Reactions from ranked influencers
5 posts
@scaling01 GLM got totally humbled by ARC though We have very little info about returns to scale on ARC, and we don't even know how many tokens Kimi is trained on, so idk still waiting for the official V4 number
@teortaxesTex that's meaningless without references and the Opus 4 reference we have scores like 5% GLM-5.2 gets ~23% @ 744B i think and that's well below Opus 4 pre-train
@teortaxesTex yeah I think I'm committing to 60-70%
@teortaxesTex it's a massive model and performs pretty well on LisanBench and DeepSWE reasoning efficiency was also good
@scaling01 Greenblatt says its pretraining power level is around Opus 4 if ARC is driven by pretraining, that's not good news