Post-95 founders max out their Musk score, but their idol is actually Mao
What drives this wave of young founders isn't money, it's "this thing is badass, so let's build it first"; they don't write BPs, don't chase leaderboards, don't do PR on their own initiative, and only accumulate reputation within the circles they care about.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AI 1.0 was tuning parameters, 2.0 is general models
Song Yacheng draws the difference between SenseTime and Minimax very clearly: the AI 1.0 era was a large number of scientists using hand-tuned parameters to solve long-tail problems one by one — face recognition, OCR, vehicle recognition, illegal dumping, one algorithm per scenario. By AI 2.0, it's the brute-force aesthetics of data driven, using one general large model to solve general problems, then growing segmented applications on top of the model. His judgment is that "it's a very black-box thing" — not covering the long tail with small algorithms plus a technical architecture, but optimizing one very general model.
— Song YachengSix months without pay, 30-plus people held on
They started the company at the end of last year, right at the point when the US was raising rates, Russia and Ukraine were at war, and everything that should have blown up and shouldn't have blown up had blown up. Ask any FA and they'd say VCs might barely invest at all in 2023. Song Yacheng put his own money in, and the company only formally started paying salaries in June; before that there were six months with no pay for anyone, and the team of roughly thirty-plus people kept going the whole time. He says this is a trait of post-95 and post-00 founders — "we are a group of very faithful people."
— Song YachengThe Musk score is demanding that essential thing to the death
Song Yacheng's definition of the "Musk score": knowing very clearly what I want, then I want these one, two, three things, and no matter what method I'll get there; if someone comes along who can help me get there, we get there, and if they can't, I swap them out. His example is going to SIGGRAPH this August to give a Keynote — the top graphics conference's fiftieth anniversary, fifty or sixty thousand attendees, with Nvidia's Jensen Huang, IBM, Unity and Sony each speaking for an hour, and them too. Never in history had a Chinese person given a Keynote at SIGGRAPH. The organizing committee at one point suspected they were a shell company; they asked partners like Blender's CEO to write recommendation letters proving the company really existed, went through two or three rounds of interviews, and in the end what he went up and said was "I'm actually not here to talk about technology."
— Song Yacheng3D data is scarce, so they scraped together 11 million entries
He breaks AI into three things: the horse trainer (algorithms), the racetrack (compute), and the fodder (data). 3D is an extremely data-scarce field — most people have written Weibo posts, posted to Moments, shot videos, but never K-ed a 3D model. At the time OpenAI's private dataset was 2 million, Nvidia's 1 million, and other big companies around 700,000. As a startup, they spent huge amounts of time working with games, animation, film and TV, model communities and databases, pestering and trading on face, and in the end got 11 million high-quality entries — 3D native data hand-K-ed one by one by people in games, animation and film. A single model costing tens or hundreds of dollars is normal; buying it was out of the question.
— Song YachengHis startup idol is Mao, not Musk
Song Yacheng says the first answer from people around him for a startup idol is basically always Elon Musk, but he's a bit different — his idol is Chairman Mao. The reason is that he cares a lot about faith: the soul of a company lies in having faith, and things done out of faith don't need you telling someone every day, they'll go do it themselves. He observes that most post-80 founders didn't go abroad in middle school but went abroad for a master's, and their idols are generally Jobs types; his generation went out earlier, and their idols are different instead. He distinguishes Jobs from Musk: Jobs pulled an MP4 out of his pocket as a kid, while what Musk makes is a sci-fi experience, "dragging nine strips and needing to run faster than nine strips."
— Song YachengTwo completely opposite kinds of people live inside an AIGC company
The technical people are relatively rational, more I-type, a bit socially anxious, don't like team building, each managing their own module like an assembly line, and are extremely credentialed — they're about to have 20 PhDs, basically all Tsinghua bachelor's, master's and PhDs. The product people need a very warm culture: eating together, talking games and animation, going to CP shows and ChinaJoy, playing games and watching anime during work hours, a cozy office environment. The two groups' social tokens and values are different: one thinks what's badass is a game that made money, a product with high DAU; the other thinks what's badass is 50 papers and 40,000 citations. How to bring these two cultures together is a problem many AIGC companies have to solve.
— Song YachengOld hands paired with young geniuses is a deliberate team design
Song Yacheng says he built it this way instinctively: every team is a "startup old hand with a lot of experience and a zero-to-one success" paired with a "very young genius full of aspiration, wanting to prove themselves, with strong execution." CTO Liang Ding was head of general models at SenseTime, an early SenseTime employee, led several hundred people, and spent nine years at SenseTime; chief scientist Cao Yanpei previously founded a 3D startup acquired by Kuaishou, then ran 3D generation at Tencent. Most of the technical staff are post-95, including Guo Wencheng, founder of 3D Stable Diffusion, born in 1997. He uses finding an oasis as an analogy: old hands can smell it and know how much provisions to bring and how to allocate them, while a genius might ride a bicycle into the desert and is more likely to discover something new.
— Song YachengAbundant capital is a curse; the trough forces productization
Song Yacheng says he's very glad to have started a company at this time. When the capital environment is especially good, a company easily raises a lot of money, so it has no urgency to ship a product or actually make money, and it doesn't focus, with lots of resources to waste, and in the end maybe three to five years pass with no product and no real money made, and the company is gone — what he calls "the curse of startup resources." Instead, companies that come out of a capital market trough think from day one about how to productize and how to commercialize, with user experience as the optimization target, and each person manages the user's expectations rather than the boss's expectations. He says the sense of crisis at the end of 2022 and start of 2023 made him clearer-headed, turning him back from the mental circuit of "startup means fundraising."
— Song YachengGPT isn't as fast as imagined, and fundraising changed with it
He analyzes two key reasons the fundraising environment worsened in the second half of this year: first, China-US relations, with some mainstream dollar institutions running into difficulties; second, OpenAI and GPT haven't developed as fast as people imagined. When ChatGPT first came out everyone off the top of their heads thought AGI would arrive immediately, AIGC could generate everything, and the world would be upended in a year or two, but by this year GPT is still GPT4, AIGC can still only replace 30% of art, and the maturity of Runway and 3D generation isn't there yet. The gap made investors conservative. The response is to change the fundraising rhythm: no more waiting three months to close round after round, but if there's 100 million you can talk, invest a round directly, short, flat and fast, flexible.
— Song YachengMiaoya and Heizhen are templates; 3D lacks a content platform
Song Yacheng thinks AIGC found quite a few paradigms this year: Miaoya stipulates you can only shoot yourself, only ID photos, which is a template; Heizhen can make Guo Degang speak English and Musk speak Chinese, also a content template. But these may be flash-in-the-pan phenomenon products, because there's no content platform to catch the traffic and creators. He judges the biggest opportunity is in 3D — because 3D has no content platform. If you go build a video content platform, TikTok already exists, and squeezing it is very hard; but 3D is a relatively blue ocean, with room to create a 3D content platform.
— Song YachengIn their own words · checked verbatim
I know very clearly what I want, then I want these one, two, three things, I want to reach these goals, and no matter what method, I just get there.
我非常清楚的知道我想要什么 然后我就要这一二三 我要达到这几个目标 不管什么样的方式 我就是达到
Song Yacheng25:45
I just want data, I don't care what kind of good deed, I just want data, and this data must be at least one quantum more than others'.
我就是要数据 我不管什么样的好事 我就要数据 而这个数据 一定要比别人 至少多一个量子
Song Yacheng34:22
What Jobs made is a product experience that makes people feel great, comfortable, high-end. What Elon Musk makes is a sci-fi experience.
乔布斯做的东西是一个让人觉得很爽 很舒服 很高档的一个产品体验 埃罗马斯克做的是一个科幻体验
Song Yacheng39:27
The biggest moat is just time, it's that in your life you take the time...
就是最大的壁垒 就是时间 就是你一生中 你把时间
Song Yacheng53:31
Instead, what we see is that those great companies all come from the trough of the capital market.
反而是我们看到那些伟大的公司 都是来自于资本市场的低谷
Song Yacheng57:32
You'd be better off from day one being a very muscular muscleman, but with no fat on him, and he only throws one punch, and that one punch is enough.
你还不如从day one 就是一个非常肌肉发达的一个肌肉男 但是他没有什么肥肉 而且他只打一拳 只要这一拳就可以
Song Yacheng58:32
Their, or our, management style is faith, that is, I think this is a very important thing, that only when you have enough faith, enough love...
他们的 或者我们的一个管理方式就是信仰 就是我觉得这是一个非常重要的事情 就是只有你有足够多的信仰 你有足够多的热爱
Song Yacheng1:12:43
Figures
| SIGGRAPH attendee scale | Fifty or sixty thousand people | 26:19 |
| SIGGRAPH history | 50th anniversary | 26:19 |
| OpenAI private 3D dataset | 2 million | 34:22 |
| Nvidia 3D dataset | 1 million | 34:22 |
| Other big companies' 3D datasets | About 700,000 | 34:22 |
| Number of PhDs at the company | Nearly 20 | 44:28 |
| Company team size | 70-plus people, of whom about 30-plus are interns | 50:31 |
| Technical team size | 30-plus people, at least 80% | 49:29 |
| Share of art AIGC can currently replace | 30% | 1:01:33 |
Glossary
- SIGGRAPH
- The world's largest top conference in computer graphics, where Nvidia and other companies give keynote talks every year.
- mesh
- A 3D model with textures and materials that can enter a professional pipeline for further editing.
- PMF
- The state where a product is genuinely validated by the market and users are willing to use it.
- MVP
- A product form that ships fast at minimum cost to validate demand.
- Citation
- A quantitative measure of academic influence, one of the standards the technical crowd uses to measure "badass."
How to listen
Founders and investors watching AI startups, AIGC applications and 3D generation, and anyone who wants to understand the real state of post-95 and post-00 founders.
The personal history from 2:13 to 8:02 can be fast-forwarded; information density is low.