166: 许华哲再次具身创业：不想错过最大的西瓜
- 01The Physical AGI Conviction: Why General-Purpose Beats Specialized Every Time
- 02AI Native Over Path Dependency: The Three Wrong Paths in Embodied AI
- 03The 18-24 Month Household Robot Inflection Point
Episode 166 | 晚点聊 LateTalk | Guest: Xu Huazhe (许华哲), Founder of PoKe Robotics (破壳机器人)
1. Key Themes
The Physical AGI Conviction: Why General-Purpose Beats Specialized Every Time
Xu Huazhe's core thesis is that most embodied AI companies are building the wrong thing. He believes the industry is too focused on narrow industrial deployment when the real prize is a general-purpose humanoid. His analogy is sharp: building an AI customer service agent versus building a general large language model.
"Should I make something that answers all questions as a smart customer service bot, or should I do what today's large language model does? I think the answer is obvious. This is why I don't reject deployment per se — I just think we should also focus on how to truly develop intelligence capabilities, making it genuinely universal. Because only becoming universal makes embodied intelligence truly capable of changing how humanity exists." [00:07:41]
He argues the economics of specialized humanoids never work — a single joint costs ~1,000 RMB, and you need seven joints just for one arm. A device-specific robot can never justify its cost. But a universal robot that can bartend, farm, and teach — all in the same body — can.
"The way I imagine it, it's not even about penetration — because ultimately what we're doing is building a person. Even though right now we're doing household robots, ultimately we want to build a person. So after work, he can be someone dancing in a bar, someone farming in a field, someone teaching in a classroom — but it's all the same robot, because it's general-purpose. That's when the economics work." [00:15:53]
AI Native Over Path Dependency: The Three Wrong Paths in Embodied AI
Xu explicitly names three legacy mindsets that will fail in embodied intelligence, and frames his company's approach as a clean break from all of them.
"Not traditional robotics. Not autonomous driving. Not prehistoric deep learning." [00:44:32]
On traditional robotics: it solves spectacular but single-purpose tasks. It cannot generalize. On autonomous driving thinking: it pursues data flywheels in narrow geographies — "get all the roads around Tsinghua, then Zhongguancun, then Haidian" — but this closed-loop approach produces brittle, non-generalizable intelligence. On small deep learning models: stacking 100 small models never equals one large model.
"The key insight is that AI is fundamentally an induction machine. The more diverse cases it sees, the better the conclusions it draws. And crucially, this cannot be patched in later. Once it's concluded 'all people use water cups,' it's very hard to correct that with new data afterward. So from the start, you should put all the data together." [00:48:29]
The 18-24 Month Household Robot Inflection Point
Xu and Eric Jang (former Chief Scientist at 1X) independently converged on the same prediction: a robot will enter a real home and do meaningful work within 18 months. He describes the trajectory as similar to LLMs — gradual, then suddenly transformative.
"I predict that within the next 18 to 24 months, there will be a robot in this world that actually walks into a home and does something — maybe serves some humans. Perhaps it won't exist as a fully stable product yet, but at least it will be able to work continuously for a long time, handle quite a few things in the home, and make a person feel — and believe — that the future has arrived." [00:05:14]
His company's specific milestone: by early 2028, robots should be in actual homes doing real work. He plans an office designed as a home, so training and real-world testing happen in the same space.
2. Contrarian Perspectives
Robot Daily Active Users (DAU) Is the Only Metric That Matters — And Nobody Tracks It
Xu introduces a concept almost no one in the industry talks about: robot utilization rate. He believes current shipment numbers are largely theater.
"I want to see a number called robot active rate. Daily active users. I want to see its DAU. Of the 5,000 robots sold, are 10 in daily use? 100? 1,000? 5,000? This framing is good. But you can't see it. It's hardware, essentially a fixed asset. Nobody will show it to you because it would look terrible." [00:124:52]
He suspects a large fraction of shipped robots are sitting idle — bought for demos, trade shows, or investor optics. This implies reported "traction" in embodied AI is far weaker than it appears.
Selling Training Data to Competitors Is Existential Self-Harm
Xu identifies a dangerous and under-discussed pattern: Chinese embodied AI companies selling proprietary robot training data to overseas buyers (including Nvidia) to generate short-term revenue.
"There are a huge number of companies in this industry collecting data and then selling it. Because Nvidia and others don't have the capacity to collect data themselves — their people are too expensive. So they buy from companies with large human operational capacity. We know this is ammunition. Yet we still sell the ammunition... It's like selling your most precious resource to your competitor. Just to make the books look better or to raise a round." [01:23:23]
He frames this as a strategic error at a civilizational level — not just a business mistake.
Scalable Post-Training for Generalization — Not Task-Specific Fine-Tuning
Most companies pre-train a large model, then fine-tune it on a specific task — effectively collapsing the model's generality. Xu argues this is the wrong approach and that his company will pursue "scaled post-training" that preserves universality.
"We want to do large-scale post-training — let the model maintain its generalizability even after post-training, and still be capable of many, many types of tasks. This is something most people doing pre-training post-training do relatively little of... Don't worry if it's mediocre on everything — because in a year it will be excellent on everything. That's better than being excellent at one thing and useless everywhere else." [01:04:32]
Product-Defined Safety Constraints Beat Algorithmic Safety Guarantees
Xu makes a product design argument that runs counter to how safety is typically discussed in robotics. Rather than trying to make the AI safe through algorithms, he argues for design constraints that make catastrophic failure structurally impossible.
"From the very first day, we said: we will not do any service involving direct physical contact with the human body — no wiping, no turning people over, no lifting, no baby-holding, no massage. All direct human contact is off the table. Because if something goes wrong there — even if the algorithm can mostly avoid it — the risk is enormous... Like telling a child: don't play with matches, don't touch the gas, don't touch sharp things. Set those rules, and the worst that happens is the robot drops your TV remote." [00:54:51]
Physical AGI Will Arrive in ~5 Years and the Window to Compete Is Closing
Xu believes true human-level general physical intelligence will arrive within 5 years, and that after a "singularity" point — where AI improves AI, and robots build robots — the competitive window closes permanently.
"When that singularity arrives, some very frightening things will happen. Large models writing large models to improve large models. Robots tightening bolts on robots to build robots. When its intelligence reaches that point, the game is over. So we must compete before that point arrives... I think full human-level general capability is maybe five years away." [01:22:22]
3. Companies Identified
Physical Intelligence (π / Pi) Leading U.S. embodied AI research company. Mentioned as the global benchmark for intelligence-first embodied AI. Xu cited their approach of shipping incrementally every three months, maintaining no explicit commercialization pressure, and staying true to their founding vision as a model for strategic discipline.
"Look at Pi — from day one, their name literally is what they do: Physical Intelligence. They just kept going and going — publish something every three months, still in the top tier. No dramatic commercial pivot. I think they have stronger strategic conviction." [01:19:12]
Generalist (Google DeepMind) Mentioned for the scale of its dataset and its demonstration of scaling in embodied AI.
"The Generalist dataset size gave everyone a scaling push." [01:55:44]
Figure U.S. humanoid robotics company. Xu watches closely but with skepticism — he wants to see their results in person to verify they are not teleoperated.
"Figure — I really pay attention to it. Whether it's real. Because Figure always radiates this aura of: maybe very impressive, maybe a bit marketing. Mainly because yesterday they released a new video — cleaning up the kitchen — it was too smooth, too impressive, which made me suspicious. I really want to go to that scene myself and move things around." [01:17:14]
Fourier Intelligence (傅利叶) Chinese humanoid robotics company. Praised specifically for product design — their GR3 robot's white, clean aesthetic is cited as well-suited for home environments.
"Domestically, there's a company called Fourier — I really like their GR3, that white robot. It fits the home environment pretty well." [01:18:12]
Fauna Robotics Mentioned for distinctive, home-friendly product design — a flat-headed robot form factor that Xu finds interesting.
"There's a robot company called Fauna — F-A-U-N-A — it has a very flat head. I also think it's quite interesting from a product design standpoint." [01:18:13]
Unitree Robotics (宇树科技) Briefly mentioned in the context of consumer-grade robots that are starting to ship and may provide visible real-world usage data.
"As consumer-grade products start coming out, we may see some better, more direct data — like Unitree, with their new generation also released." [01:29:16]
DeepMind (Google DeepMind) Cited as a case study of a company whose vision was "too far ahead" in 2010 but was preserved by Google's acquisition. Now considered Google's most critical AI unit.
"DeepMind was acquired because they started too early — 2010, before ImageNet. That they survived at all was remarkable. But now DeepMind has become one of Google's most important parts. Without Hassabis, Google's AI would have been nearly wiped out by OpenAI." [01:31:25]
4. People Identified
Eric Jang (江二十三) Former Chief Scientist at 1X (OneX). Xu recently met with him and found they independently converged on the same 18-month timeline for household robots entering homes.
"The 18-month thing — I was actually surprised he brought it up first, because I had the same judgment independently." [00:41:49]
Ilya Sutskever Co-founder of OpenAI, founder of Safe Superintelligence. Cited as the archetype of scientist-founder whose unwavering conviction changed the trajectory of AI history.
"OpenAI believed in just piling on data. Ilya just kept stacking it. Until GPT came out, people still didn't realize. Until ChatGPT — that's when people finally understood. His steadfast judgment was the single most important factor that changed the world's timeline. And that factor came from a scientist's belief." [00:29:56]
Demis Hassabis Co-founder of DeepMind and Google DeepMind CEO. Cited as a model of near-religious dedication to pursuing AGI, and his biography is what Xu is currently reading.
"What I see in Hassabis as described by Malouf is, in a sense, a microcosm of everyone pursuing AGI — or the state one should be in: a near-devout pursuit of what they're meant to pursue." [02:12:27]
Kaiming He (何凯明) AI researcher (Meta/MIT). Cited for a memorable rebuttal about deep learning safety — comparing trusting a well-trained model to trusting an experienced driver.
"Kaiming's response was great. He said: when you ride with an experienced driver, they haven't guaranteed they'll never crash. But you still trust them because they've driven for so long without crashing. The same applies to large models in embodied intelligence — as long as they haven't made serious errors over a long enough time, I can trust them." [00:53:54]
5. Operating Insights
Define "What We Will Never Do" Before Defining "What We Do"
Xu's safety and product strategy is built around explicit exclusions, not just capabilities. By publicly committing on day one to never do direct-body-contact tasks, the company eliminates entire categories of catastrophic failure modes — both technical and regulatory — before they become live risks. This is a form of negative product definition that most hardware companies skip.
"From the very first day we said: we will not do any service involving direct physical contact with the human body... You only need to have a clear 'what we will not do,' and you won't cause a catastrophe." [00:54:51]
Use Public Content as a Real-Time Product Feedback Mechanism
Xu explicitly frames his continued social media presence during his startup as a product research tool, not just personal branding. Because he wants to build a consumer (2C) product, getting unfiltered public reactions — including harsh criticism — is strategically valuable in a way it wouldn't be for a B2B company.
"Once you're online, everyone is equal — even Federer on Xiaohongshu becomes a 2.5-star player. Everyone will say what you did wrong, what you did right. Getting this kind of real feedback is incredibly rare. Especially for a consumer product, this helps us polish the product enormously." [02:08:29]
Build Your Office as Your First Test Environment
Rather than separating R&D from product design, Xu is designing the company's office as a home environment — so robot training, product testing, and daily work all happen in the same space. This collapses the sim-to-real gap at the organizational level.
"Our office will look like a home. The living room can be used for daily meetings, for robot training, and for recording podcasts. Many things can break convention." [00:38:43]
6. Overlooked Insights
Remote Haptics as a Sleeper Technology With Massive Medical and Space Applications
Xu briefly mentions a research project from his lab at IIIS (Tsinghua) that received almost no discussion time: a bidirectional tactile system where a robot in Shanghai can feel an object, and a human in Beijing can feel what the robot feels — in real time, including softness, texture, and shape. They used it to simulate remote breast cancer palpation screening.
This is significant because it represents a non-obvious commercial pathway for robotics that is entirely separate from the household assistant narrative — telemedicine, remote surgery assistance, and eventually space exploration. Xu even mentioned Mars as a use case in passing.
"We bought a breast cancer model and did remote palpation with it. And in the future, maybe a robot will go to Mars — and I want to feel what Mars feels like from Earth. Maybe we can do it this way." [02:03:34]
The technology currently captures softness, shape, and texture — but not temperature or fine surface quality. It is early, but the fact that a robotics AI lab has already demonstrated this in a medical context, unprompted, suggests this is closer to application than the market realizes. No one in the conversation flagged it as significant.
Video Data From Cooking Influencers Is an Underpriced Training Asset
Xu casually mentions that before founding his company, he was already thinking about using cooking influencer videos (specifically naming "Master Wang Gang" — a prominent Chinese chef on social media) as robot training data. He noted these videos are multi-angle, unstructured, and densely packed with dexterous manipulation sequences.
"I wanted to use cooking influencer videos — Wang Gang's videos, for example — but they have a lot of cuts because they're multi-camera. I needed to extract the actual working sequences... Learning from those videos how a robot should cook. Video data was always going to be used heavily. Teleoperation was just easier to show progress with early on." [00:24:30]
This implies that publicly available cooking video content — which exists in massive quantities across YouTube, Bilibili, Douyin, and other platforms — is already being evaluated as a pre-training data source for manipulation models. The companies that figure out how to clean, segment, and label this data at scale first will have a structural data advantage that is difficult to replicate.