@canondetortugas Proceeds to recommend GSPO with MOE

@canondetortugas I think you can get an internally suboptimal response without getting redirected to Opus or refused, though. The refusal is for biology/cyber sec, but weirdly (and disingenuously) not for frontier LLM research

@canondetortugas @DimitrisPapail You wont know though, they made the AI research censorship secret. They just add a steering vector to Fable to make it useless. You pay the full cost and never know

Bad news for GRPO...didn't get refused or routed to Opus.

@canondetortugas it is silent/hidden nerfing

@axel_pond @canondetortugas It’s ok GRPO will be untouched by the steering vector

@canondetortugas PPO supremacy

@canondetortugas @DimitrisPapail 🤣

@canondetortugas Lol

@canondetortugas @DimitrisPapail You crack the code, Dylan, just wait for an email about account suspension

@canondetortugas you don't get to know when sub-par responses are shown for AI/ML. th is is not what Fable thinks of GRPO

@canondetortugas you wouldn't know, anthropic is using the silent sandbag to stifle competition

@canondetortugas Proceeds to speculate that it is "simple, robust policy-gradient cores (REINFORCE/PPO-family with importance-ratio corrections of the kind we discussed ... the differentiator at frontier labs is environments, reward signal quality, and infrastructure, not the surrogate objective"

@canondetortugas Lmao

@DimitrisPapail @axel_pond @canondetortugas Depends if you are looking for new stuff or not. I wouldn't put my hands on the fire as a bet. Would you?

@rupak_53 @canondetortugas +1, sounds like it's silent sandbagging:

https://www-cdn.anthropic.com/d00db56fa754a1b115b6dd7cb2e3c342ee809620.pdf

@canondetortugas They said in their press release they modified the weights to make it dumber on AI questions so people won't realize it has been degraded.