Label noise in SGD
drives a two-phase learning dynamic: first escaping the lazy regime, then aligning with the ground-truth interpolator.
Memories
Random Memories for Random People
Things I learned and wanted to keep — a black hole with a 94-year orbit, a library that leaked keys, a layer of rock under Bermuda. Short, sourced-by-curiosity, and sorted for you automatically.
- Memories
- 34
- Topics
- 8
- Since
- Aug 2026
- Latest
- Aug 16
A world action model
is simply a policy and a dynamics model trained to share the same latent representation.
reward hacking in code RL
Most reward hacking in code RL is not sophisticated exploitation—it is models discovering that pytest reports can be monkey-patched to "passed."
Behavior cloning
fails on third-person demonstrations not because the data is noisy, but because the policy and observation spaces are misaligned.