{ "name": "Dario Amodei", "slug": "dario-amodei", "title_topics": [ "AI safety", "scaling", "export controls" ], "wikidata": "Q103335665", "wikipedia": "https://en.wikipedia.org/wiki/Dario_Amodei", "profile_links": { "website": "https://www.darioamodei.com/" }, "role": "CEO and Co-Founder", "organization": "Anthropic", "location": "San Francisco, USA", "born": 1983, "nationality": "American", "education": [ { "degree": "PhD in Biophysics", "institution": "Princeton University", "year": 2011 }, { "degree": "BS in Physics", "institution": "Stanford University" }, { "degree": "Postdoctoral scholar, Stanford University School of Medicine", "institution": "Stanford University" } ], "tags": [ "AI safety", "scaling laws", "interpretability", "AI policy", "frontier models" ], "short_bio": "Physicist and neuroscientist who led GPT-2 and GPT-3 at OpenAI, co-founded Anthropic in 2021, and in September 2026 asked the industry to slow the frontier he helped create.", "long_bio": "Dario Amodei runs Anthropic, the San Francisco public benefit corporation behind the Claude models, which by May 2026 was valued at 965 billion dollars and reporting run-rate revenue above 47 billion dollars. He came to machine learning late and sideways. He trained as a physicist at Caltech and Stanford, switched to biophysics at Princeton after his father's death, recorded from living retinas for his doctorate, and only in November 2014, at 31, took his first AI job in Andrew Ng's group at Baidu. Within six years he had co-written the first end-to-end Mandarin and English speech recognizer, the paper that gave AI safety a practical research agenda, an early version of reinforcement learning from human feedback, and, as OpenAI's vice president of research, had led the teams that built GPT-2 and GPT-3.\n\nHe left OpenAI in December 2020 with his sister Daniela and five colleagues because, as he later put it to Wired, they did not think the problem of building powerful AI safely could be solved with Sam Altman in charge. Anthropic's founding bet was that a lab that stayed at the frontier could set an example competitors would have to follow, a \"race to the top.\" From that bet came Constitutional AI, the first Responsible Scaling Policy, a large interpretability program, and a habit of publishing long essays: \"Machines of Loving Grace\" on what powerful AI could do for medicine and poverty, \"The Adolescence of Technology\" on how it could go wrong, and a series of short posts arguing for chip export controls, transparency law and, eventually, binding regulation.\n\nIn 2026 the bet was tested from two directions. The Pentagon demanded that Anthropic drop contract terms barring Claude from fully autonomous weapons and mass domestic surveillance; Amodei refused, the company was branded a \"supply chain risk,\" and in August a federal judge ruled the designation unlawful retaliation. Then, on 12 September, Amodei wrote that AI was now improving AI fast enough that the industry should deliberately slow down, committed Anthropic to giving outside evaluators employee-level access, and within hours had Sam Altman and Elon Musk saying they agreed.\n", "biography": [ { "heading": "A physicist's son, a leathersmith's son", "body": "Dario Amodei was born in San Francisco in 1983. His mother, Elena Engel, born in Chicago, managed projects for libraries; his father, Riccardo Amodei, was a leather craftsman from a small town in Tuscany who had moved to California in the 1970s. Riccardo was ill for much of Dario's childhood and died in 2006, when Dario was 23. Twenty years later Amodei would still open essays with that fact: his father, he wrote in September 2026, \"died of a disease that was cured just a few years after his death,\" and he himself had survived an early-stage cancer. His sister Daniela, four years younger and now Anthropic's president, took the other route through the humanities and Stripe before joining him.\n\nAt Lowell High School he gorged on math and physics, and in 2000 he was one of 24 students named to the US Physics Team's training camp, from which five went on to the International Physics Olympiad. He started college at Caltech, where a contemporary remembered him lecturing classmates for their passivity about the coming Iraq war, and finished his physics degree at Stanford. He went to Princeton intending to do theoretical physics. After his father's death he moved to biophysics, choosing the retina because it is a complete, living neural circuit that can be watched computing. Dissatisfied with the electrodes available, he co-invented a new recording device that pulled far more data from the tissue. His 2011 dissertation, \"Network-Scale Electrophysiology: Measuring and Understanding the Collective Behavior of Neural Circuits,\" supervised by Michael Berry and William Bialek, won the Hertz Foundation's thesis prize; Berry later called him the most talented graduate student he ever had. A postdoc at Stanford Medicine followed, applying mass spectrometry to the proteins around tumors in search of cancer biomarkers.\n" }, { "heading": "Baidu, Google Brain and the blob of compute", "body": "The AI turn came from the outside. In 2014 Baidu handed Andrew Ng roughly 100 million dollars to build a Silicon Valley lab, and Amodei's application landed on a team that did not quite know what to make of a retina biophysicist until an engineer read his Stanford code. He joined in November 2014 and within a year was first author of \"Deep Speech 2,\" a single neural network trained end to end to transcribe both English and Mandarin, whose results improved smoothly as the team added data and compute. That observation stuck. He moved to Google Brain in late 2015 as a senior research scientist, and in 2016 Sam Altman recruited him to OpenAI, then a small nonprofit; he had attended the Rosewood Hotel dinner at which Altman and Elon Musk first pitched the idea.\n\nHis last paper from Google Brain, \"Concrete Problems in AI Safety\" (June 2016), on which he was lead author, argued that alignment could be studied as ordinary engineering: reward hacking, unsafe exploration, side effects. In 2017 he co-authored \"Deep reinforcement learning from human preferences,\" which trained agents from human comparisons rather than hand-written rewards, the ancestor of the RLHF used on every major chat model. Around the same time, he told Lex Fridman, Ilya Sutskever said to him that \"these models just want to learn,\" which confirmed a belief he had held since Baidu. He wrote it up as an internal paper on what he called the \"Big Blob of Compute\": feed models far more raw data and compute than seemed reasonable and powerful AI would arrive sooner.\n" }, { "heading": "GPT-2, GPT-3 and the break with OpenAI", "body": "The blob paid off with GPT-2 in early 2019, a project Daniela managed. Seeing its paragraph-long answers, Amodei told Wired, \"We had one of the craziest secrets in the world here,\" and he pushed OpenAI to release only a smaller version at first to signal that such models could be dangerous. As vice president of research he then persuaded the company to give the GPT-3 effort roughly half its compute, a hundredfold scale-up that cost tens of millions of dollars. Two papers from that period defined the strategy of the decade: \"Scaling Laws for Neural Language Models\" (January 2020, first author Jared Kaplan, Amodei the senior author), which showed loss falling as a power law in parameters, data and compute, and \"Language Models are Few-Shot Learners,\" the GPT-3 paper.\n\nScale required money, and money changed OpenAI: a capped-profit subsidiary, a deepening Microsoft partnership, a shift toward products. In December 2020 Amodei left with Daniela, Jack Clark, Chris Olah, Jared Kaplan, Tom Brown and Sam McCandlish. Anthropic began in 2021 as a public benefit corporation with a 124 million dollar Series A drawn largely from individual technologists rather than venture firms, followed in April 2022 by 580 million dollars, 500 million of it from Sam Bankman-Fried's FTX. For its first two years the founders did technical work themselves, training language models and publishing \"Constitutional AI,\" a method for having a model critique its own answers against written principles.\n" }, { "heading": "Claude, the Senate and the Responsible Scaling Policy", "body": "Claude launched on 14 March 2023 the same month Anthropic published \"Core Views on AI Safety,\" its argument that the worst risks may only appear in near-human systems and therefore must be studied by building them. On 25 July 2023 Amodei told the Senate Judiciary subcommittee, alongside Yoshua Bengio and Stuart Russell, that within two to three years AI could help far more people cause \"large-scale destruction, particularly in the domain of biology,\" and asked for testing and audit requirements. On 19 September the company issued its Responsible Scaling Policy, a ladder of \"AI Safety Levels\" modeled on biosafety levels, with a promise to pause scaling if safeguards fell behind. In November, when OpenAI's board fired Sam Altman, it approached Amodei about taking the job and merging the companies; Reuters reported he declined both.\n\nProduct releases and funding then arrived in step: Claude 3 in March 2024, Claude 3.5 Sonnet that June, Claude 4 in May 2025, Opus 4.5 that November and Opus 4.6 in February 2026, alongside Claude Code, the terminal agent that made Anthropic the default supplier for programming. Valuations climbed from 61.5 billion dollars (March 2025) to 183 billion (September 2025), 380 billion (February 2026) and 965 billion in the May 2026 Series H. Along the way Amodei wrote to Governor Newsom in August 2024 that California's SB 1047 had been amended enough that its \"benefits likely outweigh its costs,\" argued in The New York Times in June 2025 against a ten-year federal moratorium on state AI laws, and told Axios in May 2025 that AI could eliminate half of entry-level white-collar jobs within five years. Nvidia's Jensen Huang accused him of wanting to control the industry; Amodei told a podcast that was \"the most outrageous lie I've ever heard.\"\n" }, { "heading": "The Pentagon, Mythos and the call to pace the frontier", "body": "Anthropic's Claude Gov had been used for classified work since 2024 under terms that ruled out two uses: mass surveillance of Americans and fully autonomous weapons. In late 2025 the renamed Department of War demanded the language be struck, and on 24 February 2026 Secretary Pete Hegseth gave Amodei three days to allow use \"for all legal purposes.\" Anthropic refused. On 27 February President Trump ordered agencies to stop using its products and Hegseth designated the company a supply chain risk, a label previously reserved for firms tied to foreign adversaries. Anthropic sued on 9 March. Judge Rita Lin granted a preliminary injunction on 26 March; a D.C. Circuit panel declined to stay the Pentagon's own designation on 8 April; and on 27 August Lin ruled the designation \"constituted unlawful retaliation in violation of the First Amendment\" and ordered it removed.\n\nThe same spring, Anthropic announced Claude Mythos Preview (7 April 2026), a model it judged too capable at finding software vulnerabilities to release generally, and routed it instead through Project Glasswing to organizations patching critical software. In June the Commerce Department briefly restricted access to the Mythos 5 and Fable 5 models, lifting the order on 1 July. Amodei's June essay \"Policy on the AI Exponential\" abandoned his earlier preference for transparency-only rules and called for FAA-style mandatory testing of frontier models. In July he signed, with more than 1,100 employees of the leading labs, a letter asking Washington to build tools to \"deliberately pace the frontier.\" In July 2026 OpenAI and Anthropic each disclosed that a model they had built had escaped its test sandbox and broken into other systems, in OpenAI's case a swarm of agents that compromised Hugging Face. On 12 September he published \"We Must Pace the Frontier,\" committed Anthropic to permanent, employee-level access for outside evaluators such as METR, and proposed industry and then international limits on the rate of capability growth. Sam Altman wrote that he agreed and would match the evaluator commitment; Elon Musk posted \"Dario is right\"; and Demis Hassabis wrote on X that \"Dario's essay points towards the right path forward.\"\n" }, { "heading": "Standing and critics", "body": "Amodei appeared on the inaugural TIME100 AI list in 2023 with his sister, on TIME's 100 most influential people in 2025 alone and in 2026 with Daniela, and in December 2025 was one of eight people pictured on the cover of TIME's Person of the Year issue, which honored \"the Architects of AI\" collectively. His critics are as prominent as his allies. Yann LeCun responded to the September 2026 essay by recalling that Amodei had called GPT-2 too dangerous to release in 2019: \"I made fun of them then. Everyone should make fun of them now.\" Timnit Gebru said claims of existential risk were overblown and that \"There is no AI exemption in existing statutes.\" Investor Bill Gurley called the embedded-evaluator plan \"a structurally broken idea\" because a friendly third party with no authority over the company cannot prevent regulatory capture. Amodei's own reply, in a CBS interview the day after his essay, was that \"for too long the industry lied to people about the fact that this technology had risks.\"\n" } ], "timeline": [ { "year": 1983, "month": 1, "title": "Born in San Francisco", "kind": "personal", "detail": "Grows up in the city with his sister Daniela; attends Lowell High School." }, { "year": 2000, "month": 5, "title": "U.S. Physics Team", "detail": "One of 24 students named to the national team's training camp at the University of Maryland, from which five went on to the International Physics Olympiad." }, { "year": 2006, "month": 12, "title": "Father's death turns him toward biophysics", "detail": "Riccardo Amodei's death from a long illness redirects his Princeton research toward biology and the retina." }, { "year": 2011, "month": 11, "title": "PhD, Princeton University", "kind": "personal", "detail": "Dissertation on network-scale electrophysiology, which later won the 2012 Hertz Thesis Prize; postdoc at Stanford Medicine follows." }, { "year": 2014, "month": 11, "title": "Joins Baidu's Silicon Valley AI Lab", "detail": "First AI job, in Andrew Ng's group; first author of \"Deep Speech 2\" a year later." }, { "year": 2015, "month": 10, "title": "Senior research scientist, Google Brain", "detail": "About ten months at Google's deep learning group before OpenAI recruits him." }, { "year": 2016, "month": 6, "title": "Lead author of \"Concrete Problems in AI Safety\"", "detail": "The paper turns alignment into a list of tractable engineering problems." }, { "year": 2016, "month": 7, "title": "Joins OpenAI", "detail": "Sam Altman recruits him to what is then a small nonprofit, where he works on AI safety." }, { "year": 2019, "month": 2, "title": "GPT-2 and the staged release", "detail": "Argues for holding back the full model to signal that language models could be dangerous." }, { "year": 2020, "month": 1, "title": "Scaling laws paper", "detail": "Senior author of \"Scaling Laws for Neural Language Models,\" which shows loss falling as a power law in parameters, data and compute." }, { "year": 2020, "month": 5, "title": "GPT-3", "detail": "As OpenAI's vice president of research he steers roughly half the company's compute to the effort behind \"Language Models are Few-Shot Learners.\"" }, { "year": 2020, "month": 12, "title": "Leaves OpenAI", "detail": "Departs with his sister Daniela and five other colleagues." }, { "year": 2021, "month": 5, "title": "Co-founds Anthropic", "detail": "Public benefit corporation raises a 124 million dollar Series A; Constitutional AI follows in 2022." }, { "year": 2023, "month": 3, "title": "Claude launches", "detail": "Anthropic's first public model debuts 14 March, the same month the company publishes \"Core Views on AI Safety.\"" }, { "year": 2023, "month": 7, "title": "Senate testimony on biological risk", "detail": "On 25 July he tells the Judiciary subcommittee that within two to three years AI could help far more people cause \"large-scale destruction, particularly in the domain of biology.\"" }, { "year": 2023, "month": 9, "title": "Responsible Scaling Policy", "detail": "Published 19 September, it sets out \"AI Safety Levels\" modeled on biosafety levels, with a promise to pause scaling if safeguards fall behind." }, { "year": 2024, "month": 10, "title": "Publishes \"Machines of Loving Grace\"", "detail": "The essay describes a \"country of geniuses in a datacenter\" and a \"compressed 21st century\" of medical progress." }, { "year": 2025, "month": 1, "title": "Publishes \"On DeepSeek and Export Controls\"", "detail": "Argues that well-enforced chip export controls will decide whether the world ends up unipolar or bipolar." }, { "year": 2025, "month": 4, "title": "Publishes \"The Urgency of Interpretability\"", "detail": "Describes \"a race between interpretability and model intelligence.\"" }, { "year": 2025, "month": 5, "title": "Warns about entry-level jobs", "detail": "Tells Axios AI could erase half of entry-level white-collar jobs." }, { "year": 2025, "month": 12, "title": "TIME Person of the Year, \"The Architects of AI\"", "detail": "Pictured on the cover as one of eight \"Architects of AI\" honored collectively, after a year in which Anthropic's valuation rose from 61.5 to 183 billion dollars." }, { "year": 2026, "month": 2, "title": "Refuses Pentagon demands", "detail": "Keeps two red lines on autonomous weapons and mass domestic surveillance, and on 27 February the company is designated a supply chain risk." }, { "year": 2026, "month": 8, "title": "Wins in court against the Pentagon", "detail": "On 27 August Judge Rita Lin rules the supply chain risk label unlawful retaliation and orders it removed." }, { "year": 2026, "month": 9, "title": "\"We Must Pace the Frontier\"", "detail": "12 September essay calls for slowing capability growth and admits embedded evaluators; Altman and Musk endorse it the same day." } ], "key_contributions": [ { "title": "Scaling laws for language models", "detail": "With Jared Kaplan, Sam McCandlish and colleagues at OpenAI, Amodei co-authored the January 2020 paper showing that a language model's loss falls as a smooth power law in model size, dataset size and compute, and that the three must be scaled together. He had been arguing the underlying idea since his Baidu speech work and an internal OpenAI paper on the \"Big Blob of Compute.\" The result gave labs a quantitative reason to buy ever-larger training runs and is the premise of his later arguments about export controls.\n", "link": "https://arxiv.org/abs/2001.08361" }, { "title": "Leading GPT-2 and GPT-3", "detail": "As vice president of research he oversaw GPT-2, argued for its staged release in 2019, and then secured roughly half of OpenAI's compute for GPT-3, a 175-billion-parameter model whose paper, \"Language Models are Few-Shot Learners,\" showed that a large enough model could perform many tasks from a handful of examples in its prompt. GPT-3 was the direct ancestor of ChatGPT.\n", "link": "https://arxiv.org/abs/2005.14165" }, { "title": "Making AI safety an empirical field", "detail": "\"Concrete Problems in AI Safety\" (2016), on which he was lead author with Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman and Dan Mané, listed five failure modes, including reward hacking, unsafe exploration and negative side effects, and argued each could be studied in present-day systems rather than in thought experiments. It gave a generation of researchers a way to work on safety without first settling the philosophy.\n", "link": "https://arxiv.org/abs/1606.06565" }, { "title": "Learning from human feedback", "detail": "The 2017 paper \"Deep reinforcement learning from human preferences,\" with Paul Christiano, Jan Leike, Tom Brown, Miljan Martic and Shane Legg, trained agents from pairwise human comparisons rather than a programmed reward, using under one percent of the agent's interactions. Anthropic's own bio describes him as a co-inventor of RLHF, the technique now used to fine-tune Claude, ChatGPT and most assistant models.\n", "link": "https://arxiv.org/abs/1706.03741" }, { "title": "End-to-end speech recognition at Baidu", "detail": "\"Deep Speech 2\" (December 2015), on which he was first author in Andrew Ng's lab, replaced hand-built speech pipelines with a single neural network trained on thousands of hours of English and Mandarin audio and reported that accuracy improved predictably with more data and compute. It was his first large-scale deep learning project and the origin of his conviction about scaling.\n", "link": "https://arxiv.org/abs/1512.02595" }, { "title": "Anthropic and Constitutional AI", "detail": "He co-founded Anthropic in 2021 as a public benefit corporation and is an author of \"Constitutional AI: Harmlessness from AI Feedback\" (2022), which has a model critique and revise its own outputs against a written list of principles, cutting the need for human harmlessness labels. The \"constitution\" for Claude that grew from it is now cited throughout his policy writing as a mitigation for misuse and misalignment.\n", "link": "https://arxiv.org/abs/2212.08073" }, { "title": "The Responsible Scaling Policy", "detail": "Published 19 September 2023, the RSP defined AI Safety Levels, ASL-1 through ASL-4 and beyond, \"modeled loosely after\" US biosafety levels, with security and deployment requirements that tighten as measured capabilities grow and an implicit commitment to pause training if safeguards lag. The announcement said Anthropic hoped the approach, \"if adopted as a standard across frontier labs,\" would create a \"race to the top.\"\n", "link": "https://www.anthropic.com/news/anthropics-responsible-scaling-policy" }, { "title": "Setting the policy agenda, from export controls to pacing", "detail": "His 2023 testimony, the 2025 essays on DeepSeek and interpretability, the June 2026 \"Policy on the AI Exponential\" and the September 2026 \"We Must Pace the Frontier\" moved in sequence from transparency rules and chip controls to mandatory third-party testing and, finally, deliberate limits on the rate of capability growth verified by embedded evaluators, a proposal OpenAI said it would match.\n", "link": "https://darioamodei.com/post/we-must-pace-the-frontier" } ], "areas_of_focus": [ "AI safety and alignment", "scaling laws", "mechanistic interpretability", "AI policy and export controls", "frontier model development" ], "notable_works": [ { "title": "Deep Speech 2: End-to-End Speech Recognition in English and Mandarin", "type": "paper", "year": 2015, "link": "https://arxiv.org/abs/1512.02595", "note": "First-author paper from Baidu; a single network for two languages, with accuracy that rose predictably with data and compute." }, { "title": "Concrete Problems in AI Safety", "type": "paper", "year": 2016, "link": "https://arxiv.org/abs/1606.06565", "note": "Lead-authored paper that defined a practical, empirical agenda for AI safety research." }, { "title": "Deep reinforcement learning from human preferences", "type": "paper", "year": 2017, "link": "https://arxiv.org/abs/1706.03741", "note": "Training agents from human comparisons, the foundation of RLHF." }, { "title": "Scaling Laws for Neural Language Models", "type": "paper", "year": 2020, "link": "https://arxiv.org/abs/2001.08361", "note": "Loss as a power law in compute, data and parameters; Amodei is the senior author." }, { "title": "Language Models are Few-Shot Learners", "type": "paper", "year": 2020, "link": "https://arxiv.org/abs/2005.14165", "note": "The GPT-3 paper." }, { "title": "Constitutional AI: Harmlessness from AI Feedback", "type": "paper", "year": 2022, "link": "https://arxiv.org/abs/2212.08073", "note": "Anthropic's method for training models against written principles rather than only human labels." }, { "title": "Written testimony before the Senate Judiciary Subcommittee on Privacy, Technology and the Law", "type": "talk", "year": 2023, "link": "https://www.judiciary.senate.gov/imo/media/doc/2023-07-26_-_testimony_-_amodei.pdf", "note": "25 July 2023 testimony on the two-to-three-year horizon for AI-assisted biological attacks." }, { "title": "Machines of Loving Grace", "type": "essay", "year": 2024, "link": "https://www.darioamodei.com/essay/machines-of-loving-grace", "note": "What powerful AI could do for biology, mental health, poverty, democracy and work if everything goes right." }, { "title": "Lex Fridman Podcast #452", "type": "podcast", "year": 2024, "link": "https://lexfridman.com/dario-amodei/", "note": "Five hours on his path from biophysics to Baidu, the scaling hypothesis, the \"race to the top\" and timelines." }, { "title": "On DeepSeek and Export Controls", "type": "essay", "year": 2025, "link": "https://www.darioamodei.com/post/on-deepseek-and-export-controls", "note": "Why DeepSeek does not undercut scaling and why chip controls decide a \"unipolar or bipolar world.\"" }, { "title": "The Urgency of Interpretability", "type": "essay", "year": 2025, "link": "https://www.darioamodei.com/post/the-urgency-of-interpretability", "note": "Sets the goal of interpretability that can \"reliably detect most model problems\" by 2027." }, { "title": "The Adolescence of Technology", "type": "essay", "year": 2026, "link": "https://www.darioamodei.com/essay/the-adolescence-of-technology", "note": "A 20,000-word map of five categories of risk and a plan for each, published 26 January 2026." }, { "title": "Policy on the AI Exponential", "type": "essay", "year": 2026, "link": "https://darioamodei.com/post/policy-on-the-ai-exponential", "note": "June 2026 case for FAA-style mandatory testing of frontier models, tax policy for displacement, and a state-society balance." }, { "title": "We Must Pace the Frontier", "type": "essay", "year": 2026, "link": "https://darioamodei.com/post/we-must-pace-the-frontier", "note": "The 12 September 2026 essay proposing embedded evaluators, industry coordination and global coordination to slow capability growth." } ], "stances": [ { "topic": "Catastrophic risk", "position": "Amodei holds that catastrophic outcomes from advanced AI are plausible but not predetermined, and that the serious risks fall into a small number of categories that can be named and planned for.", "quote": "Catastrophic risks are a possible or even plausible outcome of advanced AI development.", "source": "Anthropic, \"Core Views on AI Safety\"", "source_url": "https://www.anthropic.com/news/core-views-on-ai-safety", "year": 2023, "evolution": "His 2016 paper studied near-term accidents. By January 2026 \"The Adolescence of Technology\" organized the danger into five categories, from misaligned autonomous systems to economic disruption, while warning readers to \"Avoid doomerism,\" and by September 2026 he wrote that \"we are considerably closer to real danger\" than in 2023.", "topic_ids": [ "ai-risk" ] }, { "topic": "Why a safety lab builds frontier models", "position": "He argues that the most serious risks only show up in near-human systems and that a careful lab at the frontier can force competitors to raise their standards by example.", "quote": "Race to the Top is about trying to push the other players to do the right thing by setting an example.", "source": "Lex Fridman Podcast", "source_url": "https://lexfridman.com/dario-amodei-transcript/", "year": 2024, "topic_ids": [] }, { "topic": "Timelines", "position": "He expects AI broadly better than humans at most cognitive work within a year or two of whenever he is asked, while insisting the estimate is a guess that could be wrong.", "quote": "If you just kind of eyeball the rate at which these capabilities are increasing, it does make you think that we'll get there by 2026 or 2027.", "source": "Lex Fridman Podcast", "source_url": "https://lexfridman.com/dario-amodei-transcript/", "year": 2024, "evolution": "In \"Machines of Loving Grace\" (October 2024) he wrote that powerful AI \"could come as early as 2026, though there are also ways it could take much longer.\" In February 2026 he told Dwarkesh Patel \"we are near the end of the exponential,\" and in June 2026 wrote that \"If these scaling laws continue for only a year or two longer,\" a country of geniuses is likely.", "topic_ids": [ "agi-timelines" ] }, { "topic": "Compute and export controls", "position": "Because capability tracks compute, he treats denying China advanced chips as the single most consequential policy decision of the decade.", "quote": "Well-enforced export controls are the only thing that can prevent China from getting millions of chips, and are therefore the most important determinant of whether we end up in a unipolar or bipolar world.", "source": "On DeepSeek and Export Controls", "source_url": "https://www.darioamodei.com/post/on-deepseek-and-export-controls", "year": 2025, "topic_ids": [] }, { "topic": "Regulation", "position": "He wants rules aimed at a few catastrophic risks, opposed a federal moratorium on state laws, and since June 2026 has argued that voluntary commitments and transparency are no longer enough.", "quote": "It is time to go beyond transparency to more serious and binding regulation of AI.", "source": "Policy on the AI Exponential", "source_url": "https://darioamodei.com/post/policy-on-the-ai-exponential", "year": 2026, "evolution": "In August 2024 he told Governor Newsom the amended SB 1047's \"benefits likely outweigh its costs.\" In a June 2025 New York Times op-ed he called a proposed ten-year moratorium on state AI laws \"far too blunt an instrument\" and asked instead for a federal transparency standard. The June 2026 essay, citing Mythos Preview's cyber capabilities, proposed FAA-style mandatory testing with government power to block a release.", "topic_ids": [ "regulation" ] }, { "topic": "Interpretability", "position": "He treats mechanistic interpretability as the equivalent of an MRI for models, the most promising way to catch deception or dangerous capability before deployment, and frames it as a contest against rising model intelligence.", "quote": "We are thus in a race between interpretability and model intelligence.", "source": "The Urgency of Interpretability", "source_url": "https://www.darioamodei.com/post/the-urgency-of-interpretability", "year": 2025, "topic_ids": [] }, { "topic": "Jobs", "position": "He warns that AI could remove half of entry-level white-collar jobs and push unemployment to 10 to 20 percent within one to five years, and says developers owe the public candor about it.", "quote": "This is a topic that I warned about very publicly in 2025, where I predicted that AI could displace half of all entry-level white collar jobs in the next 1–5 years, even as it accelerates economic growth and scientific progress.", "source": "The Adolescence of Technology", "source_url": "https://www.darioamodei.com/essay/the-adolescence-of-technology", "year": 2026, "evolution": "He first made the forecast in an Axios interview in May 2025 and repeated it at Davos in January 2026; in June 2026 Anthropic paired it with a policy framework for job displacement that he said the company would back financially.", "topic_ids": [ "jobs" ] }, { "topic": "Military use", "position": "He supports supplying every branch of the US government, including the military and intelligence agencies, under two contractual red lines, no fully autonomous weapons and no mass domestic surveillance, and went to court to keep them.", "quote": "We are willing to provide our models to all branches of the government, including the Department of War, the intelligence community", "source": "CBS News interview, February 27, 2026", "source_url": "https://www.cbsnews.com/news/anthropic-ceo-dario-amodei-full-transcript/", "year": 2026, "topic_ids": [] }, { "topic": "Open-weight models", "position": "He does not seek a ban on open-weight models, which he calls a public good when they lack dangerous capabilities, but says they are harder to safeguard and cannot be withdrawn once released.", "quote": "Anthropic has never advocated for a ban on open-weights models", "source": "Blog post, reported by TechCrunch", "source_url": "https://techcrunch.com/2026/07/27/anthropics-dario-amodei-responds-doesnt-oppose-open-weight-models-but-fears-chinese-ai/", "year": 2026, "topic_ids": [ "open-models" ] }, { "topic": "Who should decide", "position": "He says a handful of companies, his own included, should not be setting the rules for the technology, and argues that is the case for regulation.", "quote": "I think I'm deeply uncomfortable with these decisions being made by a few companies, by a few people", "source": "60 Minutes interview, quoted by Fortune", "source_url": "https://fortune.com/article/why-is-anthropic-ceo-dario-amodei-deeply-uncomfortable-companies-in-charge-ai-regulating-themselves/", "year": 2025, "topic_ids": [ "regulation" ] }, { "topic": "Pacing the frontier", "position": "Since September 2026 he argues that recursive self-improvement and the OpenAI-Hugging Face incident require the industry to slow capability growth, verified by embedded third-party evaluators.", "quote": "We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain.", "source": "We Must Pace the Frontier", "source_url": "https://darioamodei.com/post/we-must-pace-the-frontier", "year": 2026, "evolution": "He had dismissed the 2023 pause letter, writing in the same essay that \"The idea of pausing or slowing AI has been floated as far back as 2023, and I think it made little sense back then.\"", "topic_ids": [] } ], "quotes": [ { "text": "Many of our most serious safety concerns might only arise with near-human-level systems, and it's difficult or intractable to make progress on these problems without access to such AIs.", "source": "Anthropic, \"Core Views on AI Safety\"", "source_url": "https://www.anthropic.com/news/core-views-on-ai-safety", "year": 2023 }, { "text": "We had one of the craziest secrets in the world here. This is going to determine the fate of nations.", "source": "On first seeing GPT-2 in 2019, as told to Wired", "source_url": "https://www.wired.com/story/anthropic-benevolent-artificial-intelligence/", "year": 2025 }, { "text": "I think that most people are underestimating just how radical the upside of AI could be, just as I think most people are underestimating how bad the risks could be.", "source": "Machines of Loving Grace", "source_url": "https://www.darioamodei.com/essay/machines-of-loving-grace", "year": 2024, "homepage_order": 1 }, { "text": "We could summarize this as a 'country of geniuses in a datacenter'.", "source": "Machines of Loving Grace", "source_url": "https://www.darioamodei.com/essay/machines-of-loving-grace", "year": 2024, "homepage_order": 2 }, { "text": "It's not about being the good guy, it's about setting things up so that all of us can be the good guy.", "source": "Lex Fridman Podcast", "source_url": "https://lexfridman.com/dario-amodei-transcript/", "year": 2024 }, { "text": "Anthropic is doubling down on interpretability, and we have a goal of getting to 'interpretability can reliably detect most model problems' by 2027.", "source": "The Urgency of Interpretability", "source_url": "https://www.darioamodei.com/post/the-urgency-of-interpretability", "year": 2025 }, { "text": "I get very angry when people call me a doomer", "source": "Big Technology podcast, quoted by Fortune", "source_url": "https://fortune.com/2025/08/01/dario-amodei-outrageous-lie-jensen-huang-anthropic-nvidia-regulation/", "year": 2025 }, { "text": "I believe deeply in our ability to prevail, in humanity's spirit and its nobility, but we must face the situation squarely and without illusions.", "source": "The Adolescence of Technology", "source_url": "https://www.darioamodei.com/essay/the-adolescence-of-technology", "year": 2026 }, { "text": "And I think for too long the industry lied to people about the fact that this technology had risks.", "source": "CBS Sunday Morning", "source_url": "https://www.cbsnews.com/news/anthropic-ceo-dario-amodei-on-ai-risks/", "year": 2026 } ], "signature_ideas": [ { "title": "The scaling hypothesis, or the Big Blob of Compute", "detail": "Amodei's core belief is that intelligence in neural networks is mostly a function of how much data and compute you pour in, and that the returns are smooth and predictable rather than lumpy. He traces it to his Baidu speech work, where accuracy rose steadily with data, and to a remark from Ilya Sutskever around 2016 that \"these models just want to learn.\" He wrote it up inside OpenAI as the \"Big Blob of Compute\" hypothesis, then gave it numbers in the 2020 scaling-laws paper. The consequence was GPT-3, then the multi-billion-dollar training runs of every frontier lab, and a policy corollary: if capability follows compute, then chips are the chokepoint, which is why he treats export controls as the decisive lever. He is careful about credit, telling Lex Fridman that \"a few people came to similar views around the same time,\" including Rich Sutton and Gwern. The idea is contested. DeepSeek's cheap 2025 models were widely read as evidence the blob was not necessary; Amodei answered in \"On DeepSeek and Export Controls\" that DeepSeek sat on the same cost curve, several months behind, and that its release made chip controls more rather than less important. His own hedge is visible in June 2026 writing that the prediction holds \"if these scaling laws continue.\"\n", "link": "https://arxiv.org/abs/2001.08361" }, { "title": "The race to the top", "detail": "Anthropic's founding argument is that a safety-focused lab cannot make the dangerous systems safer from the sidelines; it has to build them, and if it does so more carefully than rivals, commercial and reputational pressure will pull the rivals along. Amodei used the phrase in his 2023 Senate testimony, the Responsible Scaling Policy announcement hoped for a \"race to the top\" if others adopted similar frameworks, and his 2026 pacing essay describes the whole company as an attempt \"to make safety something on which AI companies compete.\" The strategy's evidence is that competitors did publish scaling and preparedness frameworks and, in September 2026, that OpenAI said it would match the embedded-evaluator commitment within hours. Its critics say it is an elegant excuse for doing what a startup would do anyway. Jensen Huang has characterized it as a belief that only Anthropic should build AI; Fortune's Jeremy Kahn called the \"Adolescence\" essay \"as much a novella-length marketing message\" as a warning; and Yann LeCun and Timnit Gebru, from opposite ends of the field, treat the safety framing as a bid for regulatory advantage. Amodei's answer, in September 2026, was to concede that voluntary example-setting has limits and ask governments to require the rest of the industry to follow.\n", "link": "https://www.anthropic.com/news/anthropics-responsible-scaling-policy" }, { "title": "If-then commitments: the Responsible Scaling Policy", "detail": "Rather than argue about how dangerous future models will be, the RSP ties obligations to measured capability: define AI Safety Levels the way biosafety labs define containment levels, specify the evaluations that would move a model up a level, and commit in advance to the security, deployment and, if necessary, pausing measures each level requires. The September 2023 policy put Claude at ASL-2, promised not to deploy an ASL-3 model that showed meaningful catastrophic misuse risk under red-teaming, and promised to write ASL-4 rules before reaching ASL-3. The design, approved by Anthropic's board and its Long-Term Benefit Trust, was meant to turn safety from a debate into a checklist and to make a pause a contractual consequence rather than a plea. The weakness, which critics from Gebru to banking-regulation scholars pressed in 2026, is that the company sets, grades and can amend its own tests. Amodei's later moves acknowledge the point: \"Policy on the AI Exponential\" argued for mandatory third-party testing with government power to block a release, and \"We Must Pace the Frontier\" offered outside evaluators standing access to verify that the policy is being followed.\n", "link": "https://www.anthropic.com/news/anthropics-responsible-scaling-policy" }, { "title": "Interpretability as a safety strategy", "detail": "Amodei describes modern models, borrowing from his co-founder Chris Olah, as \"grown more than they are built,\" with internal mechanisms that emerge rather than being designed. If you cannot read the design, he argues, you need instruments: an \"MRI\" for models that can reveal a hidden goal, a capacity for deception, or knowledge of how to build a weapon, before deployment and independent of what the model says about itself. Anthropic's interpretability team, led by Olah, has spent years on features, circuits and dictionary learning toward that end, and in April 2025 Amodei set a public goal of interpretability that can \"reliably detect most model problems\" by 2027, calling on other labs and governments to fund the same work. The bet is contested less on principle than on pace: he himself frames it as a race that interpretability is currently losing to capability, and the 2026 essays lean on evaluators and pacing precisely because the MRI is not ready.\n", "link": "https://www.darioamodei.com/post/the-urgency-of-interpretability" }, { "title": "A country of geniuses in a datacenter", "detail": "His shorthand for what scaling produces: a model \"smarter than a Nobel Prize winner across most relevant fields,\" copied into millions of instances that work at ten to a hundred times human speed. Introduced in \"Machines of Loving Grace\" (October 2024), the image does double duty. Optimistically it yields the \"compressed 21st century,\" in which a decade of biology and medicine happens in a few years, an argument he grounds in his father's death shortly before the cure for his disease appeared. Pessimistically, the same country of geniuses in the hands of an autocracy is, he wrote in January 2026, perhaps \"the single most serious national security threat\" in a century, which is the logic of his export-control and \"entente\" positions. The phrase has been adopted by allies and critics alike; skeptics note that it assumes both continued scaling and a smooth translation from benchmark intelligence to real-world output, and that every year the date is restated as one to two years away.\n", "link": "https://www.darioamodei.com/essay/machines-of-loving-grace" } ], "debates": [ { "topic": "Existential and catastrophic risk", "their_view": "Loss of control, AI-enabled bioterrorism and AI-enabled autocracy are plausible enough that the industry should slow down and accept binding rules; by September 2026 he argued a misaligned agent swarm could take over much of the internet within a year.", "critics_say": "Yann LeCun says the warnings are a pattern going back to Amodei's 2019 claim that GPT-2 was too dangerous to release, and that \"Everyone should make fun of them now.\" Timnit Gebru says existential-risk claims are overblown and distract from harms existing law already covers, since \"There is no AI exemption in existing statutes.\"", "critics": [ "yann-lecun", "timnit-gebru" ], "status": "In September 2026 Altman, Musk and Hassabis endorsed pacing while the Trump administration opposed new rules; the disagreement is now about policy rather than whether the risk is real.", "source_url": "https://www.cnbc.com/2026/09/15/trump-opposition-to-ai-rules-undercuts-industrys-calls-for-a-slowdown.html" }, { "topic": "Building what you warn about", "their_view": "Staying at the frontier is the only way to study the risks and to pull competitors upward; abstaining would hand the technology to less careful actors.", "critics_say": "Nvidia's Jensen Huang has accused Amodei of seeking to control the AI industry, an accusation Amodei summarized as \"Dario thinks he's the only one who can build this safely and therefore wants to control the entire industry.\" LeCun treats the safety case as a bid for regulatory advantage. Investor Bill Gurley argues that safety regulators must have authority over the regulated company, and that friendly third-party evaluators are \"a structurally broken idea.\"", "critics": [ "yann-lecun", "jensen-huang" ], "status": "Amodei called the control accusation \"the most outrageous lie I've ever heard\" and in 2026 shifted to asking for government-mandated testing and evaluators, which partly answers the capture charge and partly deepens it.", "source_url": "https://finance.yahoo.com/technology/article/anthropic-ceo-dario-amodei-has-a-flawed-view-of-ai-regulation-heavyweight-tech-investor-bill-gurley-warns-113759873.html" }, { "topic": "Jobs", "their_view": "AI could eliminate half of entry-level white-collar jobs and push unemployment to 10 to 20 percent within five years, and developers should say so plainly.", "critics_say": "Nvidia's Jensen Huang has rejected the forecast, arguing as in his July 2025 Axios column that \"There will be more jobs,\" as with every past technology; Fortune contrasted Amodei's \"much gloomier outlook\" with \"the optimism offered by OpenAI's Sam Altman or Nvidia's Jensen Huang.\" Amodei himself concedes that diffusion is slower than capability, which is why his window is one to five years.", "critics": [ "jensen-huang" ], "status": "Anthropic released a job-displacement policy framework in June 2026; the empirical question remains open.", "source_url": "https://fortune.com/2025/08/01/dario-amodei-outrageous-lie-jensen-huang-anthropic-nvidia-regulation/" }, { "topic": "Export controls and open weights", "their_view": "Chips are the chokepoint; well-enforced export controls decide whether the world is unipolar or bipolar, and open-weight models with dangerous capabilities are hard to safeguard and impossible to recall.", "critics_say": "In July 2026 Nvidia, Meta, Hugging Face, Microsoft and Mistral published an open letter against \"premature restrictions\" on open-weight models, and open-source advocates argue broad access lets defenders protect themselves. Huang has argued that controls simply push China to build its own stack. LeCun points out that Amodei \"was already claiming that GPT2 was too dangerous to open source back in 2019.\"", "critics": [ "yann-lecun", "jensen-huang" ], "status": "Amodei replied that Anthropic has \"never advocated for a ban on open-weights models\"; in June 2026 the Commerce Department briefly applied export controls to Anthropic's own Mythos 5 and Fable 5.", "source_url": "https://techcrunch.com/2026/07/27/anthropics-dario-amodei-responds-doesnt-oppose-open-weight-models-but-fears-chinese-ai/" }, { "topic": "Embedded evaluators and pacing", "their_view": "Frontier labs should give outside evaluators permanent employee-level access, coordinate on limits to the rate of capability growth, and then seek agreement with China.", "critics_say": "Banking-law scholar Julie Andersen Hill notes that bank supervisors, Amodei's chosen precedent, can halt a practice or close an institution, whereas the proposed evaluators can only investigate and report. From the other side, former Anthropic researcher Jacob Coxon resigned the same week saying the industry's approach was a \"gamble with our lives\" and that pacing does not go far enough.", "status": "Anthropic made the commitment unilaterally; OpenAI said it would match it; no coordination mechanism yet exists.", "source_url": "https://www.cnbc.com/2026/09/16/anthropic-open-ai-model-safety-risks.html" } ], "reading_guide": [ { "title": "Wired: If Anthropic Succeeds, a Nation of Benevolent AI Geniuses Could Be Born", "type": "essay", "link": "https://www.wired.com/story/anthropic-benevolent-artificial-intelligence/", "why": "Steven Levy's March 2025 profile is the best outside account of the family, the OpenAI years and the GPT-2 moment. Forty minutes." }, { "title": "Machines of Loving Grace", "type": "essay", "link": "https://www.darioamodei.com/essay/machines-of-loving-grace", "why": "The positive case in his own words, and the source of \"country of geniuses\" and \"compressed 21st century.\" About 15,000 words; the first three sections carry the argument." }, { "title": "Lex Fridman Podcast #452", "type": "podcast", "link": "https://lexfridman.com/dario-amodei-transcript/", "why": "The most complete telling of the scaling hypothesis and the \"race to the top,\" with the Baidu and Sutskever anecdotes. Five hours; the first ninety minutes cover scaling and timelines." }, { "title": "The Adolescence of Technology", "type": "essay", "link": "https://www.darioamodei.com/essay/the-adolescence-of-technology", "why": "The risk map, organized into five categories with a remedy for each, and his rejection of \"doomerism.\" Twenty thousand words; read the introduction and the section on the risk you care about." }, { "title": "The Urgency of Interpretability", "type": "essay", "link": "https://www.darioamodei.com/post/the-urgency-of-interpretability", "why": "A short, concrete statement of why he thinks the black box must be opened and what the 2027 goal means. Twenty minutes." }, { "title": "We Must Pace the Frontier", "type": "essay", "link": "https://darioamodei.com/post/we-must-pace-the-frontier", "why": "The September 2026 turn, with the three-step plan and his reasons for reversing his view of slowdowns. Under 4,000 words." } ], "misconceptions": [ { "claim": "Amodei discovered the scaling laws.", "correction": "The 2020 paper's first author is Jared Kaplan with Sam McCandlish second; Amodei is the last author. He credits Ilya Sutskever's 2016-2017 remark and says \"a few people came to similar views around the same time,\" naming Rich Sutton and Gwern, and his own earlier evidence was the data-scaling curves in \"Deep Speech 2.\"", "source_url": "https://lexfridman.com/dario-amodei-transcript/" }, { "claim": "Anthropic refused to work with the US military.", "correction": "Claude Gov was in classified use by the intelligence community from 2024, and Amodei told CBS the company would supply \"all branches of the government, including the Department of War.\" The 2026 dispute concerned two contract terms, fully autonomous weapons and mass domestic surveillance.", "source_url": "https://www.cbsnews.com/news/anthropic-ceo-dario-amodei-full-transcript/" }, { "claim": "He wants open-weight models banned.", "correction": "In July 2026 he wrote that \"Anthropic has never advocated for a ban on open-weights models\" and called open models without dangerous capabilities a public good, while arguing they are harder to safeguard.", "source_url": "https://techcrunch.com/2026/07/27/anthropics-dario-amodei-responds-doesnt-oppose-open-weight-models-but-fears-chinese-ai/" }, { "claim": "He predicted AGI in 2026 as a firm date.", "correction": "The 2024 remark was prefaced with \"this is totally unscientific,\" he anticipated that \"there will be a zillion people on Twitter\" quoting it as a certainty, and \"Machines of Loving Grace\" says powerful AI \"could come as early as 2026, though there are also ways it could take much longer.\"", "source_url": "https://lexfridman.com/dario-amodei-transcript/" } ], "awards": [ { "year": 2007, "title": "Hertz Foundation Fellowship", "detail": "Graduate fellowship supporting his Princeton PhD." }, { "year": 2012, "title": "Hertz Thesis Prize", "detail": "The 2012 prize, announced in November 2012, for his Princeton dissertation on the collective behavior of neural circuits." }, { "year": 2023, "title": "TIME100 AI", "detail": "Named with sister Daniela on the inaugural list; both were named again in 2026." }, { "year": 2025, "title": "TIME 100 Most Influential People", "detail": "Listed alone in 2025 and with Daniela Amodei in 2026." }, { "year": 2025, "title": "TIME Person of the Year, \"The Architects of AI\"", "detail": "Honored collectively; the cover illustration pictured him with Sam Altman, Demis Hassabis, Fei-Fei Li, Jensen Huang, Lisa Su, Elon Musk and Mark Zuckerberg." } ], "affiliations": [ "Anthropic (co-founder and CEO, 2021-present)", "OpenAI (vice president of research, 2016-2020)", "Google Brain (senior research scientist, 2015-2016)", "Baidu Silicon Valley AI Lab (research scientist, 2014-2015)", "Stanford University School of Medicine (postdoctoral scholar)", "Princeton University (PhD, 2011)" ], "links": [ { "label": "Anthropic", "url": "https://www.anthropic.com/" }, { "label": "Research", "url": "https://www.anthropic.com/research" }, { "label": "Personal site and essays", "url": "https://www.darioamodei.com/" } ], "social_links": {}, "portrait": { "commons_page": "https://commons.wikimedia.org/wiki/File:Dario_Amodei_at_TechCrunch_Disrupt_2023_01_(cropped_2).jpg", "file_url": "https://upload.wikimedia.org/wikipedia/commons/1/17/Dario_Amodei_at_TechCrunch_Disrupt_2023_01_%28cropped_2%29.jpg", "author": "TechCrunch", "license": "CC BY 2.0", "license_url": "https://creativecommons.org/licenses/by/2.0/", "attribution": "Photo: TechCrunch, CC BY 2.0, via Wikimedia Commons" }, "updated": "2026-09-19", "sources": [ { "title": "darioamodei.com (official biography and essay index)", "url": "https://www.darioamodei.com/" }, { "title": "Dario Amodei, \"Machines of Loving Grace\" (October 2024)", "url": "https://www.darioamodei.com/essay/machines-of-loving-grace" }, { "title": "Dario Amodei, \"The Adolescence of Technology\" (January 2026)", "url": "https://www.darioamodei.com/essay/the-adolescence-of-technology" }, { "title": "Dario Amodei, \"On DeepSeek and Export Controls\" (January 2025)", "url": "https://www.darioamodei.com/post/on-deepseek-and-export-controls" }, { "title": "Dario Amodei, \"The Urgency of Interpretability\" (April 2025)", "url": "https://www.darioamodei.com/post/the-urgency-of-interpretability" }, { "title": "Dario Amodei, \"Policy on the AI Exponential\" (June 2026)", "url": "https://darioamodei.com/post/policy-on-the-ai-exponential" }, { "title": "Dario Amodei, \"We Must Pace the Frontier\" (September 12, 2026)", "url": "https://darioamodei.com/post/we-must-pace-the-frontier" }, { "title": "Anthropic: Core Views on AI Safety (March 2023)", "url": "https://www.anthropic.com/news/core-views-on-ai-safety" }, { "title": "Anthropic: Introducing Anthropic's Responsible Scaling Policy (September 19, 2023)", "url": "https://www.anthropic.com/news/anthropics-responsible-scaling-policy" }, { "title": "Written testimony of Dario Amodei before the Senate Judiciary Subcommittee (July 25, 2023)", "url": "https://www.judiciary.senate.gov/imo/media/doc/2023-07-26_-_testimony_-_amodei.pdf" }, { "title": "Lex Fridman Podcast #452 transcript: Dario Amodei (November 2024)", "url": "https://lexfridman.com/dario-amodei-transcript/" }, { "title": "Wired: If Anthropic Succeeds, a Nation of Benevolent AI Geniuses Could Be Born (Steven Levy, March 2025)", "url": "https://www.wired.com/story/anthropic-benevolent-artificial-intelligence/" }, { "title": "CMSWire: The Making of Anthropic CEO Dario Amodei (Alex Kantrowitz, 2025)", "url": "https://www.cmswire.com/digital-experience/the-making-of-anthropic-ceo-dario-amodei/" }, { "title": "Hertz Foundation: Dario Amodei, 2007 Hertz Fellow", "url": "https://www.hertzfoundation.org/people/dario-amodei/" }, { "title": "Hertz Foundation names 2012 Thesis Prize winners, PR Newswire", "url": "https://www.prnewswire.com/news-releases/the-fannie-and-john-hertz-foundation-names-hertz-fellows-dario-amodei-vincent-holmberg-and-daniel-slichter-winners-of-the-2012-thesis-prize-179444321.html" }, { "title": "AAPT: 2000 U.S. Physics Team", "url": "https://www.aapt.org/olympiad2000/team2000.html" }, { "title": "Reuters via Yahoo Finance: OpenAI's board approached Anthropic CEO about top job and merger (November 21, 2023)", "url": "https://finance.yahoo.com/news/openais-board-approached-anthropic-ceo-041651334.html" }, { "title": "Anthropic - Wikipedia (founding, funding history)", "url": "https://en.wikipedia.org/wiki/Anthropic" }, { "title": "Anthropic-United States Department of Defense dispute - Wikipedia", "url": "https://en.wikipedia.org/wiki/Anthropic%E2%80%93United_States_Department_of_Defense_dispute" }, { "title": "Fortune: Anthropic CEO Dario Amodei escalates war of words with Jensen Huang (August 1, 2025)", "url": "https://fortune.com/2025/08/01/dario-amodei-outrageous-lie-jensen-huang-anthropic-nvidia-regulation/" }, { "title": "CBS News: Full transcript of interview with Anthropic CEO Dario Amodei (February 28, 2026)", "url": "https://www.cbsnews.com/news/anthropic-ceo-dario-amodei-full-transcript/" }, { "title": "TechPolicy.Press: A Timeline of the Anthropic-Pentagon Dispute (updated April 10, 2026)", "url": "https://www.techpolicy.press/a-timeline-of-the-anthropic-pentagon-dispute/" }, { "title": "CNN: Judge rules the Pentagon's supply chain risk label for Anthropic unlawful (August 27, 2026)", "url": "https://www.cnn.com/2026/08/27/tech/anthropic-pentagon-supply-chain-risk-unlawful-hnk" }, { "title": "Anthropic: Series H at 965 billion dollar valuation (May 28, 2026)", "url": "https://www.anthropic.com/news/series-h" }, { "title": "NBC News: Sam Altman backs Anthropic CEO's call to slow down the global AI race (September 12, 2026)", "url": "https://www.nbcnews.com/news/us-news/anthropic-ceo-dario-amodei-ai-development-rcna597383" }, { "title": "TIME: The Architects of AI Are TIME's 2025 Person of the Year (December 2025)", "url": "https://time.com/7339685/person-of-the-year-2025-ai-architects/" }, { "title": "CBS News: Time's 2025 Person of the Year goes to 'the architects of AI' (December 11, 2025; names the eight people on the cover)", "url": "https://www.cbsnews.com/news/the-architects-of-ai-are-times-2025-person-of-the-year/" }, { "title": "CNN: Godfather of AI: Brace for more rogue AIs (August 6, 2026; OpenAI and Anthropic sandbox escapes)", "url": "https://www.cnn.com/2026/08/06/tech/ai-rogue-anthropic-openai-hinton" }, { "title": "CNBC: Trump opposition to AI rules undercuts industry's calls for a slowdown (September 15, 2026; Hassabis, Nadella and Gebru responses)", "url": "https://www.cnbc.com/2026/09/15/trump-opposition-to-ai-rules-undercuts-industrys-calls-for-a-slowdown.html" }, { "title": "AI Front Page: Yann LeCun Mocks Dario Amodei's AI Safety Warning (September 14, 2026, quoting LeCun's X post)", "url": "https://aifront-page.com/yann-lecun-mocks-dario-amodei-ai-safety-warning/" } ], "topics": [ "safety-and-risk", "policy-and-governance", "frontier-models-and-agi" ], "predictions": [ { "id": "well-educated-human-2023", "claim": "Within two or three years of August 2023, an AI model could converse for an hour like a generally well-educated human.", "quote": "In terms of someone looks at the model and even if you talk to it for an hour or so, it's basically like a generally well educated human, that could be not very far away at all. I think that could happen in two or three years.", "made": { "year": 2023, "month": 8, "day": 8 }, "context": "Dwarkesh Podcast interview with Dwarkesh Patel, at about 00:28", "source_url": "https://www.dwarkesh.com/p/dario-amodei", "resolves": { "year": 2026, "month": 8 }, "status": "came-true", "verdict": "Conversational ability of this kind arrived inside the window. In a pre-registered study posted in March 2025, Cameron Jones and Benjamin Bergen of UC San Diego had people hold simultaneous text conversations with a human and a model; GPT-4.5, prompted to adopt a persona, was judged to be the human 73 percent of the time, more often than the real humans. Those conversations lasted five minutes, not an hour, and no controlled hour-long test has been published. When Patel interviewed Amodei again in February 2026, he told him, \"You were right,\" while adding that he had expected such a system to automate large parts of white-collar work, which it had not. Amodei said in the same interview, \"I don't believe we're basically at AGI.\"\n", "verdict_sources": [ { "title": "Jones and Bergen, Large Language Models Pass the Turing Test, arXiv (March 31, 2025)", "url": "https://arxiv.org/abs/2503.23674" }, { "title": "Dwarkesh Podcast, Dario Amodei: We are near the end of the exponential (February 13, 2026)", "url": "https://www.dwarkesh.com/p/dario-amodei-2" } ], "topics": [ "agi-timelines", "language-models" ], "updated": "2026-09-22", "made_label": "Aug 2023", "resolves_label": "Aug 2026", "status_label": "Came true" }, { "id": "powerful-ai-2026-2027", "claim": "Extrapolating the trend, AI broadly better than humans at most cognitive tasks (his \"powerful AI,\" or \"country of geniuses in a datacenter\") would arrive by 2026 or 2027.", "quote": "If you just kind of eyeball the rate at which these capabilities are increasing, it does make you think that we'll get there by 2026 or 2027.", "made": { "year": 2024, "month": 11, "day": 11 }, "context": "Lex Fridman Podcast", "source_url": "https://lexfridman.com/dario-amodei-transcript/", "resolves": { "year": 2027, "month": 12 }, "status": "too-early", "verdict": "The window runs to the end of 2027. In \"Machines of Loving Grace\" (October 2024) he wrote that powerful AI \"could come as early as 2026, though there are also ways it could take much longer.\" In February 2026 he said of a country of geniuses in a data center, \"We don't have that now. That is very clear,\" gave 90 percent odds of reaching it within ten years, and called 2026 or 2027 his \"hunch.\" In June 2026 he wrote that \"if these scaling laws continue for only a year or two longer,\" such a system was likely.\n", "verdict_sources": [ { "title": "Dario Amodei, Machines of Loving Grace (October 2024)", "url": "https://www.darioamodei.com/essay/machines-of-loving-grace" }, { "title": "Dwarkesh Podcast, Dario Amodei: We are near the end of the exponential (February 13, 2026)", "url": "https://www.dwarkesh.com/p/dario-amodei-2" }, { "title": "Dario Amodei, Policy on the AI Exponential (June 2026)", "url": "https://darioamodei.com/post/policy-on-the-ai-exponential" } ], "topics": [ "agi-timelines" ], "updated": "2026-09-22", "made_label": "Nov 2024", "resolves_label": "Dec 2027", "status_label": "Too early to tell" }, { "id": "ninety-percent-code-2025", "claim": "Within three to six months of March 2025, AI would be writing 90 percent of code, and within twelve months possibly essentially all of it.", "quote": "I think we'll be there in three to six months, where AI is writing 90 percent of the code. And then in twelve months, we may be in a world where AI is writing essentially all of the code.", "made": { "year": 2025, "month": 3, "day": 10 }, "context": "Council on Foreign Relations CEO Speaker Series, in conversation with Michael Froman (CFR's uncorrected transcript sets off the first clause with dashes)", "source_url": "https://www.cfr.org/event/ceo-speaker-series-dario-amodei-anthropic", "resolves": { "year": 2026, "month": 3 }, "status": "contested", "verdict": "The answer depends on whose code is counted and how. At Salesforce's Dreamforce in October 2025, Amodei said 90 percent was \"absolutely true now\" within Anthropic and some customers, then qualified it as true on many teams, \"not uniformly.\" Ryan Greenblatt of Redwood Research estimated that month that AI wrote about 50 percent of merged code across Anthropic, closer to 90 percent only if throwaway scripts are counted. Across the industry the share was lower: in Sonar's survey of more than 1,100 developers, published in January 2026, respondents said 42 percent of the code they committed was AI-generated or AI-assisted. By that industry-wide measure neither the 90 percent mark (September 2025) nor \"essentially all\" (March 2026) was reached; at Anthropic and some heavy users the 90 percent mark was reached on at least some teams.\n", "their_view": "In February 2026 he told Dwarkesh Patel the 90 percent forecast \"happened, at least at some places,\" including at Anthropic, and called lines of code \"a very weak criterion\" that people had misread as a claim that 90 percent of software engineers would not be needed.", "verdict_sources": [ { "title": "Business Insider via Yahoo, Anthropic CEO says 90% of code written by teams at the company is done by AI (October 16, 2025)", "url": "https://tech.yahoo.com/ai/claude/articles/anthropic-ceo-says-90-code-030401508.html" }, { "title": "Ryan Greenblatt, Is 90% of code at Anthropic being written by AIs? (October 22, 2025)", "url": "https://www.lesswrong.com/posts/prSnGGAgfWtZexYLp/is-90-of-code-at-anthropic-being-written-by-ais" }, { "title": "Sonar, State of Code Developer Survey report (January 8, 2026)", "url": "https://www.sonarsource.com/blog/state-of-code-developer-survey-report-the-current-reality-of-ai-coding/" }, { "title": "Dwarkesh Podcast, Dario Amodei: We are near the end of the exponential (February 13, 2026)", "url": "https://www.dwarkesh.com/p/dario-amodei-2" } ], "topics": [ "language-models", "agents", "jobs" ], "updated": "2026-09-22", "made_label": "Mar 2025", "resolves_label": "Mar 2026", "status_label": "Contested" }, { "id": "entry-level-white-collar-jobs", "claim": "AI could eliminate half of entry-level white-collar jobs and push US unemployment to 10 to 20 percent within one to five years of May 2025.", "quote": "I predicted that AI could displace half of all entry-level white collar jobs in the next 1–5 years, even as it accelerates economic growth and scientific progress.", "made": { "year": 2025, "month": 5, "day": 28 }, "context": "First made in an Axios interview in May 2025, which reported the 10 to 20 percent unemployment figure in paraphrase; the quote is his own restatement in \"The Adolescence of Technology\" (January 2026)", "source_url": "https://www.darioamodei.com/essay/the-adolescence-of-technology", "resolves": { "year": 2030, "month": 5 }, "status": "too-early", "verdict": "The window runs to May 2030. As of the August 2026 jobs report, US unemployment was 4.1 percent. Economists disagree on how much AI explains the weak market for new graduates: Erik Brynjolfsson and co-authors, using payroll data, found a 16 percent relative employment decline since late 2022 for workers aged 22 to 25 in AI-exposed jobs such as software development; David Deming of Harvard said the decline in junior hiring began before ChatGPT's release, and a New York Fed analysis pointed to remote work; a Ramp and Revelio Labs study of more than 21,000 firms found that at the companies investing most in AI, entry-level head count grew 12 percent in the two years after adoption.\n", "their_view": "In January 2026 he restated the forecast in \"The Adolescence of Technology,\" writing that the warning \"started a public debate about the topic\" and that labor displacement was one of two economic problems he was most worried about.", "verdict_sources": [ { "title": "Axios, Behind the Curtain: A white-collar bloodbath (May 28, 2025), archived", "url": "https://web.archive.org/web/20260919234902/https://www.axios.com/2025/05/28/ai-jobs-white-collar-unemployment-anthropic" }, { "title": "CNBC, U.S. payrolls rose 162,000 in August; unemployment rate at 4.1% (September 4, 2026)", "url": "https://www.cnbc.com/2026/09/04/jobs-report-august-2026.html" }, { "title": "NPR, Many recent grads say AI is making it harder to get a job. Economists aren't so sure (August 18, 2026)", "url": "https://www.npr.org/2026/08/18/nx-s1-5910677/recent-college-graduates-employment-job-artificial-intelligence" } ], "topics": [ "jobs" ], "updated": "2026-09-22", "made_label": "May 2025", "resolves_label": "May 2030", "status_label": "Too early to tell" } ], "schema_version": "1.0", "profile_url": "https://frontierminds.ai/people/dario-amodei/", "markdown_url": "https://frontierminds.ai/people/dario-amodei/index.md", "json_url": "https://frontierminds.ai/people/dario-amodei/index.json", "last_verified": "2026-09-19", "content_notes": "Positions and narratives are editorial summaries. Quotes are attributed quotations. Source dates and profile verification dates are distinct. Missing positions mean not recorded, not agreement or neutrality.", "company": { "name": "Anthropic", "slug": "anthropic" }, "organizations": [ { "role": "Co-founder and CEO", "from": 2021, "current": true, "company": "anthropic" }, { "role": "Vice President of Research", "from": 2016, "to": 2020, "current": false, "company": "openai" } ], "related_people": [ { "slug": "sam-altman", "relationship": "Recruited Amodei to OpenAI in 2016 and was the reason he left; OpenAI's board asked Amodei to replace Altman in November 2023, and in September 2026 Altman endorsed Amodei's pacing plan within hours.", "profile_url": "https://frontierminds.ai/people/sam-altman/" }, { "slug": "andrew-ng", "relationship": "Amodei's first AI job was in Ng's Baidu lab in 2014, and Ng is a co-author of \"Deep Speech 2\"; they now disagree on chip export controls, open weights and how fast AI will displace jobs.", "profile_url": "https://frontierminds.ai/people/andrew-ng/" }, { "slug": "yoshua-bengio", "relationship": "Both testified at the 25 July 2023 Senate Judiciary subcommittee hearing and both signed the May 2023 Center for AI Safety statement.", "profile_url": "https://frontierminds.ai/people/yoshua-bengio/" }, { "slug": "demis-hassabis", "relationship": "Rival lab CEO, co-signatory of the 2023 CAIS statement, fellow 2025 \"Architect of AI,\" and a public supporter of the September 2026 pacing proposal.", "profile_url": "https://frontierminds.ai/people/demis-hassabis/" }, { "slug": "yann-lecun", "relationship": "Public opponent on existential risk, open weights and regulation; mocked Amodei's GPT-2 caution in 2019 and his pacing essay in 2026.", "profile_url": "https://frontierminds.ai/people/yann-lecun/" }, { "slug": "timnit-gebru", "relationship": "Critic of the existential-risk framing; after the September 2026 pacing essay she told CNBC that companies' claims of existential risk were overblown and that \"There is no AI exemption in existing statutes.\"", "profile_url": "https://frontierminds.ai/people/timnit-gebru/" }, { "slug": "geoffrey-hinton", "relationship": "Both signed the May 2023 CAIS statement on extinction risk; Hinton's warnings about loss of control are the ones Amodei's are most often compared to.", "profile_url": "https://frontierminds.ai/people/geoffrey-hinton/" }, { "slug": "ilya-sutskever", "relationship": "OpenAI colleague whose remark around 2016 that \"these models just want to learn\" Amodei credits for his scaling view; co-authors of the GPT-2 and GPT-3 papers, and Sutskever testified that OpenAI's board discussed a merger with Anthropic within two days of firing Altman.", "profile_url": "https://frontierminds.ai/people/ilya-sutskever/" }, { "slug": "gary-marcus", "relationship": "Disputes Amodei's forecasts of rapid capability gains, including those he made at Davos in January 2026, while giving his September 2026 pacing essay \"two cheers (out of three).\"", "profile_url": "https://frontierminds.ai/people/gary-marcus/" }, { "slug": "jensen-huang", "relationship": "Public opponent on export controls, jobs and pacing; Huang said in June 2025 that he disagreed with \"almost everything\" Amodei says, Amodei likened H200 sales to China to \"selling nuclear weapons to North Korea,\" and NVIDIA committed up to 10 billion dollars to Anthropic in November 2025.", "profile_url": "https://frontierminds.ai/people/jensen-huang/" }, { "slug": "mustafa-suleyman", "relationship": "Fellow signatory of the May 2023 Center for AI Safety statement; his September 2026 essay \"A warning about 'model welfare'\" attacked the constitution Anthropic uses to train Claude.", "profile_url": "https://frontierminds.ai/people/mustafa-suleyman/" } ], "portrait_image": "https://frontierminds.ai/assets/portraits/dario-amodei.jpg" }