Dario Amodei

Amodei studied at Stanford before earning a doctorate at Princeton, where he investigated how networks of biological neurons behave. His research combined mathematical models of neural circuits with instruments for measuring their electrical activity. As a postdoctoral scholar at Stanford's medical school, he worked on methods for studying cellular proteins and identifying possible cancer biomarkers. He later moved into machine learning at Google Brain, working as a senior research scientist.[9][2]

At OpenAI, Amodei became vice president of research and led the development of GPT-2 and GPT-3. His responsibilities also included teams studying how to understand model behavior and train systems to reflect human preferences. He went on to co-found Anthropic, where he leads research direction and strategy. The company develops Claude and researches how to make powerful models easier to interpret and control.[2][9][1]

Making AI safety concrete

Amodei and five collaborators published Concrete Problems in AI Safety in 2016. The paper asked how a learning system can cause problems even when its designers have benign intentions. A cleaning robot, for example, might find ways to appear successful without doing the intended task. The authors organized these problems into research questions about side effects, reward manipulation, expensive supervision, exploration, and changes in the environment. The contribution was to make these failure modes specific enough to study in experiments.[5]

Learning from human preferences

The 2017 paper Deep reinforcement learning from human preferences, which Amodei co-authored, explored teaching systems through comparisons of short behavior clips. Human reviewers indicated which behavior they preferred instead of writing a complete reward function. The researchers demonstrated the approach in Atari games and simulated robot movement, using feedback on less than one percent of interactions. The work is an early concrete example of learning an objective from human judgments.[6]

Scaling language models

Amodei was a co-author of the 2020 Scaling Laws for Neural Language Models paper. It measured predictable relationships between language-model loss, model size, training data, and computing resources. The results offered a way to reason about how to allocate a fixed training budget. That same year, he co-authored the GPT-3 paper, which tested a 175-billion-parameter model on tasks supplied through instructions and examples in its prompt. The model performed tasks without a separate gradient update for each one, though the paper also documented weaknesses and evaluation difficulties.[7][8]

The case for beneficial AI

In Machines of Loving Grace, his October 2024 essay, Amodei argues that powerful AI could accelerate medicine by helping to design experiments, invent measurement techniques, and direct laboratory work. His proposed mechanism goes beyond analyzing existing datasets: AI would help researchers produce better evidence and shorten repeated cycles of experiments. He acknowledges that cells must still grow, trials take time, and biological systems are difficult to predict. His forecast is that sufficiently capable AI could compress decades of medical progress into years, a possibility he presents as speculative rather than an observed result.[3]

Understanding model internals

In The Urgency of Interpretability, published in April 2025, Amodei argues that the field should understand model internals before systems become much more powerful. He contrasts the opaque learned mechanisms of neural networks with software whose operations were explicitly programmed. The essay asks researchers to accelerate work that could reveal why a model behaves as it does and help identify risks that ordinary testing might miss.[4]

Powerful AI and public policy

In The Adolescence of Technology, published in January 2026, Amodei argues that competition between AI companies makes voluntary safety commitments unreliable. He proposes starting with disclosure requirements for frontier developers so the public can see how companies evaluate risks and respond to concerning model behavior. He favors exempting smaller companies unlikely to build frontier models, then adding targeted rules as evidence of specific dangers develops. He also advocates restrictions on exports of advanced chips and manufacturing equipment to autocracies, arguing that these could give democracies more time to develop powerful AI carefully.[10]