Ant Lingbo's 沈宇军 on Robot-Native Foundation Models and Data Scale-Up

Summary

• China’s embodied-intelligence sector is producing a new species that puts the brain first. Ant Group formally announced the establishment of Ant Lingbo as a subsidiary at the end of 2024, focused primarily on the robot brain while also making some end effectors. Its vision is that “one day every household will have a robot … and the models running on all of them will ultimately be Lingbo’s.” Chief scientist 沈宇军 argues that robot bodies are difficult to make general-purpose and that it is hard to pick the right scenarios, while “the brain may currently lag the body.” This year’s central task is to get robots doing useful work first.
• The core bet is on “embodied-native” models. After heavily modifying the 1.0 model built on a digital-world world model, the team found that many weaknesses still surfaced. In 2026, it decided to retrain everything related to vision from scratch: a vision compressor aligned with semantics and actions, a native MoE architecture, and one-way causal attention, combined into VLA 2.0. “Every parameter was trained from a completely randomized initialization.” 沈 calls it a decision he strongly backed: “If we don’t make it work, someone else will, because I think this may be the right path.”
• Robotics has not reached its GPT-1 moment, and the bottleneck is clearly data. This generation’s pretraining used 60,000 hours of data, primarily from real robots and possibly including ego data, versus 20,000 hours for the previous generation. Internet data still exceeds embodied data by “two orders of magnitude.” His benchmark is at least 100,000 hours dedicated entirely to native pretraining, with the ideal starting point at “1 million hours.” The good news is that infrastructure has entered the 100,000-GPU era, so the gap from robotics’ GPT-1 to its ChatGPT “may not need to be that long.”
• The economics of data are improving, but not as quickly as expected. Real-robot data costs have fallen by more than 3x in the past 2 years. Simulation is increasingly being reserved for evaluation and may not be used much for training because of distribution homogeneity and immature soft-body physics engines. Industry-wide data development has actually lagged expectations because nobody is truly training native foundation models: without that need, “there is no force pushing you to obtain all that data immediately.” Ego collection is a sound idea but not yet a product. During this “data gap,” China “will certainly move faster than the United States” in embodied data.
• The capability boundary is drawn with unusual candor. VLA 2.0 has made breakthroughs in position generalization and randomness: a table-based ball-game demo involving human opponents may have used only 20 data samples, whereas previously even far more data might not have been enough. Post-training data per task has fallen from roughly 100 samples to 20, and 10 tasks can now be combined into one model. But task-level zero-shot generalization “is still not very good,” and Lingbo addresses the implementation of intent, not its source. Higher-level capabilities may need to come from language or multimodal models.
• The industry structure could resemble the eventual LLM end state. One or 2 major tech companies and 2 or 3 startups may each build out their own data, body, and scenario ecosystems. The brain and body “will definitely rise in alternating cycles,” with the next wave bringing hardware companies redefined by intelligence. Sensors will also need to be redesigned around model requirements: real-time response may matter more than force accuracy. On the hardware side, some companies have already solved mass production, dexterous-hand startups are proliferating, and the industry is shifting from showing demos to putting robots into real-world scenarios.
• His startup playbook is blunt: entrepreneurship is a bet. “If a path were certain to succeed, the big companies would have already succeeded.” Lingbo’s key choices include not building a robot body, starting from real sensors, and pursuing embodied-native models. 沈 explicitly calls the sensor-first and embodied-native approaches major bets. The best-case outcome is robots entering homes running Lingbo’s model; “the worst case is simply that we don’t make it.”

Deep dive

1. From SenseTime’s East Gate to CUHK: He Wasn’t Interested in Leaderboards, but in Generation

• 沈宇军 entered the field by accident. SenseTime had just been founded at the end of 2014, and its office was by Tsinghua’s East Gate. He joined as a junior-year summer intern, stayed through graduation, and then went to The Chinese University of Hong Kong for a PhD under 汤晓鸥.
• CUHK was strong in leaderboard competition for detection, segmentation, and classification, but he admits he was “not particularly interested in the whole exercise of ultimately looking at just one number.” When GANs took off, he moved into generative work because it could “actually generate an image for everyone to look at.” Starting in 2017, he moved from image editing into pretraining and then representation learning: “From not being interested in numbers to being interested in generated content, that’s how it happened step by step.”

2. GANs Reach Their Full Stop: They Lost on Compute Efficiency, but Adversarial Training Survives

• As recently as the beginning of this year, he was still making “one last attempt at scaling up GANs,” unwilling to give up. The technology had improved substantially from prior years, but “the biggest problem is still that compute is used too inefficiently.” Diffusion can iterate many times to generate one image; given the density of video content, GANs “may not be the best choice.”
• After releasing the accumulated work to the public, he “put a full stop to it.” But he retains one conviction: adversarial training itself—training 2 models simultaneously and having them compete—will remain useful, even if it is not necessarily used for image generation.

3. One Year at ByteDance and Computer Vision’s Ceiling: “There Aren’t Enough Places to Use It”

• After finishing his PhD, he joined ByteDance to work on Douyin effects: turning men into women, aging people up or down, and showing users what they might look like 20 years later. The work drew directly on his PhD research in image editing. But chasing viral creative concepts every day left him feeling that there was still unfinished work, so he moved to Ant Group’s Institute for Technology Research in 2022. The original audio appears to say ChatGPT was what “shook him” at the end of the year after he joined Ant.
• His summary of more than a decade in computer vision is stark: the first wave was facial recognition, the second was pure-vision autonomous driving popularized by Tesla, and “apart from that, we still haven’t seen a visual application achieve large-scale adoption with strong industrial value.” Vision has “not had enough places to be used” in the digital world. “But in the physical world, without vision, a robot definitely cannot work”—the underlying reason he turned to robotics in 2024.

4. Why Use a Subsidiary for Embodied AI: Research Advances Are Not the Same as Industry Advances

• Ant’s core businesses—finance, healthcare, and mobility—are services in the digital world. At the group level, the judgment was that extending digital-world services further would “definitely require moving into the physical world.” Ant Lingbo was formally established at the end of 2024, just as 沈 proposed to the institute’s president that he wanted to work on embodied AI. The 2 decisions “happened to coincide.”
• Not keeping the team inside the research institute reflected “a fairly major decision by the group.” Research organizations can drift away from industry, and “the technology researchers consider advanced can differ from the advanced technology the industry needs.” If Ant was going to do this, it wanted to do it as a real business, not as an experimental trial-and-error team.
• He came away from Lingbo’s early days with 2 impressions. First, “we were too naive”: actually building the product was completely different from imagining it, and deployment, communications, and debugging created a technical stack far beyond what they had understood. Second, he watched the team grow from fewer than 10 people while learning alongside an industry that was changing in real time. The opportunity now exists because the technology has not converged: “Even exploring an incorrect path is a contribution.”

5. Building Only the Brain: “Every Household’s Robot Will Run Lingbo’s Model”

• The only thing clearly defined at Lingbo’s founding was that it would build the brain. The vision follows Ant’s emphasis on broad access: “We ultimately hope that one day every household will have a robot … and we hope that the models ultimately running inside those robots will all be Lingbo’s.” Robot brands may proliferate like smartphone brands, while no one knows which body form will win, so Lingbo chose not to bet on the body.
• A cooking demo at the Bund Conference in September 2025 reinforced that decision. It was a long-horizon task that exposed a large number of hardware and software problems. “Without a reasonably usable brain, the cost of replicating a new scenario is simply too high.”
• His definition of “general-purpose” goes one level beyond the industry’s. While most people mean generalization across tasks and scenarios, Lingbo also includes cross-body generalization. It introduces data from different bodies during pretraining so “the brain has already experienced those bodies during pretraining,” allowing rapid adaptation at deployment. But he insists that the body and scenario cannot be fully decoupled: the type of body suited to a given scenario is determined from the very beginning of product definition.

6. 2025’s 2 Core Claims: Learn from Video, Start from Real Sensors

• His judgment in March 2025 was that temporal and dynamic modeling were crucial: a robot has to move and interact with its environment. “At that time, the concept of a WAM, or world action model, did not yet exist,” but the team was already convinced that it had to learn from video rather than treating sampled video frames as images. 9 months later, in early 2026, it open-sourced what he believed was probably the earliest visual-action model, or VA.
• The second claim came from being taught by reality. Academia tacitly assumes relatively clean datasets, but when robots operate in the real world, “sensor noise is impossible to ignore.” Lingbo therefore insisted on starting from real visual, tactile, and other sensor inputs to build spatial intelligence. Those 2 ideas formed the company’s core technical position at the time.

7. The 2026 Decision: Throw Out All the Vision and Retrain for the Physical World

• Early on, the team had too little data and could only graft a digital-world model into the physical world as an agent. It quickly found that “the original purpose of developing a digital-world model is mismatched with the needs of the physical world.” Video generation optimizes for image quality, which is less important for robot action, but the team could not force the model’s creators to change course: “They have their own iteration schedule.”
• The major decision in 2026 was therefore to “rebuild everything based on the needs of the physical world”—all vision-related components would be trained from scratch. The LLM was left unchanged because the distinction between digital and physical worlds is not as significant at the language layer, and he has not yet seen an obvious shortcoming in directly applying an LLM there.

8. Deployment Exposes a Blind Spot: Robots Do Not Work at a Fixed Height

• The first-generation model supported only arm degrees of freedom plus a gripper, 9 configurations, and tabletop picking—the standard laboratory setup. Deployment revealed the gaps. Reaching a shelf requires raising the arm: should the head look where the arm is going? The head camera may fail, requiring the wrist camera to provide coverage. Picking something from below requires bending down, and reaching farther requires moving the base.
• VLA 2.0 responded with a major data upgrade: configurations rose from 9 to more than 20, while the degrees of freedom expanded from 2 arms to include the head, waist, base, and even dexterous hands. His method is to combine “pull from the ceiling” with “grab what lands”: the model may have 8 or 10 possible development directions, but limited attention means deployment ultimately determines which feature gets developed first.

9. 60,000 Hours and a 3B-Scale Base: Rebuilding the Data Pipeline

• The current generation was pretrained on a model with roughly 3.xB parameters, and the headline data volume rose from 20,000 hours for the first generation to 60,000 hours. But the data is not directly comparable: after the first generation’s 20,000 hours were re-cleaned under the new, stricter pipeline, “only a little over 10,000 hours may have carried over” into this run. On the higher standard, the team effectively accumulated roughly 4x the usable data.

10. Simulation Is Increasingly Reserved for Evaluation; Real-Robot Data Costs Have Fallen More Than 3x

• The data comes from real robots and possibly ego data from head-mounted cameras. Simulation is “increasingly being left for evaluation” and may not be used much for training. The issue is not volume but distribution: simulated distributions are more homogeneous, while soft-body tasks such as tearing open a plastic bag still require physics engines to iterate further.
• Real-robot economics, by contrast, are improving quickly. Personnel management and the machines have matured, while rising demand has created manufacturing scale effects. Costs have fallen “by more than 3x” over the past 2 years, which he sees as a sign of industry expansion.

11. Data Know-How: Degree-of-Freedom Mix and Cross-Body Cleaning

• The team has 2 core pieces of know-how. The first is task allocation: tasks should cover every degree of freedom as fully as possible. If head movement appears far more frequently than waist movement, the model may learn to use the head while barely moving the waist.
• The second is cleaning. Every robot body has different characteristics: camera placement and joint connections, or configurations, vary, so one data pipeline cannot process everything. Simpler operations are more generalizable, while complex tasks depend heavily on the teleoperator’s skill. “If the motion is not smooth, the model will learn that this is the correct way to operate,” and the robot will move jerkily as a result.
• Data collection combines in-house work with an ecosystem. “Our people first run through the SOP,” then data-alliance partners add volume. Correction data—where a person interrupts at the right moment and provides the correct demonstration—and real-robot reinforcement are used to improve deployment success rates, but they were not included in the open-source technical report.

12. Vision and Depth: “Seeing It Is Semantic Understanding; Not Being Able to Touch It Is Spatial Perception”

• The first-generation Depth model was built on DINOv2, a digital-world model strong at semantic understanding. Knowing that something is a flowerpot or a dog is sufficient for many digital-world applications, but it does not naturally translate into understanding space and distance. The new Vision model is pretrained entirely from geometry, modeling points, lines, planes, and edges, and then used as the base for retraining Depth. The improvement is substantial.
• He favors 2 examples. After a reflective faucet is turned on, the model can clearly see the stream of water—not merely whether the water is running, but where the stream is, since a robot must see the water before it can position a container beneath it. In another example, a cat outside a glass door is invisible to the Depth model; once the door opens, the cat gradually emerges from behind it. “Seeing it is semantic understanding; not being able to touch it is spatial perception.”
• This layer does not need an LLM at all; it relies on native visual pretraining. Lingbo’s tongue-in-cheek roadmap has only 3 items: “see more clearly, think more clearly, and work more cleanly.”

13. 3 Mismatches Between Digital and Physical Worlds: Compression, Real Time, and Dispensable Image Quality

• The first mismatch is video-generation compression. It “simply turns 100 megabytes into 30 megabytes,” without helping the downstream learning process. When training its compressor, Lingbo added semantic and action alignment so the encoding process itself would sense semantic and action information.
• The second mismatch is that the digital world does not care about efficiency. A text-to-video user can wait, “but a robot cannot wait while executing; it has to be real time.” That creates 2 requirements: activation efficiency and unidirectional modeling. A commonly used Wan model is a 14B dense model, so inference activates all 14B parameters and later acceleration has a ceiling.
• But the physical world also allows certain things to be discarded: image quality. “It’s fine if the generated output is a little pixelated … because we still have cameras.” Different species see with different levels of clarity, yet they can all live on Earth.

14. MoE: 2 Months and Dozens of Failures—“It’s a Capability Problem, Not an Attitude Problem”

• Lingbo Video uses a native MoE architecture and was trained from scratch. The challenge is not conceptual but engineering-driven: if a model has 30B parameters and activates the same 10B every time, the other 20B are barely used, making it no different from a 10B dense model. Experts must be activated with roughly equal probability, requiring choices around the loss function, sampling strategy, and regularization.
• Simply learning how to train MoE took 2 months, with an estimated dozens of failures. “You train it and discover that it does not behave the way you want.” His summary became a tagline: “Training MoE well is a capability problem; it is not an attitude problem.” The current system can now be trained stably even after switching frameworks.

15. Time Moves in One Direction: Switching from Bidirectional to Unidirectional Attention Can Destroy Pretraining

• Real-time execution requires attention to be unidirectional: “The future can see the past, but the past cannot see the future. Time moves in one direction.” Digital-world generation models are generally bidirectional because users wait for the result and both the beginning and ending are known.
• The route of pretraining bidirectionally and then switching to unidirectional attention has a fatal flaw. Knowledge and language understanding are “implicitly contained in bidirectional attention”; after forcing a unidirectional change, “a model that could previously turn 100 sentences into good videos might now manage only 20% or 30%,” effectively voiding pretraining. Lingbo therefore designed the system as unidirectional from the start, a path that took another 3 or 4 months.

16. VLA 2.0: An Embodied-Native Base Model with Every Parameter Trained from Random Initialization

• Three technologies came together: the Video track solved MoE, the World track solved causal pretraining, and the compression system added semantic-action alignment. “Putting these 3 technologies together gave us a native foundation model purely for robot control. Every parameter in this base model was trained from a completely randomized initialization,” followed by the injection of real-robot data. That is what he means by “embodied-native” itself.

17. Lingbo Implements Intent but Does Not Generate It

• He does not engage in the brain-versus-cerebellum debate over definitions; he draws a boundary. If a person gives an instruction and the robot executes it, that is Lingbo’s layer. A higher-level capability—such as the robot seeing rain after opening the curtains and proactively bringing an umbrella—belongs to language and multimodal models and is more closely tied to the digital world.
• “It only handles the implementation of intent … it does not generate its own ideas or proactively think about what you want to do.” This self-imposed limit is worth remembering: Lingbo currently stops at the execution layer.

18. A Breakthrough in Randomness: It May Have Learned to Play Against Humans from Just 20 Samples

• His aha moment was a tabletop ball-game demo against a human opponent—he called it billiards, while the host called it table football. “You never know where the object coming back from the other side will be,” the exchanges are extremely fast, and the system relies only on the robot’s own camera. There was no ring of tracking cameras around the table: just “a pair of eyes, seeing the object and then acting.”
• The data requirement was the most striking part: “We may have collected only 20 samples for this task. Previously, no amount of data might have been enough to make it work.” The same applies to the tabletop tidying demo: objects are scattered randomly, removed and rearranged midway, and even handled alongside a person, yet the robot responds dynamically. This is not the result of adding more data; “the entire model’s pretraining is different.”

19. The Limits of Generalization: Position Is Improving, but Task Generalization Still Eats Data

• He breaks generalization down cautiously. Position generalization “can now be said to be largely solved,” while task generalization—performing a new task zero-shot—“may still not be very good.” The root cause is that robot data remains insufficient: the model cannot combine familiar tasks in enough permutations to produce unfamiliar ones.
• The learning cost is falling. In the 1.0 era, each task required its own post-training model and roughly 100 samples. In 2.0, 10 tasks can be trained into one model with roughly 20 samples. “We still cannot do it with 1 sample or fewer”; that will require further improvement in the base model.

20. The GPT-1 Moment Has Not Arrived: Embodied and Internet Data Are 2 Orders of Magnitude Apart

• His direct judgment is that embodied intelligence has not reached its GPT-1 moment “because we still have not seen a good way to scale up data.” The 60,000 hours represent no order-of-magnitude increase over the previous run and remain far from sufficient. Architecturally, Lingbo has achieved embodied-native modeling, “but not in terms of data”: native pretraining still relies on large volumes of internet data for support, and internet data exceeds robot data by “2 orders of magnitude.”
• His benchmark is first reaching 100,000 hours dedicated entirely to embodied-native pretraining; the ideal starting point “should be 1 million hours.” That would bring embodied data into relative balance with internet data. The timing is difficult to estimate, but infrastructure has changed dramatically: when a 10,000-GPU cluster was impressive, whereas many US companies training language models now use 100,000 GPUs. As a result, the interval from robotics’ GPT-1 to ChatGPT “may not need to be that long.”

21. The Native Bet: “If We Don’t Make It Work, Someone Else Will”

• The decision came after the 1.0 release, and “there was actually no hesitation.” Many features had been forcibly modified into digital-world models, but “no matter how fancy the modifications, some weaknesses would still surface.” He acknowledges that it was unquestionably a bet: training for a year and spending substantial resources only to see the model fail to converge was a real risk. The failure could have meant the idea was wrong, the technology immature and adopted too early, or the team lacked sufficient resources—“trying to do a $200 job with $100.”
• His conviction is straightforward: “If the digital-world path cannot go very far, this is a necessary path. If we don’t make it work, someone else will, because I think this may be the right path.” It is also the decision he most strongly defended and the one that gives him the greatest sense of accomplishment: he did not want to remain on the extension of someone else’s path.
• Progress has exceeded expectations. Within 6 months, native modeling advanced “seeing more clearly” through Vision and “thinking more clearly” through architecture. “I didn’t expect it to come this quickly.” “Working more cleanly” still depends on native data, completing a native system that closes the loop from model to architecture to data.

22. Possibly Generalist and PI: Efficient Post-Training Versus Zero-Shot, Converging by Different Routes

• He mentions a company whose name sounds like Generalist in the audio as a reference point. Its public statements put data volume at several hundred thousand hours, and “it may not even be clear now whether it has reached 1 million.” Its data pipeline is more mature and it moved into pretraining earlier. But its technical path offers little to borrow: “We still have not seen anyone make video generation native.”
• The 2 philosophies differ. The possibly Generalist side accepts post-training and emphasizes efficient post-training, while PI places more weight on zero-shot capability: “I can do everything.” He sees the routes converging. A home robot needs to perform 80% or 90% of tasks on its own and learn the remaining 10% to 20% after being shown once. At root, both are forms of in-context learning and are not contradictory.
• The data approaches also differ. The possibly Generalist team appears to have a deep understanding of and strong execution on data that sounds like UMI in the audio, while PI is strong in real-robot teleoperation. Looking ahead, the rise of ego-centric data and the replacement of grippers by dexterous hands may change both routes. “Things come together after splitting apart, and split apart after coming together.”

23. The Independent-Discipline Debate: Language Is the Interface, and the Problem Is Demand Generalization

• At last year’s Bund Conference, the debate was whether embodied intelligence is a branch built on top of LLMs or an independent discipline. “If I had to choose, I would choose the latter.” Language models solve the problem of making a machine understand what an instruction is; embodied intelligence solves how to execute it well, and “doing it well does not require understanding language again.”
• More precisely, language models solve demand generalization. A fixed code could represent 10,000 tasks, but the system would not generalize further at the task level. “That is clearly not the core focus of embodied intelligence at this stage.” In the near term, the language system does not need to be embedded inside the embodied model; the 2 are related at the interaction layer.

24. Industry Check-In: Hardware and Deployment Beat Expectations, Data Did Not

• Progress has been “much faster than I imagined.” On the hardware side, some companies have already solved mass production, and the supply chain is maturing alongside them. “Last year everyone had grippers; this year dexterous-hand companies are appearing everywhere.” Tactile sensors are also proliferating, both to enable more complex and precise tasks and to create new methods of collecting data.
• Deployment is also advancing. Figure has demonstrated logistics sorting, and “people are no longer deploying through traditional control methods; everyone is trying to deploy with models.” At WAIC last July, the industry was still showing demos: “In 2024 it was pouring coffee, in 2025 it was folding clothes—one new action every year.” “This year there are actually fewer demos; everyone is talking about entering real scenarios.”
• The miss has been data. Everyone says data is needed, but “if you do not need to build a foundation model, you may not need that much data.” Without training a foundation model, “there is no force pushing you to obtain all this data immediately.” The fact that genuinely embodied-native systems did not emerge over the past year is rooted in that lagging data development.

25. The Brain Lags the Body: They Will Rise in Alternating Cycles, and Intelligence Will Redefine Hardware

• Why have so many brain-only companies appeared this year? Robot bodies are difficult to make general-purpose, and a non-general body requires selecting the right scenario—something “people are not very willing to say they have definitely solved.” Mass production is another barrier. More importantly, “from the industry’s current state, the brain may currently lag the body.” Someone said last year that useful robot work was a desert; this year’s central contradiction is to get the robots doing the work first.
• His dynamic view is that the brain and body “will definitely rise in alternating cycles.” As more companies build brains, the intelligence layer will accelerate and eventually surpass the body. “At that point, a group of new robot-body companies redefined by intelligence will emerge.” Hardware will no longer be designed around traditional specifications, but “to make it easier for models to control.”
• Sensors will be among the first to change. The traditional tradeoff says that a tactile sensor should be accurate in measuring force. In the AI era, “what the model needs most may not be force accuracy; it may be a sufficiently fast response.” Sensor iteration must therefore start from model requirements when deciding what to strengthen and what to de-emphasize.

26. The Next Step Is Native Data: Ego Is a Philosophy, Not Yet a Product

• After native architecture comes “moving from native architecture to native data.” Task generalization “eats data aggressively,” requiring more elegant collection methods and more rigorous collection standards. The central questions are what data to collect, how to collect it, and whether the process can be replicated at scale.
• He remains cautious on the ego route. It “undeniably proposes a way to scale up collection,” but it is still unknown whether head-mounted-camera data alone is sufficient or whether it can capture the fine hand movements needed in the dexterous-hand era. “Ego has proposed a collection philosophy, but it needs to be refined into a specific product before it can be deployed.”
• Quality and scale are naturally in tension. Models want better data and more modalities, but higher requirements make collection less scalable and turn it into customization. “It is not yet clearly defined what the model can afford to discard. Once that is clearly defined, everything downstream will accelerate.”

27. Starting a Company Is a Bet: Ant’s Expectations and the Worst-Case Outcome

• He identifies 2 firsts in his career: the first was going into industry, and the second was starting a company. Industry taught him that technical leadership must be judged on the right dimension—whether it is valuable leadership or merely an engineer’s impulse. He is now prioritizing a bet on task generalization. Entrepreneurship taught him that “starting a company is always a bet … if a path were certain to succeed, the big companies would already have succeeded,” because the organizations with the most resources and people would win. When the path is unknown, hesitation can mean producing nothing. The 2 major bets he names are sensors and native models; not building the robot body is another key choice.
• Ant positions itself as AI-first, with Lingbo carrying the group’s mission to establish the upper bound of AI in the physical world and complete the execution layer required to move digital services into that world. Management follows a startup model, with relative operating independence and shared infrastructure. Whether Lingbo should build the body has also been debated, but “the group has given Lingbo sufficient trust and respect,” and the sides have reached agreement after several rounds of discussion.
• He states the 2 ends of the outcome distribution plainly. The best case is “to really see robots enter people’s homes, with those home robots running Lingbo’s model.” “The worst case is simply that we don’t make it”—that one of the bets turns out to be wrong.

28. The End State, Robot Forms, and China’s Advantage: China Moves Faster During the Data Gap

• His end-state view mirrors large language models: “There may be a few big companies that do it well, along with 2 or 3 startups that do it well.” Lingbo sees itself as “a startup backed by a big company.” Each player will build its own ecosystem of data, bodies, and scenarios, and models will develop different strengths around those ecosystems—some suited to precise tasks, others to long-horizon tasks—before eventually converging on a general-purpose home-robot ecosystem.
• He has no fixed view on humanoid configuration. “Whether it is wheeled or legged is not the key bottleneck right now. The key bottleneck is insufficient intelligence, and the key bottleneck behind insufficient intelligence is insufficient data.” Regardless of whether the robot is legged or wheeled, “I would not let it into my home” today; there is no way to conduct a meaningful user study with a robot that cannot perform useful work.
• He puts substantial weight on China’s position. Internet data was the monetization of the internet dividend; embodied data “may be in a window period where everyone is starting from the same place.” Collection depends on hardware and sensors, so “at least in data, China will certainly move faster than the United States.” Lingbo is working with partners through a data alliance to set standards: “Anything that is to be scaled must have a standard.”

29. His Mentor 汤晓鸥 and Rapid-Fire Questions: He Wants the Robot to Be a “J Person”

• He remembers 2 things about 汤晓鸥: a humorous fellow Northeastern Chinese, and someone who could “quickly identify the central contradiction in a state of disorder and chaos.” At his dissertation defense, while others asked how the system worked, 汤 asked, “How should this thing actually be used?”
• His book of life is The Smiling, Proud Wanderer. “I especially like 令狐冲’s open-mindedness and the way he does not seem to care too much about anything.” Is he that kind of person? “I hope to become that kind of person.”
• On the robot’s personality, he is decidedly unromantic: “I hope it is a J person; it will be good at getting work done.” If forced to choose between emotional value and utility, he would “decisively choose getting the work done”—emotional value should come from people. The work that has influenced him includes the company whose name sounds like Generalist in the audio and its work that sounds like GN-1: “They are genuinely very good at scaling data,” and have moved from papers to blogs; work from the team that sounds like PI in the audio; and Google DeepMind. His ultimate goal is to make embodied systems as distributable as language models: “If one day I could hand something directly to my parents and tell them, ‘I made this,’ the sense of achievement would be on a completely different scale from anything today.”