The better L2 gets, the farther it is from L4: the self-driving route war
Pony.ai's Lou Tiancheng argues L2 and L4 have different goals: L2 users just need it to be good enough and cheaper, so the better it gets, the farther it is from L4; L4 needs 10,000 hours without an incident, and it needs a ruler to judge good from bad.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Driverless is the real shock that breaks out of the circle
Wuhan's Apollo Go drew attention, and Lou Tiancheng says fleet size and a larger trial area are only that 1%; 99% is driverless. A person sitting in a car with no one in it, even with a safety driver beside them, is a completely different kind of impact; if this opening were even 10,000 or 100,000 vehicles but there was still a person sitting there safely, it absolutely would not achieve today's effect. So why Wuhan caught fire may be related to the city, but that is definitely not the only reason. He is also asking around about why Tesla pushed Robotaxi from August 8 to October, and says traditional L2 assisted driving has no driverless component, and whether it can cross that 99% is very different.
— Lou Tiancheng100 to 1000 hours is stuck on the evaluation system
He uses 1 hour, 10 hours, 100 hours, 1000 hours, 10000 hours to describe self-driving. 1 to 10 relies on the base model, 10 to 100 relies on data collection and complex models, and 100 to 1000 is most critically about evaluation: last month you did 200, this month the new version is 220 or 180, you cannot guess, and you cannot just run for one hour; running ten thousand hours is again affected by rain, roadwork, and heavy traffic on Friday night. You have to strip out all that noise to truly judge whether you are doing well. He spends most of his energy on the contest based metric system, as it is called internally. This is a capability beyond resources; otherwise no matter how much money, data, and machines you have, you are just spinning in place.
— Lou TianchengDifferent goals are what produce different routes
On the route war, Lou Tiancheng says only the goals differ. Where you want to go and where you finally arrive differ the most; the route only determines how fast and how close. He compares lidar to a cheating student who looks straight at the answers; pure vision is like a good student, and the city thinks even an excellent good student struggles against a cheating student, so for full driverless you choose lidar. Tesla's pure vision is because it does not need full driverless, the passing bar is lower, and there are also cost and vehicle-appearance delays. After a test drive, He Xiaopeng said Tesla FSD will definitely catch up to Waymo next year; he disagrees: L2 and L4 are not the same kind of comparison, one is like football and one is like basketball, the evaluation standards differ, and the better FSD gets, the farther it is from L4.
— Lou TianchengThe better L2 gets, the farther it is from L4
L2 users only require it to be roughly V1, so an L2 product can lower its metric from 10 to 8 because of cost reduction; that is a better L2, but it is farther from L4. Because L4 needs 10,000 hours without an accident. He says incremental self-driving got a bargain: why do L2 first and then L4, rather than the reverse? L2 and L4 have different goals, and the essential technical difference means FSD definitely cannot achieve L4. Lidar is the cheating student, physically measuring distance directly with active light; L4 is an exam with ten thousand questions, and the pure-vision good student may study for a long time while lidar looks straight at the answers. Beyond safety metrics, speed and comfort are also part of the evaluation.
— Lou TianchengAfter surpassing humans, data becomes a distractor
Before 1000, the more data the better; after surpassing the human average, the more data the harder to converge. Because human driving behavior in the data varies enormously, the model does not know whom to learn from, and it cannot take an average. He mentions Edo saying that when there is a lot of data the gradient quickly drops to 0 and cannot converge. He also gives his own example: he reached 6000, and adding another 1000 of data pulls him back to 4000. Data is not to be thrown away but processed, so that only the useful part remains. End-to-end improves generalization, and more data can help generalization, but once you surpass the data's average level, data becomes a distractor, and the more there is the harder it is to surpass.
— Lou TianchengGiving up the original aspiration is not because of hardship, but unrealistic temptation
After 8 years of entrepreneurship, Lou Tiancheng says the biggest temptation is that you could absolutely build something that commercializes in one or two years. Mr. Yao Qizhi spoke of not forgetting the original aspiration, and he later understood that the second half — holding fast to the original aspiration — is the truly hard part. Most people think giving up the original aspiration is because you hit many difficulties, but the real reason is unrealistic temptation; many people can persist in the face of hardship, yet cannot hold on in the face of temptation. Many companies that came out back then are now not even known by name; they did not persist, and it was basically not because of hardship or lack of money, but because they did not withstand temptation. So he treats self-driving as the ultimate direction for several generations of technical people, and is willing to wait another three to five years.
— Lou TianchengCompetition masters hide their cards: Hold hard
Competition technique Hold hard: there are things you know how to do, but you just do not let people know, and through some method you make your opponent think you do not know either, so he gives up the idea of really going all out. For example, everyone is at 80 points, and you can actually do 90; if you let the opponent know, he will dig out his potential; so you control the score, let him think you are about the same, boil the frog in warm water, and reveal your sword at the last moment, when the opponent has no time to react. But company development is different, it needs publicity, and intermediate results, making friends, and partners all matter. In competition, leading in the middle does not necessarily mean winning in the end; a company cannot copy the card-hiding move.
— Lou TianchengAI surpassing humans is not AI's peak
Lou Tiancheng says a human is just an agent, and asking whether AI will surpass humans is too flattering to yourself — what makes you think humans can judge whether AI can surpass humans. AI itself is a fitting-and-regression process, it has not created any new logical range, and today's large models are only mimicking human behavior, far from artificial general intelligence. But he is not mystical, and he admits this approach may be very good and can surpass humans. He also says the human energy-consumption ratio is terrifying: after a meal he is still not hungry now, and the same energy given to a robot would have long failed; the heart beats so many times in a lifetime without breaking, and it cannot be built. He even thinks the world may be a simulator, and the speed of light is an important parameter.
— Lou TianchengIn their own words · checked verbatim
There is no — it is only that the goals differ. First, the biggest difference in goals: where you want to go and where you finally arrive, that is the biggest difference.
没有 这只有目标不同 首先目标不同的最大差别 想到哪去 和最终你到了哪 这是差别最大的
Lou Tiancheng17:50
When you start to surpass humans, more data is not better, because data is your distractor.
当你开始超越人类的时候 数据不是越多越好 因为数据是你的干扰项
Lou Tiancheng34:15
Say today I reached 6000, not 10000, reached 6000, fine, I add 1000 more data, and it pulls my body's 6000 back, turns it into 4000.
就是我今天达到6000了 没达到10000 达到6000了 好我把1000数据加多 就把我身体6000拉回来 变成4000
Lou Tiancheng36:16
But the real fundamental reason for giving up the original aspiration is unrealistic temptation, not hardship.
但是真正的放弃出境的根本原因是因为 不现实的诱惑 不是因为坚固
Lou Tiancheng45:18
Say there are things I know how to do, but I just do not let people know; through some method I make you think I do not know either.
就说我有些东西我知道怎么做了 但我就是不让人知道 我通过一些办法让你觉得我也不知道
Lou Tiancheng51:18
But when you ask this question, are you not flattering yourself too much? What makes you think a human can judge whether AI can surpass humans?
但是你问这个问题的时候是不是对自己太高看了 就你凭什么觉得人能够判断AI能不能超过人
Lou Tiancheng57:22
AI itself is a fitting-and-regression process; AI has not created any new logical range.
AI本身是一个拟合回归过程 AI并没有创造一些新的逻辑范围
Lou Tiancheng58:23
Figures
| Wuhan driverless time point | reached driverless by the end of 2022 | 2:02 |
| Pony car no-takeover duration | over 10,000 hours | 4:02 |
| Self-driving time scale | 1/10/100/1000/10000 hours | 5:02 |
| Reaching the 10,000 level | close to 200,000 kilometers without a problem, 10 times a human driver | 14:07 |
| Time for end-to-end to become the mainstream route | 2 years | 17:07 |
| Large-scale Robotaxi time | another three to five years | 37:16 |
| TopCoder China participation | 11 consecutive years, runner-up last year | 49:18 |
| Programming competition finals last place | 4 times | 55:20 |
| Pony.AI expectation | best case three years, worst case five years | 56:21 |
| Tesla Robotaxi launch time | said August 8 in April, later postponed to October | 3:02 |
Glossary
- contest based metric system
- Lou Tiancheng's internal name for the metric system that evaluates how good L4 is, used to judge whether a version has gotten better or worse.
- MPX
- The collective term for metrics such as Miles per Intervention, measuring how long autonomous driving goes before needing human intervention.
- Hold hard
- A programming-competition technique of hiding your true strength so your opponent misjudges the gap.
- end-to-end
- Using a single model to go directly from sensor input to driving decisions, avoiding information loss between modules.
- BEV
- A perception representation that unifies multi-sensor data into a bird's-eye view.
- Robotaxi
- An autonomous vehicle with no human driver that charges for ride-hailing service.
How to listen
Founders, investors, and engineers working on autonomous driving, Robotaxi, and L4 deployment; also for anyone who wants to see how Lou Tiancheng runs a company with a competitive-programming mindset.
If you only care about the technology and the business, the competition and personal-experience parts after 44 minutes can be skimmed.