Reactions from ranked influencers
7 posts
as predicted, Sonnet is dead, Opus is dead Fable is still ~1.5x more efficient, but still an impressive score on DeepSWE from Kimi-K3
The Sonnet tier of models was already dead before, with Kimi-K3 I expect the Opus tier to become obsolete too Frontier Labs have to go to 10T models now https://twitter.com/scaling01/status/2077030720476766559
@scaling01 Using fable high vs xhigh with those error bars is certainly a choice
Fable level can easily be reached with 1-2 more iterations of post training
@xeophon *reasoning budget
@scaling01 come on man, you know what you are doing here. you are point at literal noise, deliberately. the CI shows that kimi is within error with the fable levels. that does tell you more about the measurement than the models, but still
@xeophon would you please show me the data trend that tells you that higher reasoning performance leads to lower performance? and looking at the conservative estimates of performance (the lower CI bound) all three Fable variants outscore Kimi-K3