Announcement
@pmddomingos resurfaces @andrewho03’s claim that LLM generalization is “very poor”
In his July departure post, @andrewho03 said strong coding-test results still leave pull requests needing manual cleanup.
TLDR
Coding challenges aren’t the whole job, @andrewho03 argued in his July OpenAI departure post: he still had to manually “deslop” pull requests. He said he was starting a reinforcement-learning dataset company, initially focused on biology and statistical reasoning, and argued that scaling alone wouldn’t deliver the capabilities labs need.
Combined views
9.5K
2 Sources, first seen ago
Comments