Pinned
Diogo Almeida
@CompleteSkeptic
Sane + 🌶️ takes in an insane AI world... AI capabilities researcher: co-created RLHF/ChatGPT @ @openai now trying to right the wrong 🤭 (ceo @typesafeai)
Joined October 2014
this has always been the case! reading papers seems to be an early career thing - eventually you get the perspective that (1) most new things don't work and that (2) focus beats chasing new shiniesPSA: Most biglab people now read almost zero papers and understand ICLR/ICML/NeurIPS to be mainly full of overclaims & fraud. (but there are a few diamonds in the rough of course)- This was a problem in the 2007 Mathematical Contest in Modeling (MCM) (think of it as doing real world ML before it was cool) What makes it so fun and hard is that there are so many additional constraints: - what if a family is boarding together - what if how long a person takesboarding was solved mathematically in 2008. no airline has used it once. why?
- gonna sound like an old-timer but I don't think people realize how bad google was at instruction following at the time - their best technique (FLAN) actually hurt performanceI sometime think about that Jeff Dean interview where he said they had an internal bot before ChatGPT but didn't think it was better than just googling
- I cannot possibly disagree more! This completely ignores the 2nd most important thing to ML impact: data. TL;DR of the RLHF paper is that data >> scale (even SFT closes most of the gap)Presenting my grand unified theory of ML researcher impact: Your impact is directly proportional to how much pain you cause to infra. Fundamentally, you can only inflict pain upon infra if your approach actually works. And the better your approach works the more pain infra is
