The Pain of Frontier Tech Entrepreneurship: On IPO Day, I Felt I'd Let Down Those 12 Years
Li Zhifei, founder of Mobvoi, says that at the moment of the bell-ringing, he wasn't thinking of glory but of whether the outcome justified the 12 years of hardship — and it didn't. He sums up the fate of frontier tech startups as a block of ice that can melt at any moment.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Most frontier technologies eventually disappear
Li Zhifei proposes a "funnel model": of 100 frontier technologies, only a tiny fraction can be packaged from a technology into a product, form a business model, and have the technology investment and commercial value match to a sustainable level. Most technologies at that point in time cannot pass the funnel test. But that doesn't mean the technology is useless — it might come back 10 years later because the environment has changed, the technology has matured, demand has matured, or new demand has emerged, and the technology is pulled out and run through the funnel again. He cites Google Glass in 2013: at the time everyone literally saw the future, but the product quickly disappeared, and the entire AR and VR industry hasn't taken off in all these years; now GPT-4o and Google have spawned glasses again, and you see the future again — it's back.
— Li ZhifeiA vision without a path and rhythm plan is just耍流氓
Li Zhifei says it's easy to talk about a broad future — he said in 2014 that every company would have an AI department, and a normal person wouldn't be too far off with such a judgment. What really matters is the judgment of path and rhythm: what to do first, what to do second. He gives two competitive scenarios: if a giant and a startup reach a consensus at the same time, and that consensus is wrong, the startup will definitely die because it can't hold out; if the startup forms a cognition of the future half a year to a year ahead of the giant, and has a relatively precise plan, then there's a certain probability of success — provided you can use that one-year time gap to build some kind of barrier, otherwise the startup will also die. He adds: even if this thing looks hopeless, as long as there's an investor willing to give you a billion dollars, that can keep you alive.
— Li ZhifeiThe key to GPT-4o is unification plus end-to-end
Li Zhifei first states, "We don't know either, because they didn't publish a paper; we're all guessing." He identifies two core points: first, it's a unified single model, with all modalities in the same model; second, it's most likely based on the GPT architecture, using a language model as the cognitive foundation, then adding other perceptual modalities. He uses the analogy of teaching a monkey to cook versus teaching a six-year-old child to cook — the child has built a cognitive model of the world, so you can just describe in language how to cook; the monkey has no cognitive foundation. Why is it hard? You have to turn all modalities into tokens, just like text. He mentions that his own model workshop's voiceover model was trained last April by adding voice training on top of a language model, but the language model was too small, so it could only do voiceover, not Q&A. GPT-4o, on the other hand, can do both speech generation and understanding, and its text cognitive ability hasn't declined — it's even stronger than GPT-4 on MMLU.
— Li ZhifeiThere is voice generation, but no video generation
Li Zhifei noticed a detail: GPT-4o's assistant model does not include video generation; whether at Google or OpenAI, video generation is a separate model. He explains the reason is that Sora is not an autoregressive architecture; it probably uses DiT, a Diffusion architecture. He repeatedly emphasizes his belief: a unified model should be based on autoregression, putting understanding and generation together. He guesses that someone inside OpenAI must be working on this, but he's not sure if it's being treated as the all-in direction — "Engineers have their own beliefs too; they feel autoregression is unhappy, they feel Diffusion is happier." He also gives the order-of-magnitude difference in token length: text is a few characters per second, voice is dozens of tokens per second, video might be several hundred tokens per second, each is at least an order of magnitude increase, and token length directly determines the cost and speed of inference.
— Li ZhifeiByteDance's big price cut is inevitable; the API business model is unsustainable
Li Zhifei says ByteDance's big price cut is "inevitably like this, and it will approach zero, infinitely close to zero." He tells a story: last September or October he chatted with a large model entrepreneur who said that besides app revenue, they also had API call revenue, earning quite a bit a year. He advised the person on the spot to just shut it down, because "there's no reason you can still be collecting this money next year." He says this form appeared in the AI 1.0 era — speech recognition, TTS calls, once giants entered they made it free, there were even one-yuan bids, it's just that the economic downturn in the past few years quieted things down. His judgment: for the entire ecosystem and for users, it's a good thing — giants making large models free for everyone to use is great; but for companies whose business model is API, it's not good — you have to rethink what your business model is.
— Li ZhifeiWhen making smartwatches, I didn't think about the ecological niche
Li Zhifei reviews his strategic mistake with smartwatches: making smartwatches in 2015, from a strategic perspective, might have been wrong. Why? Five years later, under full competition, there is no so-called ecological niche — phone manufacturers will all make them, and they can do it well as a side business, because the capabilities are completely common and overlapping with their phones. He says if he had the six-element strategy method at the time, he might have worked harder to think: choosing to make smartwatches, five years later when giants are fully invested, how do I hold a position? That might have led to more differentiation, and the execution path would have been different. He hypothesizes that if on day one he had gone all in on overseas, not domestic, not so many smart hardware products, using the limited time gap to build barriers, the state might have been more successful. The core is "it forces you to think."
— Li ZhifeiZeng Ming said we are on a block of ice
At Hupan University, Li Zhifei raised as a case study the question: "In the AIGC era, three to five years from now, will we AI tools still have an ecological niche?" Zeng Ming's answer was "definitely not," but "I think you've already done the best you can." Zeng Ming's metaphor: today you are in a sea with a particularly unclear situation, you are on a block of ice, at least you're not underwater, you're on the ice; but this ice could melt or flip at any time. So your posture should be low-power operation, while very keenly observing whether there is solid ice around; once there is solid ice, you must be ready and still capable of jumping onto it. Li Zhifei says all large model companies today probably have to adopt this posture — even if you've found a business model and have user volume, you're still on a block of ice.
— Li ZhifeiThe moment I saw the startup dying, I got depressed
Li Zhifei says that in 2019 he truly felt the company would die: there were only three or four hundred million RMB left in the account, but salaries alone cost over three hundred million a year, plus 400 million in inventory, over a thousand people, and 20 offline stores. He recalls that if Google hadn't given money, they would have died too, maybe with only three or four months of money left in the account, he just didn't look; in April 2014, Volkswagen gave 180 million USD; if it had come four months later, they might have gone under. In 2020 the pandemic came, and instead of hitting bottom and rebounding, the bottom was that deep. He says, "I think I got depressed," didn't want to talk to anyone, and it showed as the CEO no longer wanting to work. His response was to continue layoffs, strategically lie flat, and calculate P&L — figuring out the cost allocation for each product line and each small business; those that were losing money, no one came to ask him for people anymore. He gives an example: at the time the company had four IT staff, and moving offices required outsourcing; now one IT person doesn't even hire an outside company for moving.
— Li ZhifeiIn their own words · checked verbatim
There's a saying: a vision without a path and rhythm plan is just耍流氓.
有一句话 没有路径和节奏规划的Vision 都是耍流氓
Li Zhifei12:38
It's inevitably like this, what's the big deal, their own big price cut, inevitably like this, and it will approach zero, infinitely close to zero.
必然就这样子嘛 这有什么 自己的大降价 必然这样子 而且会趋劝于零 无限接近于零
Li Zhifei46:46
Today you are in a sea with a particularly unclear situation, you are on a block of ice, what you're doing now is on a block of ice, but at least you're not underwater, you're on the ice, but this thing of yours could at any time — the ice could melt or flip.
你今天就是在一个局势特别不明朗的大海中 你呢 是在一块浮冰上 就你现在做的事情在一块浮冰上 但至少你不在水下 你在冰上 但是呢 你这个东西随时可能 这个冰就会化掉 或者会翻
Li Zhifei1:00:51
This might be the first time in the last two months that I've seen the sun set.
这可能是我最近两个月以来 第一次看到太阳降落的状态
Li Zhifei1:09:32
You find that entrepreneurs really sometimes — in the last month, they received ten million, and they start thinking, oh, then on average I'll probably make twenty million a month, a year has twelve months, so my revenue for the next year will be two hundred and forty million.
你发现 创业者真的是有的时候 最近这一个月 收了一千万 就开始想 哦 那我平均一个月 估计接下来做两千万 一年 有十二个月 那我就 接下来一年的收入 有两点四个亿
Li Zhifei1:13:32
Many companies, tech companies, enter a negative cycle from day one.
很多公司 就是科技公司 它从第一天就进入了负循环
Li Zhifei1:25:16
Does today's result justify my past efforts? Does it justify the hardships I've endured over these 12 years? I think it does not.
今天这个结果 对得起我过去的努力吗 对得我过去受的 这个12年受的磨难吗 我觉得是对不起的
Li Zhifei1:27:06
Figures
| Mobvoi IPO from initiation to filing | Started in March last year, filed in June, nearly a year | 2:02 |
| Number of IPO intermediaries | Forty or fifty intermediaries (private equity financing usually three or four) | 3:02 |
| Volkswagen's investment in Mobvoi | 180 million USD (April 2014) | 1:11:32 |
| Mobvoi inventory in 2019 | 400 million | 1:11:32 |
| Mobvoi employee count in 2019 | Over 1000 people | 1:11:32 |
| Employee count after layoffs in 2020 | Less than 300 people | 1:16:32 |
| Google model context length | Two million tokens | 26:13 |
Glossary
- VLA / Vision-Language-Action Model
- A model architecture based on a language model, with vision and action modalities added.
- DiT / Diffusion Transformer
- An architecture based on Diffusion, reportedly used by Sora, rather than autoregressive.
- In Context Learning
- The ability of a model to learn a new task via prompt or examples without retraining.
- codec
- Technology that compresses modalities such as sound into token representations.
- P&L / Profit and Loss Statement
- Accounting costs and revenue by product line to determine whether each business is losing money.
How to listen
Founders building large models or AI applications, agonizing over whether to go all in, and investors who want to see the real financial and strategic dilemmas of AI companies.
The IPO process details from 2:00 to 8:36 at the beginning can be fast-forwarded; low information density.