Richard Song
@XingyouSong
Followers
2K
Following
488
Media
87
Statuses
270
Research Scientist @GoogleDeepmind working on Gemini thinking and AutoML. Ex: @OpenAI, @citsecurities, @MSFTResearch.
Joined February 2018
Seeing text-to-text regression work for Google’s massive compute cluster (billion $$ problem!) was the final result to convince us we can reward model literally any world feedback. Paper: Code: Just train a simple encoder-decoder
11
69
496
Thanks so much for the repost, @_akhaliq!!.
Google presents Performance Prediction for Large Systems via Text-to-Text Regression
0
4
50
Thank you both @shaohua0116 (National Taiwan University) and @MediaTek (MediaTek Advanced Research Center) for hosting my talk on LLM reward modeling and regression and applications in e.g. chip design!. Slides here:
1
1
19
RT @yidingjiang: @MinqiJiang I think it also shows how bad the existing exploration methods are. RL works now because pretraining did the h….
0
3
0
I’ve done way better jobs as an AC than what we got for ICML…. esp. in borderline cases, I personally spend hours reading the papers and reviews in detail, overriding mediocre or wrong reviews if necessary. Papers are the lifeline of early-stage scientists, and I wouldn’t dare.
I’ve found the laziest AC or maybe they found me. Look people: if you get invited to be an AC at a top conference like… I don’t know, ICML, and then go M.I.A. during the entire process until you have ChatGPT write your decision summary, and then you don’t even bother to check.
0
1
32
Exactly why we spent last year on LLM regressors -> reward models to simulate expensive world feedback, not just humans. "To discover new ideas that go far beyond existing human knowledge, it is instead necessary to use grounded rewards: signals that arise from the environment
David Silver really hits it out of the park in this podcast. The paper "Welcome to the Era of Experience" is here:
2
46
315
Will be in Singapore for ICLR next week if you want to catch up!
0
0
10
AIME is nearly completely solved (92% on AIME 2024) 🤯 Congratulations to everybody! Very happy to be working in Gemini Thinking and Reasoning - it's been really fun so far.
1/ Gemini 2.5 is here, and it’s our most intelligent AI model ever. Our first 2.5 model, Gemini 2.5 Pro Experimental is a state-of-the-art thinking model, leading in a wide range of benchmarks – with impressive improvements in enhanced reasoning and coding and now #1 on
1
0
7
Congratulations to Gemini Thinking! 🥳.
Google Deep Research is noticeably improved today! Cool to see people's experience so far. Why is it better? A bunch of product development from the team, and the underlying model updating from 1.5 Pro --> 2.0 Flash Thinking.
0
1
82
🔥🔥🔥🔥🔥.
overall, this is decent work for a single intern's project, but i find the claims to be too grand for what is actually presented. publication pressure seems to force writing a 35 page paper on a new codebase, and the substance of the codebase could've been an add-on to SWE-Bench
0
0
4
These are exactly the type of superhuman AI problems which can benefit from the AutoML community's expertise :).
The coolest autonomous coding agent I've seen recently: use AI to write better CUDA kernels to accelerate AI. AutoML is so back! The highest leverage thing you can do with your compute resources is to increase the future productivity of the same compute. It aligns all the stars
0
0
5
In conclusion: We hope this paper can be a valuable starting point and reference for understanding next-token-prediction over numbers. What originally started out as an investigation, turned out to be very deep with connections to error-correcting codes, risk minimization
0
0
3
It's an open question of what are better tokenizations of numbers. We tried basic error-correcting codes (e.g. majority voting over repeats) and found they can reduce the error. We could consider e.g. 𝑝-adic expansions, or even the Stern–Brocot tree as well.
1
0
1