Yi Zhou (@dugu9sword) header
Yi Zhou (@dugu9sword) profile picture

LLM @ Seed. Developed SeedFold/SeedProteo, CryoSTAR/CryoFM.

Shanghai
Joined July 2014
Photo by Yi Zhou (@dugu9sword)Photo by Yi Zhou (@dugu9sword)Photo by Yi Zhou (@dugu9sword)Photo by Yi Zhou (@dugu9sword)Photo by Yi Zhou (@dugu9sword)
genius idea, awesome astra! but...
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math. rikiyat.github.io/blog/weight-sp… github.com/RikiyaT/lm_und… huggingface.co/Qwen/Qwen3.5-0… huggingface.co/RikiyaT/Qwen3.…
1
107
Yi Zhou retweeted
Many drugs work by binding to a specific target in the body and blocking or changing what it does. An important first step in the drug development process is designing a molecule that can bind tightly to its target. Traditionally, that's meant weeks or months of expert work per target, sifting through a large number of candidates to identify the few that work. We wanted to test if Claude could successfully design novel protein binders from scratch (also called de novo design). With a protein design prompt written by a human expert, Claude autonomously designed protein binders against 14 out of 15 targets. We then worked with Adaptyv Bio and Twist Bioscience, who independently built and tested the proteins Claude designed.
599
1,430
671
12,427
4,353,242
Yi Zhou retweeted
The blog from @AndrewLampinen highlights a notable feature of recent progress in LLM for math: models generate candidate proofs through informal, often natural-language reasoning, while formal systems verify them. This generation-verification distinction becomes even more interesting when mapped to biology: hypotheses can be explicit but hard to verify, so intelligence lies in maintaining competing hypotheses and continuously updating confidence as evidence increases. When we look at AI for scientific discovery, it's evident that intelligence is not in the harness alone. It is the model’s ability to wield that harness: translating hypotheses into evidence-seeking actions. The underlying model must select the right tools and experiments, interpret conflicting results, and decide what to do next. In both LLMs and human experts, a small difference in research judgment with all else held equal (e.g. tools and pipelines), can produce drastically different outcomes. Now this leads to my main point: I believe that the next leap in AI for Science will come from systems that turn biological knowledge and tools into better research decisions. This requires a closed loop among uncertain evidence, competing hypotheses, and numerous experiments. We are already getting better at scaling what models know. Now we need to scale the scientific judgment that drives discovery forward. Let's get to work!
What should we make of recent language model advancements in mathematics? In my new post, I reflect on long-standing cognitive debates about symbols and neural networks in light of this progress. (Link below)
2
9
784
We can finally talk about it: We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company. We verified that our reasoning token count matches billed API thinking tokens 1:1 for most of the prompts we queried.
382
1,957
620
13,498
3,547,536
...
don't worry. I've setup the logos so we're all included. play below
1
2
226
Back in April I implemented a Triangular Attention kernel for protein folding with Kimi-2.6. It performs insanely well, and even outperforms NVIDIA’s Cuequivariance kernel.
This quoted post is unavailable.
2
18
2,287
hybrid linear attn is scalable!!
Introducing Kimi K3: Open Frontier Intelligence 🔹 2.8 Trillion Parameters, 1 Million Context, Native Multimodal 🔹 Kimi Delta Attention enables up to 6.3x faster decoding in million-token contexts 🔹 Attention Residuals deliver ~25% higher training efficiency at <2% additional cost 🔹 Built for long-horizon agentic coding and self-evolving workflows Kimi K3 is now live on on Kimi.com, Kimi Work, Kimi Code, and the Kimi API. Open Weights by July 27, 2026. 🔗 API: platform.kimi.ai 🔗 Tech blog: kimi.com/blog/kimi-k3
9
28
1
432
25,120
the eye is essential to AI
I finally got an open model to do structural biology by itself 🔥 GLM-5.2 drives the Mol* viewer, judges its own render through Qwen3-VL, and refines until the drug pops in its pocket. Then I spun it in 3D. All open, on @huggingface. What should it build next?
2
21
3,527
so you can extend a short peptide to a length of 1T🤣
1/9 Excited to share DPLM-Evo, the latest member of the DPLM family, in our ICML 2026 paper: Towards A Generative Protein Evolution Machine with DPLM-Evo. 📍 Come find us at Hall A #3512 🗓️ Wed, Jul 8, 10:30 AM - 12:15 PM KST Paper: arxiv.org/abs/2605.00182 This is joint work by *Xinyou Wang(@Xinyou_NJU), *Liang Hong(@leungh22), Jiasheng Ye(@jsye588986), Zaixiang Zheng(@zaixiang_zheng), Yu Li, Shujian Huang, and Quanquan Gu(@QuanquanGu), with collaborators from ByteDance Seed, Nanjing University, CUHK, and Fudan University. DPLM-Evo targets a central gap in protein diffusion: protein engineering is rarely just "generate a protein-like sequence from scratch." In practice, we use evolutionary information to score mutations and to improve or reprogram existing scaffolds while preserving structure and function. Many successful diffusion PLMs, including the original DPLM line, use mask recovery as the denoising interface. DPLM-Evo asks: can the denoising steps instead be the edit operations of evolution itself: substitution, insertion, and deletion?
3
743
Sadly I had to miss ICML in Seoul at the last minute (and waste one registration along the way) but still want to share two ICML 2026 papers from our team at ByteDance Seed AI for Science. If you are at the conference, please go chat with my homies: Wesley (@chengyenhsieh), Xinyou (@Xinyou_NJU), and Liang (@leungh22)! GO grill them :D 1. Protein Autoregressive Modeling via Multiscale Structure Generation (Oral!) 📎 arxiv.org/abs/2602.04883 Presented by Wesley. Oral: 📍 Hall D2, 🕥 Tue, Jul 7, 10:30 Poster: 📍 Hall A, #1413, 🕑 2:00–3:45 PM Both are today! 2. Towards A Generative Protein Evolution Machine with DPLM-Evo 📎 arxiv.org/abs/2605.00182 Presented by Xinyou & Liang. 🕥 Wed, Jul 8, 10:30 (Tomorrow!) 📍 Poster Session 3, #3512 3. Also please DO NOT miss: @chengyenhsieh will be presenting an lighting talk at ByteDance booth, TODAY, 🕓 4:00–4:30 PM!
2
6
47
4,352
from doing great science to redefining scientists?
A bit of news: After nearly 9 years, I have decided to leave Google DeepMind and join Anthropic (after taking some time to recharge). I am incredibly grateful for my time at GDM. @demishassabis took a real chance letting me lead the AlphaFold team just six months after finishing my PhD, and the entire GDM team taught me so much about how to do great science. GDM is a special place, and I’ll still be excited to hear about what amazing things they discover next.
3
327
does it mean that if I send a fasta to codex, it will return a pdb?
We’ve also built support for OpenAI Codex and Gemini CLI so you don’t have to choose. As always: you own your inputs and outputs, and we don't train on your data.
1
1
5
1,572
Glad to see the plot showing the scaling law for biomolecular models!
Replying to @bjing2016
We also found a scaling law that explains (p=0.03) Ab-Ag accuracy across model lineages - from Boltz to AF3 and beyond. 1.8x-ing AF3 performance (to match IsoDDE) would seem to require e.g. 10x more distillation and 16x compute. (7/8)
6
1,532
a cool study that unifies structure prediction and protein design!
Replying to @mihirbafna14
@mihirbafna14 and I are excited to introduce Promera, a co-folding and design model with • best-in-class binder filtering • nanobody design with in-silico success rates matching hallucination • case studies on hantavirus epitope targeting and GPCR agonism (1/8)
13
2,645
interesting
Grok Build with Composer 2.5 visualizing predicted binders against PD-L1, a cancer immune-checkpoint protein that tumors use to switch off T-cells. Composer 2.5 in Grok Build is using ChimeraX to visualize PD-L1 binders generated by Genie 3. In this workflow, a region on PD-L1 was selected for binding, Genie 3 generated binder shapes, ProteinMPNN filled in the amino acid sequences, and Boltz folded the binders against PD-L1.
230
Yi Zhou retweeted
Today marks my last day at ByteDance Seed. Over the past 3 years, I had the opportunity to work across two of the most exciting frontiers in AI: AI for Drug Discovery and building frontier LLMs. Few opportunities in a career allow one to work simultaneously on curing disease and building frontier intelligence. I was fortunate to do both. Since joining ByteDance, I have led the AI for drug discovery effort. Together with an exceptional team and collaborators, we built SeedFold, the world's first biomolecule structure prediction model to outperform AlphaFold 3 across a broad range of benchmarks and capabilities; SeedProteo, a state-of-the-art protein binder design model surpassing AlphaProteo, RFdiffusion, Chai-2, BinderCraft, and BoltzGen ; and the DPLM series of protein language models, along with several other ambitious AI for Science projects. In early 2025, I took on a new challenge. To tackle one of the hardest problems in modern AI, reliably training and scaling frontier-scale LLMs, I joined the LLM pretraining effort and founded the LLM optimization and scaling team. Together, we built a highly scalable pretraining stack that enabled the successful training of Seed 2.0 and the subsequent frontier-scale models, significantly advancing our ability to train and iterate on frontier AI systems at scale. I'm deeply grateful to my teammates, collaborators, and leadership for an incredibly rewarding journey. The best model is yet to come. Scaling continues!
174
60
21
1,283
483,091
Most people think AI helps with research. They’re underestimating it. We’re rebuilding the entire research pipeline. EurekaClaw is just the start. 🦞
1/n 🦞 Introducing EurekaClaw💡 — a local-first AI research agent that captures your Eureka moments before they vanish. From idea → proof → experiment → paper — fully automated. Local-first. Zero data leak. 🔒 Try it: eurekaclaw.ai
12
37
6
355
67,826
Yi Zhou retweeted
Here is SeedProteo, our latest diffusion-based model for de novo all-atom protein design from ByteDance Seed! Our server is now live — feel free to give it a try! seedfold.io/dashboard/protei…
5
25
2
133
12,344
No, I think latent representation means the loss of information.
Could the new paradigm Max hints at be latent all-atom diffusion transformers? ;)
11
1,797
⚡️ A trend: prior from foundation models + likelihood from experimental data. 📖 Also suggest three related work: 1. CryoBoltz: Multiscale guidance of protein structure prediction with heterogeneous cryo-EM data (@ZhongingAlong ) 2. Inverse problems with experiment-guided AlphaFold (@AdvaithSai1 ) 3. CryoFM: A Flow-based Foundation Model for Cryo-EM Densities (from our team)
We introduce a new method, EmbedOpt, for robustly steering protein sequence-to-structure diffusion models to fit experimental data (Cryo-EM, NMR) without training. 🧬📉 @mhli41 @JiequnH @PilarCossio2 EmbedOpt tackles the brittleness of the previous coordinate-space steering methods by optimizing the conditional embedding instead. These embeddings capture rich co-evolutionary signals in protein diffusion models—unlocking a new, robust and semantically meaningful diffusion steering axis. 🚀 Result: Better fitting, wider hyperparam stability, and efficiency enabled by fewer diffusion steps 📄 Preprint: arxiv.org/abs/2602.05285
2
51
4,769