3 ms·
I think it's already out of date with verifiable reward based RL, e.g. on maths domain. When "correctness" arguments fall, the argument will probably just shift
by krackers 5mo ago
I think it's already out of date with verifiable reward based RL, e.g. on maths domain. When "correctness" arguments fall, the argument will probably just shift to whether it's just "intelligent brute force".
- TheOtherHobbes 5mo ago"stochastic genius"
- gipp 5mo agoThe set of tasks for which "correctness" is formally verifiable (in a way that doesn't put Goodharts Law in hyperdrive) is vanishingly small.