Wenting Zhao @wzhao_nlp

reasoning & llms @Alibaba_Qwen Opinions are my own wenting-zhao.github.io NYC Joined June 2013

Tweets

450
Followers

5K
Following

606
Likes

387

Dwarkesh Patel @dwarkesh_sp

2 days ago

.@RichardSSutton, father of reinforcement learning, doesn’t think LLMs are bitter-lesson-pilled. My steel man of Richard’s position: we need some new architecture to enable continual (on-the-job) learning. And if we have continual learning, we don't need a special training…

199 449 3K 1.3M 3K

Download Video

CLS @ChengleiSi

4 days ago

while we are on this, rmb we also had: - Neural Architecture Search with Reinforcement Learning arxiv.org/abs/1611.01578 - Symbolic Discovery of Optimization Algorithms arxiv.org/abs/2302.06675 - Using Large Language Models for Hyperparameter Optimization arxiv.org/abs/2312.04528 -…

$_fracapuano Profile Picture$

Francesco Capuano @_fracapuano

6 days ago

6 45 509 66K 491

Download Image

4 9 40 6K 30

Wenting Zhao @wzhao_nlp

4 days ago

Congrats to the codellama team on the release! Some real good stuff

Gabriel Synnaeve @syhw

4 days ago

Congrats to the codellama team on the release! Some real good stuff

59 296 2K 779K 1K

0 0 4 2K 0

Chetaslua @chetaslua

4 days ago

QWEN-3 MAX is so good this level of details was only generated by gemini deepthink before one shottes 3d simulation of a procedurally generated mini planet

21 76 981 122K 307

Download Video

Wenting Zhao @wzhao_nlp

5 days ago

I’m seriously so eager to learn kernel programming, but one thing I couldn’t decide is whether I should be an expert on it myself or teach AI to be really good at it…😶‍🌫️

28 10 434 54K 149

Andrej Karpathy @karpathy

6 days ago

@simonw I was a bit surprised it is less than case than I expected. Code is KING. It’s the primary means of processing digital information - long term I can’t imagine a more important domain for the AGI pilled. And it is highly valuable in the interim too - big TAM @ high salaries.…

36 35 1K 95K 219

Wenting Zhao @wzhao_nlp

6 days ago

We’ve been cooking 👩‍🍳🥘

Binyuan Hui @huybery

6 days ago

We’ve been cooking 👩‍🍳🥘

62 67 1K 397K 71

11 6 279 36K 16

Wenting Zhao @wzhao_nlp

2 weeks ago

What strikes me in the work is that as long as the data recipe is right everything can just work with RL, generalizes super well, even at 1.7B level. Even people said it’s hard to improve RL’ed qwen models, we just did it! Thanks @_akhaliq for featuring my work the third time…

AK @_akhaliq

2 weeks ago

3 38 247 99K 185

Download Image

4 39 388 68K 310

Qwen @Alibaba_Qwen

2 weeks ago

🚀 Introducing Qwen3-Next-80B-A3B — the FUTURE of efficient LLMs is here! 🔹 80B params, but only 3B activated per token → 10x cheaper training, 10x faster inference than Qwen3-32B.(esp. @ 32K+ context!) 🔹Hybrid Architecture: Gated DeltaNet + Gated Attention → best of speed &…

172 725 4K 904K 2K

Download Image

Wenting Zhao @wzhao_nlp

2 weeks ago

The online RL is so cool, guess the moat is to have a large number of users 😏

Cursor @cursor_ai

2 weeks ago

The online RL is so cool, guess the moat is to have a large number of users 😏

4 11 243 49K 152

0 0 19 3K 2

Ruiqi Zhong @ZhongRuiqi

3 weeks ago

I learned so much from this as an ML (not-system-ish) researcher. highly recommend a read!!

Thinking Machines @thinkymachines

3 weeks ago

I learned so much from this as an ML (not-system-ish) researcher. highly recommend a read!!

240 1K 8K 3.2M 5K

Download Image

3 33 677 63K 310

Wenting Zhao @wzhao_nlp

3 weeks ago

I’ve recently joined @Alibaba_Qwen! We’re building the next generation of frontier models through careful science and world-class engineering, and we are making rapid progress. Excited for what’s ahead 💜

63 22 903 114K 83

Jason Weston @jaseweston

3 weeks ago

🌀New Test-time scaling method 🌀 📝: arxiv.org/abs/2509.06870 - Use RL to train an LLM solution aggregator – Reasons, reviews, reconciles, and synthesizes a final solution -> Much better than existing techniques! - Simple new method. Strong results across 4 math benchmarks. 🧵1/5

2 123 714 68K 587

Download Image

Junxian He @junxian_he

4 weeks ago

Mirage or method? We re-assess a series of RL observations such as spurious reward, one-shot RL, test-time RL, and negative-sample training. 🧐These approaches were all proved on Qwen+Math combination originally, but do they work in other settings? If not, under which…

3 41 186 18K 89

Download Image

Ofir Press @OfirPress

3 weeks ago

Feels like everyone is slowly admitting that there's no moat in foundational models, and the only way to build a business out of AI is to build products.

Hayden Field @haydenfield

3 weeks ago

Feels like everyone is slowly admitting that there's no moat in foundational models, and the only way to build a business out of AI is to build products.

103 107 2K 218K 385

3 1 25 6K 2

Sasha Rush @srush_nlp

3 weeks ago

seems like worktrees / workspaces are going to be essential if we're going to have 20 agents going at once.

1 1 18 6K 3

Wenting Zhao @wzhao_nlp

a month ago

I've always been skeptical about PRMs, but being able to apply RL+reasoning changes the entire story for me. It was a fun ride with @weixiong_1, who has been teaching me a unified view to think about all RL methods. He'll be on the job market! It'd be so lucky to work with him.