Research Scientist@ByteDance Inc. SG, Post-Doc in MMLab@NTU, PhD@PKU. Opinions are my own.
Singapore
Joined December 2022
- Tweets336
- Following791
- Followers995
- Likes1.2K
Xiangtai Li retweeted
Gemini 4 Argon is now globally available.
Built to solve your hardest problems from start to finish, the model brings frontier capabilities across coding, enterprise workflows, and cyber defense.
Xiangtai Li retweeted
introducing muse wonder
our new frontier model. state-of-the-art at personal agents and professional agents.
our largest and most capable model yet—rivals the big bois
Xiangtai Li retweeted
Deploying DiffusionGemma-Jev (djev) just got a lot easier. You can now spin up a Jev API-compatible endpoint on Google Cloud Run using a single command.
Performance is solid: ~35-60 ms for single step latency and batch@32 is ~100-123 requests/sec.
It's a straightforward way to experiment without needing your own GPU. Runs at roughly $3/hr and drops to $0 when idle.
Get the code and instructions here: github.com/taeold/djev-run
Xiangtai Li retweeted
Interesting
5/ We also study the data-constrained, multi-epoch setting and find that looping has a regularizing effect.
At a fixed model size, val loss decreases monotonically with more loops, and the optimal number of loops grows with compute, leading to 2.2x compute efficiency over a standard transformer even with tuned weight decay.
Xiangtai Li retweeted
Code as policy, world, or editor: agents can write programs to control robots, represent worlds, and edit visual content.
My student Zunnan @zngmail2 has curated Awesome Code-as-X github.com/kkakkkka/awesome-…, connecting research across these areas through a shared idea: programs as executable, inspectable, and revisable representations. A useful reading list for those exploring this direction. Contributions welcome.
🤖 Made with AI
Xiangtai Li retweeted
The rate of improvement from original GPT to GPT-3 is impressive. If this rate of improvement continues, GPT-5 or 6 could be indistinguishable from the smartest humans. Just my opinion, not an endorsement. I left OpenAI 2 to 3 years ago. Am a neutral outsider at this point.
Xiangtai Li retweeted
Today we're releasing data on models accelerating research at OpenAI.
Recursive self-improvement could be the most important contributor to AI capabilities over the next few years, but by default it will only be seen inside a few frontier AI labs. Being transparent is more urgent than ever, so we can inform the public discussion on whether and how to pace model development. I ask other AI companies to do the same.
openai.com/index/research-ac…
omg
GPT-6 Astra recreated the Palace of Fine arts in Blender.
This is favorite building in San Francisco because it was built for the World's Fair in 1914 where the steam locomotive and telephone were showed off. It was a time when technology gave us all a deep sense of optimism for the future.
I feel like some of that sense has been lost since, but I hope models like astra can help restore it, and push us towards a more hopeful future.
Xiangtai Li retweeted
The only tutorial you should attend at #ECCV26 (kidding 😂).
But I am incredibly psyched to be doing this with the best bunch out there. I strongly feel this is a timely tutorial!
We will present general approaches alongside our learnings from (post)-training impactful models like Flux2/3.
Tutorial website: pdm-tut.github.io/
Save your calendars!
Xiangtai Li retweeted
Neoclouds have limited cybersecurity. Next time agents successfully go rouge, they'll try taking over a neocloud to run more copies. This is bad.
Thus: neoclouds should greatly strengthen their cybersecurity and every company with strong cyber models should help with that.
Xiangtai Li retweeted
注意力是这个时代最珍贵的资源，拥有无限的token不代表拥有无限的注意力。并行的agent越多，越感觉注意力不够用，现在每天都会确保一段高质量的专注时间进行单线程的知识学习。多线程并行n个agent对于提升自己的内在知识水平实在是没有帮助，只对打工做业务有帮助。
最近开了6个两百刀的账号（3*codex pro+3*claude max）渐渐意识到AI时代最重要的还是专注。
前几个月总喜欢说自己在和多少多少agent并行做几多几多事情，这一两个星期逐渐开始回归单线程，因为这样效率其实是更高的。人一天的阅读量是有限的，带宽几乎是一个硬上限，与生俱来且难以改进的。多线程看起来处理了很多事情，其实最终只是把一天的精力用不同的方式分在了不同的事情上。换言之，就算是更换线程好无消耗，其实一天下来处理的事情和单线程是没有区别的。然而事实证明更换线程本身其实是消耗精力的，最终很容易注意力涣散，变得either精神崩溃，or疲惫不堪。
说到底，人是单线程生物，只有专注，效率才是最高的。其实如果能专注把事情一件一件完成，几乎在任何情况下都是一个人能实现的最高产出的方式。从这个角度看，其实AI的作用是帮助人来实现事实多线程——在人的角度，我在专注做“一件”事情，那就是指挥AI做眼前的这个工作，然后我让AI自己设计一个workflow来帮我实现事实上的并行，并十分小心地保护自己的注意力。毕竟注意力是人的唯一的不可再生的宝贵资源。
灵台其实就是为这件事情而生的，只不过我最近发现其实用哪个harness都差不多，最最最重要的是一定要以最保护自己注意力的方式来交互。就算你和几个agent同时推进，也一定要在逻辑上是在做同一件事情的（几个不同方面），千万不能让自己做一些逻辑上是迥异的事情，那样一下子精力就会飞速消耗，开始变得烦躁了。
github.com/lingtai-ai/lingta…
Xiangtai Li retweeted
1/ today we’re releasing muse spark 1.3—available in muse code & the meta model api.
this is our most capable model yet—frontier performance almost too cheap to meter. much stronger at agentic and coding with better usability. we think users will really notice the jump.
Xiangtai Li retweeted
Here's my talk from the CVPR 2026 "Bitter Lessons" workshop earlier this summer. I've split it up into parts for the sake of discussion.
Part 1: in which I wax poetic about my youth and force the audience to look at my dissertation results.
Xiangtai Li retweeted
Our new blog presents VLANeXt: insights from 500+ controlled experiments, together with a research codebase covering VLA scaling, latent actions, predictive learning, and world-action modeling.
huggingface.co/blog/cavanloy…
Xiangtai Li retweeted
目前AI大模型后训练有两种流行方式：SFT和OPD。
SFT：学习老师的做题步骤和答案 (抄作业)。
OPD：学生先做题，然后老师给学生纠错。
对大多数人类学习而言，OPD比SFT效果好。因为人类在SFT（抄作业）时容易分心，可能没有真正理解，实质没有学到。
但是对AI大模型其实有点不一定，因为SFT一定会通过梯度更新参数。也就是说，只要老师模型水平高、数据质量好，用SFT大概率能让学生模型学到东西。
大模型的参数更新类似于人类的“领悟”，甚至可能是“醍醐灌顶”。
所以，还不能下结论说OPD会比SFT更好。
当然，不论对人类还是AI大模型而言，OPD听起来都更加珍贵：学生先暴露自己的问题，再得到老师一对一的针对性指导。直觉上OPD更好。
但有趣的是，SFT对AI也没有“抄作业”这个比喻听起来那么笨。即使面对同一份老师答案，不同学生模型因为参数不同，算出来的梯度也不同：已经会的地方改得少，不会的地方改得多。因此，SFT本身也包含一定程度的“因材施教”。
OPD真正更有针对性的地方，是连学生遇到什么问题，都是由学生自己先做出来的。
所以两者到底谁更好，其实并不确定。
更关键的问题可能是：对于最顶尖的神童人类和“神童模型”，智能的上限是不是最终要靠自学达到的？
Xiangtai Li retweeted
Try out Sheng-Yu's latest work with pip install tpips.
TPIPS is a text-prompted image perceptual similarity metric and a revisit of the classic LPIPS metric.
How similar are two images? Prior metrics (e.g., LPIPS, DreamSim) give just a single score. But actually, there are multiple *senses* of similarity (color, pose, etc.)
We introduce TPIPS -- Text-Prompted Image Perceptual Similarity
"pip install tpips"
peterwang512.github.io/TPIPS 🧵
Xiangtai Li retweeted
Is model scaling the only source of agent improvement?
We @henryqin1997 @YaxinLu1997 @VITAGroupUT @VictorKaiWang1 are glad to share our work: We reach 95.3% Raw Accuracy, or a $15 Frontier Run (88.8% with DeepSeek-v4-Flash, matching GPT-5.6 Sol Max), on Terminal-Bench 2.1 via Harness Scaling.
Human effort -> Non scalable ❌
Agent direction? Local optimization ❌
Human direction + agent action = better workflow control✅
StateM:
1. Open-source model + StateM -> Close-source frontier, everyone can try!
2. Training free, significantly improve frontier agents.
3. Zero-transfer-cost on same family of agents. [1/8]🧵
📄arxiv.org/abs/2608.15089
🌐 henryqin1997.github.io/state…