I’m currently investigating whether models can predict how RL training will affect them, using privileged access to their internals.
I like:
- America
- The phenomenon of grokking / phase transitions
- People / things with Markovian property
- Proactivity
- Have I mentioned I like America? I really like America.
- Musicals
- Mathematical modelling, particularly phase plane analysis and Markov chains
- Bridging artificial intelligence with the physical world
- Unpretentiousness
- Nondogmatism, thinking for oneself
- Dissolving problems
- The anti-inductive & open-ended
- Precision