陈哲 | 晚点聊 LateTalk Summary

1. Key Themes

Frontier labs now treat physical AI as a subproblem of AGI, not a separate industry

The central shift of the quarter is a change in research paradigm at the frontier labs. After meeting researchers in the US, 陈哲 summarized the stance: "以OpenAI为代表的前沿实验室，他们对于Physical AI的态度是通过先解决AGI，然后再把Physical AI当做一个子问题来解决。" [00:06:21] He says it is a company-wide consensus: "这是整个公司大的共识。嗯，没有见到神迹的人就是不会信的。" [00:54:37] The lineage runs from Google's Gemini Robotics, which in early 2025 shipped a VLA plus a separate spatial-reasoning model (Robotics-ER), to Anthropic's July "Claude Plays Robotics" work, and then Astra. As he puts it: "所以，语言模型进入物理世界是先从工具调用...的角度来切入的。" [00:07:54]

Astra's impact has three layers, and the third is the one that matters most

陈哲 splits the impact into three uses. (1) Direct policy: Astra outputs end-effector poses and an inverse-dynamics module converts them to motor commands. (2) Supervisor/corrector of a VLA: Galaxea (银河通用) tested Astra controlling π0.5, lifting the success rate on 10 Robo-Dojo tasks from "26%" with Astra alone to "48%". [00:19:53] (3) Astra as the robotics researcher itself: "它直接在Blender里面搭出这个仿真环境...然后自己去调试或者修改这个运动控制的策略...也就是在物理世界里面，AI可以自己去训练AI。" [00:20:56] He calls this layer "最让人头皮发麻或者是觉得最有颠覆性的一个地方" [00:21:26], and says it is already "最新的融资概念" ("Physical RSI"). [00:22:44]

Where Astra is strong and weak defines the remaining niche for physical-AI companies

Astra wins on spatial reasoning and generalization. The VLA failure mode where a changed object color or position makes the model fail is much weaker in general models, "因为他模型的体量和训练的方法，是远远超过VLA模型的。" [00:24:57] It loses on precision contact tasks: "它不擅长一些对物理的精度...比如说需要拔插一个非常精细的一个纽扣。因为这些动作在很多VLA模型里面是有专门的遥操作和后训练的储备的。" [00:25:16] Long-term, he expects a layered architecture because of latency: "未来通用机器人真正的部署，大概率还是需要在本机上面有一个实时的高速的模型本地计算。就跟我们今天看Robotaxi...一样。" [00:26:04]

Scaling law and the brain-size analogy explain why VLA/WAM architecture debates are "insignificant"

Today's VLAs and world action models are "几十个billion", while mainstream language models are at "几个trillion甚至十个trillion以上的这个量级。这就是一个初级动物跟高级动物的一个区别。" [00:36:40] He tells embodied startups to stop arguing: "大家可以停止争论这个技术路线，到底是VLA更好还是WAM更好。我觉得在一个超级智能面前，你你的这种模型架构上的这种优化其实是insignificant的。" [00:40:05] On data: "无论你是VLA还是World Action Model，如果你给同样质量、同样规模的数据，你都有可能训出一个一样好的这个这个策略模型。" [00:08:44]

The embodied-AI value chain will get squeezed in the middle, as in autonomous driving

陈哲 maps the smart-car stack (OEMs, Tier-1/algorithm suppliers, core components) onto robotics: "如果它只是停留在中间的话，它会被两端挤压...如果你是一个具身大脑公司，你比起通用大脑又很小，你在这个真正的部署上面，你可能又受限于这个硬件或者执行器的这个定价力...那你在中间怎么找到你的价值闭环？" [00:42:54] The defensible positions he names: good interfaces for general models to call, and owned closed loops such as "触觉或者这个力觉的信息", "工艺具体的这个约束", "失败任务的这种恢复" and task-specific data. [00:40:31]

Hardware/body companies are the safest beneficiary of general-model progress

The Nvidia analogy: "OpenAI再发达，其实英伟达也是受益的...相反，英伟达的需求是得到了极大的放大。" [00:59:17] He says Unitree's shape "是跟英伟达非常像的", with open-source ecosystem effects. [00:59:38] His reasoning: "今天通用机器人的硬件基本上是ready的，剩下不能解决的问题其实都是智能的问题。" [00:44:15] Dexterous hands have also turned into training infrastructure: "我有1000只手，你有100只手...我也可能训练一个模型就比你快10倍。" [01:29:23]

Neo Lab exits: the World Labs/AMD deal is an anchor for strategic acquisitions

AMD agreed to acquire Fei-Fei Li's World Labs (the transcript's hosts cite $8.2B, paid in stock). 陈哲 reads it as strategic, not a revenue valuation: "82亿这个价格可能没有反映他们在业务上面的直接贡献，但是它体现了...一个万亿市值的头部公司愿意为这个事情战略上付出的代价。" [00:50:50] He says the deal acts as an anchor for rumors that Anthropic may buy Physical Intelligence, and as a warning to other Neo Labs: "有可能中短期来看最好的出路就是尽快找到一个大厂卖身...对于大厂来说，可能只有第一个购买的公司是具有价值的。" [00:53:37] Positioning-wise, World Labs is infrastructure, not a competitor: "Astra这种通用模型更加强大过后，可能像World Labs这种专注于空间重建或者仿真环境搭建的公司可能会获得更大的价值。" [00:49:24]

In-context learning (ICL) cuts robot deployment time from days to minutes

Generalist's Gen-1.5 and Skild AI's SE-1 both show ICL. The deployment point: today it takes "几个小时，甚至几天，甚至一周" of teleoperation and post-training to adapt a model, whereas ICL could mean "一个一分钟的这个视频的这个示教". [01:13:23] Two surprises: out-of-distribution tasks can be learned from one demo, and long-sequence recombination of atomic skills works. Generalist uses 3-10 second first-person/robot-view demos, while Skild takes third-person, 10-minute videos with long-horizon skill composition. [01:17:55] He expects Chinese companies to pursue it because commercial deployment is a higher priority there. [01:18:38]

Capital markets are distorting embodied-AI company behavior

The Unitree IPO knock-on: regulators are tightening listings for embodied companies. 陈哲 warns of revenue inflation and, more dangerously, "为了迎合这个收入的需要，迫不得已去做很多其实反行业发展规律的收入。" [01:06:00] The analogy is 2018-19 pretrained-model companies (BERT, T5) building customer-service and healthcare applications: "最后你会发现所有这些投入跟解决通用AI甚至解决ChatGPT而言，都是一个背道而驰的努力。" [01:06:42] His filter: revenue that is "跟通用智能和Scaling Law正相关的", meaning better humanoid hardware, dexterous hands, automated data collection. [01:09:47]

US FCC rule is a manufacturing-origin policy that bites on supply-chain reality

The FCC extended device authorization to foreign-produced advanced robots (over 2 kg, with mobility, sensing, connectivity and control). Rules look at origin of core components, "我记得现在应该是有个65%的一个比例" [01:31:37], so reassembly in the US doesn't help. Existing authorized models (Unitree's G1, R1, H2, robot dog) can still be sold. 陈哲 sees a policy/industry-reality conflict: "中国在制造业、在供应商的这个质量跟密度是十几二十年整个电子产业、汽车产业反复迭代而形成的...很难靠短期的这个限制性政策去改变的。" [01:36:05]

2. Contrarian Perspectives

Physical-AI foundation-model valuations may rest on a collapsing premise

The dominant bull case for Pi, Generalist and others is that Physical AI is bigger than Digital AI, making them "15、16年早期的OpenAI". 陈哲 argues Astra undermines this: "Astra的出现其实一定程度上在告诉大家，其实通用模型是一个更恐怖的存在。物理AI可能只是通用模型的一个子集。如果这个假设成立的话，那么很多物理AI公司的整个估值体系或者说价值链的分布可能会得到打破。" [00:16:47] Supporting evidence: leading embodied models use "千级" GPUs versus "十万甚至百万级" at frontier labs, "差了可能两到三个量级". [00:30:02] But he adds that the market may grow: "物理AI到来的时间可能也因为通用模型的进展而大大加快了。" [00:17:25]

Deep robotics experience may be a handicap in the new paradigm

陈哲, who worked at Google next door to Everyday Robots, gives a "暴论": "这个团队如果越了解Robotics，越有Robotics相关的经验，其实会妨碍他们在通用模型达到这个能力的这个尝试...最初NLP那群人也是一开始最反对LLM，也是最不相信LLM能够work的那群人。" [01:24:25] This partly explains why Google, which invented VLA and Robotics-ER, wasn't the one to deliver the shock. OpenAI's robotics researchers are "普遍比Gemini Robotics的同事可能年轻5~10岁...基本上PhD刚毕业". [01:26:03]

The seemingly "dumb" tool-calling route beats purpose-built embodied models on the frontier

Conventional wisdom says physical intelligence needs specially trained models. Yet an LLM calling robot interfaces as a "新的形态的computer use或者说tool use" matched or beat top embodied models on Robo-Dojo. He ties it to a general principle: "对于AI来说，我调用一个MuJoCo跟我调用一个浏览器、调用Excel原理上是一样的。" [00:33:10] And on speed objections: "你很难跟持续迭代的摩尔定律去抗衡。所以以这个算力或者速度作为一个瓶颈而否定大模型这条路线的人，我觉得长期可能都会失望。" [00:15:21]

"Real revenue" pursuit by embodied startups can be actively harmful

Common investor logic rewards revenue. 陈哲's counter: if you told OpenAI at GPT-1/2 to make money, "OpenAI那个时候会投入很多资源去做各种奇奇怪怪的harness，去做各种fine-tune" [01:06:23], and that would have diverted it from the main path. Some embodied firms use traditional automation and call it embodied deployment. Companies like Unitree and DeepSeek, he argues, are valuable because their pace is "不是被资本市场或者外部环境直接奖励的", and the incentives breed "劣币驱逐良币". [01:08:16]

Frontier-lab reasoning may not transfer to everyone else

A Frontier Lab friend warned 陈哲 not to be misled by their view because it comes from huge data and compute. 陈哲: "可能这个推导对绝大部分创业公司并不成立。" [00:55:28] The Tesla V12 analogy: Chinese firms had to doubt end-to-end driving until they could replicate it; a frontier lab can fund the experiments others can't. [00:54:42]

3. Companies Identified

OpenAI (GPT-6 Astra)

Frontier AI lab. The quarter's catalyst: Astra shows strong spatial reasoning, long-horizon reasoning and robot-API tool use. Reportedly hired many Blender artists in May for 3D modeling and trajectory annotation. 陈哲: "我听说的消息是他们可能在5月份的时候就雇佣了非常多的这个Blender的这个artist，做了大量的这个建模、轨迹和这个数据的标注。" [00:27:38] Also: "我听到的所有的消息指向是他们肯定是用了[egocentric data]...我们站在Scaling Law的角度，他们没有理由不用。" [00:29:37]

Anthropic

Frontier lab; published "Claude Plays Robotics" in July, using Opus to call robot interfaces and output end-effector poses. Also rumored as a potential Physical Intelligence acquirer. 陈哲: "更近一点，说是今年7月份的时候，Anthropic做了一个研究，然后这里面也用了他们最新的这个语言模型Opus来调用机器人的一些接口。" [00:07:46]

Google DeepMind (Gemini Robotics / Robotics-ER 2)

Earliest to see the layered route, shipping a VLA plus Robotics-ER in early 2025, and Robotics 2 plus ER 2 and an on-device model at end of July. 陈哲: "Google其实就看到这样这样一个趋势，就是在复杂的空间的理解跟推理上面，一个通用的语言模型其实是有非常强的能力。" [00:07:37] And: "Google代表了一个头部的模型公司，用一个比较系统性的方法来用一个相对来说统一的架构来解决这三个问题的一个方法。" [01:23:22]

World Labs (Fei-Fei Li) / AMD

Spatial intelligence lab (Atlas model, 3D Gaussian splatting) that agreed to be acquired by AMD for $8.2B in stock. Earlier acquired SceniX, a robotics simulation/validation company founded by Li's former student. 陈哲: "AMD通过收购World Labs能够搭建一个跟这个头部公司类似的一个顶级的Physical AI相关的这种研究团队或者实验室，那我觉得这笔收购对AMD来说就是一个很划算的交易。" [00:50:28]

Unitree (宇树)

Chinese humanoid/quadruped maker, IPO'd on the STAR market. H1 2026 revenue ~RMB 1.152B, up 48%, with humanoid shipment mix rising. 陈哲 raised his conviction: "宇树从业绩来讲，一直是一个非常稳健，也是有这个全球领导力的一家公司。" [00:58:25] He previously said a RMB 100B valuation was reasonable; IPO market cap was RMB 60B+, now ~RMB 170B. Lock-up ends August next year. Released a 6B-parameter VLA in September.

Generalist AI

Embodied-model startup; Gen-1.5 shows in-context learning with 3-10 second demos from robot view. Listed with Pi among those at risk of valuation reset but also a quarter's model leader. [00:04:21]

Skild AI

Embodied-model startup; SE-1 emphasizes third-person, long-horizon (10-minute) video demos and whole-body control. 陈哲: "Skild AI更强调一个第三视角和长时序视角的这个视频输入...他们在报告里面给出的一个例子是10分钟的一个视频。" [01:18:18]

Physical Intelligence (Pi)

Embodied foundation-model company (π0.5); rumored Anthropic acquisition target. Its model was used by Galaxea as the base VLA in the Astra-supervisor experiment (26% to 48%). [00:19:53]

Galaxea (银河通用)

Chinese embodied company; its hybrid-control case showed Astra supervising π0.5 raises Robo-Dojo success from 26% (Astra alone) to 48%. [00:19:53]

Dyna Robotics

Released Dyna 2 and 2.1; 2.1 showcases a full laundry-room workflow with a semi-humanoid robot and whole-body control. 陈哲: "更多的是展示了一个完整的工作流的这个流程...他有比较细的这个工程的这个阐述跟实践。" [01:20:02] Dyna 2 used 1M+ hours of egocentric data, though scaling returns were questioned.

RoboCurve

Startup that was among the first to apply Astra to embodied tasks; ran a September comparison of Astra against an open-source embodied model (MOMU-ACT2) using pose generation. [00:23:23]

Tesla (Optimus / Robotaxi)

Optimus is deprioritized this quarter in favor of Robotaxi; production timelines slipped. The end-to-end V12 is used as an analogy for "seeing the miracle". [00:54:42]

Boston Dynamics (Hyundai)

Supported the FCC restriction but manufactures in Korea, so is seeking exemptions or an ally-country credit. [01:32:26]

Figure, Mantis (US robot vac startup), iRobot

Named as US firms that supported or are affected by the FCC rule; iRobot is the example of a US brand that falls under the rule if over 65% of components are made abroad. [01:31:37]

Nvidia

Used as the model for Unitree's position: a hardware/infrastructure leader that benefits from stronger general models and runs an open ecosystem (and its GEAR Lab). [00:59:17]

Harvey, OpenEvidence

Vertical AI companies cited as the analogy for physical-AI vertical champions that win via access barriers and data compliance, building "百亿美金的公司". [00:41:50]

Qinglang (擎朗) and Pudu (普渡)

Delivery robot leaders of the prior boom, each reportedly burned about RMB 1B in a year on subsidized share-grabbing. [01:03:37]

Huawei

Cited as an example of a Tier-1/solution provider moving up toward OEM value ("华为支架这种方案"). [00:43:04]

Scale AI

Used to illustrate that data-service businesses are phase-dependent as model capability rises. [00:31:31]

Kimi, Zhipu, DeepSeek, ByteDance Seed, Tencent Robotics X

Chinese labs and big tech expected to enter or reorganize around embodied AI. 陈哲: "所有立足于解决通用智能或者解决AGI的公司，都有能力跟意愿去解决部分的物理AI的问题。" [00:38:03]

Everyday Robots (Google X)

Where Google's decade-plus robotics tradition began; cited as an example of experience that may bind teams to the old paradigm. [01:25:07]

Model/tool names mentioned

Claude/Opus, Kimi K2.6 and Qwen 3.6 (tested in the Anthropic paper), MuJoCo and Isaac Gym (simulators), Blender, Robo-Dojo (open benchmark). Video generators SeeDance, Kling and H3 were named as comparisons for World Labs.

4. People Identified

陈哲 (Peter Chen), founding partner at Alphaist

Guest. Former Google employee, long-time observer of robotics; just returned from US/Europe and IROS 2026. His forecasts and frameworks structure the episode.

李飞飞 (Fei-Fei Li)

Founder of World Labs. Her mission: "做一个这个全模态的空间智能模型" [00:46:57], and she still wants to push research frontiers, which needs the compute AMD can provide. [00:49:50]

王兴兴 (Wang Xingxing), Unitree founder

Praised as restrained: "王兴兴在公众场合也是一直是没有那么激进判断的一个人，我觉得可能在大脑这个事情上面也是同样的态度。" [01:00:51]

石冠亚, CMU assistant professor

Ran the dinner survey at IROS of ~100+ researchers on Astra's impact, using the Three-Body framework; about one-third believed "Robotics is over". [00:13:49]

王潜 (Wang Qian), founder of X Square Robot (自变量)

Argued the data pipelines OpenAI and an embodied company need are the same, implying no advantage for labs. 陈哲 agrees with the description but disagrees with the conclusion because frontier labs have better tools, compute and experience. [01:21:39]

梁文锋 (Liang Wenfeng), DeepSeek

Reportedly said in a leaked investor meeting that AGI's next step is physical AGI. [00:37:50]

An unnamed ex-autonomous-driving technical lead turned embodied founder

Reported that long-standing AD corner-case benchmarks (e.g. wheelchair crossing, temporary traffic cones) "在Astra面前都灰飞烟灭". [00:44:34]

An unnamed head embodied-company founder

Said his company raised RMB 3B+ but won't spend it, planning to acquire teams cheaply after the bubble bursts. [00:52:21]

A Frontier Lab friend of 陈哲 (unnamed)

Cautioned him that frontier-lab conclusions rest on data/compute that most startups lack. [00:55:28]

5. Operating Insights

Build for being called, not for being the brain

陈哲 predicts "通用模型作为一个调度器，它会去充分的调用各种物理世界的各种接口和能力，就把自己想象成一个可以被反复调用的这种工具，所以接口要做好，以及留住你自己真正属于你的闭环。" [00:40:31] For an operator: expose clean APIs/tool interfaces to your hardware or workflow and decide which loop (tactile signal, process constraints, failure recovery, accumulated task data) you will own.

Draw the line between "general model can do this" and "never will" before choosing your roadmap

"首先把这个边界想清楚。" [00:40:20] Concretely: general models are weaker on high-precision contact and teleop-trained fine manipulation, stronger on spatial reasoning and generalization; plan to supply the former and stop re-solving the latter.

Buy the high-frequency/low-latency layer and rent the slow brain

Layered deployment: "有一些低频的决策计算可以在网上去做，可以接受一两秒的延时，但是大量的本地计算应该是有本地的算力、本地的模型去执行的。" [00:26:14] Use cloud general models for planning and supervision, and keep a distilled on-device model for the real-time loop.

Use the strongest model as a supervisor/data-pipeline worker even if you won't use it as the policy

Both the supervisor result (26% to 48%) and the point that embodied firms "他们都要用最聪明的大脑还是GPT和Astra的模型" for data pipelines [01:22:03] show cheap leverage: wrap your existing policy with a frontier-model checker and use frontier models to clean and label data.

Treat acquihire/exit windows as perishable

Per 陈哲, for big acquirers only the first purchase of a given category is strategically valuable, so a Neo Lab should engage buyers early. [00:53:37] Conversely, a cash-rich operator can wait for downturns to acquire cheaply.

6. Overlooked Insights

The GPT-6 paper's coverage gap: the unlock was found by the open-source community, not OpenAI

OpenAI's blog never mentioned robotics. The pipeline came from Anthropic's July paper, and the open-source community "用了一两天的时间就适配好了，就把同样的实验在Astra上面复现，结果发现它的成功率和准确率相比Claude大大地提高。" [00:11:45] The signal: a capability pipeline already existed, so any new frontier release can be benchmarked on robotics within days, and robotics gains may arrive as an unannounced side effect of general training.

The 10-year autonomous-driving benchmarks falling to a general model is a leading indicator for embodied tasks

陈哲 mentions almost in passing that corner-case benchmarks refined for a decade were "灰飞烟灭" under Astra, then asks: "如果自动驾驶出现了这样的情况，为什么在具身智能或者别的问题里面不会出现同样的情况呢？" [00:45:10] This is a cross-domain tell. Embodied startups should check whether their moat is a benchmark-defined problem that a general model can absorb overnight.