Yao Song: physical AI is won by systems, not by young geniuses
From DeePhi to Orienspace to Striding AI, Yao Song (姚颂) argues that what decides physical AI is not one or two genius algorithms but data, compute, a talent pool and commercial delivery at every milestone — and that the intelligence-first giants may not be the ones who get to pick the fruit.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Offered boring or suffering, Yao Song keeps choosing suffering
Yao Song remembers entering Tsinghua's electronic engineering department in 2011 and setting himself a rule: eleven hours of study every day, with the clock paused for toilet breaks and meals. He admits there were simply too many people at Tsinghua who were ‘both smart and hardworking’, and diligence was the only way he could buy himself any sense of safety. Not until the second semester of his freshman year did he realise that university did not have to follow a single track. A teacher later summed it up for him: life is either boring or suffering. Yao Song took the second — he cannot stand the placid life of a Silicon Valley engineer, and that fear of boredom is what keeps pushing him back into hard-tech startups.
— Song YaoTaking all three chip giants' money would have made DeePhi unsellable
DeePhi (深鉴科技) met 50 institutions before its angel round and was turned down by all of them; in the end Zhang Yutong (张予彤) and Lin Renjun (林仁俊) at GSR Ventures (金沙江创投) understood the direction and invested. What he is proudest of is a strategic choice in fundraising: bringing Xilinx into the A round while turning down NVIDIA and Intel. His reasoning ran like this — if the endgame is an IPO or an acquisition, having all three chip giants on the cap table is actually dangerous. They compete with each other, so none of them would genuinely open up business resources, and at acquisition time they could bid the price up and make trouble until the company could not be sold at all. So he picked one: Xilinx, a direct competitor of NVIDIA that also badly needed software capability.
— Song YaoThe darkest stretch came after the exit, not after any failure
In 2018 Xilinx acquired DeePhi for more than US$3 billion. Yao Song says he paced happily around the table for two minutes after signing the contract, and then his life went dark. Up to that point the direction had been set by his advisor Wang Yu (汪玉); this was the first time he had to decide his own next step, and there was no road ahead of him, only fog. At 26 he describes himself as being in something like a midlife crisis — he did not even want to scroll his phone, nothing held his interest. That loss of a sense of purpose tormented him more than failure would have, and it became the starting point for taking a part-time role at Matrix Partners China and searching for a direction systematically.
— Song YaoWhat Zhang Ying gave him was information, not cash
To find a new direction, Yao Song became a part-time venture partner at Matrix Partners China (经纬创投) at the end of 2018. Zhang Ying (张颖) told him: you don't need cash, what you need is information. Every partners' meeting and every investment committee meeting was open to him. He swept systematically through lidar, brain-computer interfaces, GPUs and controlled fusion: Li Yifan (李一凡) convinced him that the moat in lidar is engineering capability rather than the technical route; brain-computer interfaces in China have to pass through too long an approval cycle; GPUs were blocked by the non-compete terms attached to the acquisition; fusion is too far from industrialisation. In the end he waited until the commercial-space window opened.
— Song YaoReliability, payload, speed, then cost — the order cannot be reversed
Yao Song breaks Chinese commercial spaceflight into three words — China, commercial, spaceflight — and then ranks the priorities: good, many, fast, cheap. Good means a high success rate; no customer dares fly a rocket with a 60% success rate. Many means payload capacity: building out a satellite internet constellation needs at least four or five tonnes, and five or six is better. Fast means seizing the market window. Cheap comes last. That is why Orienspace (东方空间) chose the solid-strapped-onto-solid route, and Gravity-1 (引力一号) became the world's first rocket to strap solid boosters onto a solid core — and also the first launch by a Chinese private company to be livestreamed. He knew the livestream might capture a failure, and still insisted that spaceflight should not be kept so far from ordinary people.
— Song YaoOne or two young geniuses cannot solve physical AI's problems
Yao Song does not buy the narrative of ‘one or two young geniuses, or one or two young professors, founding a company alone’. He points to the history of deep learning: AlexNet came out of Hinton's group, VGG out of Oxford, GoogLeNet out of Google, ResNet was proposed by Kaiming He (何恺明), and the chain from R-CNN to RFCN was carried forward by different teams in relay. Nobody stands at the frontier forever. So competition in physical AI is a systems war over data, compute and the talent pool, not a game for a handful of clever people. He quotes something Cao Xudong (曹旭东) said years ago: 50% of an algorithm's competitiveness comes from data, 30% from compute, and only 20% from human experience and inspiration.
— Song YaoLanguage is not the world: continuous physics resists discrete semantics
Yao Song rejects ‘language is the world’. His reasoning: the digital world is discrete semantics, while the physical world is a continuous analogue signal, and you simply cannot use language to describe the tilt angle of a cup or the distance between two people precisely. A future general world model may need 50T to 100T parameters, but physical AI has to be real-time and low-latency, and in all likelihood can only run inference on-device or on a small chip. So the intelligence-first mode of OpenAI and Anthropic — stacking vast compute into a general model — may not be what picks the physical-AI fruit. That is the reason Striding AI works in latent space and keeps a physics backbone rather than generating pixels.
— Song YaoTen-zero offense is the founder's error; he plays eight-two
Yao Song draws a line between founders and entrepreneurs: a first-time founder is often ten parts offense and zero parts defense, while he calls himself eight parts offense and two parts defense. The floor on defense is: don't go to prison, don't go bankrupt, don't boast about things you cannot do. He is repelled by the industry habit of passing teleoperated video off as autonomous demos, and requires his team to label every video with whether it is fully autonomous. Pursuing success and preventing failure, he says, are two completely different paths; once your moves distort, the price of a high valuation in the short term may be a lasting loss of credibility. This is his third company, and he is running it on long-termism rather than one all-in bet.
— Song YaoIn their own words · checked verbatim
I paced around that table for two minutes, absolutely delighted. And once those two minutes were over, my life went dark.
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 push the door open and take part. Partners' meeting, investment committee, whatever project they are putting together. It 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: ‘事以密成’, things succeed by being kept close. That is, you have — there are some core strategies you simply should not talk about publicly.
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, I think one or two young professors, or one or two young geniuses, frankly 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 (正大集团) has more than 400,000 full-time employees. So this is not a resource you can simply buy with money. ... Yes, this way I can bring my data cost down to below one tenth of the normal market purchase 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 completely clear about it. I know my company went public because of the support of the Party and the state. Not necessarily because I had 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
| Daily study time Yao Song set himself as a Tsinghua freshman | 11 hours | 8:15 |
| Investment institutions that turned DeePhi down before its angel round | 50 | 32:23 |
| Market rate for buying egocentric embodied-intelligence data | 300-600 yuan/hour | 1:49:43 |
| Humanoid robot products already launched in China | more than 100 companies, more than 200 models | 2:29:35 |
| Yao Song's timeline for embodied intelligence to exceed the auto industry in output value | 10-15 years | 2:32:41 |
| Striding AI's funding and headcount six months after founding | nearly US$100 million, about 70 people | 2:38:04 |
Glossary
- WAM (world action model)
- A world/action model grown out of video generation; Yao Song thinks its pixel-generation route is a poor fit for real-time inference on-device.
- VLA (vision-language-action model)
- A model architecture that handles vision, language and robot actions in one system.
- latent space
- The low-dimensional abstract space inside a model; Yao Song uses it as shorthand for the route that learns physical laws without generating pixels.
- egocentric data
- Video data captured from a first-person point of view, used to train robot manipulation.
- teleoperation
- Operating a robot remotely; the industry often passes teleoperated video off as autonomous intelligence.
How to listen
Founders and investors preparing to move into embodied intelligence; anyone trying to understand the funding logic and cycles of Chinese hard tech; Yao Song's early years are also a useful reference for engineers thinking about a switch.
The first 25 minutes are personal-growth narrative — play it at speed. The densest stretches are commercial spaceflight after 1:13 and the physical-AI reasoning after 2:06.