Xiangtai Li

Research Scientist@Bytedance Inc. SG, Post-Doc in MMLab@NTU, PhD@PKU. Opinions are my own.

[Phoenix Yin @Phoenixyin13]

“我在田里雇了一名工程师，它叫 Codex。”
这是AI时代最浪漫且硬核的田野叙事！

最近，这位北海道农民富安弘的故事在全网刷屏。一个毫无背景的前公务员，独自跑到北海道，靠着 ChatGPT 和 Codex，实现了一人管理100公顷农场的硬核神话。

大公司在搞价值千万的智慧农业，而富安弘毅在用几美金的 API 搞平替版赛博农场。来看看他的一键DIY清单：

温室自动通风系统。用 Codex 写代码连接传感器，太热就自动开窗卷帘，甚至能通过 LINE 软件远程发指令。
卫星天眼监控。把卫星数据叠加到农田地图，坐在家里吹着空调，就能一目了然哪块地的西兰花长得不好。
物理外挂。拍一张复杂的配电盘照片直接丢给 AI。

很多人焦虑 AI 会抢走工作，是因为他们把自己定位成了执行者。
但富安弘毅把 AI 当成了24小时不拿薪水的全栈工程师。他没有被技术取代，反而用技术给自己的体力赋能，拿回了对这 1500 亩土地的绝对掌控权。

在这个人人都在卷大模型参数、卷算力、卷 PPT 的时代，这位北海道大叔坐在拖拉机上，用实际行动告诉我们，技术的终点，是脚下踏实的土地。

[Yi Ma @YiMaTweets]

Based on what happened this year at CVPR 2026, it seems that the vision community is making a full swing back to 3D vision. Just to remind young people who just got interested in this subject, my first textbook, written 20 years ago, is probably still the best introduction to this topic: https://t.co/7Ks0wYAKhY

[Phoenix Yin @Phoenixyin13]

认真读完了OpenAI 研究员 Noam Brown 今天的长帖，一个被行业严重低估的现实。
LLM 的真实能力天花板，远高于当前任何 benchmark 所显示的水平。

原因，是给它的test-time compute太少了。而随着模型迭代，这个问题会越来越突出。

GPT-5.5 发布初期，benchmark 只比 5.4 略有提升，很多人觉得没那么惊艳。
但实际体验是明显的 step-change。
为什么？因为评估时没有控制 test-time compute。
一旦把 tokens 放在 x 轴画曲线，5.5 在相同计算预算下明显更强。单个数 benchmark 把这个差距掩盖了。

更关键的观察是，更强的模型，从额外 test-time compute 中获益更大。
性能提升曲线更陡，plateau 被不断推远，甚至可能不存在。
Karpathy 的 autoresearch 实验几百次实验后性能仍在上升。
AI Security Institute 的网络攻防评测在 1 亿 tokens 后，顶级模型仍在快速进步。
这说明能力天花板可能比我们想象的远得多。

这对 AI 安全与 preparedness 的影响非常大。
当前大部分安全评测只在低 inference budget 下进行。
但真正危险的能力，尤其是长 horizon、多步规划、复杂 agent 行为，很可能只在高计算预算下才显现。
Gemini 3 Deep Think 的争议，暴露了这个问题。scaffold加上高 inference 能让普通用户接近高能力，但 system card 没反映这条曲线。

Noam 指出一个更棘手的现实，即长 horizon 任务的评估越来越困难。
有些能力只能通过实际长时间运行才能验证，而模型迭代周期可能已经短于这个验证时间。
未来 labs 可能会面临一个两难，要么大幅延迟发布，要么在不完整的安全评估下发布。

两年多过去了，从 o1 开始，行业依然习惯报单个数，安全组织依然会对多花 100x inference 就能显著变强感到惊讶。
这其实是 post-o1 时代最被低估的 paradigm shift。
未来模型之间的真实差距，更会体现在谁能更好地理解和利用 intelligence-compute 曲线。
对于做 agent、长 horizon 推理、高风险部署的人来说，重新理解这个曲线，已经成为了重中之重的必选项。

[Google Gemma @googlegemma]

Meet Gemma 4 12B!

A unified, encoder-free multimodal model designed to bring high-performance intelligence directly to your laptop, and released under an Apache 2.0 license.

Bridging the gap between edge efficiency and advanced reasoning. Here is what’s new with Gemma 4 12B: 👇

[AutoSOTA @AutoSOTA11]

Following the latest CVPR paper SAMTok: Representing Any Mask with Two Words by @xtl994 , AutoSOTA built on this inspiring work and further improved gIoU to 82.4 (+0.7) using higher MLLM image resolution from 448 to 896. We are grateful to the authors for the strong foundation that made this follow-up possible.
AutoSOTA Project: https://t.co/w3XcHD7hjI
Enhancement Details: https://t.co/t0REzJoRpB
Original paper: https://t.co/Q4QfCmdfwj

#AutoSOTA_LiveCVPR #CVPR2026 #AutoSOTA #AIScientist #AutoResearch #MachineLearning #AcademicX #ResearchAutomation

[Yossi Gandelsman @YGandelsman]

This may be a controversial take, but I think it needs to be said: the gap between computer vision research in academia and industry is widening with every conference.

A huge fraction of @CVPR papers—especially those that boil down to "we tweaked/fine-tuned/RL'ed large-scale model X to improve on task Y"—will become obsolete with the next model release. That's not where academia creates lasting value. PIs should adapt much faster to this changing reality.

Academia should focus on fundamentally new ideas, new problem formulations, explaining emergent phenomenology, or uncovering blind spots that industry can later solve with scale, compute, and data.

[Chen Change Loy @ccloy]

Proud to share our lab’s @MMLabNTU work Log-linear Sparse Attention (LLSA) - a trainable sparse attention mechanism that reduces attention complexity from O(N²) to O(N log N), making diffusion transformers much more efficient.

Also, special shout-out to the first author @zhouyifan1107 for presenting the poster in full costume - truly above and beyond. The level of dedication is impressive! 👏

#CVPR2026 #DiffusionModels #EfficientAI #SparseAttention

[#CVPR2026 @CVPR]

Announcing the official #CVPR2026 Best Paper Award Winner! 🏆 Congratulations to the authors for their landmark contributions to the field!👏👏

[Fei-Fei Li @drfeifei]

https://t.co/Kt50ttQRMJ

[Andrej Karpathy @karpathy]

@shreyansj It’s refreshing to see a company of this size successfully call bs on the whole thing to this extent. One group of MTS on a mission, clean.

[Huaizheng Zhang @zhzHNN]

Big thanks to @BanghuaZ for sharing so many honest and practical insights.

- Do the right thing at the right time with the right people
- Explore more and make more mistakes
- Let go, and embrace the new wave.
- Be kind
- Have good taste
- etc

https://t.co/Gqt2DbasrH

[Jianyuan @jianyuan_wang]

Introducing VGGT-Ω: scaling feed-forward reconstruction across static and dynamic scenes, and studying whether the learned geometric representations transfer beyond reconstruction.

[Andrej Karpathy @karpathy]

Personal update: I've joined Anthropic. I think the next few years at the frontier of LLMs will be especially formative. I am very excited to join the team here and get back to R&D. I remain deeply passionate about education and plan to resume my work on it in time.

[动物园园长 @weiyux2021]

听说有个互联网老哥为了防止被公司蒸馏成skill，在训练自己的skill时特意加了高深数学解题能力。

一图流设置好，只要公司开始蒸馏，被复活的skill就会自动推演纳维斯托克斯方程，一晚上能烧掉老东家几十万……

这思路也太绝了吧！

[禿道道🐟 @dearemon]

今天看到一句来自莫言新书的话：
你现在恐惧的一切事情，都是你三年级在校门口忘带的红领巾

[Xiangtai Li @xtl994]

very cool guide👍

[Jia-Bin Huang @jbhuang0604]

Modern Transformer - Complete Guide

Interested in learning the recent advances in transformers?

After 13 videos, I've finally completed this series!
🥳🥳🥳

Check out the course here:
https://t.co/CsujxlWigC

[Richard Sutton @RichardSSutton]

The bitter lesson in 26 words:

Don’t be distracted by human knowledge, as AI has been historically.
Instead focus on methods for creating knowledge that scale with computation, like search and learning.

[Xiangtai Li @xtl994]

@atasteoff Congratulations，Shilong！

[#CVPR2026 @CVPR]

We are grateful to all of the 17,491 reviewers who helped make #CVPR2026 possible. We are especially pleased to recognize the following Outstanding Reviewers, whose high-quality reviews (as judged by their Area Chairs) placed them among the top 5% of reviewers.

[#CVPR2026 @CVPR]

Thank you to all the #CVPR2026 Area Chairs for steering the review process.

We would like to give special recognition to the following Outstanding Area Chairs for their exceptional efforts and leadership.