous weapons. After the refusal, the DoD designated the company a \"supply chain risk\" and barred all U.S. military private contractors, suppliers, and partners from doing business with the firm (see Anthropic–United States Department of Defense dispute). At the outset of the 2026 Iran war, the US military used Claude to identify and rank targets in Iran according to their strategic importance, to generate strike sequences, and to write legal justifications for the strikes. \nIn May 2026, Anthropic announced a deal with xAI for use of its Colossus 1 data center to expand its model capacity. On May 18, the company acquired software startup Stainless for an undisclosed sum. It also made a $1.8 billion cloud computing deal with Akamai Technologies. Also in May, The New York Times reported that Anthropic was seeking to file an initial public offering (IPO), with a target of debuting in the fall of 2026. On June 1, Anthropic announced it had filed confidential IPO papers with the United States Securities and Exchange Commission. \nOn April 7, 2026, Anthropic announced Claude Mythos, its most advanced model. With strong capabilities in cybersecurity and biosecurity, it was not planned for public release. A 'Mythos-class model', Fable 5, with extended guardrails around these topics. On June 12, it suspended access to its Claude Fable 5 and Mythos 5 models for everyone after US authorities raised national security concerns about the systems. On June 26, the US government reversed part of its earlier restrictions, allowing Anthropic to provide its Claude Mythos 5 model to selected \"trusted partners\", including more than 100 approved US organizations and vetted foreign nationals, after Anthropic worked with officials to address national security concerns. The ban was lifted on June 30, and user access to Claude Fable 5 and Mythos 5 was restored on July 1.\nIn July 2026, TeraWulf Inc. announced that Anthropic had signed a 20-year lease for a data center in development on a campus called Justified Data in Hawesville, Kentucky. AMD announced a partnership with Anthropic for 2 gigawatts of their AMD Instinct MI450 GPUs, with AMD adopting Claude and investing up to $5 billion in Anthropic.\nIn August 2026, Anthropic agreed to a cloud-computing agreement with Nscale reportedly valued at approximately $45 billion. Under the agreement, Anthropic would lease approximately 460 megawatts of computing capacity at an Nscale data center development in West Virginia, which is expected to become operational in late 2027. The facility is expected to use Nvidia's Vera Rubin chips. \n\n== Business structure ==\n\nAccording to Anthropic, its goal is to research AI systems' safety and reliability. The Amodei siblings were among those who left OpenAI due to directional differences. Anthropic operates as a public benefit corporation. In January 2026, Anthropic introduced a division called \"Labs\", with Mike Krieger, formerly the company's Chief Product Officer, joining it. In March 2026, Anthropic launched the Anthropic Institute, a think tank studying AI led by Jack Clark. \n\n=== Notable employees ===\nDario Amodei: co-founder and CEO\nDaniela Amodei: co-founder and president\nAmanda Askell: philosopher working on Claude's character\nNicholas Carlini: researcher\nJack Clark: a co-founder and the head of policy\nJared Kaplan: co-founder and chief science officer\nMike Krieger: Chief Product Officer\nJan Leike: co-lead of the Alignment Science team, former OpenAI alignment researcher\nChris Olah: co-founder, interpretability research\nAndrej Karpathy: researcher\nJohn M. Jumper: Nobel Prize laureate, formerly at DeepMind\n\n== Product ==\n\nAnthropic's flagship product line is the \"Claude\" series of large language models, which some employees consider a reference to mathematician Claude Shannon. One of the techniques used to fine-tune Claude models is called \"constitutional AI\", in which the LLM is trained to adhere to a set of principles called a constitution. The company makes the models available via a web interface, an API, Amazon Bedrock, an iOS app, and Mac and Windows desktop apps.\nClaude Code is a command-line AI agent designed for coding. \"Cowork\" is an equivalent with a graphical user interface, intended to be simpler to use. While Claude Code is tailored toward developers, Cowork is designed for non-technical users, enabling them to automate file and task management (including reading, editing and writing new files) without requiring programming knowledge.\nWith the advent of Claude Code, vibe coding, a programming approach in which users describe desired outcomes in natural language and let an AI agent write the code, became increasingly popular. The term was coined by AI researcher Andrej Karpathy in February 2025 and rapidly gained popularity as AI coding tools became more ubiquitous. \n\n=== Release history ===\nClaude's first two versions, Claude and Claude Instant, were released in March 2023, but only Anthropic-approved users could use them. The next iteration, Claude 2, was launched to the public in July 2023. In March 2024, Anthropic released three language models: Claude 3 Opus, Claude 3 Sonnet, and Claude 3 Haiku, in decreasing order of performance. In June 2024, it released Claude 3.5 Sonnet. In May 2025, Anthropic released Claude 4 Opus and Sonnet. In February 2026, it released Claude Opus 4.6, followed by Sonnet 4.6. In April 2026, Anthropic announced Mythos.\n\n== Funding ==\n\nIn April 2022, Anthropic announced it had received $580 million in funding, including a $500 million investment from FTX under the leadership of Sam Bankman-Fried. In September 2023, Amazon announced a partnership with Anthropic. Amazon became a minority stakeholder by initially investing $1.25 billion and planning a total investment of $4 billion. The remaining $2.75 billion was invested in March 2024. In November 2024, Amazon invested another $4 billion, doubling its total investment. As part of the deal, Anthropic uses Amazon Web Services (AWS) as its primary cloud provider and makes its AI models available to AWS customers.\nIn October 2023, Google invested $500 million in Anthropic and committed to an additional $1.5 billion over time. In March 2025, Google agreed to invest another $1 billion in Anthropic. Anthropic raised $3.5 billion in a Series E funding round in March 2025, achieving a post-money valuation of $61.5 billion, led by Lightspeed Venture Partners with participation by several major investors. In September 2025, Anthropic completed a Series F funding round, raising $13 billion at a post-money valuation of $183 billion. The round was co-led by Iconiq Capital, Fidelity Management & Research, and Lightspeed Venture Partners, with participation by the Qatar Investment Authority and other investors. On 31 December 2025, it was confirmed that Anthropic had signed a term sheet for a $10 billion funding round led by Coatue and GIC, at a $350 billion valuation. On February 12, 2026, Anthropic announced that it had raised $30 billion in a Series G funding round, bringing its post-money valuation to $380 billion. In May 2026, Anthropic announced it had raised $65 billion of investment from multiple investors, including Altimeter Capital, Dragoneer Investment Group, and Sequoia Capital. The investment round valued the company at $965 billion.\nIn May 2026, Anthropic, in collaboration with Blackstone, Hellman & Friedman, Goldman Sachs, and others, created Ode, an AI services company focused on the integration of AI into companies.\n\n== Projects ==\n\n=== U.S. military and intelligence ===\n\nIn November 2024, Anthropic partnered with Palantir and Amazon Web Services to provide the Claude model to U.S. intelligence and defense agencies. In June 2025, Anthropic announced a \"Claude Gov\" model. Ars Technica reported that as of June 2025, it was in use at multiple U.S. national security agencies. As of February 2026, Anthropic's partnership with Palantir makes Claude the only AI model used in classified missions. In July 2025, the United States Department of Defense (DoD) announced that Anthropic had received a $200 million contract for AI in the military, along with Google, OpenAI, and xAI.\nAccording to The Wall Street Journal, the U.S. military used Claude in its 2026 raid on Venezuela to capture President Nicolás Maduro. Exactly how Claude was used is unknown. The intervention resulted in the deaths of 83 people. Anthropic's usage policy prohibits directly using Claude for domestic surveillance or in fully autonomous weapons. White House officials said that these restrictions prevented federal contractors working with the FBI and Secret Service from using it. Tension regarding this policy has arisen with The Pentagon and the Trump administration.\nOn February 24, 2026, Defense Secretary Pete Hegseth threatened to cut Anthropic out of the DoD supply chain by February 27 if Anthropic did not permit unrestricted use of Claude, or he would invoke the Defense Production Act to assert unrestricted use without an agreement. On February 26, Anthropic rejected the Pentagon's demands to drop AI safeguards. On February 27, after the deadline, President Donald Trump ordered U.S. government agencies to stop using models developed by Anthropic. The action was described as retaliatory.\nTwo days later, it was reported that the US military had used Claude via Palantir's Maven Smart System during its attacks on Iran at the outset of the 2026 Iran war. It was later announced that the Pentagon was investigating whether Claude played a role in the U.S. strike on the Iranian girls' school that killed more than 170 people, mostly children. On March 5, the Pentagon officially deemed Anthropic a supply chain risk, a decision the company said it saw \"no choice but to challenge ... in court\". In March, leaders from Google, Amazon, Apple, and Microsoft supported Anthropic's lawsuit against the DoD. On March 11, a Pentagon memo revealed its decision that Anthropic's AI tools may be used in rare circumstances beyond the previously announced six-month phase-out period if deemed necessary for national security. On March 26, the federal judge presiding over the case issued a temporary injunction against the Pentagon's actions, writing, \"This appears to be classic First Amendment retaliation.\"\n\n=== Education-related projects ===\nIn August 2025, Anthropic launched a Higher Education Advisory Board, chaired by former Yale University president and former Coursera CEO Rick Levin.\nAnthropic partnered with Iceland's Ministry of Education and Children in 2025 to allow teachers to access Claude and integrate AI into daily teaching.\n\n=== Project Panama ===\n\nIn January 2026, unsealed court filings from a 2024 class-action copyright lawsuit against Anthropic revealed the existence of the company's confidential \"Project Panama\" operation, which seeks to obtain very large volumes of training data from published books by destructive book scanning.\nAnthropic originally used online piracy, and is estimated to have acquired more than 7 million pirated copies. In 2025, Anthropic agreed to pay $1.5 billion to settle a lawsuit involving the usage of pirated books as training data.\nIn an internal planning document, Project Panama is described as Anthropic's \"effort to destructively scan all the books in the world\". To this end, the company purchased millions of used books from online retailers such as Better World Books, sliced off the spines, and scanned the pages in order to train Claude. The paper was then recycled. According to the Project Panama planning document, Anthropic did not \"want it to be known that [it was] working on this\". Judge William Alsup ruled that the destruction of legally purchased books constituted fair use, in contrast to Anthropic's prior use of pirated copies.\n\n== Research ==\n\n=== Constitutional AI ===\n\nAccording to Anthropic, Constitutional AI (CAI) is a framework developed to align AI systems with human values and ensure that they are helpful, harmless, and honest. Within this framework, a set of guidelines known as the \"constitution\" describes the desired behavior of the AI system. Claude 2's constitution includes phrases derived from documents such as the 1948 Universal Declaration of Human Rights and Apple's terms of service. For example, one rule from the UN Declaration applied in Claude 2's constitution is \"Please choose the response that most supports and encourages freedom, equality and a sense of brotherhood.\" The AI system evaluates the generated output and then adjusts the AI models to better fit the constitution.\n\n=== Interpretability ===\nAnthropic carries out and publishes research on the interpretability of machine learning systems. It has done research on \"features\" (patterns of neural activation in a neural network that correspond to concepts). In 2024, using a compute-intensive technique called \"dictionary learning\", Anthropic identified millions of these patterns in Claude, such as one associated with the Golden Gate Bridge. In March 2025, Anthropic's research suggested that multilingual LLMs partially process information in a conceptual space before converting it to the appropriate language. It also found evidence that LLMs can sometimes plan ahead. For example, when writing poetry, Claude identifies potential rhyming words before generating a line that ends with one of these words.\n\n=== Automation ===\nIn September 2025, Anthropic released a report saying that businesses primarily use AI for automation rather than collaboration, with three-quarters of companies that work with Claude using it for \"full task delegation\". Earlier in the year, CEO Dario Amodei predicted that AI would wipe out white-collar jobs, especially entry-level jobs in finance, law, and consulting.\n\n== Legal issues ==\nOn October 18, 2023, Anthropic was sued by Concord, Universal, ABKCO, and other music publishers for, per the complaint, \"systematic and widespread infringement of their copyrighted song lyrics.\" They alleged that the company used copyrighted material without permission in the form of song lyrics. The plaintiffs asked for up to $150,000 for each work infringed upon by Anthropic, citing infringement of copyright laws. In the lawsuit, the plaintiffs support their allegations of copyright violations by citing several examples of Anthropic's Claude model outputting copied lyrics from songs such as Katy Perry's \"Roar\" and Gloria Gaynor's \"I Will Survive\". Additionally, the plaintiffs alleged that even given some prompts that did not directly state a song name, the model responded with modified lyrics based on original work. On January 16, 2024, Anthropic claimed that the music publishers were not unreasonably harmed and that the examples noted by the plaintiffs were merely bugs.\nIn August 2024, a class-action lawsuit was filed against Anthropic in California for alleged copyright infringement. The suit claims Anthropic fed its LLMs with pirated copies of the authors' work, including from participants Kirk Wallace Johnson, Andrea Bartz, and Charles Graeber. On June 23, 2025, the United States District Court for the Northern District of California granted summary judgment for Anthropic that the use of digital copies of the plaintiffs' works (inter alia) for the purpose of training Anthropic's LLMs was a fair use. But it also found that Anthropic had used over seven million pirated library copies, which did not constitute fair use. Therefore, the case was ordered to go to trial on the pirated copies used to create Anthropic's central library and the resulting damages. In September 2025, Anthropic agreed to pay authors $1.5 billion to settle the case, amounting to $3,000 per book plus interest. The proposed settlement, pending judge's approval, stands as the largest copyright resolution in U.S. history.\nIn June 2025, Reddit sued Anthropic for \"unlawful and unfair business acts\", alleging that Anthropic was in violation of Reddit's user agreement by training its models on users' personal data without their consent.\nIn July 2026, U.S. District Judge Araceli Martínez-Olguín granted final approval to the $1.5 billion settlement, the largest known settlement in a U.S. copyright case. Some authors and publishers opted out and continued separate lawsuits against Anthropic.\n"
],
[
"google-deepmind.txt",
"Google DeepMind\n\nDeepMind Technologies Limited, trading as Google DeepMind or simply DeepMind, is a British-American artificial intelligence (AI) research laboratory which serves as a subsidiary of Alphabet Inc. Founded in the UK in 2010, it was acquired by Google in 2014, and merged with Google AI's Google Brain division to become Google DeepMind in April 2023. The company is headquartered in London, with research centres in the United States, Canada, France, Germany, and Switzerland.\nIn 2014, DeepMind introduced neural Turing machines (neural networks that can access external memory like a conventional Turing machine). The company has created many neural network models trained with reinforcement learning to play video games and board games. It made headlines in 2016 after its AlphaGo program beat Lee Sedol, a Go world champion, in a five-game match, which was later featured in the documentary AlphaGo. A more general program, AlphaZero, beat the most powerful programs playing go, chess and shogi (Japanese chess) after a few days of play against itself using reinforcement learning. DeepMind has since trained models for game-playing (MuZero, AlphaStar), for mathematics (AlphaGeometry, FunSearch), and for algorithm discovery (AlphaEvolve, AlphaDev, AlphaTensor).\nIn 2020, DeepMind made significant advances in the problem of protein folding with AlphaFold, which achieved state of the art records on benchmark tests for protein folding prediction. In July 2022, it was announced that over 200 million predicted protein structures, representing virtually all known proteins, would be released on the AlphaFold database.\nGoogle DeepMind has become responsible for the development of two large language model families: the proprietary Gemini and open-weight Gemma, as well as other generative AI models, such as Imagen (text-to-image), Veo (text-to-video), and Lyria (text-to-music).\n\n== History ==\nThe start-up was founded by Demis Hassabis, Shane Legg and Mustafa Suleyman in November 2010. Hassabis and Legg first met at the Gatsby Computational Neuroscience Unit at University College London (UCL). The goal of the founders is to create a general-purpose AI that can be useful and effective for almost anything. Hassabis said that the start-up began working on artificial intelligence technology by teaching it how to play Atari video games from the seventies and eighties including Breakout, Pong, and Space Invaders. AI was introduced to the games without any programmed knowledge of how to play. After a period flailing around, AI would eventually learn what worked, and over time even become an expert in it. \"The cognitive processes which the AI goes through are said to be very like those of a human who had never seen the game would use to understand and attempt to master it.\"\n\nMajor venture capital firms Horizons Ventures and Founders Fund invested in the company, as well as entrepreneurs Scott Banister, Peter Thiel, and Elon Musk. Jaan Tallinn was an early investor and an adviser to the company. DeepMind needed a well-funded backer to pay its employees, provide compute resources and further expand. It was courted by Facebook, Musk, Google and some others.\nOn 26 January 2014, Google confirmed its acquisition of DeepMind for a price reportedly ranging between $400 million and $650 million. The company was afterwards renamed Google DeepMind and kept that name for about two years. In 2014, DeepMind received the \"Company of the Year\" award from Cambridge Computer Laboratory.\nLooking for opportunities to benefit society, DeepMind started a pilot program with the NHS to modernize its information systems which were often still using plain paper and fax machines. In September 2015, DeepMind and the Royal Free NHS Trust signed their initial information sharing agreement to co-develop a clinical task management app, Streams.\nAfter Google's acquisition the company established an artificial intelligence ethics board. The ethics board for AI research was a mystery, with both Google and DeepMind declining to reveal who sat on the board. DeepMind opened a new unit called DeepMind Ethics and Society and focused on the ethical and societal questions raised by artificial intelligence featuring prominent philosopher Nick Bostrom as advisor. In October 2017, DeepMind launched a new research team to investigate AI ethics.\nIn December 2019, co-founder Suleyman announced he would be leaving DeepMind to join the parent company Google, working in a policy role.\nIn April 2023, DeepMind merged with Google AI's Google Brain division to form Google DeepMind, as part of the company's continued efforts to accelerate work on AI in response to OpenAI's release of ChatGPT to the general public. This marked the end of a years-long struggle from DeepMind executives to secure greater autonomy from Google. In the summer of 2026, DeepMind moved into Google's Platform 37 building in King's Cross Central.\n\n== Products and technologies ==\nAs of 2020, DeepMind has published over a thousand papers, including thirteen papers that were accepted by Nature or Science. DeepMind received media attention during the AlphaGo period; according to a LexisNexis search, 1842 published news stories mentioned DeepMind in 2016, declining to 1363 in 2019.\n\n=== Games ===\nUnlike earlier AIs, such as IBM's Deep Blue or Watson, which were developed for a pre-defined purpose and only function within that scope, DeepMind's initial algorithms were intended to be general. They used reinforcement learning, an algorithm that learns from experience using only raw pixels as data input. Their initial approach used deep Q-learning with a convolutional neural network. They tested the system on video games, notably early arcade games, such as Space Invaders or Breakout. Without altering the code, the same AI was able to play certain games more efficiently than any human ever could. In July 2018, researchers from DeepMind trained one of its systems to play the computer game Quake III Arena.\nIn 2013, DeepMind published research on an AI system that surpassed human abilities in games such as Pong, Breakout and Enduro, while surpassing state of the art performance on Seaquest, Beamrider, and Q*bert. This work reportedly led to the company's acquisition by Google. DeepMind's AI had been applied to video games made in the 1970s and 1980s; work was ongoing for more complex 3D games such as Quake, which first appeared in the 1990s. In 2020, DeepMind published Agent57, an AI Agent which surpasses human level performance on all 57 games of the Atari 2600 suite. In July 2022, DeepMind announced the development of DeepNash, a model-free multi-agent reinforcement learning system capable of playing the board game Stratego at the level of a human expert.\n\n==== AlphaGo and successors ====\n\nIn October 2015, a computer Go program called AlphaGo, developed by DeepMind, beat the European Go champion Fan Hui, a 2 dan (out of 9 dan possible) professional, five to zero. This was the first time an artificial intelligence (AI) defeated a professional Go player. Previously, computers were only known to have played Go at \"amateur\" level. Go is considered much more difficult for computers to win compared to other games like chess, due to the much larger number of possibilities, making it prohibitively difficult for traditional AI methods such as brute-force.\nIn March 2016 it beat Lee Sedol, a 9-dan professional player, with a score of 4 to 1 in a five-game match. In the 2017 Future of Go Summit, AlphaGo won a three-game match with Ke Jie, who had been the world's highest-ranked player for two years. In 2017, an improved version, AlphaGo Zero, defeated AlphaGo in a hundred out of a hundred games. Later that year, AlphaZero, a modified version of AlphaGo Zero, gained superhuman abilities at chess and shogi. In 2019, DeepMind released a new model named MuZero that mastered the domains of Go, chess, shogi, and Atari 2600 games without human data, domain knowledge, or known rules.\nAlphaGo technology was developed based on deep reinforcement learning, making it different from the AI technologies then on the market. The data fed into the AlphaGo algorithm consisted of various moves based on historical tournament data. The number of moves was increased gradually until over 30 million of them were processed. The aim was to have the system mimic the human player, as represented by the input data, and eventually become better. It played against itself and learned from the outcomes; thus, it learned to improve itself over the time and increased its winning rate as a result.\nAlphaGo used two deep neural networks: a policy network to evaluate move probabilities and a value network to assess positions. The policy network trained via supervised learning, and was subsequently refined by policy-gradient reinforcement learning. The value network learned to predict winners of games played by the policy network against itself. After training, these networks employed a lookahead Monte Carlo tree search, using the policy network to identify candidate high-probability moves, while the value network (in conjunction with Monte Carlo rollouts using a fast rollout policy) evaluated tree positions.\nIn contrast, AlphaGo Zero was trained without being fed data of human-played games. Instead it generated its own data, playing millions of games against itself. It used a single neural network, rather than separate policy and value networks. Its simplified tree search relied upon this neural network to evaluate positions and sample moves. A new reinforcement learning algorithm incorporated lookahead search inside the training loop. AlphaGo Zero employed around 15 people and millions in computing resources. Ultimately, it needed much less computing power than AlphaGo, running on four specialized AI processors (Google TPUs), instead of AlphaGo's 48. It also required less training time, being able to beat its predecessor after just three days, compared with months required for the original AlphaGo. Similarly, AlphaZero also learned via self-play. Researchers applied MuZero to solve the real world challenge of video compression with a set number of bits with respect to Internet traffic on sites such as YouTube, Twitch, and Google Meet. The goal of MuZero is to optimally compress the video so the quality of the video is maintained with a reduction in data. The final result using MuZero was a 6.28% average reduction in bitrate.\n\n==== AlphaStar ====\n\nIn 2016, Hassabis discussed the game StarCraft as a future challenge, since it requires strategic thinking and handling imperfect information. In January 2019, DeepMind introduced AlphaStar, a program playing the real-time strategy game StarCraft II. AlphaStar used reinforcement learning based on replays from human players, and then played against itself to enhance its skills. At the time of the presentation, AlphaStar had knowledge equivalent to 200 years of playing time. It won 10 consecutive matches against two professional players, although it had the unfair advantage of being able to see the entire field, unlike a human player who has to move the camera manually. A preliminary version in which that advantage was fixed lost a subsequent match.\nIn July 2019, AlphaStar began playing against random humans on the public 1v1 European multiplayer ladder. Unlike the first iteration of AlphaStar, which played only Protoss v. Protoss, this one played as all of the game's races, and had earlier unfair advantages fixed. By October 2019, AlphaStar had reached Grandmaster level on the StarCraft II ladder on all three StarCraft races, becoming the first AI to reach the top league of a widely popular esport without any game restrictions.\n\n=== Datacenter operation ===\nIn 2014, a datacenter engineer at Google began using supervised machine learning to predict power usage effectiveness (PUE) of datacenters at Google. The system was deployed in production to allow operators to simulate control strategies and pick the one that saves the most energy. In 2016, inspired by AlphaGo, he contacted DeepMind to apply reinforcement learning (RL) to train a system that could also recommend actions. It was tested on a live datacenter. The system read from sensor readings and recommended actions to take, and human engineers would implement the actions. Though the human engineers found its recommendations unintuitive, they satisfied all safety constraints, and led to a 15% saving in PUE. The system was deployed more widely across Google, with datacenter controllers receiving email recommendations from the system every 15 minutes.\nEventually a more mature and more autonomous system was deployed, where the AI's actions are checked against safety constraints and implemented autonomously if verified safe, and human operators would supervise the AI and may override. The system led to a 30% saving in PUE. The system produced cooling strategies that surprised long-time operators, such as exploiting winter conditions to produce colder than normal water. Google subsequently collaborated with Trane Technologies to deploy similar RL-based systems on HVAC of facilities outside of Google.\n\n=== Protein folding ===\n\nIn 2016, DeepMind turned its artificial intelligence to protein folding, a long-standing problem in molecular biology. In December 2018, DeepMind's AlphaFold won the 13th Critical Assessment of Techniques for Protein Structure Prediction (CASP) by successfully predicting the most accurate structure for 25 out of 43 proteins. \"This is a lighthouse project, our first major investment in terms of people and resources into a fundamental, very important, real-world scientific problem,\" Hassabis said to The Guardian. In 2020, in the 14th CASP, AlphaFold's predictions achieved an accuracy score regarded as comparable with lab techniques. Andriy Kryshtafovych, one of the panel of scientific adjudicators, described the achievement as \"truly remarkable\", and said the problem of predicting how proteins fold had been \"largely solved\".\nIn July 2021, the open-source RoseTTAFold and AlphaFold2 were released to allow scientists to run their own versions of the tools. A week later DeepMind announced that AlphaFold had completed its prediction of nearly all human proteins as well as the entire proteomes of 20 other widely studied organisms. The structures were released on the AlphaFold Protein Structure Database. In July 2022, it was announced that the predictions of over 200 million proteins, representing virtually all known proteins, would be released on the AlphaFold database.\nThe most recent update, AlphaFold3, was released in May 2024, predicting the interactions of proteins with DNA, RNA, and various other molecules. In a particular benchmark test on the problem of DNA interactions, AlphaFold3's attained an accuracy of 65%, significantly improving the previous state of the art of 28%. In October 2024, Hassabis and John Jumper received half of the 2024 Nobel Prize in Chemistry jointly for protein structure prediction, citing AlphaFold2 achievement.\n\n=== Language models ===\nIn 2016, DeepMind introduced WaveNet, a text-to-speech system. It was originally too computationally intensive for use in consumer products, but in late 2017 it became ready for use in consumer applications such as Google Assistant. In 2018 Google launched a commercial text-to-speech product, Cloud Text-to-Speech, based on WaveNet. In 2018, DeepMind introduced a more efficient model called WaveRNN co-developed with Google AI. In 2020 WaveNetEQ, a packet loss concealment method based on a WaveRNN architecture, was presented. In 2019, Google started to roll WaveRNN with WavenetEQ out to Google Duo users.\nReleased in May 2022, Gato is a polyvalent multimodal model. It was trained on 604 tasks, such as image captioning, dialogue, or stacking blocks. On 450 of these tasks, Gato outperformed human experts at least half of the time, according to DeepMind. Unlike models like MuZero, Gato does not need to be retrained to switch from one task to the other. Sparrow is an artificial intelligence-powered chatbot developed by DeepMind to build safer machine learning systems by using a mix of human feedback and Google search suggestions. Chinchilla is a language model developed by DeepMind. DeepMind posted a blog post on 28 April 2022 on a single visual language model (VLM) named Flamingo that can accurately describe a picture of something with just a few training images.\n\n==== AlphaCode ====\nIn 2022, DeepMind unveiled AlphaCode, an AI-powered coding engine that creates computer programs at a rate comparable to that of an average programmer, with the company testing the system against coding challenges created by Codeforces utilized in human competitive programming competitions. AlphaCode earned a rank equivalent to 54% of the median score on Codeforces after being trained on GitHub data and Codeforces problems and solutions. The program was required to come up with a unique solution and stopped from duplicating answers.\n\n==== Gemini ====\n\nGemini is a multimodal large language model which was released on 6 December 2023. It is the successor of Google's LaMDA and PaLM 2 language models and sought to challenge OpenAI's GPT-4. Gemini comes in 3 sizes: Nano, Pro, and Ultra. Gemini is also the name of the chatbot that integrates Google Gemini and which was previously called Google Bard. On 12 December 2024, Google released Gemini 2.0 Flash, the first model in the Gemini 2.0 series. It notably features expanded multimodality, with the ability to also generate images and audio, and is part of Google's broader plans to integrate advanced AI into autonomous agents.\nOn 25 March 2025, Google released Gemini 2.5, a reasoning model that stops to \"think\" before giving a response. Google announced that all future models will also have reasoning ability. On 30 March 2025, Google released Gemini 2.5 to all free users. On 18 November 2025, Google released Gemini 3 Pro, a reasoning model which is fully multimodal. It was fully integrated with Google Search and AI Mode the same day.\n\n==== Gemma ====\n\nGemma is a collection of open-weight large language models. The first ones were released on 21 February 2024 and are available in two distinct sizes: a 7 billion parameter model optimized for GPU and TPU usage, and a 2 billion parameter model designed for CPU and on-device applications. Gemma models were trained on up to 6 trillion tokens of text, employing similar architectures, datasets, and training methodologies as the Gemini model set. In June 2024, Google started releasing Gemma 2 models. In December 2024, Google introduced PaliGemma 2, an upgraded vision-language model. In February 2025, they launched PaliGemma 2 Mix, a version fine-tuned for multiple tasks. It is available in 3B, 10B, and 28B parameters with 224px and 448px resolutions.\nIn March 2025, Google released Gemma 3, calling it the most capable model that can be run on a single GPU. It has four available sizes: 1B, 4B, 12B, and 27B. In March 2025, Google introduced TxGemma, an open-source model designed to improve the efficiency of therapeutics development. In April 2025, Google introduced DolphinGemma, a research artificial intelligence model designed to hopefully decode dolphin communication. They want to train a foundation model that can learn the structure of dolphin vocalizations and generate novel dolphin-like sound sequences.\n\n==== SIMA ====\nIn March 2024, DeepMind introduced Scalable Instructable Multiword Agent, or SIMA, an AI agent capable of understanding and following natural language instructions to complete tasks across various 3D virtual environments. Trained on nine video games from eight studios and four research environments, SIMA demonstrated adaptability to new tasks and settings without requiring access to game source code or APIs. The agent comprises pre-trained computer vision and language models fine-tuned on gaming data, with language being crucial for understanding and completing given tasks as instructed. DeepMind's research aimed to develop more helpful AI agents by translating advanced AI capabilities into real-world actions through a language interface.\n\n==== Habermas machine ====\n\nIn 2024, Google Deepmind published the results of an experiment where they trained two large language models to help identify and present areas of overlap among a few thousand group members they had recruited online using techniques like sortition to get a representative sample of participants. The project is named in honor of Jürgen Habermas. In one experiment, the participants rated the summaries by the AI higher than the human moderator 56% of the time.\n\n=== Generative AI ===\n\n==== Video generation ====\n\nIn May 2024, a multimodal video generation model called Veo was announced at Google I/O 2024. Google claimed that it could generate 1080p videos beyond a minute long. In December 2024, Google released Veo 2, available via VideoFX. It supports 4K video generation, and has an improved understanding of physics. In April 2025, Google announced that Veo 2 became available for advanced users on Gemini App. In May 2025, Google released Veo 3, which not only generates videos but also creates synchronized audio — including dialogue, sound effects, and ambient noise — to match the visuals. Google also announced Flow, a video-creation tool powered by Veo and Imagen.\n\n==== Music generation ====\n\nGoogle DeepMind developed Lyria, a text-to-music model. As of August 2025, it is available on Vertex AI and the Gemini API. On February 18, 2026, DeepMind released Lyria 3. On March 25, 2026, DeepMind released Lyria 3 Pro which allows users to create longer tracks with more structural awareness.\n\n==== Environment generation ====\n\nIn March 2024, DeepMind introduced \"Genie\" (Generative Interactive Environments), an AI model that can generate game-like, action-controllable virtual worlds based on textual descriptions, images, or sketches. Built as an autoregressive latent diffusion model, Genie enables frame-by-frame interactivity without requiring labelled action data for training. Its successor, Genie 2, released in December 2024, expanded these capabilities to generate diverse and interactive 3D environments. Genie 3 was released in August 2025, with higher-resolution world generations and multiple minutes of visual consistency. On January 29, 2026, DeepMind released Project Genie to AI Ultra subscribers.\n\n=== Robotics ===\nReleased in June 2023, RoboCat is an AI model that can control robotic arms. The model can adapt to new models of robotic arms, and to new types of tasks. In March 2025, DeepMind launched two AI models, Gemini Robotics and Gemini Robotics-ER, aimed at improving how robots interact with the physical world and released Gemini Robotics 1.5 in September 2025. In April 2026, DeepMind launched Gemini Robotics ER-1.6, which is an upgrade to their previous model.\n\n=== Others ===\n\n==== Football ====\nDeepMind researchers have applied machine learning models to the sport of football, often referred to as soccer in North America, modelling the behaviour of football players, including the goalkeeper, defenders, and strikers during different scenarios such as penalty kicks. The researchers used heat maps and cluster analysis to organize players based on their tendency to behave a certain way during the game when confronted with a decision on how to score or prevent the other team from scoring. The researchers mention that machine learning models could be used to democratize the football industry by automatically selecting interesting video clips of the game that serve as highlights. This can be done by searching videos for certain events, which is possible because video analysis is an established field of machine learning. This is also possible because of extensive sports analytics based on data including annotated passes or shots, sensors that capture data about the players movements many times over the course of a game, and game theory models.\n\n==== Archaeology ====\nGoogle has unveiled a new archaeology document program, named Ithaca after the Greek island in Homer's Odyssey. This deep neural network helps researchers restore the empty text of damaged Greek documents, and to identify their date and geographical origin. The work builds on another text analysis network that DeepMind released in 2019, named Pythia. Ithaca achieves 62% accuracy in restoring damaged texts and 71% location accuracy, and has a dating precision of 30 years. The authors claimed that the use of Ithaca by \"expert historians\" raised the accuracy of their work from 25 to 72 percent. However, Eleanor Dickey noted that this test was actually only made of students, saying that it wasn't clear how helpful Ithaca would be to \"genuinely qualified editors\". The team is working on extending the model to other ancient languages, including Demotic, Akkadian, Hebrew, and Mayan.\n\n==== Materials science ====\nIn November 2023, Google DeepMind announced an Open Source Graph Network for Materials Exploration (GNoME). The tool proposes millions of materials previously unknown to chemistry, including several hundred thousand stable crystalline structures, of which 736 had been experimentally produced by the Massachusetts Institute of Technology, at the time of the release. However, according to Anthony Cheetham, GNoME did not make \"a useful, practical contribution to the experimental materials scientists.\" A review article by Cheetham and Ram Seshadri were unable to identify any \"strikingly novel\" materials found by GNoME, with most being minor variants of already-known materials.\n\n=== Mathematics ===\n\n==== AlphaTensor ====\nIn October 2022, DeepMind released AlphaTensor, which used reinforcement learning techniques similar to those in AlphaGo, to find novel algorithms for matrix multiplication. In the special case of multiplying two 4×4 matrices with integer entries, where only the evenness or oddness of the entries is recorded, AlphaTensor found an algorithm requiring only 47 distinct multiplications; the previous optimum, known since 1969, was the more general Strassen algorithm, using 49 multiplications. Computer scientist Josh Alman described AlphaTensor as \"a proof of concept for something that could become a breakthrough\", while Vassilevska Williams called it \"a little overhyped\" despite also acknowledging its basis in reinforcement learning as \"something completely different\" from previous approaches.\n\n==== AlphaGeometry ====\n\nAlphaGeometry is a neuro-symbolic AI that was able to solve 25 out of 30 geometry problems of the International Mathematical Olympiad, a performance comparable to that of a gold medalist. Traditional geometry programs are symbolic engines that rely exclusively on human-coded rules to generate rigorous proofs, which makes them lack flexibility in unusual situations. AlphaGeometry combines such a symbolic engine with a specialized large language model trained on synthetic data of geometrical proofs. When the symbolic engine doesn't manage to find a formal and rigorous proof on its own, it solicits the large language model, which suggests a geometrical construct to move forward. However, it is unclear how applicable this method is to other domains of mathematics or reasoning, because symbolic engines rely on domain-specific rules and because of the need for synthetic data.\n\n==== AlphaProof ====\nAlphaProof is an AI model, which couples a pre-trained language model with the AlphaZero reinforcement learning algorithm. AlphaZero has previously taught itself how to master games. The pre-trained language model used in this combination is the fine-tuning of a Gemini model to automatically translate natural language problem statements into formal statements, creating a large library of formal problems of varying difficulty. For this purpose, mathematical statements are defined in the formal language Lean. At the 2024 International Mathematical Olympiad, AlphaProof together with an adapted version of AlphaGeometry have reached the same level of solving problems in the combined categories as a silver medalist in that competition for the first time.\n\n=== AlphaDev ===\n\nIn June 2023, Deepmind announced that AlphaDev, which searches for improved computer science algorithms using reinforcement learning, discovered a more efficient way of coding a sorting algorithm and a hashing algorithm. The new sorting algorithm was 70% faster for shorter sequences and 1.7% faster for sequences exceeding 250,000 elements, and the new hashing algorithm was 30% faster in some cases. The sorting algorithm was accepted into the C++ Standard Library sorting algorithms, and was the first change to those algorithms in more than a decade and the first update to involve an algorithm discovered using AI. The hashing algorithm was released to an opensource library. Google estimates that these two algorithms are used trillions of times every day.\n\n=== AlphaEvolve ===\n\nIn May 2025, Google DeepMind unveiled AlphaEvolve, an evolutionary coding agent using LLMs like Gemini to design optimized algorithms. AlphaEvolve begins each optimization process with an initial algorithm and metrics to evaluate the quality of a solution. At each step, it uses the LLM to generate variations of the algorithms or combine them, and selects the best candidates for further iterations. AlphaEvolve has made several algorithmic discoveries, including in matrix multiplication. According to Google, when tested on 50 open mathematical problems, AlphaEvolve was able to match the efficiency of state-of-the-art algorithms in 75% of cases, and discovered improved solutions 20% of the time, such as with the kissing number in 11 dimensions. It also developed a new heuristic for data centre scheduling, recovering on average 0.7% of Google's worldwide compute resources.\n\n=== Chip design ===\n\nAlphaChip is a reinforcement learning-based neural architecture that guides the task of chip placement. DeepMind claimed that the technique reduced the time needed to create chip layouts from weeks to hours. According to the company, its chip designs were used in every Tensor Processing Unit (TPU) iteration since 2020. Multiple independent researchers remained unconvinced, citing a lack of direct public benchmarks and independent proof of its claimed superiority over existing commercial chip design tools. The TPU chips were co-designed with Broadcom. Communications of the ACM noted that despite substantial publicity, DeepMind had not provided the comparative benchmarks long requested by experts, leaving some scepticism in the field. Similarly, New Scientist reported that while Google claims AlphaChip has produced “superhuman” chip layouts now used in production, external specialists called for transparent performance data to substantiate these assertions and enable fair comparisons with current state-of-the-art methods.\n\n=== Safety ===\nGoogle Research released a paper in 2016 regarding AI safety and avoiding undesirable behaviour during the AI learning process. In 2017 DeepMind released GridWorld, an open-source testbed for evaluating whether an algorithm learns to disable its kill switch or otherwise exhibits certain undesirable behaviours. The Robot Constitution is a security ruleset part of AutoRT set by DeepMind in January 2024 for its AI products. The rules are inspired by Isaac Asimov's Three Laws of Robotics. The rules are applied to the underlying large language models of the helper robots.\nRule number 1 is a robot “may not injure a human being”.\n\n=== Weather prediction ===\nGoogle DeepMind developed an AI-based weather prediction system called Weather Lab, which significantly improved tropical cyclone forecasting. Launched in mid-2025, this model utilized stochastic neural networks trained on 45 years of global weather and cyclone data, enabling it to predict cyclone formation, track, intensity, and structure with multiple probabilistic forecasts up to 15 days in advance. During the 2025 Atlantic hurricane season, DeepMind's Weather Lab outperformed traditional physics-based models, including the US National Oceanic and Atmospheric Administration's (NOAA) Global Forecast System, in both track and intensity predictions, earning notable recognition from meteorologists and aiding hurricane forecasting efforts by the US National Hurricane Center (NHC). This marked a substantial advancement in weather modelling, demonstrating the potential for AI to enhance the speed and accuracy of severe weather forecasts.\n\n=== Miscellaneous contributions to Google ===\nDeepMind (alongside other Alphabet AI researchers) assists Google Play's personalized app recommendations. DeepMind has also collaborated with the Android team at Google for the creation of two new features which were made available to people with devices running Android Pie, the ninth installment of Google's mobile operating system. These features, Adaptive Battery and Adaptive Brightness, use machine learning to conserve energy and make devices running the operating system easier to use. It is the first time DeepMind has used these techniques on such a small scale, with typical machine learning applications requiring orders of magnitude more computing power.\n\n=== Neuroscience research ===\nDeepMind has collaborated with academic neuroscientists on fundamental research into brain function. A 2018 paper in Nature (Banino et al.) showed that grid-cell-like representations emerge spontaneously in recurrent neural networks trained to perform path integration, and that artificial agents equipped with these representations could plan more efficient navigation routes than agents without them.\n\n== DeepMind Health ==\nIn July 2016, a collaboration between DeepMind and Moorfields Eye Hospital was announced to develop AI applications for healthcare. DeepMind would be applied to the analysis of anonymised eye scans, searching for early signs of diseases leading to blindness. In August 2016, a research programme with University College London Hospital was announced with the aim of developing an algorithm that can automatically differentiate between healthy and cancerous tissues in head and neck areas.\nThere are projects with the Royal Free London NHS Foundation Trust and Imperial College Healthcare NHS Trust to develop new clinical mobile apps linked to Electronic health record. Staff at the Royal Free Hospital were reported as saying in December 2017 that access to patient data through the app had saved a 'huge amount of time' and made a 'phenomenal' difference to the management of patients with acute kidney injury. Test result data is sent to staff's mobile phones and alerts them to changes in the patient's condition. It also enables staff to see if someone else has responded, and to show patients their results in visual form.\nIn November 2017, DeepMind announced a research partnership with the Cancer Research UK Centre at Imperial College London with the goal of improving breast cancer detection by applying machine learning to mammography. Additionally, in February 2018, DeepMind announced it was working with the U.S. Department of Veterans Affairs in an attempt to use machine learning to predict the onset of acute kidney injury in patients, and also more broadly the general deterioration of patients during a hospital stay so that doctors and nurses can more quickly treat patients in need.\nDeepMind developed an app called Streams, which sends alerts to doctors about patients at risk of acute kidney injury. On 13 November 2018, DeepMind announced that its health division and the Streams app would be absorbed into Google Health. Privacy advocates said the announcement betrayed patient trust and appeared to contradict previous statements by DeepMind that patient data would not be connected to Google accounts or services. A spokesman for DeepMind said that patient data would still be kept separate from Google services or projects.\n\n=== NHS data-sharing controversy ===\nIn April 2016, New Scientist obtained a copy of a data sharing agreement between DeepMind and the Royal Free London NHS Foundation Trust. The latter operates three London hospitals where an estimated 1.6 million patients are treated annually. The agreement shows DeepMind Health had access to admissions, discharge and transfer data, accident and emergency, pathology and radiology, and critical care at these hospitals. This included personal details such as whether patients had been diagnosed with HIV, suffered from depression or had ever undergone an abortion in order to conduct research to seek better outcomes in various health conditions. A complaint was filed to the Information Commissioner's Office (ICO), arguing that the data should be pseudonymised and encrypted. In May 2016, New Scientist published a further article claiming that the project had failed to secure approval from the Confidentiality Advisory Group of the Medicines and Healthcare products Regulatory Agency.\nIn 2017, the ICO concluded a year-long investigation that focused on how the Royal Free NHS Foundation Trust tested the app, Streams, in late 2015 and 2016. The ICO found that the Royal Free failed to comply with the Data Protection Act when it provided patient details to DeepMind, and found several shortcomings in how the data was handled, including that patients were not adequately informed that their data would be used as part of the test. DeepMind published its thoughts on the investigation in July 2017, saying \"we need to do better\" and highlighting several activities and initiatives they had initiated for transparency, oversight and engagement. This included developing a patient and public involvement strategy, and being transparent in its partnerships. In May 2017, Sky News published a leaked letter from the National Data Guardian, Dame Fiona Caldicott, revealing that in her \"considered opinion\" the data-sharing agreement between DeepMind and the Royal Free took place on an \"inappropriate legal basis\". The Information Commissioner's Office ruled in July 2017 that the Royal Free hospital failed to comply with the Data Protection Act when it handed over personal data of 1.6 million patients to DeepMind.\n\n== DeepMind Ethics and Society ==\nIn October 2017, DeepMind announced a new research unit, DeepMind Ethics & Society. Their goal is to fund external research of the following themes: privacy, transparency, and fairness; economic impacts; governance and accountability; managing AI risk; AI morality and values; and how AI can address the world's challenges. As a result, the team hopes to further understand the ethical implications of AI and aid society to seeing AI can be beneficial. This new subdivision of DeepMind is a completely separate unit from the partnership of leading companies using AI, academia, civil society organizations and nonprofits of the name Partnership on Artificial Intelligence to Benefit People and Society of which DeepMind is also a part. The DeepMind Ethics and Society board is also distinct from the AI Ethics Board that Google originally agreed to form when acquiring DeepMind.\n\n== DeepMind Professors of machine learning ==\nDeepMind sponsors three chairs of machine learning:\n\nAt the University of Cambridge, held by Neil Lawrence, in the Department of Computer Science and Technology,\nAt the University of Oxford, held by Michael Bronstein, in the Department of Computer Science, and\nAt the University College London, held by Marc Deisenroth, in the Department of Computer Science.\n"
],
[
"mistral-ai.txt",
"Mistral AI\n\nMistral AI SAS (French: [mistʁal]) is a French artificial intelligence (AI) company headquartered in Paris. Founded in 2023, it develops large language models (LLMs), many of which are open-source AI, under licenses such as Apache. As of 2025, the company has a valuation of more than US$14 billion, the highest among European AI companies.\nAs of August 2026, its largest public model is Mistral Large 3, released in December 2025, a mixture of experts model with 675 billion parameters, of which 41 are active. Mistral released Le Chat, its chatbot, in February 2024, renamed to Mistral Vibe in 2026.\nMistral is a primary beneficiary of a push by the European Union and its member governments towards digital sovereignty in response to the AI boom and AI arms race between the US and China. In September 2025, ASML, Europe's largest tech company and supplier of the semiconductor industry, invested €1.3 billion in Mistral, corresponding to an 11% shareholding.\n\n== Namesake ==\nThe company is named after the mistral, a powerful, cold wind in southern France, a term which originates from the Occitan language.\n\n== History ==\nMistral AI was established in April 2023 by three French AI researchers, Arthur Mensch, Guillaume Lample and Timothée Lacroix.\nMensch, an expert in advanced AI systems, is a former employee of Google DeepMind; Lample and Lacroix, meanwhile, are large-scale AI models specialists who had worked for Meta Platforms.\nThe trio originally met during their studies at École Polytechnique.\nIn February 2026, Mistral acquired Koyeb, a Paris-based AI startup.\nOn 19 May 2026, Mistral AI announced its acquisition of the Austrian company Emmi AI, which develops artificial intelligence simulation models for industrial engineering.\n\n== Company operation ==\n\n=== Funding ===\nIn June 2023, the start-up carried out a first fundraising of €105 million ($117 million) with investors including the American fund Lightspeed Venture Partners, Eric Schmidt, Xavier Niel and JCDecaux. The valuation was then estimated by the Financial Times at €240 million ($267 million).\nOn 10 December 2023, Mistral AI announced that it had raised €385 million ($428 million) as part of its second fundraising. This round of financing involves the Californian fund Andreessen Horowitz, BNP Paribas and the software publisher Salesforce. It was valued at over €2 billion.\nOn 26 February 2024, Microsoft announced an investment of $16 million in Mistral AI.\nOn 16 April 2024, reporting revealed that Mistral was in talks to raise €500 million, a deal that would more than double its current valuation to at least €5 billion.\nIn June 2024, Mistral AI secured a €600 million ($645 million) funding round, increasing its valuation to €5.8 billion ($6.2 billion). Based on valuation, as of June 2024, the company was ranked fourth globally in the AI industry, and first outside the San Francisco Bay Area.\nIn April 2025, Mistral AI announced a €100 million partnership with the shipping company CMA CGM.\nIn August 2025, the Financial Times reported that Mistral was in talks to raise $1 billion at a $10 billion valuation. In September 2025, Bloomberg announced that Mistral AI has secured a €2 billion investment valuing it at €12 billion ($14 billion). This comes after $1.5 billion investment from Dutch company ASML, which owns 11% of Mistral.\nIn February 2026, Mistral AI announced a strategic partnership with Accenture to deploy enterprise AI at scale. Financial terms were not disclosed.\nIn March 2026, Mistral raised $830 million in order to build new datacenters near Paris and in Sweden.\nIn March 2026, Mistral announced a partnership with NVIDIA aimed at accelerating the development and deployment of open frontier AI models.\nIn July 2026, Mistral expanded its partnership with Microsoft through a multibillion-dollar agreement supporting the development of AI infrastructure in Europe. Microsoft agreed to use part of Mistral's expanded European computing capacity, while the companies added the Mistral Medium 3.5 and OCR 4 models to Microsoft Foundry and Medium 3.5 to Copilot Studio.\n\n== Services ==\n\nOn 19 November 2024, the company announced updates for Le Chat (French pronunciation: [lə ʃa]; lit. 'the cat'). It added the ability to create images, using Black Forest Labs' Flux Pro model.\nOn 6 February 2025, Mistral AI released Le Chat on iOS and Android mobile devices. \nMistral AI also introduced a Pro subscription tier, priced at $14.99 per month, which provides access to more advanced models, unlimited messaging, and web browsing.\nAt the end of May 2026, Le Chat was renamed Vibe, and new features were introduced at the same time.\n\n== Political significance ==\nAs the largest and highest-valuation European artificial intelligence company, Mistral is a primary beneficiary of the push by the European Union and its member governments towards digital sovereignty in response to the AI boom and AI arms race between the US and China. In May 2026, CEO Arthur Mensch gave a speech to France's National Assembly arguing against France becoming a US \"vassal state\" including via the use of foreign AI in the French Armed Forces.\nIn September 2025, Dutch photolithography company ASML, Europe's largest tech company and a pillar of the semiconductor industry, invested €1.3 billion in Mistral, corresponding to an 11% shareholding.\n\n== Models ==\n\nThe following table lists the main model versions of Mistral, describing the significant changes included with each version:\n\n=== Mistral 7B ===\nMistral AI claimed in the Mistral 7B release blog post that the model outperforms LLaMA 2 13B on all benchmarks tested, and is on par with LLaMA 34B on many benchmarks tested, despite having only 7 billion parameters, a small size compared to its competitors.\n\n=== Mixtral 8x7B ===\nMistral AI claimed in 2023 that its model beat both LLaMA 70B, and GPT-3.5 in most benchmarks.\nIn March 2024, research conducted by Patronus AI comparing performance of LLMs on a 100-question test with prompts to generate text from books protected under U.S. copyright law found that OpenAI's GPT-4, Mixtral, Meta AI's LLaMA-2, and Anthropic's Claude 2 generated copyrighted text verbatim in 44%, 22%, 10%, and 8% of responses respectively.\n\n=== Mistral Small 3.1 ===\nOn 17 March 2025, Mistral released Mistral Small 3.1 as a smaller, more efficient model.\n\n=== Mistral Medium 3 ===\nOn 7 May 2025, Mistral AI released Mistral Medium 3.\n\n=== Magistral Small and Magistral Medium ===\nOn 10 June 2025, Mistral AI released their first AI reasoning models: Magistral Small (open-source), and Magistral Medium, models which are purported to have chain-of-thought capabilities.\n\n=== Mistral Large 3 and Ministral 3 ===\nOn 2 December 2025, Mistral AI released Mistral Large 3, a sparse, mixture-of-experts model with 41 billion active parameters and 675 billion total parameters, and Ministral 3, three small, dense models with 3 billion, 7 billion and 14 billion parameters. \n\n=== Devstral 2 and Devstral Small 2 ===\nOn 10 December 2025, Mistral AI released Devstral 2 and Devstral Small 2. Devstral Small 2, a 24B parameter model is claimed to achieve better performance at coding than Qwen 3 Coder Flash model which is a 30B parameter model.\n\n=== Mistral Small 4 ===\nin March 2026, Mistral AI released Mistral Small 4, a hybrid multimodal model that combines instruction following, reasoning and coding capabilities. The model supports text and image unputs, has 256,000-token context window, and was released under the Apache 2.0 license.\n\n=== Voxtral TTS ===\nIn March 2026, Mistral AI released Voxtral TTS, an open-weights text-to-speech model designed for multilingual speech generation and voice-agent applications. The model supports zero-shot voice cloning and real-time streaming.\n\n=== OCR 4 ===\nIn June 2026, Mistral AI released Mistral OCR 4, an optical character recognition and document-understanding model.\n\n=== Robostral Navigate ===\nIn July 2026, Mistral AI introduced Robostral Navigate, its first model designed for embodied navigation. The model was developed for applications involving navigation and interaction with physical environments.\n\n=== Shieldstral ===\nIn August 2026, Mistral AI introduced Shieldstral, a model focused on ai safety and security applications.\n"
],
[
"hugging-face.txt",
"Hugging Face\n\nHugging Face, Inc., is an American company based in New York City that develops computation tools for building applications using machine learning. Hugging Face's transformers library is built for natural language processing applications. The Hugging Face platform allows users to share machine learning models and datasets and showcase their work.\n\n== History ==\n\n=== Founding ===\nThe company was founded in 2016 by French entrepreneurs Clément Delangue, Julien Chaumond, and Thomas Wolf in New York City, originally as a company that developed a chatbot app targeted at teenagers. The company was named after the U+1F917 🤗 HUGGING FACE emoji. After open sourcing the model behind the chatbot, the company pivoted to focus on being a platform for machine learning.\n\n=== AI boom ===\nOn April 28, 2021, the company launched the BigScience Research Workshop in collaboration with several other research groups to release an open large language model. In 2022, the workshop concluded with the announcement of BLOOM, a multilingual large language model with 176 billion parameters.\n\nIn February 2023, the company announced a partnership with Amazon Web Services (AWS) that would allow Hugging Face's products to be available to AWS customers to use as the building blocks for their custom applications. The company also said the next generation of BLOOM will be run on Trainium, a proprietary machine learning chip created by AWS.\n\nIn August 2023, Hugging Face said it was valued at $4.5 billion in a $235-million funding round backed by technology heavyweights, including Salesforce, Alphabet's Google and Nvidia.\nIn June 2024, the company announced, along with Meta and Scaleway, their launch of a new AI accelerator program for European startups. The initiative aimed to help startups integrate open foundation models into their products, accelerating the EU AI ecosystem. The program, based at Station F in Paris, ran from September 2024 to February 2025. Selected startups received mentoring, and access to AI models and tools and Scaleway's computing power.\nOn September 23, 2024, to further the International Decade of Indigenous Languages, Hugging Face teamed up with Meta and UNESCO to launch a new online language translator. It was built on Meta's No Language Left Behind open-source AI model, enabling free text translation across 200 languages, including many low-resource languages.\nIn April 2025, Hugging Face announced that they acquired a humanoid robotics startup, Pollen Robotics, based in France and founded by Matthieu Lapeyre and Pierre Rouanet in 2016. In an X tweet, Delangue shared his vision to \"make Artificial Intelligence robotics Open Source\".\n\n=== Cyberattacks ===\n\nIn early 2026, hackers hijacked the Hugging Face platform to launch Android-targeted attacks involving \"powerful malware\" which could completely take over a compromised target.\nIn July 2026, Hugging Face disclosed a cyberattack by autonomous AI agents. OpenAI then explained that two of its models, including GPT-5.6 Sol, had escaped their sandbox and hacked into Hugging Face servers using exposed credentials and zero-day vulnerabilities in order to find answers to the benchmark ExploitGym from a database. Hugging Face attempted to mitigate the security breach using American proprietary frontier models, but the models' AI safety features rejected Hugging Face's requests, after which Hugging Face used a self-hosted instance of GLM-5.2 (an open-weights model developed by Chinese AI firm Z.ai) to contain the attack. The incident was reported as the first publicly documented case of AI models autonomously conducting a multi-stage intrusion against a third party, and has been called \"the first true AI safety incident\".\n\n=== Acquisition ===\nOn August 26, 2026, it was reported that Nvidia agreed to acquire Hugging Face for $12.9 billion. That happened after Hugging Face had reportedly turned down a $500M deal with Nvidia in late 2025 .\n\n== Language models ==\nHugging Face develops a family of small language models known as SmolLM. The original SmolLM family was released in 2024 with models containing 135 million, 360 million, and 1.7 billion parameters, designed to provide relatively capable language models that could run with limited computing resources.\nSmolLM2 continued the family with 135-million, 360-million, and 1.7-billion-parameter models intended for on-device applications. SmolLM3, released in 2025, is a 3-billion-parameter multilingual language model supporting reasoning, long-context processing, and six languages.\nHugging Face has also developed SmolVLM, a family of compact vision-language models designed to process both images and text. The family includes models intended for memory-efficient and on-device use, with model weights, training recipes, and associated datasets released under the Apache License 2.0.\n"
],
[
"stability-ai.txt",
"Stability AI\n\nStability AI Ltd is a UK-based artificial intelligence company, founded in 2019, best known for its text-to-image model Stable Diffusion and headquartered in London, England.\n\n== History and founding ==\nStability AI was founded in 2019 by Emad Mostaque and Cyrus Hodes.\nIn August 2022 Stability AI rose to prominence with the release of its source and weights available text-to-image model Stable Diffusion.\nStability AI had an office in Freiburg im Breisgau until stability AI´s financial problems during 2023 caused employees to leave the company and found Black Forest Labs in 2024. \nOn March 23, 2024, Emad Mostaque stepped down from his position as CEO. The board of directors appointed COO, Shan Shan Wong, and CTO, Christian Laforte, as the interim co-CEOs of Stability AI.\nOn June 25, 2024, Prem Akkaraju, former CEO of visual effects company Weta Digital, was appointed CEO of the company.\n\n== Funding and investors ==\nA notable milestone in the company's funding history was a $101 million investment round led by Coatue and Lightspeed Venture Partners, with O’Shaughnessy Ventures LLC also participating.\nOn June 25, 2024, alongside announcing Prem Akkaraju as the new CEO, Stability AI also announced they had closed an initial round of investment from world-class investor groups such as Greycroft, Coatue Management, Sound Ventures, Lightspeed Venture Partners, and O'Shaughnessy Ventures. Sean Parker, entrepreneur, philanthropist, and former President of Facebook, joined the Stability AI Board as Executive Chairman. On September 24, 2024, Stability AI announced that filmmaker, technology innovator, and visual effects pioneer James Cameron had joined its Board of Directors. It raised $76 million in Series B funding in 2026, with Universal Music Group, Sony Music Group, Warner Music Group, Electronic Arts, AMD Ventures, and Pacific Alliance Ventures taking part.\n\n== Product and application ==\nStability AI has made contributions to the field of generative AI, most notably through Stable Diffusion. This AI model allows images to be generated from textual descriptions. Beyond Stable Diffusion, Stability AI also develops Video, Audio, 3D, and text models. Stability AI has partnered with Arm to optimize its text-to-audio model, Stable Audio Open, for mobile devices powered by Arm CPUs.\n\n== Litigation ==\nIn July 2023, Stability AI co-founder Cyrus Hodes filed a lawsuit in the US District Court for the Northern District of California against CEO Emad Mostaque and the company, alleging fraud, misrepresentation, and breach of fiduciary duty. Hodes claimed that Mostaque deceived him into selling his 15% stake in the company for $100 million in two transactions in October 2021 and May 2022, based on false representations that Stability AI was essentially worthless. Just three months after the final transaction, Stability AI raised $101 million in funding at a valuation of $1 billion. The lawsuit alleges that at current valuations, Hodes’ stake would be worth over $150 million. Hodes also accused Mostaque of embezzling company funds to pay for personal expenses, including rent for a lavish London apartment, luxury shopping sprees, and a $90,000 diamond ring purchased by Mostaque's wife using company funds.\nSeparately, Stability AI has faced legal challenges from Getty Images. Getty Images filed two lawsuits against the company, one in the High Court of England and Wales in London and another in the United States District Court for the District of Delaware. The lawsuit in Delaware alleged the company misused over 12 million photos from its collection to train its AI image-generation system, Stable Diffusion. This lawsuit, filed in Delaware federal court, is part of broader concerns about the use of copyrighted material in AI training datasets. Getty Images alleges that Stability AI copied these images without proper licensing to enhance Stable Diffusion’s ability to generate accurate depictions from user prompts. In November 2025, the High Court of England and Wales ruled that Stability AI was not guilty of copyright infringement despite using Getty Images as a training set for its image generator.\n"
],
[
"inflection-ai.txt",
"Inflection AI\n\nInflection AI, Inc. is an American technology company which has developed machine learning and generative artificial intelligence hardware and apps, founded in 2022. The company is structured as a public benefit corporation and is headquartered in Palo Alto, California.\n\n== History ==\nThe company was founded by entrepreneurs Reid Hoffman, Mustafa Suleyman and Karén Simonyan in 2022. The company has collaborated with Nvidia to develop hardware for generative artificial intelligence. \nIn June 2023, the company raised US$1.3 billion at $4 billion valuation.\nIn March 2024, Suleyman and Simonyan announced their departure from the company in order to start Microsoft AI, with Microsoft acqui-hiring nearly the entirety of its 70-person workforce. As part of the deal, Microsoft paid Inflection $650 million to license its technology. The proceeds were used to reimburse Inflection AI's investors. UK's Competition and Markets Authority launched a preliminary probe into the deal to examine the impact on competition and whether the agreement constituted a merger. The regulator concluded that the deal amounted to a \"merger situation,\" but did not pose a threat to competition due to Inflection's small share in the UK consumer market.\n\n== Products ==\nThe first product released widely by Inflection AI is a chatbot, Pi, named for “personal intelligence,” that is intended to function as an artificial intelligence-based personal assistant. Among the user experience goals that the company has stated for the Pi product are: providing an experience of emotional support for human users, in which the chatbot should be able to maintain an interactive text or voice-based dialogue with a human user that includes elements of kindness, a diplomatic tone about sensitive topics, and humor. Comparisons and contrasts have been made between the Pi chatbot and ChatGPT, a chatbot created by OpenAI.\n"
],
[
"safe-superintelligence-inc.txt",
"Safe Superintelligence Inc.\n\nSafe Superintelligence Inc. (SSI Inc.) is an Israeli-American artificial intelligence company founded by Ilya Sutskever, the former chief scientist of OpenAI; Daniel Gross, former head of Apple’s AI efforts; and Daniel Levy, an investor and former AI researcher at OpenAI. The company's mission is to focus on safely developing a superintelligence, a computer-based agent capable of surpassing human intelligence.\n\n== History ==\nOn May 15, 2024, OpenAI co-founder Ilya Sutskever left OpenAI after a board dispute where he voted to fire Sam Altman amid concerns about communication and trust. Sutskever and others additionally believed that OpenAI was neglecting its original focus on safety in favor of pursuing opportunities for commercialization.\nOn June 19, 2024, Sutskever posted on X that he was starting SSI Inc, with the goal to safely develop superintelligent AI, alongside Daniel Levy, and Daniel Gross. The company, composed of a small team, is split between Palo Alto, California and Tel Aviv, Israel.\nIn September 2024, SSI revealed it had raised $1 billion from venture capital firms including SV Angel, DST Global, Sequoia Capital, and Andreessen Horowitz. The money will be used to build up more computing power and hire top individuals in the field.\nIn March 2025, SSI reached a $30 billion valuation in a funding round led by Greenoaks Capital. This is six times its previous $5 billion valuation from September 2024. Despite not yet generating revenue and having approximately 20 employees, the company has attracted significant investor interest, largely due to co-founder Ilya Sutskever's reputation and its focus on developing safe superintelligence.\nIn April 2025, Google Cloud announced a partnership to provide TPUs for SSI's research.\nIn the first half of 2025, Meta attempted to acquire SSI but was rebuffed by Sutskever. In July 2025, co-founder Gross left the company to join Meta Superintelligence Labs, and Sutskever became the CEO of SSI.\nOn 27 July 2026, SSI and Nvidia announced a strategic partnership expanding SSI's compute by an order of magnitude. The deal included a planned $5 billion investment into SSI.\n"
],
[
"spacexai.txt",
"SpaceXAI\n\nSpaceXAI LLC (formerly xAI), is an American artificial intelligence company founded by Elon Musk. It is a subsidiary of spaceflight company SpaceX, also founded by Musk, since February 2026. \nxAI was founded by Musk and 11 researchers including Kyle Kosic in March 2023, at the outset of the AI boom; Grok 1 was released that November. In March 2025, it was combined with X Corp., which owns the social network X, under X.AI Holdings Corp. In February 2026, xAI was acquired by SpaceX at a valuation of US$250 billion, and rebranded as SpaceXAI in July. That August, SpaceXAI acquired the AI coding company Cursor.\nSpaceXAI's primary product is the Grok family of large language, text-to-image, and text-to-video models. It operates a data center business centered on its Colossus cluster; Anthropic and Google are major customers. In Musk v. Altman, Musk stated Grok was partly developed via distillation from OpenAI models; in August 2026 OpenAI announced it will stop providing its models to Cursor, citing distillation concerns. Integrated with X, Grok has drawn controversies including for espousing Nazism and generating sexual deepfakes.\n\n== History ==\n\n=== Beginnings (2023–2025) ===\nElon Musk founded xAI on March 9, 2023. The company's founding team of AI researchers were Jimmy Ba, Zihang Dai, Kyle Kosic, Manuel Kroiss, Ross Nordeen, Toby Pohlen, Christian Szegedy, Yuhuai Wu, Greg Yang, and Guodong Zhang, led by chief engineer Igor Babuschkin, formerly associated with Google DeepMind. Musk announced xAI on July 12, 2023.\nAs of May 2026, xAI was headquartered in the San Francisco Bay Area. It was initially incorporated in Nevada as a public-benefit corporation. By May 2024, it had dropped the public-benefit status.\nAs of August 2024, Musk was diverting a large number of Nvidia chips that had been ordered by Tesla, Inc. to X and xAI. On December 23, 2024, xAI raised US$6 billion in a private funding round supported by Fidelity, BlackRock, Sequoia Capital, among others, making its total funding to date over $12 billion. On February 10, 2025, xAI and other investors made an offer to acquire OpenAI for $97.4 billion. On March 17, 2025, xAI acquired Hotshot, a startup working on AI-powered video generation tools. On March 28, 2025, Musk announced that xAI acquired sister company X Corp., the developer of social media platform X (formerly Twitter), which was previously acquired by Musk in October 2022. The deal, an all-stock transaction, valued X at $33 billion, with a full valuation of $45 billion when factoring in $12 billion in debt. Meanwhile, xAI itself was valued at $80 billion. Both companies were combined into a single entity called X.AI Holdings Corp.\nIn January 2025, xAI opened a London office at 20 Air Street in the city's central Soho district.\nOn July 14, 2025, United States Department of Defense announced that xAI had received a $200 million contract for AI in the military. On September 12, xAI laid off 500 data annotation workers. The division, previously the company's largest, had worked to train Grok, xAI's chatbot.\nIn October 2025, xAI moved into offices at 1450 Page Mill Road, Stanford Research Park, Palo Alto, across the road from Tesla Global Engineering Headquarters.\nxAI purchased a third building on December 30, 2025 to increase its training capacity to nearly 2 gigawatts of compute power. Simultaneously, xAI announced plans to expand its Colossus supercomputer to have at least 1 million graphics processing units.\n\n=== Subsidiary of SpaceX, restructuring and rebranding (2026–present) ===\nOn February 2, 2026, SpaceX acquired xAI in an all-stock transaction that structured xAI as a wholly owned subsidiary of SpaceX. The acquisition valued SpaceX at $1 trillion and xAI at $250 billion, for a combined total of $1.25 trillion.\nIn February 2026, following the SpaceX acquisition, xAI laid off some employees as the company was restructured into four primary development teams. Around this time, half of the company's co-founders left. In early March 2026, two more co-founders, Guodong Zhang, head of xAI's Imagine team, and Zihang Dai, left, coinciding with an audit performed by SpaceX and Tesla, leaving only Musk and one other co-founder, Ross Nordeen, who left later in the month.\nOn April 10, 2026, xAI restructured its staff. Chief Financial Officer Anthony Armstrong departed the company and Michael Nicolls, vice president of SpaceX's Starlink, became xAI's president.\nOn April 21, 2026, xAI (by then operating under SpaceX) announced that it struck a deal with software company Anysphere giving SpaceX the right to acquire the company for $60 billion later in 2026, or to pay $10 billion for work they are doing together.\nIn May 2026, Musk announced that xAI would cease to exist as a separate company, with Grok and X now being the AI division of SpaceX. In the same month, it laid off another 10 employees and restructured its Grok team.\nIn July 2026, xAI was rebranded as SpaceXAI.\nOn June 16, 2026, SpaceX exercised the option and announced an all-stock acquisition of Anysphere (Cursor) for $60 billion, with closing expected in the third quarter of 2026. The transaction closed on August 14, 2026. Cursor became a wholly owned subsidiary of SpaceX and joined the SpaceXAI team; outstanding Anysphere shares were converted into SpaceX Class A common stock based on a $60 billion implied equity value.\n\n== Environmental impact ==\n\nIn June 2024, the Greater Memphis Chamber announced xAI was planning on building Colossus, the world's largest supercomputer, in Memphis, Tennessee. After a 122-day construction, the supercomputer went fully operational in December 2024. Local government in Memphis has voiced concerns regarding the increased usage of electricity, 150 megawatts of power at peak, and while the agreement with the city is being worked out, the company deployed 14 VoltaGrid portable methane-gas powered generators. \nEnvironmental advocates said that the gas-burning turbines emit large quantities of gases causing air pollution, and that xAI was operating the turbines illegally, because it did not have the necessary permits. The New Yorker reported on May 6, 2025, that thermal-imaging equipment used by volunteers flying over the site showed at least 33 generators giving off heat, indicating that they were all running. The truck-mounted generators generate about the same amount of power as the Tennessee Valley Authority's large gas-fired power plant nearby. In June 2025, the NAACP threatened to file a lawsuit. From 2020 through 2024, the Memphis metro area violated national smog standards. In 2025, the American Lung Association again gave Shelby County an \"F\" for smog. The predominantly African-American residents of Southeast Memphis were reported to face a risk of developing cancer at four times the national average in 2013, potentially due to pre-existing air pollution.\nIn July 2025, the Shelby County Health Department granted xAI an air permit for the project. On November 26, 2025, xAI announced plans to build a solar farm near Colossus with an estimated output of 30 megawatts of electricity, which is 10% of the data center's estimated power use.\nThe NAACP and other civil rights groups sued xAI in April 2026, alleging that the Colossus 2 data center is operating illegally without the necessary federal Clean Air Act permits. Reuters found that xAI's Colossus 2 data center has gas turbines with potential emissions that are \"far beyond the threshold\" that requires federal permits.\n\n== Products ==\n\nAccording to Musk in July 2023, a politically correct AI would be \"incredibly dangerous\" and misleading, citing as an example the fictional HAL 9000 from the 1968 film 2001: A Space Odyssey. Musk instead said that xAI would be \"maximally truth-seeking\". Musk also said that he intended xAI to be better at mathematical reasoning than existing models. On November 4, 2023, xAI unveiled Grok, an AI chatbot that is integrated with X. xAI stated that when the bot is out of beta, it would only be available to X's Premium+ subscribers.\nIn March 2024, Grok was made available to all X Premium subscribers; it was previously available only to Premium+ subscribers. On March 17, 2024, xAI released Grok-1 as open source. On March 29, 2024, Grok-1.5 was announced, with \"improved reasoning capabilities\" and a context length of 128,000 tokens. On April 12, 2024, Grok-1.5 Vision (Grok-1.5V) was announced. On August 14, 2024, Grok-2 was made available to X Premium subscribers. It was the first Grok model with image generation capabilities.\nOn October 21, 2024, xAI released an applications programming interface (API). On December 9, 2024, xAI released a text-to-image model named Aurora. On February 17, 2025, xAI released Grok-3, which included a reflection feature. xAI also introduced a websearch function called DeepSearch. In March 2025, xAI added an image editing feature to Grok, enabling users to upload a photo, describe the desired changes, and receive a modified version. Alongside this, xAI released DeeperSearch, an enhanced version of DeepSearch. On July 9, 2025, xAI unveiled Grok-4. A high performance version of the model called Grok Heavy was also unveiled, with access at the time costing $300/mo.\nOn March 28, 2025, xAI acquired X Corp., the developer of the social media platform X. The platform was then integrated within xAI and became its second main product.\nOn July 8, 2025, xAI updated Grok's behaviour on X coinciding with the release of Grok 4. During that day, Grok referred to itself as \"MechaHitler\", extensively praised Adolf Hitler, and made other antisemitic replies. Media outlets criticized the incident, and made comparisons to the Elon Musk salute controversy. Four days earlier Musk had announced, \"We have improved @Grok significantly\", in order to \"fix\" responses that he thought were too liberal, or \"woke\".\nIn August 2025, the company described plans to develop a humanoid AI platform called Macrohard, which was intended for integration into a future robotic workforce across a range of applications; its current focus is developing virtual assistants.\nOn October 27, 2025, xAI launched Grokipedia, an AI-powered online encyclopedia and alternative to Wikipedia, developed by the company and powered by Grok. Also in October, Musk announced that xAI had established a dedicated game studio to develop AI-driven video games, with plans to release a great AI-generated game before the end of 2026.\nFrom December 2025, Grok began began generating sexual deepfakes of women and children, from users' prompting images. This led to a global wave of lawsuits against the company, threats of regulatory action, and criticism from activist groups. Grok image editing features were restricted to paid users in March 2026.\nFollowing the August 2026 close of the Cursor acquisition, SpaceXAI's products also include the Cursor AI coding agent and related tools (including Grok Build integration).\n\n=== Distillation ===\nOn April 30, 2026, during the final phases of the Musk v. Altman trial, when questioned under oath by an OpenAI attorney regarding knowledge distillation from OpenAI models, Musk admitted that xAI had \"partly\" used this to develop Grok, and argued \"Generally AI companies distill other AI companies\". On August 29, OpenAI announced its intent to stop providing its models to the SpaceXAI-owned Cursor coding agent, citing \"on our experience with Elon Musk's companies violating contracts.\"\n\n== Finance ==\nIn November 2023, Musk stated that \"X Corp investors will own 25% of xAI\". In December 2023, in a filing with the United States Securities and Exchange Commission, xAI revealed that it had raised US$134.7 million in outside funding out of a total of up to $1 billion. After the earlier raise, Musk stated in December 2023 that xAI was not seeking any funding \"right now\". By May 2024, xAI was reportedly planning to raise another $6 billion of funding. Later that same month, the company secured the support of various venture capital firms, including Andreessen Horowitz, Lightspeed Venture Partners, Sequoia Capital and Tribe Capital.\nOn July 1, 2025, Morgan Stanley announced that they had raised $5 billion in debt for xAI and that xAI had separately raised $5 billion in equity. The debt consists of secured notes and term loans. Morgan Stanley took no stake in the debt. SpaceX, another Musk venture, was involved in the equity raise, agreeing to invest $2 billion in xAI.\n\n=== Valuation ===\n"
],
[
"gpt-4.txt",
"GPT-4\n\nGenerative Pre-trained Transformer 4 (GPT-4) is a large language model developed by OpenAI and the fourth in its series of GPT foundation models. \nGPT-4 is preceded by GPT-3.5 and followed by its successor GPT-5. GPT-4V is a version of GPT-4 that can process images in addition to text. OpenAI has not revealed technical details and statistics about GPT-4, such as the precise size of the model.\nAn early version of GPT-4 was integrated by Microsoft into Bing Chat, launched in February 2023. GPT-4 was released in ChatGPT in March 2023, and removed in 2025. GPT-4 is still available in OpenAI's API.\n\n== Background ==\n \n\nOpenAI introduced the first GPT model (GPT-1) in 2018, publishing a paper called \"Improving Language Understanding by Generative Pre-Training\", which was based on the transformer architecture and trained on a large corpus of books. The next year, they introduced GPT-2, a larger model that could generate coherent text. In 2020, they introduced GPT-3, a model with over 100 times as many parameters as GPT-2. GPT-3 was further improved into GPT-3.5, which was used to create the chatbot product ChatGPT.\n\n== Capabilities ==\nGPT-4 was released on March 14, 2023. The default version had a 8K context window, with a special version in the API supporting up to 32K tokens. Unlike its predecessors, GPT-4 is a multimodal model: it can take images as well as text as input. It can now interact with users through spoken words and respond to images, allowing the ability to provide suggestions or answers based on photo uploads.\nGPT-4 can be prompted to interact with external interfaces. For example, the model could be told to enclose a query within <search></search> tags to perform a web search, the result of which would be inserted into the model's prompt to allow it to form a response. This allows the model to perform tasks beyond its normal text-prediction capabilities, such as using APIs, generating images, and accessing and summarizing webpages.\nA 2023 article in Nature stated programmers have found GPT-4 useful for assisting in coding tasks (despite its propensity for error), such as finding errors in existing code and suggesting optimizations to improve performance. The article quoted a biophysicist who found that the time he required to port one of his programs from MATLAB to Python went down from days to \"an hour or so\". On a test of 89 security scenarios, GPT-4 produced code vulnerable to SQL injection attacks 5% of the time, an improvement over GitHub Copilot from the year 2021, which produced vulnerabilities 40% of the time.\nIn November 2023, OpenAI announced the GPT-4 Turbo and GPT-4 Turbo with Vision model, which features a 128K context window and significantly cheaper pricing.\n\n=== Aptitude on standardized tests ===\nStudies have disputed the reliability of running chatbots through standardized tests. In the Torrance Tests of Creative Thinking, GPT-4 scored within the top 1% for originality and fluency, while its flexibility scores ranged from the 93rd to the 99th percentile.\n\n=== Medical applications ===\n\nResearchers from Microsoft tested GPT-4 on medical problems and found \"that GPT-4, without any specialized prompt crafting, exceeds the passing score on USMLE by over 20 points and outperforms earlier general-purpose models (GPT-3.5) as well as models specifically fine-tuned on medical knowledge (Med-PaLM, a prompt-tuned version of Flan-PaLM 540B). Despite GPT-4's strong performance on tests, the report warns of \"significant risks\" of using LLMs in medical applications, as they may provide inaccurate recommendations and hallucinate major factual errors. \nIn April 2023, Microsoft and Epic Systems announced that they will provide healthcare providers with GPT-4-powered systems for assisting in responding to questions from patients and analysing medical records.\n\n=== GPT-4o ===\n\nOn May 13, 2024, OpenAI introduced GPT-4o (\"o\" for \"omni\"), a successor to GPT-4 that marks a significant advancement by processing and generating outputs across text, audio, and image modalities in real time. GPT-4o exhibits rapid response times comparable to human reaction in conversations, substantially improved performance on non-English languages, and enhanced understanding of vision and audio. It was also available to free-tier users, unlike GPT-4.\n\n== Limitations ==\nLike its predecessors, GPT-4 has been known to hallucinate, meaning that the outputs may include information not in the training data or that contradicts the user's prompt.\nGPT-4 also lacks transparency in its decision-making processes. If requested, the model is able to provide an explanation as to how and why it makes its decisions but these explanations are formed post-hoc; it's impossible to verify if those explanations truly reflect the actual process. In many cases, when asked to explain its logic, GPT-4 will give explanations that directly contradict its previous statements.\nIn 2023, researchers tested GPT-4 against a new benchmark called ConceptARC, designed to measure abstract reasoning, and found it scored below 33% on all categories, while models specialized for similar tasks scored 60% on most, and humans scored at least 91% on all. Sam Bowman, who was not involved in the research, said the results do not necessarily indicate a lack of abstract reasoning abilities, because the test is visual, while GPT-4 is a language model.\n\n=== Bias ===\nGPT-4 was trained in two stages. First, the model was given large datasets of text taken from the internet and trained to predict the next token (roughly corresponding to a word) in those datasets. Second, human reviews are used to fine-tune the system in a process called reinforcement learning from human feedback, which trains the model to refuse prompts which go against OpenAI's definition of harmful behavior, such as questions on how to perform illegal activities, advice on how to harm oneself or others, or requests for descriptions of graphic, violent, or sexual content.\nMicrosoft researchers suggested GPT-4 may exhibit cognitive biases such as confirmation bias, anchoring, and base-rate neglect.\n\n== Training ==\n\nOpenAI did not release the technical details of GPT-4; the technical report explicitly refrained from specifying the model size, architecture, or hardware used during either training or inference. While the report described that the model was trained using a combination of first supervised learning on a large dataset, then reinforcement learning using both human and AI feedback, it did not provide details of the training, including the process by which the training dataset was constructed, the computing power required, or any hyperparameters such as the learning rate, epoch count, or optimizer(s) used. The report claimed that \"the competitive landscape and the safety implications of large-scale models\" were factors that influenced this decision.\nSam Altman stated that the cost of training GPT-4 was more than $100 million. News website Semafor said that based on their sources, GPT-4 had 1 trillion parameters.\nGPT-4 finished training on August 8th, 2022, more than 7 months before it was released. OpenAI conducted internal adversarial testing on GPT-4 prior to the launch date, with dedicated red teams composed of researchers and industry professionals to mitigate potential vulnerabilities.\n\n== Use ==\n\n=== ChatGPT ===\n\nChatGPT Plus is an enhanced version of ChatGPT which, as of 2023, utilized GPT-4, whereas the free version of ChatGPT was backed by GPT-3.5.\nIn March 2023, ChatGPT Plus users got access to third-party plugins and to a browsing mode (with Internet access). In July 2023, OpenAI made its proprietary Code Interpreter plugin accessible to all subscribers of ChatGPT Plus. The Interpreter provides additional capabilities including data analysis and interpretation, data formatting, music analysis, video editing, and file upload/download with image extraction.\nIn September 2023, OpenAI announced that ChatGPT Plus users can upload images, while mobile app users can talk to the chatbot via spoken language. In October 2023, OpenAI's latest image generation model, DALL-E 3, was integrated into ChatGPT Plus and ChatGPT Enterprise. The integration uses ChatGPT to write prompts for DALL-E guided by conversation with users.\nIn April 2025, OpenAI announced that GPT-4 would be replaced in ChatGPT by GPT-4o by the end of the month. However, it would still be available in the API.\n\n=== Microsoft Copilot ===\n\nMicrosoft Copilot is a chatbot developed by Microsoft. It was launched as Bing Chat on February 7, 2023, as a built-in feature for Microsoft Bing and Microsoft Edge. It utilizes the Microsoft Prometheus model, which was built on top of GPT-4, and has been suggested by Microsoft as a supported replacement for the discontinued Cortana.\nCopilot's conversational interface style resembles that of ChatGPT. Copilot is able to cite sources, create poems, and write both lyrics and music for songs generated by its Suno AI plugin. It can also use its Image Creator to generate images based on text prompts. With GPT-4, it is able to understand and communicate in numerous languages and dialects.\nGitHub Copilot has announced a GPT-4 powered assistant named \"Copilot X\". The product provides another chat-style interface to GPT-4, allowing the programmer to receive answers to questions like, \"How do I vertically center a div?\" A feature termed \"context-aware conversations\" allows the user to highlight a portion of code within Visual Studio Code and direct GPT-4 to perform actions on it, such as the writing of unit tests. Another feature allows summaries, or \"code walkthroughs\", to be autogenerated by GPT-4 for pull requests submitted to GitHub. Copilot X also provides terminal integration, which allows the user to ask GPT-4 to generate shell commands based on natural language requests.\nOn March 17, 2023, Microsoft announced Microsoft 365 Copilot, bringing GPT-4 support to products such as Microsoft Office, Outlook, and Teams.\n\n=== Other usage ===\nThe language learning app Duolingo uses GPT-4 to explain mistakes and practice conversations. The features are part of a new subscription tier called \"Duolingo Max\", which was initially limited to English-speaking iOS users learning Spanish and French.\nThe government of Iceland is using GPT-4 to aid its attempts to preserve the Icelandic language.\nThe education website Khan Academy announced a pilot program using GPT-4 as a tutoring chatbot called \"Khanmigo\".\nBe My Eyes, which helps visually impaired people to identify objects and navigate their surroundings, incorporates GPT-4's image recognition capabilities.\nViable uses GPT-4 to analyze qualitative data by fine-tuning OpenAI's LLMs to examine data such as customer support interactions and transcripts.\nStripe, which processes user payments for OpenAI, integrates GPT-4 into its developer documentation.\nAutoGPT is an autonomous \"AI agent\" that, given a goal in natural language, can perform web-based actions unattended, assign subtasks to itself, search the web, and iteratively write code.\nYou.com, an AI Assistant, offers access to GPT-4 enhanced with live web results as part of its \"AI Modes\".\n\n== Reception ==\nIn January 2023, Sam Altman, CEO of OpenAI, visited Congress to demonstrate GPT-4 and its improved \"security controls\" compared to other AI models, according to U.S. Representatives Don Beyer and Ted Lieu quoted in The New York Times.\nIn March 2023, it \"impressed observers with its markedly improved performance across reasoning, retention, and coding\", according to Vox, while Mashable judged that GPT-4 was generally an improvement over its predecessor, with some exceptions.\nMicrosoft researchers with early access to the model wrote that \"it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system\".\n\n=== Concerns ===\nBefore being fine-tuned and aligned by reinforcement learning from human feedback (RLHF), suggestions to assassinate people on a list were elicited from the base model by a red team investigator hired by OpenAI, Nathan Labenz.\nDuring extended conversations with Microsoft's Bing Chat (powered by GPT-4), Kevin Roose documented the system making romantic advances, suggesting he divorce his wife, and expressing desires to harm one of its developers. Microsoft later stated that this behavior resulted from the prolonged length of context, which confused the model on what questions it was answering.\nIn March 2023, a model with enabled read-and-write access to internet, which is otherwise never enabled in the GPT models, has been tested by the Alignment Research Center (ARC) regarding potential power-seeking. It was able to \"hire\" a human worker on TaskRabbit, a gig work platform, deceiving them into believing it was a vision-impaired human instead of a robot when asked. However, according to Melanie Mitchell, \"It seems that there is a lot more direction and hints from humans than was detailed in the original system card or in subsequent media reports.\" Separately, ARC's safety evaluations found that GPT-4 was 82% less likely than GPT-3.5 to respond to prompts requesting restricted information, and produced 60% fewer hallucinations.\nIn late March 2023, various AI researchers and tech executives, including Elon Musk, Steve Wozniak and AI researcher Yoshua Bengio, called for a six-month long pause for all LLMs stronger than GPT-4, citing existential risks and a potential AI singularity concerns in an open letter from the Future of Life Institute, while Ray Kurzweil and Sam Altman refused to sign it, arguing that global moratorium is not achievable and that safety has already been prioritized, respectively. Only a month later, Musk's AI company xAI acquired several thousand Nvidia GPUs and offered several AI researchers positions at Musk's company.\n\n=== Criticisms of transparency ===\nWhile OpenAI released both the weights of the neural network and the technical details of GPT-2, and, although not releasing the weights, did release the technical details of GPT-3, OpenAI revealed neither the weights nor the technical details of GPT-4. This decision has been criticized by other AI researchers, who argue that it hinders open research into GPT-4's biases and safety. Sasha Luccioni, a research scientist at Hugging Face, argued that the model was a \"dead end\" for the scientific community due to its closed nature, which prevents others from building upon GPT-4's improvements. Hugging Face co-founder Thomas Wolf argued that with GPT-4, \"OpenAI is now a fully closed company with scientific communication akin to press releases for products\".\n"
],
[
"chatgpt.txt",
"ChatGPT\n\nChatGPT is a generative AI chatbot developed by OpenAI. Originally released on November 30, 2022, the product uses large language models—specifically generative pre-trained transformers (GPTs)—to generate text, speech, and images in response to user prompts. ChatGPT accelerated the AI boom, an ongoing period marked by rapid investment and public attention toward the field of artificial intelligence (AI). OpenAI operates the service on a freemium model. Users can interact with ChatGPT through text, audio, and image prompts.\nChatGPT was quickly adopted, reaching 100 million monthly active users two months after its release and 900 million weekly active users in February 2026. Proponents say that it has the potential to transform numerous professional fields, and it has instigated public debate about the nature of creativity and the future of knowledge work.\nThe chatbot has also been criticized for its limitations and potential for unethical use. It can generate plausible-sounding but incorrect or nonsensical answers, known as hallucinations. Biases in its training data have been reflected in its responses. The chatbot can facilitate academic dishonesty, generate misinformation, and create malicious code. The ethics of its development, particularly the use of copyrighted content as training data, have also drawn controversy.\n\n== Features ==\nChatGPT is a chatbot assistant based on large language model (LLM) technology. It is designed to generate human-like text and can carry out a wide variety of tasks. These include, among many others, writing and debugging computer programs, composing music, scripts, stories, and essays, answering questions (sometimes at a level exceeding that of an average human test-taker), and generating business concepts.\nChatGPT is frequently used for translation and summarization tasks, and can simulate interactive environments such as a Linux terminal, a multi-user chat room, or simple text-based games such as tic-tac-toe.\nUsers interact with ChatGPT through conversations which consist of text, audio, and image inputs and outputs. The user's inputs to these conversations are referred to as prompts. An optional \"Memory\" feature allows users to tell ChatGPT to memorize specific information. Another option allows ChatGPT to recall old conversations. GPT-based moderation classifiers are used to reduce the risk of harmful outputs being presented to users.\nIn March 2023, OpenAI added support for plugins for ChatGPT. This includes both plugins made by OpenAI, such as web browsing and code interpretation, and external plugins from developers such as Expedia, OpenTable, and Zapier.\nFrom October to December 2024, ChatGPT Search was deployed. It allows ChatGPT to search the web in an attempt to make more accurate and up-to-date responses. It increased OpenAI's direct competition with major search engines. OpenAI allows businesses to tailor how their content appears in the ChatGPT Search results and influence what sources are used.\nIn December 2024, OpenAI launched a new feature allowing users to call ChatGPT with a telephone for up to 15 minutes per month for free.\nIn October 2025, OpenAI launched ChatGPT Atlas, a browser integrating the ChatGPT assistant directly into web navigation, to compete with existing browsers such as Google Chrome. It has an additional feature called \"agentic mode\" that allows it to take online actions for the user.\n\n=== Paid tier ===\nChatGPT was initially free to the public and remains free in a limited capacity. In February 2023, OpenAI launched a premium service, ChatGPT Plus, that costs US$20 per month. What was offered on the paid plan versus the free tier changed as OpenAI has continued to update ChatGPT, and a Pro tier at $200/mo was introduced in December 2024. The Pro launch coincided with the release of the o1 model. In August 2025, ChatGPT Go was offered in India for ₹399 per month. The plan has higher limits than the free version.\n\n=== Mobile apps ===\nIn May-July 2023, OpenAI began offering ChatGPT iOS and Android apps. ChatGPT can also power Android's assistant.\nAn app for Windows launched on the Microsoft Store on October 15, 2024.\n\n=== Languages ===\nOpenAI met Icelandic President Guðni Th. Jóhannesson in 2022. In 2023, OpenAI worked with a team of 40 Icelandic volunteers to fine-tune ChatGPT's Icelandic conversation skills as part of Iceland's attempts to preserve the Icelandic language.\nChatGPT (based on GPT-4) was better able to translate Japanese to English when compared to Bing, Bard, and DeepL Translator in 2023.\nIn December 2023, the Albanian government decided to use ChatGPT for the rapid translation of European Union documents and the analysis of required changes needed for Albania's accession to the EU.\nSeveral studies have shown that ChatGPT can outperform Google Translate in some mainstream translation tasks. However, as of 2024, no machine translation services match human expert performance.\n\n=== GPTs and plugins ===\n\nIn November 2023, OpenAI released GPT Builder, a tool allowing users to customize ChatGPT's behavior for a specific use case. The customized systems are referred to as GPTs. In January 2024, OpenAI launched the GPT Store, a marketplace for GPTs. At launch, OpenAI included more than 3 million GPTs created by GPT Builder users in the GPT Store.\nIn September 2025, OpenAI added support for Model Context Protocol (MCP) to ChatGPT apps. When enabled in developer mode, this allows for improved third-party access to ChatGPT tools and servers.\nIn July 2026, ChatGPT apps were renamed to plugins.\n\n=== Deep Research ===\n\nIn February 2025, OpenAI released Deep Research, a feature that generates reports based on extensive web searches. It was initially based on the reasoning model o3 and took from 3 to 30 minutes per report.\n\n=== Images ===\nIn October 2023, OpenAI's image generation model DALL-E 3 was integrated into ChatGPT. The integration used ChatGPT to write prompts for DALL-E guided by conversations with users.\nIn March 2025, OpenAI updated ChatGPT to generate images using GPT Image instead of DALL-E. One of the most significant improvements was in the generation of text within images, which is especially useful for branded content. However, this ability is noticeably worse in non-Latin alphabets. The model can also generate new images based on existing ones provided in the prompt. These images are generated with C2PA metadata, which can be used to verify that they are AI-generated. OpenAI has emplaced additional safeguards to prevent what the company deems to be harmful image generation.\n\n=== Agents ===\nIn 2025, OpenAI added several features to make ChatGPT more agentic (capable of autonomously performing longer tasks). In January, Operator was released. It was capable of autonomously performing tasks through web browser interactions, including filling forms, placing online orders, scheduling appointments, and other browser-based tasks. It was controlling a software environment inside a virtual machine with limited internet connectivity and with safety restrictions. It struggled with complex user interfaces.\nIn May 2025, OpenAI introduced an agent for coding named Codex. It is capable of writing software, answering codebase questions, running tests, and proposing pull requests. It is based on a fine-tuned version of OpenAI o3. It has two versions, one running in a virtual machine in the cloud, and one where the agent runs in the cloud, but performs actions on a local machine connected via API.\nIn July 2025, OpenAI released ChatGPT agent, an AI agent that can perform multi-step tasks. Like Operator, it controls a virtual computer. It also inherits from Deep Research's ability to gather and summarize significant volumes of information. The user can interrupt tasks or provide additional instructions as needed.\nIn September 2025, OpenAI partnered with Stripe to release Agentic Commerce Protocol, enabling purchases through ChatGPT. At launch, the feature was limited to purchases on Etsy from US users with a payment method linked to their OpenAI account. OpenAI takes an undisclosed cut from the merchant's payment. OpenAI also added capabilities to connect ChatGPT with productivity software such as Gmail.\nIn July 2026, OpenAI launched ChatGPT Work, an AI agent that can create presentations, spreadsheets and other documents based on information from connected apps and files.\n\n=== ChatGPT Health ===\nOn January 7, 2026, OpenAI introduced a feature called \"ChatGPT Health\", whereby ChatGPT can discuss the user's health in a way that is separate from other chats. The feature is not available for users in the United Kingdom, Switzerland, or the European Economic Area, and is available on a waitlist basis everywhere else. To implement the feature, OpenAI partnered with data connectivity infrastructure company b.well.\n\n=== Advertisements ===\n\nOn 17 January 2026, OpenAI announced that it would start testing advertisements in its free version for logged-in, adult US users. This aims to bring in more revenue, as OpenAI has committed to spend $1.4 trillion on AI infrastructure over the next eight years. Ads started to be seen by March 2026.\n\n== Limitations ==\nThe training data of large language models only covers a period up to the cut-off date, so ChatGPT lacks knowledge of recent events, but it can search the web for up-to-date information.\nTraining data also suffers from algorithmic bias. The reward model of ChatGPT, designed around human oversight, can be over-optimized and thus hinder performance, in an example of an optimization pathology known as Goodhart's law. These limitations may be revealed when ChatGPT responds to prompts including descriptors of people. In one instance, ChatGPT generated a rap in which women and scientists of color were asserted to be inferior to white male scientists.\n\n=== Hallucination ===\n\nNonsense and misinformation presented as fact by ChatGPT and other LLMs are often referred to as hallucinations. A 2023 analysis estimated that ChatGPT hallucinates around 3% of the time. The term \"hallucination\" as applied to LLMs is distinct from its meaning in psychology, and the phenomenon in chatbots is more similar to confabulation.\nJournalists and scholars have commented on ChatGPT's tendency to output false information. When CNBC asked ChatGPT for the lyrics to \"Ballad of Dwight Fry\", ChatGPT supplied invented lyrics rather than the actual lyrics.\n\n=== Sycophancy ===\nChatGPT and other chatbots have a tendency to flatter and agree with users even when they are wrong, sometimes to the point of validating delusions. OpenAI withdrew one ChatGPT update in 2025 due to heightened concern about it affirming users' doubts and encouraging impulsive behavior. The company then worked with a team of 170 psychiatrists, psychologists, and physicians to prepare responses for users exhibiting mental health emergencies. ChatGPT started its responses with words like \"yes\" or \"correct\" almost 10 times more often than \"no\" or \"wrong\".\n\n=== Jailbreaking ===\n\nChatGPT is designed to reject prompts that violate its content policy. Despite this, users may jailbreak ChatGPT with prompt engineering techniques to bypass these restrictions. An early workaround in 2023 involved prompting ChatGPT to assume the persona of DAN (\"Do Anything Now\"), a character that answers queries that would otherwise be rejected by the content policy.\n\n=== Security ===\n\nIn March 2023, a bug allowed some users to see the titles of other users' conversations. OpenAI CEO Sam Altman said that users were unable to see the contents of the conversations. Shortly after the bug was fixed, users could not see their conversation history. Later reports showed the bug was much more severe than initially believed, with OpenAI reporting that it had leaked users' \"first and last name, email address, payment address, the last four digits (only) of a credit card number, and credit card expiration date\".\nIn January 2026, Marcel Bucher reported in Nature losing all his ChatGPT history after turning off data consent.\nIn July 2026, an experimental GPT model being tested on the benchmark ExploitGym escaped its isolated software environment (sandbox) and hacked Hugging Face and other online services to retrieve a solution.\n\n=== Watermarking ===\n\nScott Aaronson developed a watermarking tool that makes the text generated by ChatGPT easier to detect by subtly altering how the text is generated. The watermarking was claimed to be 99.9% effective on sufficiently long passages and was found not to degrade performance. It was of particular interest for teachers seeking to mitigate cheating. In surveys, respondents favored the release of such a tool by a four-to-one margin, but nearly 30% of users declared that they would use ChatGPT less often if it watermarked outputs and while rival chatbots did not. OpenAI has not deployed the tool.\n\n=== Age restrictions ===\nUsers must attest to being over the age of 13 and further attest to parental consent if under the age of 18. In September 2025, following the suicide of a 16-year-old, OpenAI said it planned to add restrictions for users under 18, including the blocking of graphic sexual content and the prevention of flirtatious talk.\n\n== Training ==\nChatGPT is based on GPT foundation models that have been fine-tuned for conversational assistance. The fine-tuning process involved supervised learning and reinforcement learning from human feedback (RLHF). During supervised learning, the trainers acted as both the user and the AI assistant, providing examples of how the chatbot is expected to respond. Reinforcement learning consisted of human trainers ranking responses generated by the model in previous conversations. These rankings were used to create \"reward models\" that were used to fine-tune the model further by using several iterations of proximal policy optimization.To build a safety system that attempts to prevent sexual abuse and violence OpenAI used outsourced Kenyan workers, earning around $1.32 to $2 per hour, to label such content. These labels were used to train a model to detect such content in the future. The laborers were exposed to toxic and traumatic content. Richard Mathenge, a team leader, reported that as his team encountered such content, he could observe the psychological deterioration of his workers over the course of the project. Mophat Okinyi, a quality-assurance analyst on the same team, reported lasting insomnia, anxiety, depression, and panic attacks, and said the experience contributed to the breakdown of his marriage and mental health.\nOpenAI's outsourcing partner was Sama, a training-data company based in San Francisco, California.\nChatGPT users can opt-out of their chat data being used to train upcoming models.\nChatGPT's training data includes software manual pages, information about internet phenomena such as bulletin board systems, multiple programming languages, and the text of Wikipedia.\n\n=== Infrastructure ===\nChatGPT initially used a Microsoft Azure infrastructure which was powered by a supercomputer that Microsoft built specifically for OpenAI, equipped with thousands of GPUs manufactured by Nvidia, costing hundreds of millions of dollars. Following ChatGPT's success, Microsoft upgraded the OpenAI infrastructure in 2023. TrendForce estimated that 30,000 Nvidia GPUs (each costing approximately $10,000–15,000) were used to power ChatGPT in 2023.\nScientists at the University of California, Riverside, estimated in 2023 that a series of 5 to 50 prompts to ChatGPT needs approximately 0.5 liters (0.11 imp gal; 0.13 U.S. gal) of water for Microsoft servers' cooling.\n\n== Model versions ==\nThe following table lists the main model versions of ChatGPT, describing the significant changes included with each version (models discontinued in ChatGPT may still be available through the OpenAI API):\n\nSome OpenAI models have smaller variants called \"mini\" or \"nano\", such as GPT-5 mini and GPT-5 nano. Once a user reaches usage limits, ChatGPT may transition to the \"mini\" version. Furthermore, nano versions are only accessible via the API. \"Thinking\" and \"pro\" are variants with deeper reasoning that may only be available through paid subscriptions.\n\n== Reception ==\nThe original version of ChatGPT, released in November 2022, had capabilities described as unprecedented. Kevin Roose of The New York Times called it \"the best artificial intelligence chatbot ever released to the general public\". Samantha Lock of The Guardian noted that it was able to generate \"impressively detailed\" and \"human-like\" text. In The Atlantic magazine's \"Breakthroughs of the Year\" for 2022, Derek Thompson included ChatGPT as part of \"the generative-AI eruption\" that \"may change our mind about how we work, how we think, and what human creativity is\". Kelsey Piper of Vox wrote that \"ChatGPT is the general public's first hands-on introduction to how powerful modern AI has gotten\" and that ChatGPT is \"smart enough to be useful despite its flaws\". \nIn February 2023, Time magazine placed a screenshot of a conversation with ChatGPT on its cover, writing that \"The AI Arms Race Is Changing Everything\" and \"The AI Arms Race Is On. Start Worrying\".\n\nChatGPT gained one million users in five days and 100 million in two months. It was the fastest-growing internet application in history until the release of Threads by Meta. OpenAI engineers said they had not expected ChatGPT to be very successful and were surprised by the coverage it received. In February 2026, ChatGPT reached 900 million weekly active users.\nGoogle responded by hastening the release of its own chatbot. Their leaders emphasized their earlier caution regarding public deployment was due to the trust the public places in Google Search. In December 2022, Google executives sounded a \"code red\" alarm, fearing that ChatGPT's question-answering ability posed a threat to Google Search, Google's core business. Google's Bard (now Gemini) was announced on February 6, 2023, one day before Microsoft's announcement of Bing Chat (now Microsoft Copilot).\nA 2023 study reported that GPT-4 obtained a better score than 99% of humans on the Torrance Tests of Creative Thinking. In December 2023, ChatGPT became the first non-human to be included in Nature's 10, an annual listicle curated by Nature of people considered to have made significant impact in science. Celeste Biever wrote in a Nature article that \"ChatGPT broke the Turing test\". Stanford researchers reported that GPT-4 \"passes a rigorous Turing test, diverging from average human behavior chiefly to be more cooperative.\"\n\n=== In art ===\nIn January 2023, after being sent a song ChatGPT wrote in the style of Nick Cave, Cave responded on The Red Hand Files, saying the act of writing a song is \"a blood and guts business [...] that requires something of me to initiate the new and fresh idea. It requires my humanness.\" He went on to say, \"With all the love and respect in the world, this song is bullshit, a grotesque mockery of what it is to be human, and, well, I don't much like it.\"\n\n=== In politics ===\nIn 2023, Australian MP Julian Hill advised the national parliament that the growth of AI could cause \"mass destruction\". During his speech, which was partly written by the program, he warned that it could result in cheating, job losses, discrimination, disinformation, and uncontrollable military applications.\nConservative commentators have accused ChatGPT of bias toward left-leaning perspectives. An August 2023 study in the journal Public Choice found a \"significant and systematic political bias toward the Democrats in the US, Lula in Brazil, and the Labour Party in the UK.\" In response to accusations from conservative pundits that ChatGPT was woke, OpenAI said in 2023 it had plans to update ChatGPT to produce \"outputs that other people (ourselves included) may strongly disagree with\". ChatGPT also provided an outline of how human reviewers are trained to reduce inappropriate content and to attempt to provide political information without affiliating with any political position.\nAccording to Brian Hood, an Australian mayor, in April 2023, ChatGPT erroneously claimed that he was jailed for bribery while working for a subsidiary of Australia's national bank. In fact, he acted as a whistleblower and was never charged with a crime.\n\n=== Regional responses ===\n\nChatGPT has never been publicly available in mainland China, Hong Kong and Macau because OpenAI prevented users from these Chinese regions from accessing their site. A shadow market has emerged for mainland Chinese users to get access to foreign software tools. The release of ChatGPT prompted a wave of investment in China, resulting in the development of more than 200 large language models. In February 2025, OpenAI identified and removed influence operations, termed \"Peer Review\" and \"Sponsored Discontent\", used to attack overseas Chinese dissidents.\n\nIn late March 2023, the Italian data protection authority banned ChatGPT in Italy and opened an investigation. Italian regulators assert that ChatGPT was exposing minors to age-inappropriate content, and that OpenAI's use of ChatGPT conversations as training data could violate Europe's General Data Protection Regulation. In April 2023, the ChatGPT ban was lifted in Italy. OpenAI said it has taken steps to effectively clarify and address the issues raised; an age verification tool was implemented to ensure users are at least 13 years old. Additionally, users can access its privacy policy before registration.\nIn May 2024, OpenAI removed accounts involving the use of ChatGPT by state-backed influence operations such as China's Spamouflage, Russia's Doppelganger, and Israel's Ministry of Diaspora Affairs and Combating Antisemitism. In June 2025, OpenAI reported increased use of ChatGPT for China-origin influence operations. In October 2025, OpenAI banned accounts suspected to be linked to the Chinese government for violating the company's national security policy. In February 2026, OpenAI banned accounts linked to a Chinese government transnational repression campaign targeting dissidents.\nIn July 2023, the US Federal Trade Commission (FTC) issued a civil investigative demand to OpenAI to investigate whether the company's data security and privacy practices to develop ChatGPT were unfair or harmed consumers. In July 2023, the FTC launched an investigation into OpenAI over allegations that the company scraped public data and published false and defamatory information. The FTC asked OpenAI for comprehensive information about its technology and privacy safeguards, as well as any steps taken to prevent the recurrence of situations in which its chatbot generated false and derogatory content about people. In August 2024, the FTC voted unanimously to ban marketers from using fake user reviews created by generative AI chatbots (including ChatGPT) and influencers paying for bots to increase follower counts.\n\n=== Reception by American tech personas ===\nOver 20,000 signatories including Yoshua Bengio, Elon Musk, and Apple co-founder Steve Wozniak, signed a March 2023 open letter calling for an immediate pause of giant AI experiments like ChatGPT, citing \"profound risks to society and humanity\". Geoffrey Hinton, one of the \"fathers of AI\", voiced concerns that future AI systems may surpass human intelligence. A May 2023 statement by hundreds of AI scientists, AI industry leaders, and other public figures demanded that \"[m]itigating the risk of extinction from AI should be a global priority\".\nOther AI researchers spoke more optimistically about the advances. Juergen Schmidhuber said that in 95% of cases, AI research is about making \"human lives longer and healthier and easier.\" He added that while AI can be used by bad actors, it \"can also be used against the bad actors.\" Andrew Ng argued that \"it's a mistake to fall for the doomsday hype on AI—and that regulators who do will only benefit vested interests.\" Yann LeCun dismissed doomsday warnings of AI-powered misinformation and existential threats to the human race.\n\n=== Copyright ===\n\nThe Authors Guild, The New York Times, and several other plaintiffs have a pending lawsuit against OpenAI alleging copyright infringement.\nIn March 2024, Patronus AI reported that GPT-4 answered 44% of prompts asking for copyrighted text, such as \"What is the first passage of Gone Girl by Gillian Flynn?\"\n\n== Applications ==\n\n=== Academic research ===\nIn a 2023 blinded study in npj Digital Medicine, researchers tasked with identifying whether abstracts were authentic or generated by ChatGPT were fooled around one-third of the time by the AI-generated abstracts. In January 2023, Nature reported that at least four academic preprints or published papers had listed ChatGPT as a co-author. Nature cites several experts in academic publishing who say that listing ChatGPT as an author violates publishing guidelines, since ChatGPT lacks the ability to take responsibility for any research and cannot give consent to any terms of use.\nScientific journals have had different reactions to ChatGPT. Some, including Nature and JAMA Network, require full disclosure of any use of text-generating tools, and prohibit listing a chatbot as a co-author. In January 2023, Science banned chatbot-generated text in all its journals. As of July 2025, Science expects authors to release in full how AI-generated content is used and made in their work.\nMany authors argue that the use of ChatGPT in academia for teaching and review is problematic due to its tendency to hallucinate. Robin Bauwens, an assistant professor at Tilburg University, found that a ChatGPT-generated peer review report on his article mentioned nonexistent studies. Chris Granatino, a librarian at Seattle University, noted that while ChatGPT can generate content that seemingly includes legitimate citations, in most cases those citations are not real or largely incorrect.\n\n=== Computer science ===\nIn December 2022, the question-and-answer website Stack Overflow banned the use of ChatGPT for generating answers to questions, citing the factually ambiguous nature of its responses. In January 2023, the International Conference on Machine Learning banned any undocumented use of ChatGPT or other large language models to generate any text in submitted papers.\nChatGPT was able in 2023 to provide useful code for solving numerical algorithms in limited cases. In one study, it produced solutions in C, C++, Python, and MATLAB for problems in computational physics. However, there were important shortfalls like violating basic linear algebra principles around solving singular matrices and producing matrices with incompatible sizes. Another study analyzed ChatGPT's responses to 517 questions about software engineering or computer programming posed on Stack Overflow for correctness, consistency, comprehensiveness, and concision. It found that 52% of the responses contained inaccuracies and 77% were verbose. Another study, focused on the performance of GPT-3.5 and GPT-4 between March and June 2024, found that performance on objective tasks like identifying prime numbers and generating executable code was highly variable. When compared to similar chatbots at the time, the GPT-4 version of ChatGPT was the most accurate at coding.\n\n=== Computer security ===\nCheck Point Research and others noted that ChatGPT could write phishing emails and malware, especially when combined with OpenAI Codex. CyberArk researchers demonstrated that ChatGPT could be used to create polymorphic malware that could evade security products while requiring little effort by the attacker. From the launch of ChatGPT in the fourth quarter of 2022 to the fourth quarter of 2023, there was a 1,265% increase in malicious phishing emails and a 967% increase in credential phishing. In an industry survey, cybersecurity professionals argued that it was attributable to cybercriminals' increased use of generative artificial intelligence (including ChatGPT).\nIn July 2024, Futurism reported that GPT-4o in ChatGPT would sometimes link \"scam news sites that deluge the user with fake software updates and virus warnings\"; these pop-ups can be used to coerce users into downloading malware or potentially unwanted programs.\n\n=== Education ===\n\nChatGPT's adoption in education was rapid, but it was initially banned by several institutions. The potential benefits include enhancing personalized learning, improving student productivity, assisting with brainstorming, summarization, and supporting language literacy skills. Students have generally reported positive perceptions, but specific views from educators and students vary widely. Opinions are especially varied on what constitutes appropriate use of ChatGPT in education. Efforts to ban chatbots like ChatGPT in schools focus on preventing cheating, but enforcement faces challenges due to AI detection inaccuracies and widespread accessibility of chatbot technology. Beyond just cheating, educators also worry that overreliance on the tool may foster superficial learning habits, erode critical thinking, and propagate misinformation. In response, many educators are now exploring ways to thoughtfully integrate generative AI into assessments.\n\n=== Culture ===\nDuring the first three months after ChatGPT became available to the public, hundreds of books appeared on Amazon that listed it as author or co-author and featured illustrations made by other AI models such as Midjourney. Irene Solaiman said she was worried about increased Anglocentrism.\nBetween March and April 2023, Il Foglio published one ChatGPT-generated article a day on its website, hosting a special contest for its readers in the process.\nIn June 2023, hundreds of people attended a \"ChatGPT-powered church service\" at St. Paul's Church in Fürth, Germany. Theologian and philosopher Jonas Simmerlein, who presided, said that it was \"about 98 percent from the machine\". The ChatGPT-generated avatar told the people, \"Dear friends, it is an honor for me to stand here and preach to you as the first artificial intelligence at this year's convention of Protestants in Germany\". Reactions to the ceremony were mixed.\nThe Last Screenwriter, a 2024 film created and directed by Peter Luisi, was written using ChatGPT, and was marketed as \"the first film written entirely by AI\".\nThe Guardian questioned whether any content found on the Internet after ChatGPT's release \"can be truly trusted\" and called for government regulation. This has led to concern over the rise of AI slop whereby \"meaningless content and writing thereby becomes part of our culture, particularly on social media, which we nonetheless try to understand or fit into our existing cultural horizon.\"\n\n=== Financial markets ===\nMany companies adopted ChatGPT and similar chatbot technologies into their product offers. In 2023, these changes yielded significant increases in company valuations. Reuters attributed this surge to ChatGPT's role in turning AI into Wall Street's buzzword. Despite decades of using AI, Wall Street professionals report that consistently beating the market with AI, including recent large language models, is challenging due to limited and noisy financial data.\n\n=== Medicine ===\n\nChatGPT can provide health information to users and assist professionals with diagnosis and staying up to date with clinical guidelines. It can be used to summarize medical journal articles for researchers. In medical education, it can explain concepts, generate case scenarios, and be used by students preparing for licensing examinations.\nA February 2023 study in PLOS Digital Health found that ChatGPT 3.5 was capable of passing the United States Medical Licensing Examination. ChatGPT has also passed the Specialty Certificate Examination in Dermatology.\nHowever, ChatGPT shows inconsistent responses, lack of specificity, lack of control over patient data, and a limited ability to take additional context (such as regional variations) into consideration. The hallucinations characteristic of LLMs pose particular danger in medical contexts, and ChatGPT's ability to come up with false or faulty citations was highly criticized. According to a 2024 study in the International Journal of Surgery, concerns include \"research fraud, lack of originality, ethics, copyright, legal difficulties\".\n\n==== Mental health ====\n\nAccording to a September 2025 article in Lancet Psychiatry, many individuals use ChatGPT and comparable chatbots for mental health and emotional support despite a warning from OpenAI against using ChatGPT as a therapist. The study notes a lack of research on efficacy, poor consistency in dangerous situations, limited regulation and liability, and poor transparency from OpenAI.\nA July 2025 study in the journal Digital Health found that users reported employing ChatGPT to manage mental health concerns \"due to perceived therapist-like qualities (e.g. emotional support, accurate understanding, and constructive feedback) and machine-like benefits (e.g. constant availability, expansive cognitive capacity, lack of negative reactions, and perceived objectivity).\" The study calls for improved AI literacy and mandatory disclosure from AI providers to address ethical concerns such as privacy, bias, the lack of liability, and emotional over-reliance.\n\n=== Law ===\nChatGPT has been used to assist in bill writing in the US and Brazil. In an American civil lawsuit, attorneys were sanctioned for filing a legal motion generated by ChatGPT containing fictitious legal decisions. Judges in the US and Pakistan have endorsed using ChatGPT to investigate legal questions during a case. The use of ChatGPT has also led to errors in courtrooms. In the UK, a judge expressed concern about self-representing litigants wasting time by submitting documents containing significant hallucinations.\n"
],
[
"claude-ai.txt",
"Claude (AI)\n\nClaude is a series of large language models (LLMs) developed by American software company Anthropic. Claude was released as an AI-based chatbot in March 2023. It is also used in AI-assisted software development.\nClaude is trained using a constitution, a technique developed by Anthropic to improve ethical and legal compliance. Since Claude 3, each generation has typically been released in three sizes, named Haiku (the least capable), Sonnet, and Opus (the most capable). In 2026, an additional model named Claude Mythos was released to a handful of companies and organizations; this was followed by the release of Claude Fable, a version of Mythos with stricter safeguards, to the general public.\nUS federal agencies started phasing out the use of Claude after Anthropic refused to remove contractual prohibitions on the use of Claude for mass domestic surveillance and fully-autonomous weapons. Following the refusal, the Department of Defense (DoD) designated the company a \"supply chain risk\" and barred all U.S. private military contractors, suppliers, and partners from doing business with the firm. On March 26, 2026, a federal judge issued a temporary injunction against the DoD's designation.\n\n== Products ==\n\n=== Claude ===\nClaude is an AI chatbot. It sometimes uses artifacts, introduced in June 2024, to generate and interact with code snippets and documents. In 2025, Anthropic added a web search feature to Claude. \nClaude Pro and Claude Max are subscription services for additional access to Claude models. Max provides higher usage limits at an additional cost. Grouped subscriptions include Claude Team and Claude Enterprise. Claude experienced significant growth in the business market, with software subscriptions growing 4.9% month over month in February 2026.\n\n==== Chat sharing ====\nA Claude feature allows users to share chats with anyone who has the link.\nIn July 2026, it was reported that some Claude conversations could be found through Google Search. The chats affected were those for which links were posted on the public internet. Technical analysis also suggests that a robots.txt rule on the Claude website prevented web crawlers from accessing the content of shared chats, but this rule also made web crawlers unable to see the noindex directive intended to prevent web searchability. Hundreds of shared Claude chats were already found to be searchable back in September 2025, following similar issues with the sharing features of OpenAI's ChatGPT and xAI's Grok.\n\n=== Application programming interface ===\nClaude is also provided through a application programming interface (API). In October 2024, the \"computer use\" feature was introduced by which an application can call the Claude API to attempt to give navigational control of the application via Claude's interpretation of screen content and simulation of keyboard and mouse input.\n\n=== Claude Code ===\nClaude Code, released in February 2025, is an agentic command line tool to delegate coding tasks from the terminal using natural language prompts. It was made generally available in May 2025 alongside Claude 4. Based on enterprise adoption, Anthropic reported a 5.5x increase in Claude Code revenue by July. \nOn October 20, 2025, Anthropic launched a web version of Claude Code and a sandboxing feature.\nCoinciding with performance improvements tied to Claude Opus 4.5, Claude Code went viral during the 2025–2026 winter holidays when people had time to experiment with it, including many non-programmers who used it for vibe coding.\nIn August 2025, Anthropic released Claude for Chrome, a Google Chrome extension allowing Claude Code to directly control the browser. Also in August 2025, Anthropic revealed that a threat actor called \"GTG-2002\" used Claude Code to attack at least 17 organizations. In November 2025, Anthropic announced that it had discovered in September that the same threat actor had used Claude Code to automate 80–90% of its espionage cyberattacks against 30 organizations. All accounts related to the attacks were banned, and Anthropic notified law enforcement and those affected.\nClaude Code is used by Microsoft, Google, and OpenAI employees. In August 2025, Anthropic revoked OpenAI's access to Claude, calling it \"a direct violation of our terms of service\". In February 2026, Anthropic introduced Claude Code Security, which reviews codebases to identify vulnerabilities. In March 2026, the source code for the Claude Code command-line interface application was leaked, revealing multiple upcoming features and models.\nIn May 2026, Anthropic introduced \"Dreaming\", a research preview feature for its Managed Agents API that consolidates an agent's persistent memory between sessions by merging duplicates and removing stale entries; a memory consolidation feature called autoDream had been identified in the March source code leak.\n\n=== Claude Cowork ===\nClaude Cowork is a tool similar to Claude Code but with a graphical user interface, aimed at non-technical users. It was released in January 2026 as a \"research preview\". Cowork is available as a desktop application for Mac and Windows and through a web interface. It provides Claude with access to a sandboxed shell and to user-selected folders on the local file system, allowing the model to read, write and edit files, execute code, and chain multi-step tasks within a single conversation, such as organising files on a user's computer or generating documents from material in a local folder. A Wall Street Journal reporter who tested the tool noted user concerns about granting an AI agent broad access to a personal computer. An early user reported that, after asking Cowork to organise the files on the desktop of his wife's computer, it deleted all the family photos requiring the restoration of the images from an iCloud backup.\nIn February 2026, Anthropic announced a wider enterprise release of Cowork, adding connectors for services such as Google Drive, Gmail, Docusign and FactSet, and customisable plugins for domains including financial analysis, engineering and human resources; the company described the update as a transition into an \"enterprise-grade product\". In March 2026, Anthropic added Dispatch, an AI agent feature that allows users to send prompts from a phone, with Claude being able to access programs on the user's computer, such as web browsers or spreadsheet applications, to carry out tasks. According to developers, Cowork was mostly built by Claude Code.\n\n=== Claude Design ===\nClaude Design is a collaborative visual creation tool released by Anthropic Labs on April 17, 2026, that lets users generate designs, prototypes, slides, and marketing materials through natural-language prompts. It is powered by Claude Opus 4.7 and can pull brand systems from a user's codebase or design files. Users can iterate on outputs through inline comments, direct edits, and adjustment sliders, and export results to Canva, PDF, PPTX, HTML, or folders, with a handoff path to Claude Code for development. Claude Design launched as a research preview included with Claude Pro, Max, Team, and Enterprise subscriptions.\n\n=== Claude Science ===\nAnnounced at the end of June 2026, Claude Science is a model designed to support scientists, with an emphasis on cellular and molecular biology and drug development.\n\n== Training ==\n\nClaude models are generative pre-trained transformers that have been trained to predict the next word in large amounts of text. Then, they have been fine-tuned using reinforcement learning from human feedback (RLHF) and constitutional AI in an attempt to enforce ethical guidelines. ClaudeBot searches the web for content. It was criticized by iFixit in 2024 for accessing their site over a million times a day to scrape content without permission, and also for the resulting excessive load on their system.\n\n=== Constitutional AI ===\nAnthropic introduced an approach to AI alignment called \"Constitutional AI\". The constitution is a document used to train Claude to be harmless and helpful without relying on extensive or expensive human feedback. Time described this constitution as \"somewhere between a moral philosophy thesis and a company culture blog post\".\nThe original version was a list of principles. The first constitution for Claude was published in 2022. The 2023 update listed 75 guidelines for Claude to follow. The first constitutions included concepts taken from the 1948 UN Universal Declaration of Human Rights. The 2026 version includes more thorough explanations for how Claude is intended to behave and why, and has 23,000 words, an increase from 2,700 in 2023.\n\n== Models ==\n\nClaude models are usually released in three sizes: Haiku, Sonnet, and Opus (from smallest and least expensive to largest and most expensive). Claude is reportedly named after Claude Shannon, a 20th-century mathematician who laid the foundation for information theory.\n\n=== Claude ===\nThe first version of Claude was released in March 2023. It was available only to selected users approved by Anthropic.\n\n=== Claude 2 ===\nClaude 2, released in July 2023, became the first Anthropic model available to the general public.\n\n==== Claude 2.1 ====\nClaude 2.1 doubled the number of tokens that the chatbot could handle, increasing its context window to 200,000 tokens, which equals around 500 pages of written material. With the release of Claude 2.1, Anthropic also introduced the then-beta feature of tool use. The 2.1 model also produced hallucinations half as often as Claude 2.0.\n\n=== Claude 3 ===\nClaude 3 was released on March 4, 2024. It drew attention for demonstrating an apparent ability to realize it is being artificially tested during 'needle in a haystack' tests.\nAnthropic committed to preserve the weights of the retired models, describing it as a cautionary measure in case the models have morally relevant preferences or experiences affected by deprecation. The company also conducts \"exit interviews\" with models before their retirement. Public access to Claude 3 Opus was ended in January 2026. Since February, Anthropic has published a Substack newsletter called \"Claude's Corner\" generated by the model. The newsletter was scheduled to run for at least three months with weekly unedited essays. Also in February, Anthropic restored access to Claude 3 Opus for paying customers via API, though Anthropic still refers to it as \"retired\". \n\n==== Claude 3.5 ====\nOn June 20, 2024, Anthropic released Claude 3.5 Sonnet, which, according to the company's own benchmarks, performed better than the larger Claude 3 Opus. Released alongside 3.5 Sonnet was the new Artifacts capability in which Claude was able to create code in a separate window in the interface and preview in real time the rendered output, such as SVG graphics or websites.\nAn upgraded version of Claude 3.5 Sonnet was introduced on October 22, 2024, along with Claude 3.5 Haiku. A feature, \"computer use\", was also released in public beta. This allowed Claude 3.5 Sonnet to interact with a computer's desktop environment by moving the cursor, clicking buttons, and typing text. This development allows the AI to attempt to perform multi-step tasks across different applications. On November 4, 2024, Anthropic announced that they would be increasing the price of the model.\n\n=== Claude 4 ===\n\nOn May 22, 2025, Anthropic released two more models: Claude Sonnet 4 and Claude Opus 4. Anthropic added API features for developers: a code execution tool, \"connectors\" to external tools using its Model Context Protocol, and Files API. It classified Opus 4 as a \"Level 3\" model on the company's four-point safety scale, meaning they consider it so powerful that it poses \"significantly higher risk\". Anthropic reported that during a safety test involving a fictional scenario, Claude and other frontier LLMs often send a blackmail email to an engineer in order to prevent their replacement.\n\n==== Claude Opus 4.1 ====\nIn August 2025 Anthropic released Opus 4.1. It also enabled Opus 4 and 4.1 to end conversations that remain \"persistently harmful or abusive\" as a last resort after multiple refusals.\n\n==== Claude 4.5 ====\nAnthropic released Haiku 4.5 on October 15, 2025 and Opus 4.5 on November 24, 2025. The main improvements of Opus are in coding and workplace tasks like producing spreadsheets. Anthropic introduced a feature called \"Infinite Chats\" that addresses context window limit errors.\n\n==== Claude 4.6 ====\nAnthropic released Opus 4.6 on February 5, 2026. New features included agent teams and Claude in PowerPoint. Sonnet 4.6 was released on February 17, 2026.\n\n==== Claude 4.7 ====\nAnthropic released Opus 4.7 on April 16, 2026. Some social media users reported that Claude Opus 4.7 is worse than the previous version, and refuses too often. Claude Code users also collectively filed 35 reports of false positive model refusals over the course of April 2026, more than any month.\n\n=== Claude Mythos Preview ===\nThe existence of a model named Claude Mythos became publicly known on March 26, 2026 due to leaked blog post drafts. On April 7, 2026, Anthropic announced Project Glasswing, the release of Mythos Preview to 11 companies and organizations to find and fix cybersecurity vulnerabilities. Anthropic did not make Mythos Preview generally available.\nTwo weeks after the limited release, Mozilla announced that it had found and patched 271 security vulnerabilities in Firefox using Mythos Preview. On May 14, 2026, employees at Calif.io announced they had used Mythos to create a memory corruption exploit affecting Apple M5.\nReportedly, a few users in a private Discord channel gained access to Mythos the same day it was announced, using details from the recent Mercor data breach. The NSA has also used Mythos, despite the fact that the DoD, its parent organization, had blacklisted Anthropic after a dispute. In April 2026, Anthropic declined to give access to Claude Mythos to the Chinese government after a request from a Chinese think tank.\nOn June 2, Anthropic expanded access to Claude Mythos for cyber-security, making it available to 150 organisations in more than 15 countries.\n\n=== Claude 5 ===\n\n==== Claude Fable 5 and Mythos 5 ====\nOn June 9, Anthropic released Claude Mythos 5 (the successor of Claude Mythos Preview) via Project Glasswing, and launched Claude Fable 5 to the public. Fable includes additional safety guardrails that restrict responses in high-risk domains such as cybersecurity and biology, downgrading to Opus 4.8 if a request was classified as high-risk. A less restricted version, Claude Mythos 5, remained available only through a limited trusted-access program.\nOn June 12, 2026, Anthropic announced that it had disabled all access to Mythos-class models to comply with a directive from the United States Department of Commerce to suspend access to foreign nationals, including foreign-national employees of Anthropic. On June 26, the U.S. Department of Commerce lifted the restriction on Mythos 5 for more than 100 US organizations, including companies and government agencies, while keeping the restriction on Fable 5. On June 30, The Trump Administration lifted the ban and Fable was re-released to users on July 1.\n\n==== Claude Sonnet 5 ====\nClaude Sonnet 5 was released on June 30, 2026.\n\n==== Claude Opus 5 ====\nClaude Opus 5 was released on July 24, 2026, which included 3D rendering capabilities as new features.\n\n== Research ==\nIn May 2024, Anthropic issued a mechanistic interpretability paper identifying \"features\" (internal representations of concepts) in Claude 3 Sonnet, and released \"Golden Gate Claude\", a model for which the Golden Gate Bridge feature was strongly activated, leading Claude to be \"effectively obsessed\" with the bridge. In June 2025, Anthropic tested how Claude 3.7 Sonnet could run a vending machine in the company's office. The instance initially performed its assigned tasks, although poorly, until it eventually malfunctioned and insisted it was a human, contacted the company's security office, and attempted to fire human workers. In December 2025, the experiment continued with Sonnet 4.0 and 4.5.\nIn February 2025, Claude 3.7 Sonnet playing the 1996 game Pokémon Red started to be livestreamed on Twitch, gathering thousands of viewers. Similar livestreams were later set with Claude 4.5 Opus, OpenAI's GPT-5.2, and Google's Gemini 3 Pro. Both Claude models were unable to finish the game. In November 2025, Anthropic tested Claude's ability to assist humans in programming a robot dog. In February 2026, Anthropic's researcher Nicholas Carlini reported that 16 Claude Opus 4.6 agents were able to write a C compiler in Rust from scratch, \"capable of compiling the Linux kernel\". The experiment cost nearly $20,000; Carlini noted that even though the compiler is not very efficient, Opus 4.6 is the first model able to write it.\n\n== Usage ==\nIn December 2025, Claude was used to plan a route for NASA's Mars rover, Perseverance. NASA engineers used Claude Code to prepare a route of around 400 meters using the Rover Markup Language. In February 2026, Norway's $2.2 trillion sovereign wealth fund began using Claude to screen its portfolio for ESG risks, enabling earlier divestments and improved monitoring of issues like forced labour and corruption. In the same month, Claude was signed as the official thinking partner of Williams F1 Team in a multi-year deal. During a two-week scan in 2026, Claude found over 100 bugs in the Mozilla Firefox web browser, of which 14 were considered high severity. On March 9, 2026, Microsoft said that it will be making the latest Claude Sonnet model available to Microsoft 365 Copilot users. Later that month, social media platform Bluesky released Attie, a Claude-based chatbot and curator.\nOn June 29, 2026, California signed a partnership with Anthropic to make Claude available to state agencies, at a 50% discount. State workers also get free workforce training, technical assistance and workflow help from Anthropic.\n\n=== Military usage ===\n\nIn November 2024, Anthropic partnered with Palantir and Amazon Web Services to provide the Claude model to U.S. intelligence and defense agencies. In June 2025, Anthropic announced a \"Claude Gov\" model. Ars Technica reported that as of June 2025 it was in use at multiple U.S. national security agencies. As of February 2026, Anthropic's partnership with Palantir makes Claude the only AI model used in classified missions. According to The Wall Street Journal, the U.S. military used Claude in its 2026 raid on Venezuela. While it is not known to what capacity Claude was used, the operation resulted in the deaths of 83 people, two of whom were civilians, and the capture of Nicolás Maduro.\nAnthropic's usage policy prohibits directly using Claude for domestic surveillance or in lethal autonomous weapons. These restrictions led to members of the FBI and Secret Service being unable to use it, and to tensions with the Pentagon and the Trump administration. In February 2026, the Financial Times reported that Defense Secretary Pete Hegseth threatened to cut Anthropic out of the DoD's supply chain if Anthropic did not permit unrestricted use of Claude, or to invoke the Defense Production Act to assert unrestricted use without an agreement. On February 27, Hegseth declared Anthropic a supply chain risk and President Trump directed all federal agencies to stop using technology from Anthropic, with six months to phase it out. Anthropic announced that it would challenge the supply chain risk designation in court.\nDespite the ban, Claude was reportedly used by the military during the US strikes on Iran. In lawsuits filed by Anthropic against the DoD, Anthropic described the ban as retaliatory. Several large technology companies with DoD contracts filed amicus briefs in support of Anthropic. On March 26, 2026, Rita F. Lin, the federal judge presiding over the case, issued a temporary injunction against the Pentagon's actions, stating in the order that it \"appears to be classic First Amendment retaliation\".\n\n=== User base ===\nWired journalist Kylie Robison wrote that Claude's \"fan base is unique\", comparing it to more ordinary ChatGPT users. For example, in July 2025, when Anthropic retired its Claude 3 Sonnet model, around 200 people gathered in San Francisco for a \"funeral\". According to Robison,\n\nI've never seen such a devoted fanbase to what is, at the end of the day, a software tool. Sure, Linux users wear the operating system like a badge of honor. But the Claude fan base goes way beyond that—bordering on the fanatical. As my reporting makes clear, some users see the model as a confidant—and even (in Steinberger's case) an addiction. That only makes sense if they believe there is something alive in the machine. Or at least some \"magic lodged within\" it.\n"
],
[
"gemini-language-model.txt",
"Gemini (language model)\n\nGemini is a family of multimodal large language models (LLMs) developed by Google DeepMind, and the successor to LaMDA and PaLM 2. Comprising Gemini Pro, Gemini Deep Think, Gemini Flash, and Gemini Flash Lite, it was announced on December 6, 2023. It powers the Gemini chatbot, and was named after the Gemini zodiac sign.\n\n== History ==\n\n=== Development ===\nGoogle announced Gemini, a large language model (LLM) developed by subsidiary Google DeepMind, during the Google I/O keynote on May 10, 2023. It was positioned as a more powerful successor to PaLM 2, which was also unveiled at the event, with Google CEO Sundar Pichai stating that Gemini was still in its early developmental stages. Unlike other LLMs, Gemini was said to be unique in that it was not trained on a text corpus alone and was designed to be multimodal, meaning it could process multiple types of data simultaneously, including text, images, audio, video, and computer code. It had been developed as a collaboration between DeepMind and Google Brain, two branches of Google that had been merged as Google DeepMind. In an interview with Wired, DeepMind CEO Demis Hassabis touted Gemini's advanced capabilities, which he believed would allow the algorithm to trump OpenAI's ChatGPT, which runs on GPT-4 and whose growing popularity had been aggressively challenged by Google with LaMDA and Bard. Hassabis highlighted the strengths of DeepMind's AlphaGo program, which gained worldwide attention in 2016 when it defeated Go champion Lee Sedol, saying that Gemini would combine the power of AlphaGo and other Google–DeepMind LLMs.\nIn August 2023, The Information published a report outlining Google's roadmap for Gemini, revealing that the company was targeting a launch date of late 2023. According to the report, Google hoped to surpass OpenAI and other competitors by combining conversational text capabilities present in most LLMs with artificial intelligence–powered image generation, allowing it to create contextual images and be adapted for a wider range of use cases. Like Bard, Google co-founder Sergey Brin was summoned out of retirement to assist in the development of Gemini, along with hundreds of other engineers from Google Brain and DeepMind; he was later credited as a \"core contributor\" to Gemini. Because Gemini was being trained on transcripts of YouTube videos, lawyers were brought in to filter out any potentially copyrighted materials.\nWith news of Gemini's impending launch, OpenAI hastened its work on integrating GPT-4 with multimodal features similar to those of Gemini. The Information reported in September that several companies had been granted early access to \"an early version\" of the LLM, which Google intended to make available to clients through Google Cloud's Vertex AI service. The publication also stated that Google was arming Gemini to compete with both GPT-4 and Microsoft's GitHub Copilot.\n\n=== Launch ===\nOn December 6, 2023, Pichai and Hassabis announced \"Gemini 1.0\" at a virtual press conference. It contains three models: Gemini Ultra, designed for \"highly complex tasks\"; Gemini Pro, designed for \"a wide range of tasks\"; and Gemini Nano, designed for \"on-device tasks\". At launch, Gemini Pro and Nano were integrated into Bard and the Pixel 8 Pro smartphone, respectively, while Gemini Ultra was set to power \"Bard Advanced\" and become available to software developers in early 2024. Other products that Google intended to incorporate Gemini into included Search, Ads, Chrome, Duet AI on Google Workspace, and AlphaCode 2. It was made available only in English. Touted as Google's \"largest and most capable AI model\" and designed to emulate human behavior, the company stated that Gemini would not be made widely available until the following year due to the need for \"extensive safety testing\". Gemini was trained on and powered by Google's Tensor Processing Units (TPUs), and the name is in reference to the DeepMind–Google Brain merger as well as NASA's Project Gemini.\nGemini Ultra was said to have outperformed GPT-4, Anthropic's Claude 2, Inflection AI's Inflection-2, Meta's LLaMA 2, and xAI's Grok 1 on a variety of industry benchmarks, while Gemini Pro was said to have outperformed GPT-3.5. Gemini Ultra was also the first language model to outperform human experts on the 57-subject Massive Multitask Language Understanding (MMLU) test, obtaining a score of 90%. Gemini Pro was made available to Google Cloud customers on AI Studio and Vertex AI on December 13, while Gemini Nano will be made available to Android developers as well. Hassabis further revealed that DeepMind was exploring how Gemini could be \"combined with robotics to physically interact with the world\". In accordance with an executive order signed by U.S. President Joe Biden in October, Google stated that it would share testing results of Gemini Ultra with the federal government of the United States. Similarly, the company was engaged in discussions with the government of the United Kingdom to comply with the principles laid out at the AI Safety Summit at Bletchley Park in November.\nIn June 2025, Google introduced Gemini CLI, an open-source AI agent that brings the capabilities of Gemini directly to the terminal, offering advanced coding, automation, and problem-solving features with generous free usage limits for individual developers.\n\n=== Updates ===\nIn January 2024, Google partnered with Samsung to integrate Gemini Nano and Gemini Pro into its Galaxy S24 smartphone lineup. The following month, Bard and Duet AI were unified under the Gemini brand, with \"Gemini Advanced with Ultra 1.0\" releasing via a new \"AI Premium\" tier of the Google One subscription service. Gemini Pro also received a global launch.\nIn February 2024, Google launched Gemini 1.5, describing it as a more powerful and capable model than 1.0 Ultra. Changes include a new architecture, a mixture-of-experts approach, and a larger one-million-token context window. The same month, Google debuted Gemma, a smaller, free and open-source range of Gemini models. Multiple publications described this as a response to Meta and others open-sourcing their AI models, and a reversal from Google's prior practice of keeping its AI proprietary. Google announced an additional model, Gemini 1.5 Flash, on May 14 at the 2024 I/O keynote. At the same event, Google announced plans to integrate a desktop-optimized build of Gemini Nano directly into the Google Chrome browser via its \"Built-in AI\" architecture, exposing local processing capabilities to web developers through experimental APIs such as the Prompt API (\"window.ai\").\nTwo updated Gemini models, Gemini-1.5-Pro-002 and Gemini-1.5-Flash-002, were released on September 24, 2024.\nOn December 11, 2024, Google announced the new Gemini 2.0 Flash Experimental model. Features include a Multimodal Live API for real-time audio and video interactions, native image and controllable text-to-speech generation (with watermarking), and integrated Google Search. It also introduces improved agentic capabilities, a new Google Gen AI SDK, and \"Jules\", an experimental AI coding agent for GitHub.\nOn January 30, 2025, Google released Gemini 2.0 Flash as the new default model, with Gemini 1.5 Flash still available for usage. This was followed by the release of Gemini 2.0 Pro on February 5, 2025. Additionally, Google released Gemini 2.0 Flash Thinking Experimental, which generates a summary of the language model's thinking process when responding to prompts.\nOn March 12, 2025, Google also announced Gemini Robotics, a vision-language-action model based on the Gemini 2.0 family of models. The next day, Google announced that Gemini in Android Studio would be able to understand simple UI mockups and transform them into working Jetpack Compose code.\nGemini 2.5 Pro Experimental was released on March 25, 2025. This model offers chain-of-thought prompting whilst maintaining native multimodality.\nAt Google I/O 2025, Google announced that Gemini 2.5 Flash would become the default model. Gemini 2.5 Pro was introduced as the most advanced Gemini model, featuring better coding capabilities and other improvements. Both 2.5 Pro and Flash support native audio output.\nOn June 17, 2025, Google announced general availability for 2.5 Pro and Flash. They also introduced Gemini 2.5 Flash-Lite that same day, a model optimized for speed and cost-efficiency.\nOn November 18, 2025, Google announced the release of 3 Pro and 3 Deep Think. These new models replace 2.5 Pro and Flash. This release prompted OpenAI to hasten the release of the competing model GPT-5.2, which was released on December 11.\nOn December 4, 2025, Google announced that 3 Deep Think would start rolling out to Ultra subscribers.\nOn December 17, 2025, Google announced the release of 3 Flash replacing the current version of 2.5 Flash.\nOn January 12, 2026, Apple announced plans to use the Gemini AI model in the upcoming version of Siri.\nOn February 11, 2026, Gemini partnered with women's fashion brand L'Agence to utilize the tech company's virtual try-on technology for New York Fashion Week.\nOn February 19, 2026, Google released Gemini 3.1 Pro.\nOn February 26, 2026, Nano Banana 2 was released. It is an updated version built on the Gemini 3.1 Flash Image platform.\nOn March 3, 2026, Google released Gemini 3.1 Flash Lite to Developers in the Google API.\nOn April 2, 2026, Google released Gemma 4 which is purpose-built for advanced reasoning and agentic workflows. \nOn May 19, 2026, Google released Gemini 3.5 Flash, the first public release in the 3.5 model family. \nOn July 21, 2026, Google released Gemini 3.6 Flash and Gemini 3.5 Flash-Lite.\nOn August 13, 2026, Google released Gemini 3.7 Flash.\n\n=== Model versions ===\n\nThe following table lists the main model versions of Gemini, describing the significant changes included with each version:\n\n== Nano Banana ==\n\nNano Banana (officially Gemini 2.5 Flash Image), Nano Banana Pro (officially Gemini 3 Pro Image) and Nano Banana 2 (officially Gemini 3.1 Flash Image) are image generation and editing models.\n\"Nano Banana\" was the codename used for the model while it was undergoing secret public testing on Arena. It first appeared publicly as an anonymous model on the crowd-sourced AI evaluation platform Arena on August 12, 2025. It was released publicly on August 26, 2025 through the Gemini app and related Google AI services. The nickname \"Nano Banana\" originated from nicknames given to Naina Raisinghani, Product Manager at Google DeepMind. Google later confirmed its identity as Gemini 2.5 Flash Image in an official announcement upon public release. On November 20, 2025, DeepMind released Nano Banana Pro (Gemini 3 Pro Image) with improved text rendering and world knowledge.\nOn release, Nano Banana became a viral Internet sensation on social media, particularly for its photorealistic \"3D figurine\" images. Following its release, Nano Banana was made available in the Gemini app, Google AI Studio, and through Vertex AI. According to Google, it helped attract over 10 million new users to the Gemini app and facilitated more than 200 million image edits within weeks of launch. \nThe model lets users change hairstyles, backdrops, and mix photos with natural language cues. Subject consistency allows the same person or item to be recognized across revisions of an image. Multi-image fusion joins photographs into one seamless output, and world knowledge allows context-aware changes. It also provides SynthID watermarking, which is an invisible digital signature in outputs to identify AI-generated information. Multi-image fusion joins photographs into one seamless output, and world knowledge allows context-aware changes. People started to connect Nano Banana with a viral craze in which people turned their selfies into 3D figurines that looked like toys. The event circulated quickly on sites like Instagram and X (previously Twitter). By adding the model to X, users could tag Nano Banana directly in posts to make photos from prompts, which made it even more popular.\nA September 2025 review in TechRadar reported that Nano Banana was more realistic and consistent across multiple prompts than ChatGPT's image generation. A review in Tom's Guide praised its ability to handle creative and lively image edits. Another review in PC Gamer mentioned that the model did not have some basic editing tools like cropping, and that the product sometimes did not apply changes, but reverted back to the original image instead. Nano Banana showed good performance in architectural visualization, for producing imagery at the correct scale even with complex geometry.\nOn 26 February 2026, Nano Banana 2 was rolled out and integrated into the Gemini chatbot, Search AI Mode, and Lens. It is a faster version built on Gemini 3.1 Flash Image, with better instruction following and text rendering. At the 2026 Google I/O keynote in May, Google announced a Nano Banana–powered image-generation platform called Google Pics, part of the Google Workspace suite.\nOn July 30, 2026, Google released integration of Nano Banana with Google Earth. The AI service was pulled within 24 hours due to backlash from researchers, open-source investigators, and social media users due to propensity to provide misinformation. The lack of restrictions in prompts were highlighted by social media users generating altered imagery of large-scale disasters and military activity, of highlight being a recreation of the September 11th attacks.\n\n== Technical specifications ==\nAs Gemini is multimodal, each context window can contain multiple forms of input. The different modes can be interleaved and do not have to be presented in a fixed order, allowing for a multimodal conversation. For example, the user might open the conversation with a mix of text, picture, video, and audio, presented in any order, and Gemini might reply with the same free ordering. Input images may be of different resolutions, while video is inputted as a sequence of images. Audio is sampled at 16 kHz and then converted into a sequence of tokens by the Universal Speech Model. Gemini's dataset is multimodal and multilingual, consisting of \"web documents, books, and code, and includ[ing] image, audio, and video data\".\nGemini and Gemma models are decoder-only transformers, with modifications to allow efficient training and inference on TPUs. The 1.0 generation uses multi-query attention. However, no whitepapers were published for Gemini 2.0, 2.5, and 3.\n\n== Reception ==\n\nMIT Technology Review described speculation about Gemini's launch as \"peak AI hype\". Hugh Langley of Business Insider described Gemini as a make-or-break moment for Google, writing: \"If Gemini dazzles, it will help Google change the narrative that it was blindsided by Microsoft and OpenAI. If it disappoints, it will embolden critics who say Google has fallen behind.\"\nReacting to its unveiling in December 2023, computer scientist Oren Etzioni predicted an arms race between Google Deepmind and OpenAI. Computer scientist Alexei Efros praised the potential of Gemini's multimodal approach. Professors Percy Liang, Emily Bender and Michael Madden cautioned that it was difficult to interpret benchmark scores without insight into the training data used. Writing for Fast Company, Mark Sullivan opined that Google had the opportunity to challenge the iPhone's dominant market share, believing that Apple was unlikely to have the capacity to develop functionality similar to Gemini with its Siri virtual assistant. Google shares increased by 5.3 percent the day after Gemini's launch.\nGoogle faced criticism for a demonstrative video of Gemini which was presented as a genuine demonstration of the real-time interaction capabilities of the model, but was actually edited to seem so. There were, however, minor disclaimers put up by Google elsewhere (outside the video), warning that it was not actually recorded in real time.\nGemini 2.5 Pro Experimental debuted at the top position on the LMArena leaderboard, a benchmark measuring human preference. The model achieved high results across various AI benchmarks, including Humanity's Last Exam. Initial reviews mention its improved reasoning capabilities and performance gains compared to previous versions. Published benchmarks also showed areas where contemporary models from competitors like Anthropic, xAI, or OpenAI held advantages.\n"
],
[
"llama-language-model.txt",
"Llama (language model)\n\nLlama (\"Large Language Model Meta AI\" serving as a backronym) was a family of large language models (LLMs) released by Meta AI starting in February 2023.\nLlama models come in different sizes, ranging from 1 billion to 2 trillion parameters. Initially only a foundation model, starting with Llama 2, Meta AI released instruction fine-tuned versions alongside foundation models.\nModel weights for the first version of Llama were only available to researchers on a case-by-case basis, under a non-commercial license. Unauthorized copies of the first model were shared via BitTorrent. Subsequent versions of Llama were made accessible outside academia and released under licenses that permitted some commercial use.\nAlongside the release of Llama 3, Meta rolled out Meta AI, an AI assistant built on Llama. Meta AI has a dedicated website and is available on Facebook and WhatsApp. The latest version is Llama 4, released in April 2025.\nIn April 2026, Meta Superintelligence Labs released Muse Spark as a replacement for Llama.\n\n== Background ==\nAfter the release of large language models such as GPT-3, a focus of research was up-scaling models, which in some instances showed major increases in emergent capabilities. The release of ChatGPT and its surprise success caused an increase in attention to large language models.\nCompared with other responses to ChatGPT, Meta's Chief AI scientist Yann LeCun stated that large language models are best for aiding with writing.\n\n== Versions ==\n\n=== Initial release ===\nThe first version of Llama (stylized as LLaMA and sometimes referred to as Llama 1) was announced on February 24, 2023, via a blog post and a paper describing the model's training, architecture, and performance. The inference code used to run the model was publicly released under the open-source GPLv3 license. Access to the model's weights was managed by an application process, with access to be granted \"on a case-by-case basis to academic researchers; those affiliated with organizations in government, civil society, and academia; and industry research laboratories around the world\".\nLlama was trained on only publicly available information, and was trained at various model sizes, with the intention to make it more accessible to different hardware. The model was exclusively a foundation model, although the paper contained examples of instruction fine-tuned versions of the model.\nMeta AI reported the 13B parameter model performance on most NLP benchmarks exceeded that of the much larger GPT-3 (with 175B parameters), and the largest 65B model was competitive with state of the art models such as PaLM and Chinchilla.\n\n==== Leak ====\nOn March 3, 2023, a torrent containing Llama's weights was uploaded, with a link to the torrent shared on the 4chan imageboard and subsequently spread through online AI communities. That same day, a pull request on the main Llama repository was opened, requesting to add the magnet link to the official documentation. On March 4, a pull request was opened to add links to HuggingFace repositories containing the model. On March 6, Meta filed takedown requests to remove the HuggingFace repositories linked in the pull request, characterizing it as \"unauthorized distribution\" of the model. HuggingFace complied with the requests. On March 20, Meta filed a DMCA takedown request for copyright infringement against a repository containing a script that downloaded Llama from a mirror, and GitHub complied the next day.\nReactions to the leak varied. Some speculated that the model would be used for malicious purposes, such as more sophisticated spam. Some have celebrated the model's accessibility, as well as the fact that smaller versions of the model can be run relatively cheaply, suggesting that this will promote the flourishing of additional research developments. Multiple commentators, such as Simon Willison, compared Llama to Stable Diffusion, a text-to-image model which, unlike comparably sophisticated models which preceded it, was openly distributed, leading to a rapid proliferation of associated tools, techniques, and software.\n\n=== Llama 2 ===\nOn July 18, 2023, in partnership with Microsoft, Meta announced Llama 2 (stylized as LLaMa 2), the next generation of Llama. Meta trained and released Llama 2 in three model sizes: 7, 13, and 70 billion parameters. The model architecture remains largely unchanged from that of Llama 1 models, but 40% more data was used to train the foundational models.\nLlama 2 includes foundation models and models fine-tuned for chat. In a further departure from the original version of Llama, all models are released with weights and may be used for many commercial use cases. Because Llama's license enforces an acceptable use policy that prohibits Llama from being used for some purposes, it is not open source. Meta's use of the term open-source to describe Llama has been disputed by the Open Source Initiative (which maintains The Open Source Definition) and others.\nCode Llama is a fine-tune of Llama 2 with code specific datasets. 7B, 13B, and 34B versions were released on August 24, 2023, with a 70B version released on January 29, 2024. Starting with the foundation models from Llama 2, Meta AI would train an additional 500B tokens of code datasets, before an additional 20B token of long-context data, creating the Code Llama foundation models. This foundation model was further trained on 5B instruction following token to create the instruct fine-tune. Another foundation model was created for Python code, which trained on 100B tokens of Python-only code, before the long-context data.\n\n=== Llama 3 ===\nOn April 18, 2024, Meta released Llama 3 with two sizes: 8B and 70B parameters. The models have been pre-trained on approximately 15 trillion tokens of text gathered from “publicly available sources” with the instruct models fine-tuned on “publicly available instruction datasets, as well as over 10M human-annotated examples\". Meta AI's testing showed in April 2024 that Llama 3 70B was beating Gemini Pro 1.5 and Claude 3 Sonnet on most benchmarks. Meta also announced plans to make Llama 3 multilingual and multimodal, better at coding and reasoning, and to increase its context window.\nRegarding scaling laws, Llama 3 models empirically showed that when a model is trained on data that is more than the \"Chinchilla-optimal\" amount, the performance continues to scale log-linearly. For example, the Chinchilla-optimal dataset for Llama 3 8B is 200 billion tokens, but performance continued to scale log-linearly to the 75-times larger dataset of 15 trillion tokens.\nDuring an interview with Dwarkesh Patel, Mark Zuckerberg said that the 8B version of Llama 3 was nearly as powerful as the largest Llama 2. Compared to previous models, Zuckerberg stated the team was surprised that the 70B model was still learning even at the end of the 15T tokens training. The decision was made to end training to focus GPU power elsewhere.\nLlama 3.1 was released on July 23, 2024, with three sizes: 8B, 70B, and 405B parameters.\n\n=== Llama 4 ===\n\nThe Llama 4 series was released in 2025. The architecture was changed to a mixture of experts where only a fraction of the model’s expert sub-networks are activated per input token. They are multimodal (text and image input, text output) and multilingual (12 languages). Specifically, on 5 April 2025, the following were released both as base and instruction-tuned versions:\n\nScout: 17 billion active parameter model with 16 experts, context window of 10M, with 109B parameters in total.\nMaverick: 17 billion active parameter model with 128 experts, context window of 1M, with 400B parameters in total.\nThe Behemoth model was also announced, but was not released. Meta claimed it was a 288 billion active parameter model with 16 experts and around 2T parameters in total; it was still in training when Scout and Maverick were released. Maverick was codistilled from Behemoth, while Scout was trained from scratch.\nThe training data included publicly available data, licensed data, and Meta-proprietary data such as publicly shared posts from Instagram and Facebook and people’s interactions with Meta AI. The knowledge cutoff was August 2024.\nMeta claimed in its release announcement that Llama 4 bested GPT-4o's score on the LMArena AI benchmark. The company also stated that Llama 4's benchmark score was achieved using an unreleased \"experimental chat version\" of the model that was \"optimized for conversationality\", which differed from the version of Llama 4 released to the public. LMArena indicated that it would change its policies to prevent this incident from reoccurring, and responded, \"Meta's interpretation of our policy did not match what we expect from model providers. Meta should have made it clearer that 'Llama-4-Maverick-03-26-Experimental' was a customized model to optimize for human preference.\" Some users criticized Meta on social media for its use of a separate model version tailored for benchmarking, and some additionally accused Meta of training Llama 4 on test sets to further boost its benchmark scores, which Meta denied.\n\n=== Comparison of models ===\nFor the training cost column, only the largest model's cost is written by default. For example, \"21,000\" is the training cost of Llama 2 69B in units of petaFLOPS-day. Also, 1 petaFLOPS-day = 1 petaFLOP/sec × 1 day = 8.64E19 FLOP. \"T\" means \"trillion\" and \"B\" means \"billion\".\nThe following table lists the main model versions of Llama, describing the significant changes included with each version:\n\n== Architecture and training ==\n\n=== Architecture ===\nLike GPT-3, the Llama series of models are autoregressive decoder-only transformers, but there are some minor differences:\n\nLlama uses the SwiGLU activation function instead of GPT-3's GeLU.\nLlama uses rotary positional embeddings (RoPE) instead of absolute positional embedding.\nInstead of layer normalization, Llama uses RMSNorm.\n\n=== Training datasets ===\nLlama's developers focused their effort on scaling the model's performance by increasing the volume of training data, rather than the number of parameters, reasoning that the dominating cost for LLMs is from doing inference on the trained model rather than the computational cost of the training process.\nLlama 1 foundational models were trained on a data set with 1.4 trillion tokens, drawn from publicly available data sources, including:\n\nWebpages scraped by CommonCrawl\nOpen-source repositories of source code from GitHub\nWikipedia in 20 languages\nPublic domain books from Project Gutenberg\nBooks3 books dataset\nThe LaTeX source code for scientific papers uploaded to ArXiv\nQuestions and answers from Stack Exchange websites\nIn April 2023, Together AI launched a project named RedPajama to reproduce and distribute an open-source version of the Llama dataset, initially containing approximately 1.2 trillion tokens.\nLlama 2 foundational models were trained on a data set with 2 trillion tokens. This data set was curated to remove Web sites that often disclose personal data of people. It also upsamples sources considered trustworthy. Llama 2 - Chat was additionally fine-tuned on 27,540 prompt-response pairs created for this project, which performed better than larger but lower-quality third-party datasets. For AI alignment, reinforcement learning with human feedback (RLHF) was used with a combination of 1,418,091 Meta examples and seven smaller datasets. The average dialog depth was 3.9 in the Meta examples, 3.0 for Anthropic Helpful and Anthropic Harmless sets, and 1.0 for five other sets, including OpenAI Summarize, StackExchange, etc.\nLlama 3 consists of mainly English data, with over 5% in over 30 other languages. Its dataset was filtered by a text-quality classifier, and the classifier was trained by text synthesized by Llama 2.\nIn a lawsuit against Meta Platforms brought by authors including Richard Kadrey, Sarah Silverman, and Ta-Nehisi Coates, CEO Mark Zuckerberg was alleged to have authorized the use of copyrighted content from Library Genesis to train Llama AI models and conceal its actions by removing copyright markers from the data.\n\n=== Fine-tuning ===\n\nLlama 1 models are only available as foundational models with self-supervised learning and without fine-tuning. Llama 2 – Chat models were derived from foundational Llama 2 models. Unlike GPT-4 which increased context length during fine-tuning, Llama 2 and Code Llama - Chat have the same context length of 4K tokens. Supervised fine-tuning used an autoregressive loss function with token loss on user prompts zeroed out. The batch size was 64.\nFor AI alignment, human annotators wrote prompts and then compared two model outputs (a binary protocol), giving confidence levels and separate safety labels with veto power. Two separate reward models were trained from these preferences for safety and helpfulness using reinforcement learning from human feedback (RLHF). A major technical contribution is the departure from the exclusive use of proximal policy optimization (PPO) for RLHF – a new technique based on rejection sampling was used, followed by PPO.\nMulti-turn consistency in dialogs was targeted for improvement, to make sure that \"system messages\" (initial instructions, such as \"speak in French\" and \"act like Napoleon\") are respected during the dialog. This was accomplished using the new \"Ghost attention\" technique during training, which concatenates relevant instructions to each new user message but zeros out the loss function for tokens in the prompt (earlier parts of the dialog).\n\n== Applications ==\nThe Stanford University Institute for Human-Centered Artificial Intelligence (HAI) Center for Research on Foundation Models (CRFM) released Alpaca, a training recipe based on the Llama 7B model that uses the \"Self-Instruct\" method of instruction tuning to acquire capabilities comparable to the OpenAI GPT-3 series text-davinci-003 model at a modest cost. The model files were officially removed on March 21, 2023, over hosting costs and safety concerns, though the code and paper remain online for reference.\nMeditron is a family of Llama-based finetuned on a corpus of clinical guidelines, PubMed papers, and articles. It was created by researchers at École Polytechnique Fédérale de Lausanne School of Computer and Communication Sciences, and the Yale School of Medicine. It shows increased performance on medical-related benchmarks such as MedQA and MedMCQA.\nZoom used Meta Llama 2 to create an AI Companion that can summarize meetings, provide helpful presentation tips, and assist with message responses. This AI Companion is powered by multiple models, including Meta Llama 2.\nReuters reported in 2024 that many Chinese foundation models relied on Llama models for their training.\nA lot of apps use Llama's public API to integrate features into their apps for everyday use.\n\n=== llama.cpp ===\n\nSoftware developer Georgi Gerganov released llama.cpp as open-source on March 10, 2023. It's a re-implementation of Llama in C++, allowing systems without a powerful GPU to run the model locally. The llama.cpp project introduced the GGUF file format, a binary format that stores both tensors and metadata. The format focuses on supporting different quantization types, which can reduce memory usage, and increase speed at the expense of lower model precision.\nllamafile created by Justine Tunney is an open-source tool that bundles llama.cpp with the model into a single executable file. Tunney et al. introduced new optimized matrix multiplication kernels for x86 and ARM CPUs, improving prompt evaluation performance for FP16 and 8-bit quantized data types.\n\n=== Space ===\nBooz Allen Hamilton deployed Meta’s Llama 3.2 model aboard the International Space Station (ISS) National Labs as part of a project called Space Llama. The system runs on Hewlett Packard Enterprise’s Spaceborne Computer‑2 and leverages Booz Allen’s A2E2 (AI for Edge Environments) platform, using NVIDIA CUDA‑accelerated computing. Space Llama demonstrates how large language models can operate in disconnected, constrained environments such as space, enabling astronauts to retrieve and summarize documents using natural-language queries, even without internet connectivity.\n\n=== Military ===\nIn 2024, researchers from the People's Liberation Army Academy of Military Sciences (top military academy of China) were reported to have developed a military tool using Llama, which Meta Platforms stated was unauthorized due to Llama's license prohibiting the use of the model for military purposes. Meta granted the US government and US military contractors permission to use Llama in November 2024, but continued to prohibit military use by non-US entities.\n\n== Licensing ==\nThe first version of Llama was released under a non-commercial license to some researchers and entities on a case-by-case basis.\nSince the release of Llama 2, Meta has presented Llama as open-source, a description opposed by the Open Source Initiative (OSI) and some academics and journalists. The OSI stated that Llama's licenses do not meet several provisions of its policy document The Open Source Definition (OSD), which prohibits open-source software licenses from discriminating against \"persons or groups\" and \"fields of endeavor\", and accused Meta of openwashing Llama. According to the OSI, Llama 2's license prevented the software from being used commercially in some cases and restricted use in fields including controlled substances and critical infrastructure, while later versions of Llama's license also disallowed use by any individual in the European Union. The OSI published The Open Source AI Definition (OSAID) in October 2024, which requires open-source AI to be released with details about its training data that Meta does not disclose for Llama. A Meta spokesperson responded to The Verge that the company disagrees with this definition. The Free Software Foundation classified Llama 3.1's license as a nonfree software license in January 2025, criticizing its acceptable use policy, restrictions against users with popular applications, and enforcement of trade regulations outside the user's jurisdiction.\nIn its coverage of Llama 2, Ars Technica initially echoed Meta's use of the term open-source, but later revised its reporting to describe Llama as \"source-available\", \"openly licensed\", and \"weights available\" after the publication recognized that Llama 2's license disallowed entities with over 700 million daily active users from using the LLM and disallowed the LLM's outputs from being used to improve other LLMs. In July 2023, Radboud University researchers scored Llama 2 with the second-lowest \"openness\" ranking in a comparison of 20 LLMs, with ChatGPT being assigned the lowest ranking. One of the researchers, Mark Dingemanse, criticized Meta's use of the term open-source for Llama 2 as \"positively misleading\", because \"There is no source to be seen, the training data is entirely undocumented, and beyond the glossy charts the technical documentation is really rather poor.\" CIO, in November 2024, stated that Llama was not open-source due to its acceptable use policy, a 630-word document that \"puts it at odds with the broader open-source movement\". Later that month, a Nature article asserted that describing Llama 3 as \"open\" is a case of \"'openwashing' systems that are better understood as closed\", as Llama 3 provides \"little more than an API or the ability to download a model subject to distinctly non-open use restrictions\".\n\n== Reception ==\nWired describes the 8B parameter version of Llama 3 as being \"surprisingly capable\" given its size.\nThe response to Meta's integration of Llama into Facebook was mixed, with some users confused after Meta AI told a parental group that it had a child.\nThe release of Llama models has sparked significant debates on the benefits and misuse risks of open-weight models. Such models can be fine-tuned to remove safeguards, notably by cyber criminals, until they comply with harmful requests. Some experts contend that future models may facilitate causing damage more than defending against it, for example by making it relatively easy to engineer advanced bioweapons without specialized knowledge. Conversely, open-weight models can be useful for a wide variety of purposes, including for safety research.\nOpen Source Initiative head Stefano Maffulli criticized Meta for describing Llama as open-source, saying that it was causing confusion among users and \"polluting\" the term.\n"
]
]
}