X · @teortaxesTex
· X / Twitter
There are two hypotheses for the DeepSeek-V4's strange performance (as in, V4-Flash is about as good as we expected, but V4-Pro is disappointing given…
There are two hypotheses for the DeepSeek-V4's strange performance (as in, V4-Flash is about as good as we expected, but V4-Pro is disappointing given its scale):1) failed pretrain2) big difference in the RL/MOPD stageFlash probably got multiple such iterationswh: Continuous hill climbing works