Seungwook Kim

PhD @ POSTECH, prev intern @ ByteDance Seed

[Sahil Bloom @SahilBloom]

I believe that 99% of success is just the ability to outlast uncertainty. The one who can tolerate the most uncertainty will eventually win.

[Remy Li🚀 @theRemyLi]

斯坦福这门课程(cs329z)太硬核了，已经开始教授在实际中怎么搭建AI Agent，涉及到RAG、Tool Use、MCP、Memory、Multi-Agent、Eval、Coding Agent……

不管是大学生还是业务工作者，都强烈建议学这门课。
我自己也打算跟着看看，就是不知道时间精力是否跟得上。😂

https://t.co/GE9JvEfGED

[Katja Schwarz @K_S_Schwarz]

Atlas + Ian = Magic ✨

[Emanuele Rossi @emaros96]

🐬 OpenWhistle is Spotlight @ NeurIPS2026 (top 1.3%)!

We release the largest open dataset of dolphin whistles (114h), enabling ML at scale to better understand one of the most complex non-human communication systems.

Led by @FMustun & @chiara_semenzin
📄 https://t.co/jeTvlmQQnk

[Ramith Hettiarachchi @ramith__]

did a group presentation today with @3blue1brown's manim style transitions. Sharing the template for anyone to build upon

Brought to life with @NarekMaloyan's manim-web + reveal.js + claude + codex

view template: https://t.co/LOSKmrtlXa
view repo: https://t.co/gTb9xxYp7F

[Vincent Sitzmann @vincesitzmann]

I have talked much about the "bitter lesson" of computer vision, arguing that many of the intermediate tasks that vision has solved traditionally - 3D reconstruction, segmentation, tracking, etc - are not necessary to build AI that understands the world through vision... (1/n)

[Lucas Beyer (bl16) @giffmana]

Love this type of work, nice one:

[Ren @Ryrenz]

⛰️ 太惊艳了，微软的 TRELLIS.2 能把一张平面图直接变成能上手用的 3D 资产。

GitHub 上 1.2 万 star，一千四百多个 fork，背后有论文（arXiv 2512.14692），40 亿参数的模型权重和在线演示都放了出来。

做 3D 的朋友最清楚，以前拿一张参考图起模型，要先建基础形体，再手动缝拓扑、补 UV，最后一张张贴材质，一个道具磨一天不稀奇。TRELLIS.2 用的是一种叫 O-Voxel 的稀疏体素结构，开放表面、非 manifold 的几何、内部封闭结构都能直接生成，不用先做有损转换；出来的东西带完整 PBR 材质，粗糙度、金属度、透明度都在里面，渲染出来有透明效果也不用返工。

速度写得很实在：在 H100 上 512³ 分辨率约 3 秒，1024³ 约 17 秒，1536³ 约 60 秒。门槛也没藏着，目前只在 Linux 上测过，显卡至少要 24GB 显存，A100 和 H100 上验证过。

模型和代码按 MIT 许可放出来，这种大方程度不多见。

GitHub：https://t.co/iyhFIjcBbX

[Ziwei Liu @liuziwei7]

🎉 Native, encoder-free VLMs are an important step toward more unified multimodal intelligence.

We’re also pioneering this paradigm with our NEO & SenseNova-U series:

- NEO: https://t.co/qooHtOdaM6
- SN-U: https://t.co/U2uOLI1ISL

Excited to see the field move forward together!

[Akshay 🚀 @akshay_pachaar]

Google just open-sourced Harness².

(turn any harness into a self-improving system)

a harness is everything wrapped around the model, like the system prompt, tools, memory, skills, and subagents. today a person designs it once, and the agent runs every new task with that same fixed setup.

Harness² lets the agent refine that setup for each new task on its own, with the model weights left untouched.

on every task, the agent proposes edits to its harness, runs a few candidate versions, and composes a better harness by comparing how those runs behaved.

because it learns from the differences between runs, it needs no correct answer or grader for the task.

that covers a single task. Harness² also carries what it learns from one task into the next.

after each task, it saves the harness it settled on along with the lessons that held up across the runs. the next time it refines a harness, it starts from those notes instead of from scratch.

this means every finished task makes the next round of edits more informed.

Google tested this with Gemini 3.5 Flash and GLM-5 on legal, office, and professional work benchmarks. one round of edits was enough to beat their original harnesses, and giving the agent more rounds pushed scores higher.

those extra rounds can go wide or deep. the agent either tries several candidate harnesses at once or keeps refining a single one step by step, using OpenCode or Codex to carry out the work.

you can find the Harness² repo here: https://t.co/i00WLeeFpw

I wrote about a similar idea for agents already running in production, where the harness should repair itself after a failure instead of waiting for someone to debug the trace by hand.

that piece walks through how @Cometml's Opik (open-source) closes this loop, going from a failing trace to a proposed fix to a regression test that stops the same failure from coming back. it's quoted below.

[Jay Yang @Jayyanginspires]

CEO of Stripe's advice to young ambitious people:

[Zayn Hao @ZaynHao]

绝对是本周最值得读的文章之一，来着 Claude 团队的 Felix 的个人博客。

一篇帮助你理解如何更好地与 AI 协作，如何更聪明地使用 AI。

在这个能力每隔几周就突飞猛进的世界里，我们需要一些技巧，让我们能够做到以前从未想过或认为不可能的事情。

其中有个对我有启发的点：

关于如何把一件事情做到极致的指令或思路：坚持使用业内顶尖专家使用的工具；选择一些看似不切实际的目标，然后告诉模型这些目标非常合理；挑战你的认知，创造新知识，使用外部评判者。

甚至还搭配了一个提示词：

我想要的是一个更深入的版本，而不是一个更大的版本。在重写题目之前：（1）请说明该领域的顶尖专家会做哪些我没有要求的事情；（2）请列出他们会从哪些最原始的素材入手，而不是我已经整理好的素材；（3）请列出他们会使用的具体方法和工具；（4）除了你我之外，还有什么可以证明结果是正确的。然后，请修改我的题目，使其完全符合上述要求。

[lifcc @mylifcc]

OpenAI 前首席科学家 Ilya Sutskever 曾给传奇程序员 John Carmack 开过一份神秘书单。

Ilya 说：“如果你真正吃透这 30 篇论文，你就掌握了如今 AI 领域 90% 的精髓。”

但纯啃英文 PDF 和数学公式太反人类了。最近有位硬核小哥用 30 天时间，把这 30 篇神作全部手搓复现成了极简 Python 代码，并完整开源！

想从底层彻底看懂现代 AI？这个宝藏项目建议先马后看 🧵👇

[Yuandong Tian @tydsh]

A great read. The blog post captures years of progress in mechanistic interpretability and offers an optimistic vision: perhaps we really can open the black box and understand how neural networks work.

I share that optimism, but with some reservations.

I'm skeptical of explanations that stop at individual neurons, accompanied by compelling visualizations that seem to "explain everything."

Humans are remarkably good at convincing themselves. It's easy to cherry-pick evidence that supports a hypothesis, especially when the visualizations look convincing.

I'm also not convinced that "Recursive Generalization" is fundamentally different from hierarchical representation learning. As representations become more abstract, they discard details while preserving invariances. Isn't that essentially what hierarchical representations have always been about?

IMHO, we need to move beyond a largely descriptive, behavioral approach toward a principled mathematical framework. And we have a natural starting point: gradient descent.

Rather than just asking what a network has learned, we should understand why learning dynamics produce these structures in the first place.

Of course, this is extraordinarily difficult. Look at OpenAI's 719 mathematical proofs. We see plausible proofs tackling problems at the Fields Medal level, yet no comparable breakthrough that cracks deep learning theory. That contrast is striking.

Part of the difficulty is that there are countless plausible theories of how deep networks work. Proving even one can be mathematically formidable. And a proof alone isn't enough. Each theory needs empirical validation, often requiring substantial GPU compute. It's a daunting combination of mathematical and experimental challenges.

Still, I believe this is the way forward. Interpretability gives us observations and hypotheses. But ultimately, we need a theory that explains why these phenomena emerge from the learning process itself.

That's how we turn compelling stories about neural networks into actual understanding. And hopefully recursive self-improvement can help with that 😀

[virgilxbt @virgilxbt]

This is f*cking gold..

A Chinese lab just dropped "The End of Documentation" and it reads like an obituary for every README on GitHub.

The claim in one line: documentation is a rumor about the program, the artifact is the truth.

The text was written once, at release. The code moved on the next morning. Every agent built on the docs is built on a memory.

The fix is one reading procedure the model runs on the artifact itself. Six steps: inventory, structure, behavior, data flow, rebuild, verify until the diff is empty.

Across 26 systems the recovered spec beat the official docs in 23 and surfaced 61 behaviors the docs never mentioned.

So you stop being the person who reads the manual. You become the person who points: you choose what gets opened and you check the diff.

Docs turn disposable. The artifact turns into the job.

That is the exact system I run with REA + Kimi K3 and documented below: one prompt, 14 binaries, 14 rebuildable specs, zero docs opened.

The whole paper in one line: the program is the spec, the model is the reader ↓

[Liang Zheng @LiangZheng_06]

Representation conditioning is better than representation alignment.

The two recent papers reported similar observations. Time for some new trials?

RECON: https://t.co/Zc5HEnU4BN
Dino forcing: https://t.co/yBW8IgNy4f

[Chris Choy, Ph.D. @realChrisChoy]

S4VY is an incredible advancement in the field of real-time 4D joint segmentation and depth prediction! We’ll be releasing the code and checkpoints soon, so stay tuned for more updates.

[Miles Cranmer @MilesCranmer]

Here is the full talk:

[Max For AI @MaxForAI]

有点震撼。。。我们可能真的快要搞清楚大模型内部到底是怎么运作的了。

OpenAI早期机械可解释性研究的参与者，昨天突然提出了一个很大胆的结论：

我们可能一直误解了神经网络。

昨天， @nickcammarata 发了一篇很有意思的文章，叫《Neural Networks Are Already Clean》。

这哥们早年就在OpenAI做机械可解释性研究，2023 年还牵头启动了@OpenAI 第一个大规模自动化神经元解释项目。

当时他们发现，现有工具只能解释大约 2% 的神经元。

Nick因此得出了一个很悲观的判断：神经网络内部存在大量纠缠，想搞清楚它们到底在干什么，可能非常困难。

结果三年后，他带着更先进的工具和Agent重新研究，发现情况完全不同。

在他们测试的现代视觉模型中，最后一层 99%～100% 的神经元都能找到相当清晰的解释。

甚至在Qwen3-32B这样的语言模型中，初步研究也发现，约90%的神经元都有一个可以识别的主要功能。

最有意思的是，他还重新研究了过去的一些经典案例。

比如曾经有一个神经元，既会对猫脸产生反应，也会对汽车前脸产生反应。

以前大家觉得，这是典型的Superposition，多个不相关的概念纠缠在了同一个神经元里。

结果重新研究后发现，这个神经元可能只是在识别一种共同的V形视觉结构。

猫脸和汽车前脸恰好都有这种特征。

全世界的家人们折腾了这么多年，可能连最基础的观察方式都有问题。

而且我觉得这篇文章最值得关注的，其实是他们使用 Agent的方式。

他们让Agent自动观察神经元、提出功能假设、设计实验、分析电路，再通过新的输入验证假设，甚至自动审查了10 万张神经元归因图。

这其实也是我最近一直关注的一个方向：让 AI 自己去做科学研究，把大量观察、假设、验证的工作自动化。

过去需要研究员花大量时间逐个分析的神经元，现在已经可以让Agent成规模地开展研究了。

当然，目前这些结果还处于早期阶段，距离完整解释大模型的内部机制，还有很多工作要做。

但我觉得Nick提出的这个观点非常有意思：

神经网络内部可能一直都很有规律，只是过去我们的工具太原始，根本看不清楚。

如果这个方向最终被验证，机械可解释性领域过去很多基础假设，恐怕都需要重新审视了。

[Graham Neubig @gneubig]

Very ambitious work by Seungone on trying to figure out if LLMs can actually write strong follow-up papers across the natural sciences.

Based on our previous work on automated paper review, we apply rigorous standards to the generated papers and find that even with frontier models, none are hallucination-free and meet the bar of the scientists working in each domain.

We believe that this poses a challenge for upcoming LMs, can they actually do an end-to-end research work and generate an artifact that could move science forward?