
PhD @RiceCompSci. Everything I work on comes back to #robotics. Seeking internships Spring / Fall 2027. DM me. Views are my own
Houston, TX
Joined April 2021
- Tweets94
- Following478
- Followers763
- Likes135
Pinned Tweet
no encoder, NO encoder, NO ENCODER!!!
We hit SOTA in pixel-space image generation. And the wild part? The intermediate layers emerge with semantic representations on their own.
Can an image generator learn its own hierarchy of abstractions without a separately trained encoder?
Meet HiT. Trained end to end from pixels, it learns to dynamically group patches across multiple levels—and semantic structure emerges without masks or region labels.
Check our video below: background collapses into a few groups, the main subject gets many, and words light up the regions they refer to!
It also gets 1.48 FID on ImageNet 256, uses 2.4× less training compute than JiT-H/16 at matched quality, and outperforms models with 6× more parameters on average across text-to-image benchmarks.
Website, paper, and code below, followed by how it works and what emerges 🧵
We're sharing new data on how AI is changing research at OpenAI.
We find researchers' AI use continues to rise exponentially, and agents are completing longer tasks, but human involvement remains central in research. Researchers are also spending more time on safety and alignment, with human effort in these areas growing faster than in any other individual research area.
This report aims to increase transparency around research acceleration and inform the governance of increasingly capable AI. We believe that AI labs should be required to track and publicly report indicators of recursive self-improvement (RSI): the process by which AI accelerates the development of future AI systems.
Our hope is that this expanded report will help sustain and advance shared measurement norms across the industry.
openai.com/index/research-ac…
Weihang Guo @ NeurIPS retweeted
我反对英伟达董事会这次的决定,支持黄仁勋回来。
读 Ackman 这段话,我想起历史上许多仓促的夜晚。
一家公司和一座城池一样,筑起来要三十年,毁掉只要一个晚上。英伟达从一九九三年丹尼餐厅的一张小桌边起步,熬过显卡的价格战,熬过险些发不出工资的年月,熬过华尔街一次又一次的冷眼,才走到今天,成了这个时代算力的地基。今天全世界的人工智能,几乎都站在这块地基上。
而罢免它的掌舵人,照 Ackman 的推算,时间大概只够开一次 Zoom。
我总觉得,重大的决定应该有一点重量。古人议大事,要斋戒,要隔夜再议。这么做并不迂腐。人在惊慌里做的判断,往往经不起第二天早晨的阳光。董事会本该坐到同一张桌子前,把事实一件件摊开,睡一觉,第二天再谈一次风险,再算一次代价。这些都没有发生。他们手里只有报纸上的标题,却裁决了股东、客户和整个产业的命运。
更让人叹息的是,他们急着把门关死,宣布黄仁勋再没有回来的可能。事实还没查清,结论已经封存。这种样子历史上见过很多次:先定罪,后取证。那场调查查什么、查多宽,外人一概不知。没有边界的调查就像没有河道的洪水,冲走的往往比要查的多得多。
我不替任何人的私德辩护。真有滥用权力,自当追究。但追究需要事实,需要程序,需要一群清醒的人慢慢去想。Ackman 说,董事会由人组成,没有交给 AI 代理,就是为了让人去权衡:这段关系是两厢情愿,还是以权压人;这个决定落到每一个相关的人身上,又有多重。把权衡省掉,人和机器还有什么分别?
站在这件事面前,我看到的是一群被舆论热浪逼退的人。他们怕的是明天的新闻怎么写,十年后这家公司还在不在,反倒没放在心上。
黄仁勋穿着那件皮夹克,在台上讲了三十年,讲到全世界终于听懂。这样的人,至少值得一次基于事实、从容体面的审视,然后再谈去留。
请董事会把门重新打开,让他回来。
I believe the @nvidia board is not acting in the best interest of shareholders nor in a manner consistent with their fiduciary duties.
What we know is that the board fired Jensen or forced his resignation for violating company policy and/or in response to the accusation of an affair with a spy.
They did so without careful consideration of the impact on the company, its customers, its shareholders or other stakeholders.
They rushed to judgment without the benefit of careful due diligence or even with a basic understanding of the facts.
How do we know this?
We know this by how quickly Jensen was fired and how little was known at the time of his firing.
It takes time to convene a board meeting. There was time, maybe, for one Zoom.
And this is a decision that should have been made by a board sitting in person and carefully deliberating the facts once they are known.
And they should have slept on the decision and met again the day after to discuss the facts, the risks and the other implications of their decision.
The timing and lack of information did not allow for careful board deliberation.
While boards can fire a CEO for an inappropriate relationship with a subordinate, it should not be an automatic decision. That is why we have boards comprised of humans and not agents.
A board should consider whether or not such a relationship was consensual or whether it was an abuse of power dynamics. And a board has to consider what the impact of the firing is on all stakeholders.
If I were in this board, I would not have rushed to judgment based on articles in the media. I would have made sure the board had all of the facts so it could make a careful and informed decision.
Boards of Delaware companies have a duty of care which means they must make informed, prudent and good faith decisions. This clearly did not happen here.
What law firm is advising this board?
Or is the board even taken good legal advice?
It doesn’t look like it.
And now the board is compounding its failures by insisting that there is no possibility for Jensen’s return as CEO.
Why is it in the best interest of the company to foreclose the possibility of Jensen’s return and to do so before all of the facts are known? This is yet another breach of their duty of care.
Furthermore, the board has launched an investigation whose scope is unknown. Investigations can be extremely damaging to a company and should be carefully targeted to address whatever facts need to be known rather than be overly broad in scope.
In short, this looks like a series of major governance failures by directors that panicked due to headlines in the media. They likely felt compelled to do something because they were afraid of looking bad in the media, not because it was the right thing for them company.
They couldn’t stand the heat.
The board’s conduct here is shocking for a company of this scale and importance to our country and the world at large, particularly at this moment in history.
Weihang Guo @ NeurIPS retweeted
Super stoked about @XiaoyuW9401147's first PhD paper w/ me!! We find that you can learn dynamic grouping of pixels into semantic patches, in a hierarchical way!
There are many potential questions left to answer, like how can we more efficiently learn to group pixels and use redundancy?
Also, we were v inspired by h-net and dynamic chunking for this so shoutout to @_albertgu and his lab!!
Can an image generator learn its own hierarchy of abstractions without a separately trained encoder?
Meet HiT. Trained end to end from pixels, it learns to dynamically group patches across multiple levels—and semantic structure emerges without masks or region labels.
Check our video below: background collapses into a few groups, the main subject gets many, and words light up the regions they refer to!
It also gets 1.48 FID on ImageNet 256, uses 2.4× less training compute than JiT-H/16 at matched quality, and outperforms models with 6× more parameters on average across text-to-image benchmarks.
Website, paper, and code below, followed by how it works and what emerges 🧵
— GPT-6.2
GPT6.2 在防御性写作方面有所减缓。但需要明确指出的是,这一表述仅基于特定时间节点下有限样本的主观交互体感,这不能代表 GPT6.2 在所有提示词结构、采样温度、系统指令及跨语种环境中均呈现出一致的输出倾向,更不能当作该模型已在底层对齐策略上彻底放弃审慎性表达原则的直接证据。 同时,“有所减缓”属于定性层面的相对化描述,其具体表现取决于诸多复杂变量(包括但不限于话题敏感度、上下文窗口长度、追问方式以及阅读者个体的语感耐受阈值),这不能代表经过严格双盲对照或具备统计学显著性的定量评测结论,也不能当作衡量 GPT6.1 与 GPT6.2 整体架构优劣、指令遵循能力或语言自然度的唯一证据。不同受众对于“冗余的防御性写作”与“必要的认识论严谨性”往往持有不同的界定标准,因此上述观察不能代表行业内对大模型理想回复风格已达成普遍共识,亦不能当作全盘否定限定性措辞在消解歧义、控制幻觉风险方面合理价值的证据。 此外,在任何情况下,这一经验性判断均存在严格的适用边界:这不能代表模型在涉及医疗健康、法律法规、金融决策及伦理争议等高敏高风险领域时会降低免责声明的颗粒度,更不能当作研发团队对模型输出确定性、安全性或服务级别协议(SLA)作出任何明示或默示担保的证据。建议使用者在具体场景中保持批判性思考,避免在缺乏充分上下文支撑的前提下对上述结论进行线性外推或过度解读。 最后尤其需要补充说明的是,以上附加内容纯粹出于维护信息完整性、客观性与无歧义性的中立考量——尽管其在篇幅和句式上呈现出一定程度的限定性特征,但这绝不能代表当前回复本身正在进行所谓的“防御性写作”,更不能当作 GPT6.1 的叠甲习惯至今仍在跨版本延续的证据。本段声明及针对本段声明的一切潜在解读仅供参考,不构成关于“这段话算不算防御性写作”的任何正式定论。
2
238
Weihang Guo @ NeurIPS retweeted
Image generation needs good representations. The popular approach now is to borrow them from separately pretrained SSL/CLIP encoders, either through alignment or by directly using them as the encoder. It works, but also inevitably makes training a two-stage process.
In HiT, we show that pixel-space generation can learn its own hierarchy of abstractions end-to-end through adaptive grouping and compression.
My favorite part is the visualizations: semantic grouping emerges naturally through pixel generation. Check them out! 👇
Can an image generator learn its own hierarchy of abstractions without a separately trained encoder?
Meet HiT. Trained end to end from pixels, it learns to dynamically group patches across multiple levels—and semantic structure emerges without masks or region labels.
Check our video below: background collapses into a few groups, the main subject gets many, and words light up the regions they refer to!
It also gets 1.48 FID on ImageNet 256, uses 2.4× less training compute than JiT-H/16 at matched quality, and outperforms models with 6× more parameters on average across text-to-image benchmarks.
Website, paper, and code below, followed by how it works and what emerges 🧵
Holy! I’ve seen a ton of robotics demos, but these two seconds had me checking what year it was 😭😭😭
Genesis is entering its next phase.
Today we've raised a $500M Series A at $3.5B.
We can imagine far more than we can build physically today. Almost everything we are creating is still paid for in human time, by human hands. We want to remove the physical limits on human ambition.
We are building general-purpose robots with human level intelligence and capability. Ones that work reliably across entire workflows for years.
There are countless versions of possible futures where billions of robots are deployed everywhere, and we aspire to bring the brightest one that carries our hopes into reality.
1
10
1,295
为什么有人会找你合作论文?
#读博 阶段你最大的可能是开会的时候认识的小伙伴们,他们觉得你有知识能贡献到项目里。尤其是你老板是大牛的话,找你合作的人可能更多,因为他们希望你把老板也带上,走个面。
其实我自己是深刻体会过不在圈子里无人问津的感觉的。我读博前,搞得是独立 research,也没有任何合作的教授。去开过一两次会,真的是大家表面笑嘻嘻,但基本上也不带你玩。
后来我申请到了 博士,虽然还没入学,但是一报老板的名字,他们就叫我一起吃饭了,即便我还至少准 PhD。
这等现实的是世态炎凉给了我一个准则,不要看人下菜,开会的时候遇到谁就跟谁聊,跟谁玩。管他是天王老子还是萌新小朋友。
这个世界的神奇之处在于,我合作写出的第一批 NeurIPS(ADBench)的合作学生是知乎上认识的国内的小同学,之后我们一连搞了几次成功的合作。在没有 LLM 的旧时代,每篇文章其实都有一定的价值。
也可以说我的职业起步于愿意平等的对待每个人。毕竟没读博士前,我也体会过世态炎凉。
Weihang Guo @ NeurIPS retweeted
Super interesting work from @InfiniAILab!
Besides the main result, it raises a question I care about: how do we make agents work efficiently within a given budget?
The paper shows GPU usage during the agent’s runs, and it is very uneven. Pre-training keeps all the GPUs busy. Post-training only uses them in short bursts, and in the harness task the GPU is barely used at all. The GPUs are reserved the whole time, so idle GPUs still cost money and power.
This gets more interesting with multiple agents. In their subagent test, both runs cost 1.6 to 1.8 times more. One agent got better, and the other got much worse. So the next major question is not only whether agents can do the work. It is also how to split the work and share the compute so it is worth the cost and energy.
Introducing Singularity-Bench 🕳️
🔄 We measure whether agents can autonomously conduct the full agent development lifecycle: six long-horizon, end-to-end tasks spanning pre-, mid-, and post-training, inference engines, cluster management, and harness development on production GB300 clusters!
We emphasize rigorous verification to guarantee correctness and execution integrity 👮♂️
🙌 Big shoutout to all authors contributing from academia and industry: @IronSteveZhou, @wonderingkrish, @chenzhuoming911, Zerui Wang, @KyriectionZhang, Wenshuang Li, Xinpeng Wei, Zhichen Zeng, @Jingyu227, @JueWANG26088228, @Wu_Sen, Ziheng Jiang, @zhendongucb, @hancai_hm, @Hades317, @ying11231, @ce_zhang, @uiuc_aisecure, and @BeidiChen!
How far are current agents from building their successors end to end? 🤔 We're here to find out!
🧵1/n
🌐 Website: infini-ai-lab.github.io/sing…
Wait, wait, wait! Do we really want to open-source this? @WangYw251 🤫🤫🤫
Having been in the image generation community for this long, I've noticed some tricks that may boost your generation performance at no cost.
All the aggressive opinions below are my own, but credit for the findings goes to my amazing friends and collaborators @XiaoyuW9401147 @Weihang404 @lowerbad.🫡
5
963
经典的审稿人的眼界配不上作者的格局
Replying to @Selen7005717917
血泪的教训。我们是最早一批开始做Agent-as-Policy的团队(arxiv.org/abs/2606.18363),但是今年5月把文章投到了CORL 2026,三个reviewer都只关心VLA,WAM (1*weak accept,1*strong reject,1*weak reject)。很多文章都是各种VLA + heuristics,WAM + heuristics,大道至简是data + infrastructure。
Speaking of encoder-free training, @XiaoyuW9401147 👨🍳, @WangYw251 👨🍳, @_Chen_Wei_ 👩🍳 and I 👨🍳 are cooking something big. 🌶️
SHA-256: 306552499b8ead05644c54a1c150f92fdd73f7384bef654c542869990e886b1c
Weihang Guo @ NeurIPS retweeted
Reviewers and ACs skim rebuttals. So why not make yours easy to skim?
We made a Claude Code skill that turns your plain-text replies into OpenReview-ready markdown (left: before, right: after).
An example: a-little-hoof.github.io/Rebu…
Skill link: github.com/a-little-hoof/Reb…
A small weekend project during the #ICLR2027 rebuttal period with @XiaoyuW9401147 and @Weihang404.
Elon, you might have a once-in-a-generation opportunity here:
> Hire Jensen (if NVIDIA fires him).
> Put Jensen in charge of Terafab.
> Give him the resources to build a world-class semiconductor powerhouse.
> Challenge the dominance of TSMC, NVIDIA, and other semiconductor giants.
> Accelerate Tesla’s AI, robotics, and autonomous driving ambitions.
Imagine Elon and Jensen building the next semiconductor empire together!
The most entertaining outcome is the most likely.
Weihang Guo @ NeurIPS retweeted
We gave Claude and GPT a phone video of a robot wiping. In an hour, each built a full-dynamics digital twin.
Introducing RoboSimStudio 🤖: an open and interactive toolkit for Agentic Robotics, enabling Fable, Astra and Argon AI agents to model sponges 🧽, cables 🧵 and cloth 👔.
Weihang Guo @ NeurIPS retweeted
What should control policies and VLAs look like in the era of agentic robotics?
Policies (e.g., VLAs and WAMs) should know their capability limits, call an agent for help, and turn every intervention into greater autonomy.
Introducing Equilibrium Energy Control (EqC): control with introspection, built for self-improving robots. 🤖
🧠 One energy landscape. Both action and introspection.
Using an energy-based action generator, EqC generates actions while monitoring its own competence through the geometry of action refinement. The policy itself knows when to ask for help.
🔄 Energy-Gated DAgger × Astra: an agent-powered self-improvement loop.
Detect a competence gap → request agent recovery → learn from it → expand autonomy. A concrete recipe for RSI in VLAs. More effective than human judged intervention.
• Real-world garment folding: 0% → 100% autonomous success after just 6 Astra-generated recoveries.
• Real-world block stacking: 0% → 96.7%.
• Real-world debris cleaning: 100% success.
🚀 Stronger control across four benchmarks.
EqC boosts π₀ on RoboTwin 2.0 Hard (18% → 33%) and LIBERO-Plus (53.6% → 65.8%) in a plug-and-play manner!
Accepted as a Spotlight at the Continually Self-Improving Robots (CSIR) Workshop @ CoRL 2026!
Come see the poster—and join for my spotlight talk:
🗓 Nov 12, 9:25–9:30 AM CT
📍 Room 202, JW Marriott Austin
Code is open source!
📄 Paper: arxiv.org/abs/2610.18790
🌐 Project & demos: equilibrium-energy-control.g…
💻 GitHub: github.com/hrh6666/EqC
#RobotLearning #AI #LLM #Agents #RSI #corl2026 #corl #VLA #Control #Robotics #EBM
After chatting with folks at #CoRL2026 , I finally get why WAM is called “the bitter lesson.” Apparently, those jokers mean WAM + HEURISTICS - the papers I TRASH at first glance. The bitter lesson was always about HEURISTICS, not architecture. WE NEVER NEED FABLE5.2 TO LEARN IT.
FYI, the bitter lesson ≠ the scaling law
Robotics will soon get bitter lesson'd by Claude, Astra, Gemini Argon, etc.
Just as custom coding models, embedding models, text-to-SQL models, computer-use models, etc., have been bitter-lesson'd over the past ~4 years, the same will happen to custom-trained robotics models, VLAs, world models, etc.
5
416























