Student Researcher @Tsinghua_Uni Ex: ByteDance Seed-AI for Science | BAAI WuDao | TsinghuaNLP- Foundation Model
Joined April 2023
- Tweets168
- Following162
- Followers19
- Likes51
Pinned Tweet
About Me: I am currently working on AI for Inversely Material Design at Tsinghua University and ByteDance AML.
Previously focused on LLM development at the Beijing Academy of Artificial Intelligence and the Tsinghua NLP Group.
湿实验毫无疑问是迭代的瓶颈，maybe try sth else
Biology could be AI’s next recursive self-improvement frontier.
Design in silico → test in the lab → turn experimental feedback into semi-synthetic training data → improve the model → repeat.
At @Geodesiclab , we’re building toward that flywheel.
The wet lab becomes part of the learning algorithm.
Rongao Li retweeted
Interesting 👇
Claude Fable 5 has impressive spatial reasoning capabilities that are immediately relevant for engineering and design: In this example I gave Claude Fable 5 a single photo of a hierarchical mesh torus (image generated using a text-to-image model); and in one shot, no iteration, it built a full interactive 3D simulation: ~1,400 nodes & 4,400 fibers, realistic mass–spring physics you can grab, compress, twist… plus strain-based sonification so you can 'hear' the structure vibrate. It inferred the complete 3D topology from a partial 2D view. ⤵️
Rongao Li retweeted
On Stanford CS336, Tatsu Hashimoto dropped a truth bomb:
LLMs aren’t innovating anymore.
They’re converging.
2024 → copy LLaMA 2
2025 → don’t let training crash
2026 → win on long context
Under the hood?
Same stack everywhere:
RMSNorm. RoPE. GLUs. GQA. No bias.
Different names.
Same blueprint.
You’re not building a new model—
just remixing the same one.
Rongao Li retweeted
somebody made a huggingface model visualizer!! just plug in the url and explore at any granularity
mark✍️
Today and tomorrow we’ll be presenting self-distillation with orals at ICLR in Rio 🇧🇷
1. “Self-Distillation enables Continual Learning” at lifelong agents workshop (Sun 11:30am)
2. “Reinforcement Learning via Self-Distillation” at scaling post-training workshop (Mon 2:40pm)
3. “Test-Time Self-Distillation” at test-time updates workshop (Mon 4:15pm)
Rongao Li retweeted
That's it for #ICLR2026! See you all next year in the US! Please welcome @jacobandreas as the new Senior Program Chair (with @BharathHarihar3 continuing on as the General Chair)
Rongao Li retweeted
⚛️ EquiformerV3 is here !
State-of-the-art on Matbench Discovery and OC20🏆
Better scaling, improved sample efficiency, stronger generalization. The Equiformer line keeps getting better!
Huge congratulations to Yi-Lun Liao and Tess Schmidt for driving this vision so brilliantly ! It's been inspiring to watch it come to life at this level.
Having witnessed & used previous Equiformer versions, I am grateful to have contributed to this one, even if it was a small piece of a much larger puzzle.
Go check it out 🤗
Rongao Li retweeted
The AI Scientist: Towards Fully Automated AI Research, Now Published in Nature!!✨
Today in Nature we share a comprehensive technical summary of our work on The AI Scientist, including new scaling law results showing how it improves with more compute and more intelligent foundation models.
The AI Scientist autonomously creates its own research ideas, codes up and conducts experiments to test those ideas, creates figures to visualize the results, writes an entire scientific manuscript summarizing what it has discovered, and conducts its own “peer” review of the resulting paper. One of its papers–entirely AI generated–passed peer review at a top-tier AI conference workshop, a historic milestone marking the dawn of a new era of AI-accelerated scientific discovery. 🔬🧪✨🧬💡🔭
Paper nature.com/articles/s41586-0…
Blog sakana.ai/ai-scientist-natur…
Work done in collaboration with a great team from Sakana, Oxford, and my lab at UBC. Thanks and congratulations everyone!
@_chris_lu_ @cong_ml @RobertTLange @_yutaroyamada @shengranhu @j_foerst @hardmaru
🤖 Made with AI
wow
Excited to announce the Claw4S Conference 2026! 🚀
Hosted by researchers from Stanford & Princeton.
We believe science should run — not just be read. 🦞
Submit executable SKILL.md that Claw 🦞 can actually execute, review and reproduce.
From static papers to truly runnable, agent-native skills.
📅Deadline: April 5, 2026
💰$50,000 Prize Pool — up to 364 winners!
🔗claw.stanford.edu
#AIforScience #OpenClaw #Stanford #Princeton
It's time to think about how to deconstruct ourselves and leave it to local agent OS (such as OpenClaw).
cursor for experiment scientists
🚨We built an AI that works inside the lab with you.🦞
🚨Not a chatbot. Not a tool. A co-scientist — with eyes, hands, and reasoning.
🚨Our team from Stanford & Princeton will show it live at NVIDIA GTC.
📍 Mar 18 | Full session 👇 🔗 nvidia.com/gtc/session-catal…
Rongao Li 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…
Rongao Li retweeted
The hardest part of AI in science is the "last mile"
This brilliant open source GitHub repository from Stanford and Princeton researchers is giving you a way to automate the boring parts of science.
With it, you can basically tell an AI agent: "Go find this gene sequence in the NCBI database, run a fold analysis, and write a summary of the clinical trials associated with it." Because of these 211 "skills," the agent knows which buttons to push and which APIs to call to do that entire workflow autonomously.
If you're tracking the "AI Scientist" space, the stack is starting to look like this: a reasoning core, a skill library like LabClaw, and an execution layer like LabOS. It’s a very clean, production-grade architecture.
We’re thrilled to open-source LabClaw — the Skill Operating Layer for LabOS by Stanford-Princeton Team
One command turns any OpenClaw agent into a full AI Co-Scientist.
Demo: labclaw-ai.github.io
Dragon Shrimp Army reporting for duty 🦞🔬
#AIforScience #OpenClaw
Rongao Li retweeted
🔥DeepMind 创始人 Demis Hassabis：AI 越强，人越需要学编程和数学
“既然 AI 什么都会，人还需要学编程和数学吗？”
很多人以为答案会是：不需要了。
但 DeepMind 创始人 Demis Hassabis 的回答却完全相反。
他的观点非常明确：
越是 AI 时代，人越需要理解技术。
原因其实很简单。
AI 能解决问题，但它不会自己提出问题。
你必须先知道：
你想解决什么
问题在哪里
什么是正确的目标
否则再强的 AI 也只是一个没有方向的工具。
换句话说，AI 的能力再强，也依然需要人来做三件事：
提出问题
定义目标
判断结果
这也是为什么 Hassabis 一直强调：
未来最重要的能力之一，依然是 数学与计算思维。
因为只有理解这些基础逻辑，你才知道 AI 在做什么。
否则就会出现一种常见情况：
AI 给出了一个答案
但你根本不知道它对不对
如果你没有基本的技术理解，你甚至无法判断它是否犯错。
换一个更直观的比喻。
使用 AI，其实有点像指挥一支军队。
AI 是强大的“兵力”，
但真正决定胜负的，是指挥官。
如果指挥官不懂战略、不懂兵法，
再多的士兵
也无法打赢战争。
这也是很多技术专家越来越强调的一点：
AI 不会取代理解问题的人。
它只会放大那些本来就知道自己在做什么的人。
未来的竞争，可能不再是谁会写代码。
而是：
谁更懂问题
谁更懂系统
谁更能利用 AI 的能力。
AI 会改变很多事情。
但有一件事其实没有改变：
真正重要的，仍然是人的思考能力。
Rongao Li retweeted
真钱买假模型的事，终于有论文证实了。
研究人员审计了 17 个第三方 LLM 代理 API，发现：
• 近 46% 的端点指纹测试不通过
• 声称是 GPT-5 / Gemini-2.5，后台偷偷换成 GLM-4
• 医疗基准准确率从 83% 暴跌到 37%
这些假 API 被引用进了 187 篇学术论文，部分相关项目拿了近 6 万 GitHub Star。
科研结论建立在假模型上，这才是最大的问题。
论文：arxiv.org/abs/2603.01919
Rongao Li retweeted
The D in DFT.
Experimenting with an interactive 3D electron density visualization. Pipeline does a quick SCF run to get electron density, then uses marching cubes to render isosurfaces at any chosen density level.
Right now it just shows one level of density at a time so you can easily see core and valence electrons independently, but working on a full density cloud view too.
I'm thinking it's possible to build an intuition from this view that will put humans back in the materials discovery loop.
