Embodied AI hasn't reached its GPT-1 moment — it's stuck because the data hasn't scaled up
Shen Yujun of Ant Lingbo judges that embodied AI has not yet reached its GPT-1 moment; the bottleneck is that the data hasn't scaled up: internet data is two orders of magnitude larger than real-robot data, and ideal pretraining needs a million hours as a starting point. To get there, the team completely redid visual pretraining, insisting on starting from the sensor.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Giving up GAN: compute is used too inefficiently
Shen Yujun has worked on generative models since 2017, and the team kept pushing GAN scaling. Even after Diffusion took off, they made one last attempt early this year. It improved on previous years, but GAN's biggest problem is that compute is used too inefficiently: Diffusion can iterate many times to generate one image, while GAN struggles with complex content, especially video. They open-sourced the technology accumulated on GAN and closed the book on it, but believe the know-how from adversarial training — training two models at once to compete against each other — won't go to waste.
— Shen YujunDigital-world models don't fit the physical world
The first-generation model relied on digital-world models for pretraining, but they found the development motives of the digital world don't match the needs of the physical world: video generation chases image quality, while robot actions don't need high image quality. The digital world has its own iteration cadence and can't be pushed backward to change their model, so in '26 they decided to redo everything based on physical-world needs, training vision-related understanding and generation from scratch, with only the language model continuing to reuse the LLM. Another position is to insist on starting from real sensors, because real sensors are noisy, which runs naturally counter to academia's clean datasets.
— Shen YujunDepth estimation restarts from geometry
The first-generation depth model used DINO v2 (a digital-world semantic model), but semantic understanding isn't enough for the physical world — knowing there's a dog still requires knowing the distance. This time they dropped DINO and redid pretraining entirely from a geometric angle (points, lines, planes, edges), then iterated out a new depth estimation model. Examples: a reflective faucet shows where the water stream is; a cat behind a glass door is completely invisible in the new depth model, because the cat is blocked by the glass and only gradually appears once the door is pulled open. In semantic understanding the cat is always there, but in spatial perception you have to hit the door first. This explains why semantic pretraining from the digital world can't transfer directly to the physical world.
— Shen YujunPhysical-world native: unidirectional modeling and MOE
Generating video in the digital world can wait 30 seconds, so bidirectional attention works; but robot execution must be real-time, and history cannot see the future, so the model must be unidirectionally modeled. They spent three or four months training unidirectional attention from scratch — forcibly converting a bidirectional model to unidirectional causes forgetting of pretrained knowledge, so that route had to be abandoned. Another technical point is native MOE: every expert must be activated with equal probability, and only sparse-but-uniform counts as truly sparse; just getting MOE trained properly took two months and dozens of failures. The physical world can sacrifice image quality, which gives room for efficiency optimization.
— Shen YujunThe GPT-1 moment hasn't arrived; the bottleneck is data
Shen Yujun judges that embodied AI today has not reached its GPT-1 moment, because no good way to scale up data has appeared. This time they used 60,000 hours of real-robot data, but that's no order-of-magnitude improvement and far from enough. Embodied-native pretraining still relies on large amounts of internet data, and internet data is two orders of magnitude larger than real-robot data. The architecture is already native; the data is not yet native. When embodied data itself can scale up — at least reaching the same order of magnitude as the internet — that will be his GPT-1 moment.
— Shen YujunWhy build only the brain, not the body
Many companies turned this year to building only the brain, and Shen Yujun sees two reasons: first, a single body is hard to make general-purpose, and if it isn't general-purpose you have to pick the right scenario for specialized design, otherwise costs are bound to be high; second, the main contradiction now is that the brain lags the body and still isn't good enough at doing work. Brain and body definitely rise in alternation; right now the brain lags, and as more people work on brains the intelligence layer runs faster and will one day overtake the body, at which point a batch of body companies redefined by intelligence will emerge. The previous generation of hardware won't necessarily fit the next generation of models.
— Shen YujunStartups are a gamble
Shen Yujun's first startup taught him two lessons: on execution, he used to focus on technical leadership, but now focuses on where the technology leads — you have to pick valuable leadership; on startups, nobody knows which path will definitely succeed, and with limited resources you have to dare to gamble. If everyone knew which path succeeds, the big companies would definitely succeed and it wouldn't be your turn. He admits to three big bets: sensors, a native model, and not building the body. Native is a huge gamble — training might fail to converge halfway through, or the approach might be wrong, or you enter too early, or resources run short — but it's a path you must take.
— Shen YujunThe landscape: two or three startups plus one or two big companies
Competition in robot brains will resemble large models: one or two players from the big companies, two or three from startups, each running out an ecosystem that includes data, body and scenarios. Because once there's an ecosystem, scenarios plus bodies bring data, and models will gradually differentiate to suit that ecosystem. Before entering the home, each company's model will have different characteristics — some good at delicate tasks, some at long-horizon tasks — and ultimately they'll converge toward a general ecosystem.
— Shen YujunIn their own words · checked verbatim
You can see it, but you can't touch it — seeing it is semantic understanding, not being able to touch it is spatial perception.
你能看到，但你摸不到，嗯，能看到是语义理解，摸不到是空间感知。
Shen Yujun45:07
What I'm saying is that training MOE well is a capability problem, not an attitude problem.
我我说就是MOE训把MOE训好是一个能力问题，他不是一个态度问题。
Shen Yujun55:55
Jushen will definitely have its own model — a model designed for Jushen.
巨深一定会有自己的模型，嗯，就是他为巨深设计的模型。
Shen Yujun1:15:55
The best expectation is that I really see robots entering the home, and the robots entering the home run Lingbo's model — that's my best expectation.
最好的预期就是我真的看到机器人进家庭了，然后进家庭的机器人跑的是凌波的模型，这是我最好的预期。
Shen Yujun1:37:19
But at least on the data side, I think China will definitely move faster than the US.
但至少在数据这方面，我觉得中国一定是比美国跑的快的。
Shen Yujun1:50:45
Figures
| Second-generation pretraining data volume | 60,000 hours | 33:29 |
| First-generation pretraining data volume (actually retained after cleaning) | about 10,000+ hours (originally 20,000 hours after cleaning) | 34:00 |
| Drop in real-robot data cost over the past two years | at least 3x | 35:36 |
| Data volume needed for post-training | from about 100 samples down to about 20 samples | 1:06:31 |
| Internet data volume relative to real-robot data volume | two orders of magnitude larger | 1:08:35 |
| Ideal pretraining data volume | a million hours as a starting point | 1:22:04 |
| Data collected for post-training on the billiards task | about 20 samples | 1:11:40 |
| Time to get MOE trained properly | about 2 months | 54:46 |
| Time to train unidirectional attention from scratch | about three or four months | 56:56 |
| Generalist public data volume | several hundred thousand hours (possibly close to a million by now) | 1:12:44 |
Glossary
- Ego / first-person-view data
- A person wears a head-mounted camera to capture manipulation video from their own viewpoint, used to scale up robot data.
- WAM / world action model
- Learns from video how the world responds to actions, and can serve as a robot's world simulator or executor.
- VLA / vision-language-action model
- The mainstream robot manipulation route: takes visual and language instructions as input and outputs actions.
- MOE / mixture of experts
- A sparse-activation architecture that activates only part of the parameters per inference, lowering compute demand.
How to listen
Embodied AI founders, robot algorithm engineers, and investors watching AI in the physical world — especially teams torn between building the brain and building the body.
The first 16 minutes of personal academic background can be fast-forwarded; the substance begins with the origins of Ant Lingbo.