Dario Amodei

CEO and co-founder of Anthropic. Former VP of Research at OpenAI. Background in physics (undergrad) and biophysics (PhD). One of the earliest advocates of the scaling hypothesis, first observing it during speech recognition work at Baidu under Andrew Ng in 2014.

Background

Joined AI research at Baidu (2014) under Andrew Ng, working on speech recognition with recurrent neural networks. Became convinced of scaling laws through direct experimentation. Moved to OpenAI, where he worked as VP of Research. Left in 2021 with Daniela Amodei and others to co-found Anthropic, whose mission is the responsible development of AI for the long-term benefit of humanity.

Author of the essays Machines of Loving Grace (2024), which argues powerful AI could compress decades of scientific and economic progress, and The Adolescence of Technology (2025), on the near-term national-security and economic stakes.

Appearances in this wiki

Key positions

• The scaling hypothesis will continue to hold; inductive inference across 10 years of evidence is sufficient prior. His 2017 Big Blob of Compute Hypothesis — a handful of ingredients, cleverness barely matters — predates Sutton’s Bitter Lesson, and he holds that RL now scales ‘the same’ way as pre-training.
• ‘Near the end of the exponential’ — close to a country of geniuses in a data center (hunch 1–3 years; 90% within ten); a soft takeoff between AI-stagnation and FOOM.
• Two exponentials — capability and diffusion, both ‘fast but not infinitely fast’; the binding constraint on buying compute is demand-prediction/bankruptcy risk, not vision.
• Frontier-AI economics are a Cournot oligopoly (cloud-like, 3–4 differentiated players); the API model is durable.
• What’s hard is distribution, not growth — political freedom, equitable benefit, and preventing authoritarian entrenchment are the real problems; favours chip export controls and transparency-first regulation (opposing a 10-year state moratorium with no federal plan).
• ASL framework: if-then capability-triggered safety commitments are the right regulatory structure — for industry self-governance and external regulation alike
• Two AI risk categories: catastrophic misuse (near-term, manageable with filters) and model autonomy (longer-term, requires interpretability)
• Race to the Top: Anthropic’s competitive strategy is to raise the safety floor for the whole industry, not to win by being uniquely responsible