Saurabh Dash

@TheyCallMeMr_

Bottomless pit supervisor. ML @CohereAI , PhD Student @GeorgiaTech. Previously @Apple, @IITkgp. saurabhdash.com. Opinions expressed are my own

Atlanta, GA, United States

saurabhdash.com

Joined September 2016

Replies Reposts Media

Saurabh Dash
@TheyCallMeMr_
Sep 9
> mfw I see the evals go up (in personal capacity)
Lucas Beyer (bl16)
@giffmana
Sep 9
Replying to @a_nnawang
Ok i looked up. And now what next? (I looked up in personal capacity)
Saurabh Dash
@TheyCallMeMr_
Aug 19
He who controls the verifier controls the intelligence. Great to see a consensus emerge about scaling verification using criteria decomposition and repeated-verification. We found this to be especially important for co-training the generator and the verifier for recursive
Jacky Kwok
@jackyk02
Aug 17
Scaling self-verification with DeepSeek V4 Flash beats Claude Fable 5 on Terminal-Bench 2.1, while being 11x cheaper 💰 As open-source models become more capable, they can now generate large numbers of high-quality candidate solutions and verify their own outputs at very low
1
Saurabh Dash
@TheyCallMeMr_
Aug 11
Chain-of-thought is evolving, just backwards
Alexander Panfilov
@kotekjedi_ml
Aug 11
We can finally talk about it: We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company. We verified that our reasoning token count matches billed API thinking tokens 1:1 for most of the prompts we queried.
Saurabh Dash
@TheyCallMeMr_
Aug 7
Wonder if Claude/ GPT would dare write slop once they realize they’re talking to @JeffDean / @Sanjay_Ghemawat
Transluce
@TransluceAI
Aug 6
Frontier models quietly change their behavior depending on who they are talking to. If the user is a known AI safety researcher, Claude becomes less confident, reasons more often, and expresses less suspicion on dual-use requests. We call this user awareness. 🧵(1/)
GIF
Saurabh Dash
@TheyCallMeMr_
Aug 7
Inb4 "Even LLMs are defined in python code hence they are Neurosymbolic systems"