Yue Wu
@FrankYueWu1
Followers
2K
Following
335
Media
9
Statuses
61
Scaling RL @xAI | Prev. Postdoc @Princeton, CS PhD @UCLA. BSc @PKU1898.
Joined October 2019
FP16 can have a smaller training-inference gap compared to BFloat16, thus fits better for RL. Even the difference between RL algorithms vanishes once FP16 is adopted. Surprising!
11
69
2K
🎉 Thrilled to share that our paper “DISC: Dynamic Decomposition Improves LLM Inference Scaling” has been accepted to #neurips2025 ! 🚀 Here’s how we push reasoning and inference scaling to new heights 🧵 1/n
4
6
18
This is mostly how I imagine postagi life
nothin better than kicking off a couple codex jobs and going back to watching a physics lecture i don't understand (from Prof. Coskun Kocabas talk on topological effects in graphene at @periodiclabs)
12
38
715
If RL has 1 bit per rollout (0/1 reward) what is the amount of information of one rollout in SFT/pretraining?
best paper or blog i've read in a while, highly recommend! John is brilliant and his research sets an example for the rest of us. recently i too have been thinking deeply about how many bits might be learned via one step of RL or SFT.. if you're thinking about this too, lmk!
5
0
28
I don’t often tweet on technical topics but I may have an opposite opinion here…
10
8
384
Hiring for a new team building computer control agents. Join us to build Grok5 / macrohard later this year. DM me! Will send out a job post soon too.
528
1K
7K
Grok4 Fast maximizing intelligence density.
Introducing Grok 4 Fast, a multimodal reasoning model with a 2M context window that sets a new standard for cost-efficient intelligence. Available for free on https://t.co/AnXpIEOhOD,
https://t.co/53pltypvkw, iOS and Android apps, and OpenRouter. https://t.co/3YZ1yVwueV
38
63
752
Join us to build world's largest RL system.
Come join us to work on 🔢🧮 for 🔄! https://t.co/f3elwtEHxM
6
12
192
Remarkable that TML achieves 0-KL RL.
Today Thinking Machines Lab is launching our research blog, Connectionism. Our first blog post is “Defeating Nondeterminism in LLM Inference” We believe that science is better when shared. Connectionism will cover topics as varied as our research is: from kernel numerics to
0
0
18
We built FutureX, the world’s first live benchmark for real future prediction — politics, economy, culture, sports, etc. Among 23 AI agents, #Grok4 ranked #1 🏆 Elon didn’t lie. @elonmusk your model sees further 🚀🍀 LeaderBoard: https://t.co/fwck0NROHZ
229
206
1K
Is X data the key to mastering AI predictions? FutureX transforms prediction markets into a dynamic, real-time benchmark for agents—and Grok4 tops the leaderboard, #1 among 25 models! 🏆Good job @grok ! Check it out: https://t.co/xisDgWdSjw
We built FutureX, the world’s first live benchmark for real future prediction — politics, economy, culture, sports, etc. Among 23 AI agents, #Grok4 ranked #1 🏆 Elon didn’t lie. @elonmusk your model sees further 🚀🍀 LeaderBoard: https://t.co/fwck0NROHZ
2
3
25
You asked for it! Grok4 for free.
Grok 4 is now free for all users worldwide! Simply use Auto mode, and Grok will route complex queries to Grok 4. Prefer control? Choose "Expert" anytime to always use Grok 4. For a limited time, we are rolling out generous usage limits so you can explore Grok 4’s full
60
115
1K
The problem with all these agent companies/products is that since you don’t have access to the underlying weights, the bet you’re making is that your scaffolds are better than the labs. This is hard because: 1) The labs can bake the scaffolds into the model (Claude Code) 2)
36
20
487
We are hiring! Interested in optimizing/scaling RL framework for pretrain scale RL? DM me or apply here:
job-boards.greenhouse.io
Palo Alto, CA
7
85
491
My thoughts on Grok 4 Heavy after 12hrs: Crazy good! “Create an animation of a crowd of people walking to form “Hello world, I am Grok” as camera changes to birds-eye.” And it 1-shotted the *entire* thing. No other model comes close. Watch the full clip.
527
1K
10K
Grok 4 (Thinking) achieves new SOTA on ARC-AGI-2 with 15.9% This nearly doubles the previous commercial SOTA and tops the current Kaggle competition SOTA
243
725
5K