Dario Amodei

Dario Amodei is the CEO of Anthropic and a coauthor of research on practical AI safety problems, reinforcement learning from human preferences, and language-model scaling laws. Previously, he served as OpenAI’s Vice President of Research, where he led development of GPT-2 and GPT-3. His work encompasses both collaborative technical research and leadership of efforts to build more reliable, steerable AI systems.[1][3][4][5][2]

Anthropic’s May 28, 2021 announcement identified Amodei as CEO and his sister Daniela Amodei as President. He described its goal as making fundamental research advances toward more capable, general, reliable AI systems, then deploying them to benefit people. The initial agenda combined model reliability, interpretability, and tighter integration of human feedback into development and deployment.[2]

On February 26, 2026, Amodei publicly refused a request that, in his account, would remove Anthropic’s safeguards against mass domestic surveillance and fully autonomous weapons in its dealings with the Department of War. He supported defense applications of AI, but argued that these two uses raised distinct concerns about civil liberties, reliability, and oversight. His weapons position was qualified: he allowed that fully autonomous weapons might eventually be important for national defense, while judging contemporary frontier AI insufficiently reliable. In his statement, he wrote: “Regardless, these threats do not change our position: we cannot in good conscience accede to their request.”[9]

In Machines of Loving Grace, dated October 2024, Amodei outlined a conditional optimistic future rather than a description of existing capabilities. His hypothetical “country of geniuses in a datacenter” would combine broad expertise, autonomous work over hours or weeks, access to virtual interfaces and physical tools, and many collaborating instances. He argued that such systems could accelerate biology, neuroscience, and economic development, while explicitly describing his predictions as guesses.[6]

His biological argument concerns the whole research process: AI could design and direct experiments, invent measurement techniques, and improve methods as well as analyze data. He speculated that powerful AI could compress 50–100 years of biological and medical progress into 5–10 years. The qualification is central to his framework: intelligence remains limited by experimental times, missing data, intrinsic complexity, human constraints, and physical laws. As he put it, “Intelligence may be very powerful, but it isn’t magic fairy dust.”[6]

In The Adolescence of Technology, dated January 2026, Amodei organized his concerns into five categories: autonomy risks, destructive misuse, misuse for seizing power, economic disruption, and indirect effects. He rejected the argument that AI misalignment is inevitable or probable from first principles, while maintaining that unpredictable behavior combined with greater intelligence and agency creates a serious risk. His proposed defenses combined alignment training, mechanistic interpretability, monitoring and disclosure, and societal rules.[7]

His regulatory recommendations became more demanding during 2026. In January, he advocated starting with transparency requirements and strengthening rules as concrete evidence emerged. In Policy on the AI Exponential, dated June 2026, he called for mandatory third-party testing of frontier models and government authority to block deployment when risks were unacceptable. He proposed limiting that authority to cybersecurity, biological weapons, loss of control, and automated research that could accelerate those risks, with protections against arbitrary decisions and political favoritism.[7][8]

Before Anthropic’s 2021 announcement, Amodei had coauthored research on safety, human supervision, and scaling. Those earlier papers provide technical context for the research priorities he later articulated as CEO.[3][4][5][2]

Amodei earned a doctorate in biophysics at Princeton as a Hertz Fellow and was a postdoctoral scholar at Stanford University School of Medicine. Before joining OpenAI, he worked at Google Brain as a Senior Research Scientist.[1]

In June 2016, Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané submitted Concrete Problems in AI Safety. They framed accidents as unintended harmful behavior arising from poorly designed systems and organized five practical research problems by their source: wrong objectives produce side effects and reward hacking; objectives too expensive to evaluate require scalable supervision; and learning itself raises problems of safe exploration and distributional shift.[3]

Amodei also coauthored Deep reinforcement learning from human preferences, first submitted in June 2017, with Paul Christiano, Jan Leike, Tom B. Brown, Miljan Martic, and Shane Legg. The approach communicates goals through non-expert human comparisons between pairs of trajectory segments. The paper reports solving Atari and simulated robot-locomotion tasks without access to the reward function, using feedback on less than one percent of agent interactions; it separately reports training complex novel behaviors with about an hour of human time. These results concern the reported reinforcement-learning experiments.[4]ⁿ¹

In January 2020, Amodei coauthored Scaling Laws for Neural Language Models with Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, and Jeffrey Wu. The team reported power-law relationships between language-model cross-entropy loss and model size, dataset size, and training compute, with some measured trends spanning more than seven orders of magnitude. Within the paper’s experimental assumptions, the relationships supported allocating a fixed compute budget to very large models trained on relatively modest data and stopped well before convergence.[5]