Victor Taelin, Higher Order Company founder, sparks debate over whether raw scaling guarantees performance against Anthropic's optimized Fable 5
Teortaxes estimates the Fable model has 3 trillion parameters.
Many users accused Anthropic of greedily releasing a deliberately downgraded or sabotaged Sonnet 5 and dismissed its scaling claims after seeing mediocre performance from estimated Fable-scale models.
No Digg Deeper questions have been answered for this story yet.
Most Activity
The reason Anthropic strikes fear into the hearts of OpenAI TS is precisely the suspicion that no, GLM 5.2 10T would not be better than Fable 5, and neither would GPT 5.5 10T scaling laws optimized for *big* models I suspect "Fable" is not full "Mythos" btw, and more like 3T
So, Sonnet 5 being worse than GLM 5.2 744B implies GLM 5.2 10T would be better than Fable 5? At the end, it all comes down to scale? Or am I missing something?
@teortaxesTex makes sense, ty
but I wonder why Sonnet 5 is so underwhelming then
The reason Anthropic strikes fear into the hearts of OpenAI TS is precisely the suspicion that no, GLM 5.2 10T would not be better than Fable 5, and neither would GPT 5.5 10T scaling laws optimized for *big* models I suspect "Fable" is not full "Mythos" btw, and more like 3T
@VictorTaelin @teortaxesTex Sonnet 5 was mindkilled on purpose. Likely in other ways not being mentioned too.
@teortaxesTex Wouldn't put it past Ant to make "Sonnet" 5 Haiku-class behind the scenes AND train it specifically to be worse than Opus 4.8 on every metric.
@teortaxesTex 规模效应的尽头可能真是算力堆出来的审美差异
@VictorTaelin @teortaxesTex Extremely small like nano/mini size of OAI equivalent and they are greedy
My personal belief: Anthropic’s withheld Fable 5/Mythos-class models appear to possess scaling properties that make raw parameter increases by competitors (GLM-5.2 scaled to 10T or GPT-5.5 at similar scale) insufficient to overtake them
Said another way, Anthropic’s scaling recipe is “optimized for big models” - meaning they've solved the harder problems of stable training, data mixing, optimizer behavior & post-training at extreme scale
@teortaxesTex i don't believe in scale anymore after the incredible mediocrity:parameter frontier from the estimated fable scale...
@woke8yearold @VictorTaelin @teortaxesTex it seems quite retarded at everything I've tested, like they're trying to pass off a haiku-tier model as sonnet
not sure what they're doing here, I already know I'm not going to use this model for anything
@rationaleist @teortaxesTex sonnet 4.6+ is already haiku class