Dario Amodei's research

Key papers

• 2015
  Deep Speech 2 (Baidu)
  End-to-end deep learning speech recognition for English and Mandarin.
• 2016
  Concrete Problems in AI Safety (Google Brain)
  An influential agenda describing practical safety problems in machine learning systems.
• 2017
  Deep RL from Human Preferences (OpenAI)
  Early work on learning from human feedback, the basis of RLHF.
• 2020
  Scaling Laws for Neural Language Models
  Showed that model performance improves predictably with more compute, data and parameters.
• 2020
  Language Models are Few-Shot Learners (GPT-3)
  Co-author of the GPT-3 paper.
• 2022
  Constitutional AI (Anthropic)
  Training a model to be harmless using a written set of principles and AI feedback.

The common thread: scaling

Dario's central research bet is that larger models trained on more compute keep getting more capable in predictable ways. He held that view early, helped quantify it in the scaling laws work, and built Anthropic around it. The same conviction is why he takes AI risk seriously: if capabilities keep scaling, safety research has to keep pace.

Read more about scaling laws and Constitutional AI on the Work & Research page. Paper links are on Resources.

Frequently asked questions

What papers has Dario Amodei written?

Notable co-authored papers include Deep Speech 2, Concrete Problems in AI Safety, Deep Reinforcement Learning from Human Preferences, Scaling Laws for Neural Language Models, the GPT-3 paper and Constitutional AI.

Did Dario Amodei invent RLHF?

He co-authored the 2017 paper Deep Reinforcement Learning from Human Preferences, one of the foundational papers behind RLHF, together with colleagues at OpenAI and DeepMind.

Did Dario Amodei work at Google?

Yes. He was a research scientist at Google Brain before joining OpenAI in 2016.

More about Dario Amodei