Creator

Quotes

All goal seeking can be well thought of as the maximization of a single, scalar, externally received signal… As part of achieving that we pose many sub-problems … that are really useful for us to solve as part of achieving the overall goal.

The amazing thing about the reward hypothesis is that there’s one, tiny, little scalar juice that you’re trying to maximize… Out of that low level thing… comes, “I want to raise a family”, “I want to have a career where I’m a successful research scientist…”

Slogans

  • Approximate the solution, not the problem (no special cases)
  • Drive from the problem
  • Take the agent’s point of view
  • Don’t ask the agent to achieve what it can’t measure
  • Don’t ask the agent to know what it can’t verify
  • Set measurable goals for subparts of the agent
  • Discriminative models are usually better than generative models
  • Work by orthogonal dimensions. Work issue by issue
  • Work on ideas, not software
  • Experience is the data of AI

Interviews

Writing