Entropix creator _xjdr argues reinforcement learning compute, not raw model capability, is the primary bottleneck in AI development

> the real limitation being RL flops right now I wonder what the limits of parallelization with MOPD are obviously, 1000 "experts" with 20 RL steps each are ≈useless compared to 10 experts with 2000 steps. But what about 40 experts @ 500 steps, merging@4 before OPD?

@xandykati98 http://Code.noumena.com and http://noumena.com

@Orwelian84 @jmbollenbacher @_xjdr I spent 2 weeks optimizing metal kernels with Opus struggling to get ahead and 2 days with fable finished the job

@_xjdr the difference between fable and opus 4.8 is way way bigger, qualitatively so, than between Opus 4.6 and 4.8

@_xjdr what are mostly working on rn?

@Rafa_Schwinger @Orwelian84 @_xjdr bigger. for sure. claude 3.5 -> 4 wasnt nearly as big as opus->mythos.

i think its more akin to gpt-3.5 -> gpt-4. definitely on that level.

@jmbollenbacher @Rafa_Schwinger @_xjdr That's very interesting to hear. I did not try fable, so I have no point of reference right now. I look forward to getting to see the stepwise change, though

@Rafa_Schwinger @_xjdr Bigger or smaller than the difference between 3.5 and 4?

@Orwelian84 @_xjdr idk, I wasn't really using them seriously at the time

@_xjdr yeah and no, I mean, we are not in horrible place atm if big labs shut down

still, better model is just better model, often surprises you by picking better path than you expected, but chances are we are hitting cognitive limits

@Rafa_Schwinger @jmbollenbacher @_xjdr Well damn

@_xjdr what's launching?

@teortaxesTex Smart scaling always beats brute force when the engineering gets tough