Yao Song: The Decisive Battle in Physical AI Isn't a Genius, It's a Systems War
From DeePhi to Orienspace to Striding AI, Yao Song judges that the decisive battle in physical AI is not one or two genius algorithms, but data, compute, talent pools, and commercial landing at every milestone. Even intelligence-first giants may not necessarily pick the fruit.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Tsinghua's 11 hours and boring/suffering
Yao Song recalls that after entering Tsinghua's Department of Electronic Engineering in 2011, he set a rule for himself: he must study 11 hours every day, pausing the timer even for bathroom breaks and meals. He admits that Tsinghua has too many people who are "both smart and hardworking," so he could only exchange diligence for a sense of security. It was not until the second semester of his freshman year that he realized there was more than one path through university. Later, a teacher summed it up: life is either boring or suffering. Yao Song chose the latter—he could not endure the idle life of a Silicon Valley engineer, and this fear of boredom drove him to jump into hard-tech entrepreneurship again and again.
— Song YaoRejecting Nvidia and Intel
Before its angel round, DeePhi Technology met 50 institutions and was rejected by all of them, until Zhang Yutong and Lin Renjun of GSR Ventures understood the direction and invested. What he is most proud of is the strategic choice in financing: bringing Xilinx into the A round and rejecting Nvidia and Intel. His reasoning was that if the endgame is an IPO or acquisition, having three chip giants as shareholders simultaneously is actually dangerous—they compete with each other, won't truly provide business resources, and in a future acquisition might bid up the price and cause trouble, making the company unsellable. So he chose Xilinx, which directly competed with Nvidia and urgently needed to strengthen its software capabilities.
— Song YaoAfter signing, falling into darkness
In 2018, DeePhi was acquired by Xilinx for over 3 billion USD. Yao Song says that after signing the contract he happily paced around the table for two minutes, and then his life fell into darkness. Because the direction had previously been set by his advisor Wang Yu, and for the first time he had to decide the next step himself, but there was no path before him, only a fog. At 26, he describes himself as having a midlife crisis, not even wanting to scroll on his phone, interested in nothing. This chaos of losing a sense of purpose was more tormenting than failure, and it became the starting point for him to work part-time at Matrix Partners and systematically search for direction.
— Song YaoZhang Ying gave information, not cash
To find a new direction, at the end of 2018 Yao Song worked part-time as a venture partner at Matrix Partners China. Zhang Ying told him: you don't need cash, you need information. All partner meetings and investment committee meetings at Matrix were open to him. He systematically scanned lidar, brain-computer interfaces, GPUs, and controllable nuclear fusion: Li Yifan convinced him that the barrier in lidar lies in engineering capability rather than technical route; brain-computer interfaces in China would require too long an approval cycle; GPUs were restricted by non-compete clauses from mergers; nuclear fusion was too far from industrialization. Eventually he waited for the commercial space window to appear.
— Song YaoGood, many, fast, cheap—the order cannot be wrong
Yao Song breaks down China's commercial space into three keywords: China, commercial, space, and then prioritizes: good, many, fast, cheap. Good means high success rate—no customer dares to use a rocket with a 60% success rate; many means carrying capacity—satellite internet constellations need at least four to five tons, preferably five to six tons of capacity; fast means seizing the market window; cheap is placed last. Therefore Orienspace chose the solid-on-solid route, Gravity-1 became the world's first solid-fuel rocket with solid boosters, and also conducted China's first private rocket launch livestream. He knew the livestream might capture a failure, but insisted that space should not be too far from ordinary people.
— Song YaoOne or two little geniuses can't solve the problem
Yao Song is not optimistic about the narrative of "one or two little geniuses or young professors starting a company alone." He cites the history of deep learning: AlexNet came from Hinton's group, VGG from Oxford, GoogLeNet from Google, ResNet proposed by He Kaiming, R-CNN to RFCN relayed by different teams—no one stands at the forefront forever. So competition in physical AI is a systems war of data, compute, and talent resource pools, not a game for a few smart people. He quotes Cao Xudong's early words: 50% of algorithmic competitiveness comes from data, 30% from compute, and only 20% from human experience and inspiration.
— Song YaoLanguage is the world? Impossible
Yao Song opposes "language is the world." His reason: the digital world is discrete semantics, while the physical world is continuous analog signals; using language to precisely describe the tilt angle of a cup or the distance between two people is simply impossible. A future general world model might need 50T to 100T parameters, but physical AI must be real-time and low-latency, and very likely can only infer on the edge or on small chips. Therefore intelligence-first models like OpenAI and Anthropic that pile up general models with massive compute may not necessarily pick the fruit of physical AI—this is why Striding AI adopts latent space and retains a physical backbone rather than generating pixels.
— Song YaoTen attack, zero defense vs. eight attack, two defense
Yao Song distinguishes between a startup founder and an entrepreneur: a first-time founder is often ten parts offense, zero parts defense, while he calls himself eight parts offense, two parts defense. The bottom line of defense is not going to prison, not going bankrupt, and not boasting about things you can't do. He dislikes the industry using teleoperation videos to pass off as autonomous demos, and requires his team to label every video whether it is fully autonomous. He says pursuing success and preventing failure are two completely different paths; once actions become distorted, the cost of short-term high valuation may be long-term loss of credibility. He is now on his third startup, relying on long-termism rather than going all in.
— Song YaoIn their own words · checked verbatim
I was very happy, pacing around this table for 2 minutes. After pacing for two minutes, my life fell into darkness.
我非常高兴的围绕着这个桌子踱步了2分钟。踱过了两分钟以后，我的人生就陷入了灰暗。
Song Yao57:20
Any meeting at Matrix Partners China, you can just push open the door and participate. Whether it's a partner meeting or an investment committee meeting, or any project they're building, this gives you a privilege.
经纬中国的任何的会，你都可以直接推开门去参与。不管是合伙人会还是投决会，还是他们在建任何的项目，这个给你一个特权。
Song Yao1:04:44
I think there's a saying: things succeed through secrecy. That is, you have some core strategies that should not be spoken externally.
我觉得有一句话叫做事以密成。就是你有。有一些核心的战略就是不应该对外讲的。
Song Yao1:29:05
But frankly speaking, I think one or two young teachers or one or two little geniuses, frankly speaking, cannot solve these problems.
但是坦率讲，我觉得一两个年轻老师或者一两个小天才，坦率讲是无法去解决这些问题的。
Song Yao1:47:39
Charoen Pokphand Group has over 400,000 full-time employees. So this is not a resource that money can simply buy. ... In this way, I can reduce my data cost to less than one-tenth of the normal market procurement price.
正大集团有40多万的全职员工。所以这个不是钱能简单买来的资源。…对我可以通过这种方式把我的数据成本下降到正常市场采购价的10分之1以下。
Song Yao1:50:44
He said, I am very clear, I know that my company's IPO is because of the support of the Party and the country. It doesn't necessarily mean I have already done well enough.
他说我很清楚，我知道我公司的上市是因为党和国家的支持。并不一定我已经做的足够好了。
Song Yao2:46:48
Figures
| Yao Song's self-imposed daily study hours during freshman year at Tsinghua | 11 hours | 8:15 |
| Number of investment institutions that rejected DeePhi Technology before its angel round | 50 | 32:23 |
| Market price for purchasing egocentric data for embodied intelligence | 300-600 yuan/hour | 1:49:43 |
| Statistics on humanoid robot products launched in China | Over 100 companies, over 200 models | 2:29:35 |
| Yao Song's estimated period for embodied intelligence output value to surpass the auto industry | 10-15 years | 2:32:41 |
| Striding AI's funding and team size half a year after founding | Nearly 100 million USD, about 70 people | 2:38:04 |
Glossary
- WAM / World Action Model
- A world/action model developed from video generation; Yao Song believes its pixel-generation route is not suitable for real-time edge deployment.
- VLA / Vision-Language-Action Model
- A model architecture that uniformly processes vision, language, and robot actions.
- latent space
- A low-dimensional abstract space inside a model; Yao Song uses it to refer to a path that does not generate pixels but only learns physical laws.
- egocentric data
- Video data collected from a first-person perspective, used to train robot manipulation capabilities.
- teleoperation
- Remotely controlling a robot; the industry often passes off teleoperation videos as autonomous intelligence.
How to listen
Founders and investors preparing to enter embodied intelligence; those who want to understand the logic and cycles of hard-tech financing in China; Yao Song's early experience also offers reference for engineers transitioning careers.
The first 25 minutes are personal growth narrative, can be sped up; after 1:13 commercial space and after 2:06 physical AI deductions are the densest.