you know RL doesn’t really generalize even at frontier scale bc lab employees are always like “let us know what the model isn’t good at and we’ll fix it for the next one”
you know RL doesn’t really generalize even at frontier scale bc lab employees are always like “let us know what the model isn’t good at and we’ll fix it for the next one”
Users dismissed frontier RL as bug reports or whack-a-mole patching rather than genuine generalization, viewing the manual fixes as proof it fails to scale.
Based on 6 visible X reactions from 34 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@willccbb That's the quiet admission that RL is mostly targeted patching, not capability gain. If it generalized you wouldn't need a human to name the weakness, the model would transfer. Tell-us-what's-broken is whack-a-mole dressed as training. That's the benchmark-vs-generalization gap.
@willccbb frontier rl is just bug reports with a trillion dollar training budget
@willccbb rlhf is just bug reports with extra steps
@willccbb Whole loop is just whack a mole tbh
you know RL doesn’t really generalize even at frontier scale bc lab employees are always like “let us know what the model isn’t good at and we’ll fix it for the next one”
Users dismissed frontier RL as bug reports or whack-a-mole patching rather than genuine generalization, viewing the manual fixes as proof it fails to scale.
Based on 6 visible X reactions from 34 accounts; directional sample.
Ask a question below.
Published answers will appear here.
