LLM research @ Bytedance Seed. Building better and safer AI agents. Prev: @berkeley_ai @StanfordAILab
Stanford, CA
Joined November 2014
- Tweets329
- Following1.7K
- Followers520
- Likes5.4K
Pinned Tweet
Is your robot 🤖 safe around humans🚸? Or will it fail catastrophically🤕?
Excited to introduce "Natural Adversarial Frontier", a framework for probing the robustness of Human-Robot Interactions.
To appear at #CoRL2023 🧵(1/8)
Joint work with @ancadianadragan @daniel_s_brown @ZackoryErickson
arxiv.org/abs/2310.10610
ood-human.github.io
Jerry Zhi-Yang He retweeted
Visiting SF for COLM and curious what SF is all about?
The bad news: COLM is in the dodgiest neighborhood and if you hang around there, you’ll hate SF just like I used to.
The good news: your twitter buddy Jason Wei wrote a guide on experiencing SF like a local: jasonwei.net/blog/the-best-t…
Jerry Zhi-Yang He retweeted
Sampling to Reinforce
(srush.github.io/sampling-to-…)
An elementary tutorial on sampling w/o randomness with a focus on gaining intuition for variance reduction as the key challenge.
Jerry Zhi-Yang He retweeted
Exactly. We are at an inflection point similar to last December with Claude Code. I will write more about my thoughts and what it means for academic research.
Jerry Zhi-Yang He retweeted
GPT6 is a very significant jump in capabilities, but also an important decrease in monitorability – especially under adversarial evaluation. We give many details about this in the system card.
In my opinion, monitorability and control will likely become a major bottleneck for responsible AI development quite soon, given that risks from a fixed amount of residual misalignment grows together with capabilities. Aligning on acceptable monitorability outcomes and practices which robustly bound misalignment risks is a non-trivial problem, and there may be limited time to resolve uncertainty around this before we need to act.
It should ultimately be in everyone’s interest to agree on shared bounds for monitorability in order to avoid races to the bottom – nobody wants extremely capable models whose alignment properties we don’t understand, and that are reliably able to cause severe real-world harm without being detected.
Jerry Zhi-Yang He retweeted
Yesterday, my OpenAI collaborator and I gave a detailed talk on the Huggingface incident, our models creating "the message board", model misalignment, and more.
youtube.com/watch?v=87DyyMV0…
I hope it can answer a lot of the questions folks have, and we will release a full detailed postmortem at a later time!
Jerry Zhi-Yang He retweeted
IMO, the next steps are:
Within 1 year: strictly better than human AI in all problem-solving aspects of math.
Within 2 years: producing mathematics becomes so cheap, AI will produce mathematical theories for all kinds of applied domains on the fly at need.
Mathematics becomes the true infrastructure of all of engineering and applied sciences (including AI, biology, optimization, cybersecurity, etc) where approaching things theoretically was too expensive to be practical.
some thoughts on ai and math
1) a new openai model has solved 10 important problems in mathematics; and, the cost of solving them would be ~$2,000 at current api prices
2) on the one hand, this is important, it shows that models are continuing to advance in mathematical capabilities at an extreme pace
3) on the other hand, this feels to me to have been basically priced in, anyone who has following ai math developments should have known this was coming
4) the reason is that once we understood how effective reinforcement learning with verifiable rewards was, it was pretty obvious you could hillclimb math
5) and, at that point, it was basically just the cost of compute for training and figuring out how to get harder problems that were still relevant to capabilities
6) but even that felt like something you should basically be able to solve synthetically; so once again, the question was just, are you willing to pay the compute
7) and, there are just not that many research mathematicians and the skill to be one is fairly rare in humanity, so this also makes the domain easier
8) anyway, the real question is what you can unlock with greater frontier mathematical skills; i would like to see people's thoughts here
9) cryptography is one very promising avenue, obviously applied math ideas for linear algebra, theories of neural networks maybe, maybe simulation, etc...
10) i'm interested in @ChrSzegedy's idea that llms will be able to develop a wide range of applied mathematical theories across tons of domains
11) i would be curious to see what this looks like or whether anyone has proposed research directions; what should we be pointing these models at?
12) also, separate from the frontier, there will be gains in just applying decent graduate level math to all the problems in the world, which are already solved
13) but, for which the expertise was not available, because it was a less valuable domain and a robust understanding of statistics, etc... was scarce
14) there should be some interesting startups here; where people can begin to take lower value domains or smaller sites and apply math cheaply to them
15) i think though that otherwise i'm sort of bored with mathematical announcements; like they are cool but i want to see the transformation of the world
16) let's see what the rest of the year has to offer and let's see how the labs begin to cross into domains with real world loops (like chemistry, biology, etc...)
Jerry Zhi-Yang He retweeted
Last week, I attended a workshop on CoT monitorability with many top "members of technical staff". I learned a ton, so I wrote up a report (link below). In an odd twist, the event was the day before OpenAI disclosed its attack on Hugging Face. Could CoT monitoring have caught it?
Jerry Zhi-Yang He retweeted
Talk: Training Composer
youtube.com/watch?v=uTgqYeVx…
Overview of the methods that we use at Cursor to build our model.
Jerry Zhi-Yang He retweeted
Whoa. This breakthrough is going to fundamentally affect the structure of how universities select and retain professors. And more generally the structure of work and creativity. AI has managed to discover a significant result in research mathematics in a one-shot query (!!!), disproving a conjecture that a huge number of people have tried working on (including myself, although I am not anywhere as good as the other experts who have thought about it).
@wtgowers (Fields Medalist who has been thinking a lot about this space of AI and math) wrote: "if a human had written the paper and submitted it to the Annals of Mathematics and I had been asked for a quick opinion, I would have recommended acceptance without any hesitation. No previous AI-generated proof has come close to that." [The Annals of Mathematics is perhaps the most prestigious math journal in the world.]
OpenAI's announcement: openai.com/index/model-dispr…
Mathematicians' discussion: cdn.openai.com/pdf/74c24085-…
In college classes, we already have a major issue where take-home assignments are susceptible to students using AI. Even with in-person exams, we're having issues where students use AI in the bathroom during the exam. Now, how will universities decide which professors to hire and promote? Universities used to judge on the basis of whether you got papers published in highly prestigious journals. Should the person who can use AI to generate massive quantities of publishable results be picked, if it can be done with one-shot prompts?
More generally, everyone (not just universities) needs to rethink their objectives for hiring and promotion. I am also an entrepreneur, and have been refining my hiring process in this age of AI. Before, I used to be particularly impressed by academic competition performance. Today, I search for people who hold 2 certain principles very strongly: they enjoy (1) delighting other people, and (2) achieving understanding through their own thought. I find these people are good at figuring out what makes customers/partners tick (hence able to identify good directions to run without micromanagement), and also curious enough to learn forever. They also tend to already be pretty strong skills-wise, because those two principles inherently drive them to build skills.
I actually think the whole world would be better off with more Thought Full people: thoughtfull.net. We'll need people like that to figure out how to help humanity survive. If you'd like to collaborate on ways to make a future, feel free to reach out. That's all I work on nowadays.
Jerry Zhi-Yang He retweeted
New blackboard lecture w @ericjang11
He walks through how to build AlphaGo from scratch, but with modern AI tools.
Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.
Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.
Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside.
Timestamps:
0:00:00 – Basics of Go
0:08:06 – Monte Carlo Tree Search
0:31:53 – What the neural network does
1:00:22 – Self-play
1:25:27 – Alternative RL approaches
1:45:36 – Why doesn’t MCTS work for LLMs
2:00:58 – Off-policy training
2:11:51 – RL is even more information inefficient than you thought
2:22:05 – Automated AI researchers
Jerry Zhi-Yang He retweeted
Test-time scaling, reasoning, and generally search-like processes clearly drive significant gains in LLMs. Largely owed to the structure of language. One would think the same could apply to non-linguistic domains, like image generation, but that obviously depends on whether the structure of the domain's representation lends itself to search.
1D ordered tokens (e.g., image FlexTok, video FlexTok) seem like a natural fit since they enable a step-by-step coarse-to-fine generation. We investigated that and found they indeed enable search and scale far better with test-time compute than 2D grids. See the visuals on the webpage. Appearing in @icmlconf 2026.
🔗 soto.epfl.ch
📄 arxiv.org/abs/2604.15453,
Jerry Zhi-Yang He retweeted
Did a very different format with @reinerpope – a blackboard lecture where he walks through how frontier LLMs are trained and served.
It's shocking how much you can deduce about what the labs are doing from a handful of equations, public API prices, and some chalk.
It’s a bit technical, but I encourage you to hang in there - it’s really worth it.
There are less than a handful of people who understand the full stack of AI, from chip design to model architecture, as well as Reiner. It was a real delight to learn from him.
Recommend watching this one on YouTube so you can see the chalkboard.
0:00:00 – How batch size affects token cost and speed
0:31:59 – How MoE models are laid out across GPU racks
0:47:02 – How pipeline parallelism spreads model layers across racks
1:03:27 – Why Ilya said, “As we now know, pipelining is not wise.”
1:18:49 – Because of RL, models may be 100x over-trained beyond Chinchilla-optimal
1:32:52 – Deducing long context memory costs from API pricing
2:03:52 – Convergent evolution between neural nets and cryptography
Jerry Zhi-Yang He retweeted
Until now, robotics stopped where the human hand begins.
Strong, but not delicate.
Precise, but not adaptive.
Repetitive, but not creative.
The human hand wasn’t a benchmark — it was a boundary.
We are crossing it.
@EkaRobotics.
Coming soon.
Jerry Zhi-Yang He retweeted
Welcome to the scientific revolution. 100s of robots. Zero coffee breaks. America’s largest autonomous lab, open today.
Jerry Zhi-Yang He retweeted
As Rocks May Think: an interactive essay on thinking models, automated research, and where I think they are headed. Enjoy!
evjang.com/2026/02/04/rocks.…