Analysis of OpenRouter API outputs reveals Fable5 shares high behavioral similarity with Chinese LLMs like Kimi and Qwen

@teortaxesTex I think this just measures RL intensity. Smaller model RL more because it's easier

V4-Flash and V4-Pro being basically entirely dissimilar is weird as hell too

one of the problems with the "evil distillation" narrative is the everyone is already distilling each other. not even necessarily on purpose - but there's now a LOT of potential tokens on the internet from Chinese models than whenever Opus 4.8 base model was trained. Anthropic sucks them up, trains on them - maybe they shift the distribution? Maybe that's why if you prompt Anthropic models in Chinese they often think they're Deepseek or Qwen? Every model lineage is a descendent of all the previous generation, regardless of using each other to synth datasets deliberatly

@max_paperclips would be funny if every accusation was an admission, and they've been squeezing every set of open weights dry in base model modes to recover private datasets in other people's pretrains @teortaxesTex

@teortaxesTex Can be explained by RL saturation.

4.6 is the same base as 4.8 with less RL, same way as Mythos / Fable 5 will be relative to 5.x

@teortaxesTex 确实4.8那个手感真的怪怪的

@teortaxesTex Terrible test constructed around Opus. It seems to have very low variance on the task so you only need a few samples to determine its distribution. 300 samples is insufficient for every other model, OAI and DS seem to actually get more random with more RL.

@teortaxesTex Opus 4.7 was 20% closer to GPT 5.2 in temperament than Opus 4.6 and Opus 4.8 was another five percentage points closer still.

Hallucinating errors from the user that were never made. Plus the paranoia.

So far Fable 5 has lacked all of that.

@teortaxesTex Hmmm this seems like, not reliable eval. I kinda want it to be tbh

@teortaxesTex It is the stupid rlhf opus has that makes it incompetent and unpleasantly bitter

@max_paperclips In summary IP law is lame and doesnt stand the test of time. Libertarians talk about this.

@tensecorrection @teortaxesTex not impossible either tbh

@teortaxesTex Opus 4.6/Fable are dogs, meanwhile 4.7 and 4.8 are cats

@valeriyproklov This has nothing to do with spiky, this is a verbal pattern fingerprint

@teortaxesTex Confusing.

Either they are literally distilling unreleased models via some movie-esque corporate espionage, or they are using near identical data sources and RL pipelines.

@teortaxesTex BET 4.6 Jailbroken performs similar to Fab5 with the same system instructions.

@teortaxesTex 4.6 and fable are generic application models. 4.7/8 are 'safety' and benchmark maxxed.

@teortaxesTex Less saturation = less spiky intelligence perhaps.