Summary

First interview after the leadership change in Xiaopeng’s autonomous driving unit: on October 9, 2025, Li Liyun stepped down and Liu Xianming, head of the world foundation model, took over as the fourth person to lead autonomous driving after Gu Junli, Wu Xinzhou, and Li Liyun. Asked about the speculation, he replied only, “There aren’t that many stories behind it—just look at it positively,” and redefined Xiaopeng as a “physical AI company”: a chip-based brain plus devices that interact with the physical world, with cars, robots, and flying cars all part of the same physical AI system.
The interview’s most aggressive technical call: “Language is poison.” Using an open-source language model as the foundation, as Li Auto does with DeepSeek, is the simplest shortcut, “but it will inevitably become like a poison—you will depend on it more and more.” Language is a discrete, redundant form of expression, while physical-world inputs are continuous sensor signals and outputs are continuous control values; inserting L in the middle is simply inefficient. Xiaopeng’s VLA 2.0 takes joint VL corpora as input and outputs actions directly, trained on 270,000-300,000 hours of pure video. After removing L, there was basically no obvious regression, while scaling became immediately visible. Long-standing modality confusion also improved materially.
“Strip” is the verb running through the entire interview: since last year they have stripped out lidar, planning-and-control rules, end-to-end, and language—“basically everything.” The current vehicle version outputs control signals directly from the model; it retains not even the rules for scenarios such as ETC. The philosophy came from his PhD adviser: “Simple is beautiful; the good things in the world are always simple.” Rules and loss functions were the hardest to remove, and he has slammed his hand on the table over them.
“Next year we will launch L4 in Guangzhou,” and he hopes to run Robotaxi well. The same architecture will handle more complex problems through more compute on the vehicle, more inference time, and larger models. Generalization has been unexpectedly strong—“generalization in overseas markets is actually very strong”—opening the question of taking the system global. Data is growing 30-40% per quarter, all of it real-world data.
His direct answer to余凯’s claim that automakers will not develop autonomous driving in-house: “Data is a gold mine,” and only automakers can close the loop at scale. Autonomous driving will standardize, but “one takeover per month and one takeover every twenty kilometers are completely different experiences.” The AI race among automakers is already white-hot, and “those who cannot do it well will basically be eliminated.” Xiaopeng’s 2025 AI investment of $4.5B is “more or less all spent”; when the senior brother approves budgets, his reply is one word: “Okay.”
The hidden through-line is organizational theory: AI is not a flash of genius but “necessarily a rigorously systematized thing”—a small, concentrated team plus a large-scale engineering system, with technology built bottom-up and the North Star set top-down. Meta struggled with AI because semiannual reorganizations and performance reviews forced a “break things and move fast” culture; Xiaopeng is radically flat, with Liu reviewing frontline engineers’ code directly and sleeping at 2 or 3 a.m. His key bet: “scaling in physical AI.”
Waymo is already handling 250,000 orders a week in San Francisco and has become part of everyday life rather than a tourist attraction. The first time he took his child in a Waymo, “sitting in the car of a former competitor, watching it perform so well, I felt extremely unwilling to accept it.” His milestone vision: “One day I’ll be able to put our kids in the car and say, okay, go to school by yourself.”

Deep dive

1. No hidden story behind the leadership change: a hard-headed leader spending his first three weeks subtracting

The interview was recorded on October 30, 2025, three weeks after he took the job. His entire response to the personnel change was: “There aren’t that many stories behind it—just look at it positively.” Asked about the respective missions of his three predecessors, his answer carried an engineer’s dry humor: “Maybe one day I’ll do this so well that I feel like I’ve somehow eliminated myself. That’s it.”
He accepts the labels outsiders attach to him: some call him “very stubborn,” some say he micromanages, and some call him hard-headed—“people inside the company say that too.” His defense is methodological: to make something happen, you have to accept risk, make forward-looking judgments, and keep risk within an acceptable range. “Without a hard-headed personality, many things simply cannot be pushed through, and you probably would not see the result.”
His most important decision in the first three weeks was simplification: no listening to reports, but going straight to the issues raised in the daily debugging or problem-solving meetings. “Simplify the processes, simplify the R&D steps, simplify what needs to be done.” Merge repeated work; downgrade or pause unnecessary work. “I still hope the entire center becomes an extremely simple organizational structure.”

2. From Facebook satellite mapping to Cruise: a mission-driven career and the “cruise coin” methodology

He graduated with a PhD from UIUC in 2016 and joined Facebook Connectivity Lab, working on high-altitude drone laser communications and satellite-image analysis. One project at one point used more than 16% of Facebook’s computing resources. After the Haiti earthquake, Sheryl demanded pre- and post-earthquake population and building distributions within 3 days; the team delivered in under 2 days for Red Cross rescue operations. The line that persuaded him when choosing a job was: “how many people have chance to change one billion people’s life.” He joined at the same time as a colleague named “开明”: “There was no way I could out-paper him… despair.”
A video from a truck driver’s driver-monitoring system triggered his move into autonomous driving. The driver had been driving for hours, becoming so tired that his eyes closed. “They have families too, wives and children waiting for them at home.” He chose Cruise over Waymo for a candid reason: “Joining the underdog in a comeback story is always incredibly exciting.” Once inside, he was equally blunt about the gap between expectations and reality: Cruise’s technology stack was “even worse than I thought,” and turning a demo into a scaled product was “far more difficult than imagined.”
Cruise left him with 3 things that still shape his route today: extreme simplification, large-scale infrastructure—its infrastructure team had several hundred people, one of the largest in autonomous driving at the time—and a “continuous learning machine.” The internal joke was that everyone could lie on a beach in Hawaii waiting for cruise coin to fall from the trees: “Data iteration solves every problem, and we can make money lying down.” His summary: every physical AI company today is still doing essentially these 2 things—scaling up with larger infrastructure and using data to solve problems.

3. A 1-hour offer decision: he joined an automaker for one reason—the data gold mine

In January 2024, he met the senior brother for 1 hour at a Silicon Valley office and, on leaving, told HR, “Hey, prepare the offer.” He joined in March. The conversation was more like a conspiracy than an interview. The senior brother asked, “I want to pull far ahead of the current generation of competitors. How should we do it?” He needed someone who understood how to make it happen, while Liu needed a boss who would support and understand him.
The core logic for joining an automaker fits in 1 sentence: “The core reason I came to Xiaopeng was data. Data is a gold mine.” If the future of physical AI is large-scale data, rapid iteration, and a closed infrastructure loop, “then the place I had to find was an automaker, because only automakers have this kind of data and a controllable data pipeline.”
There was also an element of personal obsession: “If it ultimately proves I was wrong, I accept that. But if it proves I was right and I was not on this track, that this had nothing to do with me—I might not be able to take it.”

4. Four generations of software stacks and the “cloud factory”: autonomous driving is becoming a foundation model

His technical periodization: software 1.0, from the DARPA Urban Challenge around 2009 through 2016-17, used lidar clustering plus traditional mathematical planning and control; 1.5, in the Mobileye mold, was half model and half rules—“No matter how strong your model is, your ceiling is capped by those rule codes”; 2.0, the pure-neural-network-plus-data-iteration approach proposed by Karpathy, shifted the bottleneck to network architecture and data volume; today’s VLA, VLM, and world models are still essentially end-to-end, supported by more chip compute for more parameters, which in turn require more data.
The “most reliable way forward” today is a cloud factory: vehicle hardware cannot support a large model, so a massive model is trained in the cloud and then deployed through distillation, quantization, and pruning. “Once the factory is trained, it will continuously generate models that can be deployed on different hardware.”
He deliberately provokes the naming debate: “People really don’t need to invent so many terms. What everyone is doing is essentially quite similar.” The most accurate term is physical AI. As for the models, “everyone is essentially building end-to-end models with different multimodal data.”

5. “Language is poison”: the complete case for stripping out L

The starting point was a temptation they had also tested: using an open-source model such as DeepSeek as the L foundation, as Li Auto does. “It is the simplest and most direct path, but in reality it will inevitably become like a poison—you will depend on it more and more,” because “that capability is not yours” and has a ceiling. Internally, they say: “A model is actually a machine, and its fuel comes from data. A machine by itself is useless; you have to get the data flowing. Once the data flow is mixed with any language supervision, the entire process becomes extremely inefficient.”
L has 3 technical failings. First, it must generate language supervision signals in the middle; even if a large model labels the data, humans still need to inspect it, introducing human supervision and reducing data efficiency. Second, language occupies a discrete codebook space, while the physical-world model takes continuous visual and sensor signals as input and produces continuous control values—longitudinal acceleration, kappa, yaw angle, and so on. “Generating discrete tokens with language and then translating them into continuous control values is obviously inefficient.” Third, deployment makes the problem even clearer: “Adding a very small amount of signal requires adding hundreds of tokens. That makes absolutely no sense.”
The turn came “when we added data and saw the bottleneck—when we could not add any more.” Stripping out L ran against intuition and the consensus of published papers—“Everything public tells you that language is important”—but the decision was unusually fast: “We discussed it a little. It seemed we did not discuss it for very long. We felt this was what we should do, so let’s try it.” He also cited the DeepSeek OCR paper as supporting evidence: there is no need to align images with the text-token space; “this is essentially also anti-formula.”

6. Strip everything to the end: the hardest things to remove were not lidar but rules and loss functions

The timeline: last year they removed lidar and planning-and-control rules. There was “a lot of debate” internally, until the senior brother decided that it “would definitely be more beneficial in the long term.” In the first half of this year they removed language. “We even removed end-to-end… basically everything.” The version now running on production vehicles outputs control signals directly from the model; even the rules for entering ETC scenarios “have also been removed.”
The greatest resistance came from removing rules and loss functions. “When the model performs poorly in certain situations, the most natural instinct is to add a specific loss function or rule. Can you withstand the pressure not to do that, and believe that I am going to pursue scaling and extreme simplification?” He admits he lost his temper over it: “I have slammed my hand on the table… and said many things I should not have said in meetings.” Outsiders assume end-to-end has long been a consensus; he corrected them: “Actually doing it inside a company is still extremely difficult. When you have real business objectives and accidents can really happen on the road, it is hard.”
The underlying philosophy came from a line by his PhD adviser: “Simple is beautiful; the good things in the world are always simple.” The common mistake is continually adding complexity to make a model look good with little data and little compute. He accepts the lesson of The Bitter Lesson: handcrafted features and hand-written reward functions may work in the short term, but “in the long run they are all a retreat from artificial intelligence.” What can the next generation remove? “It seems there is not much left to remove.” But he refuses to call this the final architecture: “You have to be able to abandon things you previously believed were correct in pursuit of something better. You need the courage and ability to do that.”

7. VLA 2.0 as a world model with decoding removed: 270,000-300,000 hours of video, with basically no obvious regression after stripping out L

The contrast with Li Auto is clearest in the training recipe. Li Auto trains L corpora first, then vision, then joint VL; he says that is merely “a good way to initialize the network parameters, but it cannot fundamentally solve autonomous driving.” Xiaopeng feeds in only joint VL corpora and outputs actions directly, with no pure L corpus. To address the “curse of dimensionality” between high-dimensional inputs and dozens of control values, it first trains a world model to understand how the world operates—effectively replacing textual CoT with latent CoT. “Let CoT exist directly as a latent space, do generation and diffusion, and see what kind of video it generates in the middle” to test whether it understands.
The world model and VLA are not 2 separate things. “Remove all the computationally complex decoding from the world model, let it produce only the intermediate reasoning and the final action, and it becomes the VLA 2.0 we have released.” Robots reuse the same architecture. At the same time, they are developing an in-house foundation model code-named Turing—it “may not necessarily” be a language model.
The empirical result surprised even them. At a data scale of 270,000-300,000 hours of pure video, “after removing it, our first reaction was to worry that it would not work—for example, traffic lights and waiting-to-turn areas. But none of that happened. It just went through naturally, completely beyond our expectations.” There was basically no obvious regression, and “we immediately saw obvious scaling.” Long-standing modality confusion—overdependence on L and inflexible actions—also improved materially. His retrospective strips away the mystery: “A lot of the time, you simply feel intuitively that something is right, dare to take that step, and push scale like a madman, only to discover that it works.”

8. Emergence and generalization: problems unsolvable in San Francisco are handled easily in Guangzhou

The concrete version of emergence: “In San Francisco, homeless people are lying everywhere and there is trash everywhere. You feel this problem cannot be solved. But go take a look at Guangzhou—this is not even the same order of magnitude. Problems you once thought were unsolvable are being solved very easily.” On the narrow roads of Guangzhou’s urban villages, the car drives “extremely smoothly, very much like a person but very safe.” This is not simple imitation learning; “it is actually thinking.” It waits for an oncoming car to clear a single-lane road and decides when to smoothly nudge based on tiny changes in the other vehicle.
Generalization is the biggest surprise and next year’s KPI: “Generalization overseas or in other markets has not degraded as much as you imagine. It is actually very strong.” That has created a new proposition: use this model to enter global markets. Data is growing 30-40% per quarter, all of it real-world data. Corner cases are the hardest data to obtain, and obtaining them first requires a zero-to-one system that tells you what a corner case is—shadow mode, user-takeover data, and data-distribution trigger models. “Without this system, much of what you collect will be useless data.”
The route has “coincidentally converged” with Tesla’s. After seeing Tesla’s recent talk, “everyone’s first reaction was: Damn, we collided again.” His interpretation carries a sense of inevitability: “There may be only one path to the truth when something is right. In the end, everyone may step onto this road.” Asked how many people in the industry are stripping out L as he is, he replied: “I don’t think there are that many… focus on doing what you should do well.”

9. Launching L4 in Guangzhou next year: the safety floor is the final barrier, and Waymo remains the thorn in his side

“Next year we will launch L4 in Guangzhou.” The same architecture will solve more complex problems with more compute on the vehicle, more testing-time scaling, and larger models. The essence of commercialization is “reducing the probability of system errors, thereby reducing the frequency of cloud takeovers and lowering costs.” There is only 1 clear route: a physical AI model plus extreme scaling—more resources, better infrastructure, and a larger data scale. He is equally candid about the unclear parts: how long scaling can continue, how to maintain the safety floor—safety alignment “is being attempted but has not formally started”—and how much software and hardware customization will matter once scaling takes off.
His engineering realism is worth recording: long context “is a very unfriendly solution for production,” because slower inference directly degrades vehicle control. “When a pedestrian suddenly emerges from behind an obstacle or an electric scooter suddenly darts out, you have to respond quickly. This will always be a trade-off.” Long-horizon user memory is “something everyone is exploring; there is currently no particularly good method.”
The Waymo passage is the most emotional moment in the interview. He took his child in a Waymo for the first time on a return trip to San Francisco—Cruise prohibited children under 13 from riding, so the child has called every one a robot car ever since. “You are sitting in the car of someone you once considered a competitor, and it performs very well. You feel extremely unwilling to accept it.” Waymo now handles 250,000 orders per week in San Francisco and has become part of everyday life rather than a tourist attraction. His technical judgment is that “it is not a generational gap; the next-generation technology stack is completely different.” His milestone is “running Robotaxi well in Guangzhou next year.” The end point is: “One day I’ll put our kids in the car and say, okay, go to school by yourself.”

10. The organizational theory of AI transformation: split, combine, split; $4.5B invested; those who cannot do it will “basically be eliminated”

The transformation was set in motion at Xiaopeng’s 10th anniversary last year, when the senior brother declared that Xiaopeng was an AI company: building cars, robots, flying cars, and chips ultimately means doing 2 things—building a chip-based brain and building devices that interact with the physical world. The organizational rhythm is from split to combine and back to split. For now, the company is concentrating its efforts on a physical AI foundation; once the model becomes strong, applications will emerge naturally, similar to “a very powerful model behind Claude—you do not need many people to build an application.” The company’s $4.5B AI and autonomous-driving investment for 2025 is “more or less all spent.” When he requests a budget, the senior brother says “okay,” at most asking whether it is enough. “So far, I have never been rejected.” The sharp question he was asked most recently was why he was spending so much money. He “still has not replied with a specific amount”; later, face-to-face, he said, “No more is needed.”
Responding to余凯’s claim that automakers will not develop autonomous driving in-house and should leave it to suppliers, he first laughed: “Let’s see where we are in 5 years.” Then he gave the direct answer: third parties “have a hard time obtaining the data you want at scale.” An automaker can automatically identify where data is missing and have the fleet collect it the next day, closing the loop. Autonomous driving will standardize, but “one car being taken over once a month and one car being taken over once every 20 kilometers are completely different experiences.” Strong capability is what gives a company the right to define the product. The AI competition among automakers is “quite white-hot,” and “those who cannot do it well will basically be eliminated.”
Meta is his repeatedly cited negative example. Reorganizations and performance reviews every 6 months meant that “after taking out holidays, you have 3 months left to do things,” producing the internet philosophy of “break things and move fast.” Meta’s innovation also relied on flashes of genius from individuals. “But AI is necessarily a rigorously systematized thing”—a small, concentrated team, sufficient resources, and “invisible underlying engineering capabilities,” including data, training infrastructure, fault tolerance, and system efficiency. Technical breakthroughs must come bottom-up, while top-down leadership must provide a clear North Star and reasonable timing expectations. He attributes Google’s improvement to Larry Page and Brin returning: ultimately, the test is still the founders’ style.
His own management mirrors that philosophy. First, he built the infrastructure team; second, he flattened the organization—team leaders write code and debug directly, while he reviews frontline engineers’ experimental results. He goes to bed at 2 or 3 a.m. and gets up at 7 or 8. He has no fixed recurring meeting with the senior brother; “whenever we think of a question, we just discuss it.” In less than a month, they have discussed 4 things: the production plan, overseas markets, generalization, and Robotaxi. The shift from rules to models followed the same pattern: at some point, model performance suddenly overtook the rules, and “the testers applauded.” 张小珺 noted that Li Auto had to “start a major fire” before removing its rules; Xiaopeng’s version is to “wait until the moment it surpasses them, then let it happen naturally.” In the rapid-fire questions, he placed only 1 key bet: “scaling in physical AI.”

The $2B Company Cutting AI Costs By 60% | Tuhin Srivastava Market Overview — November 18, 2025