🤺Less Than 20 Chinese LLMs Will Make it to 2024, Says Didi and Xiaohongshu's Early Investor
Zhu Xiaohu, an early investor of Didi Chuxing, Xiaohongshu, and Eleme, told Phoenix Finance that only 10-20 LLMs in China will make it to the semifinals in 2024.
In June 2023, Zhu Xiaohu, a prominent early investor in Didi, Xiaohongshu, and Eleme, complained on his social media that “ChatGPT is not friendly to startups. Please give up on seeking funding for the next 2-3 years.”
Fu Sheng, CEO of Cheetah Mobile, countered, “Half of the startups in Silicon Valley are now revolving around ChatGPT, yet our investors still remain blissfully unaware.”
Zhu believes that GPT will take 99% of the value, limiting opportunities for startups, while Fu argued that significant companies have consistently emerged in past tech shifts. This debate sparked widespread discussion within the Chinese AI community.
In a recent Phoenix Finance interview, Zhu shared his perspectives on generative AI and LLMs. While I hold different opinions on certain aspects, such as his underestimation of AI-to-C value and his intense focus on industry-specific LLM applications, the interview provided valuable insights.
China's market now has over 200 Large Language Models (LLMs), a situation more hyped than previous tech trends like group buying and bike-sharing.
In 2024, only about 10-20 LLMs in China might make it to the semifinals.
Advances in LLMs, like a potential GPT-5, could significantly enhance their intelligence and commercial viability.
Zhu advises startup founders to find fault-tolerant verticals where minor errors are acceptable.
Below is the full translation of this interview (the part irreverent to AI was not included; you can find the original interview here, which was conducted before OpenAI’s drama).
China now has over 200 LLMs, which is more hyped than the past “battle of hundreds of group buying companies”
Phoenix Finance: The debate you had with Fu Sheng about LLMs recently prompted a lot of discussion, and you both reconciled on the show. What common ground have you reached?
Zhu Xiaohu: Actually, our views are fairly consistent. LLMs are not particularly suitable for startup entrepreneurs; instead, they should seek opportunities in specific vertical applications.
Phoenix Finance: However, some entrepreneurs feel that there are more people creating products for these vertical applications than there are consumers. Is this a misidentification of demand?
Zhu Xiaohu: Precisely. The current landscape in China, with its 200+ LLMs, is even more hyped than during the “battle of hundreds of group buying companies.” Back then, we only saw dozens of group-buying companies and a handful of bike-sharing ones, but now the competition among LLMs is even fiercer. Indeed, creators outnumber consumers.
Identifying viable vertical application scenarios is not so straightforward at present. Why? Because today’s LLMs aren’t sufficiently intelligent. They’re seen as intriguing novelties in many vertical scenarios but don’t yet achieve commercial-grade quality. At this stage, users may sample out of curiosity, but for sustained usage, the models need further development.
Phoenix Finance: OpenAI’s user growth is showing signs of decline; could this be a process of disillusionment? Is OpenAI potentially not as intelligent as we’ve imagined?
Zhu Xiaohu: At the current GPT-4, ChatGPT’s intelligence for commercial scenarios is quite limited. However, we’re uncertain if it is being trained further. There’s speculation that GPT-5 could enhance its intelligence tenfold, potentially achieving commercial viability in various applications.
Thus, the primary risk for investors and entrepreneurs is that the evolution of LLMs is unpredictable and could be revolutionary, unlike the iterative iPhone upgrades. If a model advances from version 4 to 5, it could experience a more than tenfold increase in capability. Such a leap could disrupt current efforts and invalidate extensive work, as seen when OpenAI’s DALL-E 3 posed a significant threat to Midjourney by nullifying its image optimization efforts due to DALL-E’s superior multimodal outcomes.
Phoenix Finance: With such rapid development, is there a viable way to keep pace?
Zhu Xiaohu: For entrepreneurs, the essence is to capture user scenarios, clients, and the data loop. You evolve as the LLMs evolve. If a model improves tenfold by next year, you leverage the new model to provide enhanced services to customers.
Since no LLM can cater to every scenario or client, the focus should be on securing users, scenarios, and the data loop first.
Next year, the battle of LLMs in China might enter the semifinal stage.
Phoenix Finance: With so many models competing over the past half year, are there any feasible scenarios that have emerged?
Zhu Xiaohu: Certainly. In fact, Chinese entrepreneurs are extremely powerful in terms of application-level innovation. For instance, a few companies we've invested in have seen remarkable growth. One company, which had a monthly revenue of only several hundred thousand last year, has now surged to nearly 10 million per month. This is particularly true in the sales and marketing sector. The deployment in sales and marketing is very down-to-earth.
Because today's LLMs may still have defects in intelligence and might also generate illusions, entrepreneurs must find highly fault-tolerant scenarios. A few minor errors or slight illusions do not impact the final outcome in such scenarios, making it relatively easy to implement.
Currently, the cost-effectiveness of LLMs in sales and marketing is very clear. Customers are willing to pay only if the product offers a performance improvement of more than ten times. In the past few years, the growth of enterprise software in China has been very tough. But today, we have proved that Chinese companies are not unwilling to pay for software; they just won't pay for minor value creation.
If your software can save costs by 20%, 50%, or even 100%, customers may still be reluctant to pay. But once it achieves a tenfold performance improvement or 1000% cost savings, they are very willing to pay.
So, our assessment of whether an AI company meets market needs is whether it can secure customer contracts. If it does, it shows that the product truly addresses customer pain points and creates sufficient cost savings.
Phoenix Finance: When you talk about signing contracts, you refer to enterprise services. But you recently said that in the next two or three years, enterprise services should not expect fundraising opportunities.
Zhu Xiaohu: Yes, traditional enterprise services are very difficult to fund today. However, enterprise software empowered by AIGC has new opportunities.
Phoenix Finance: So, AIGC is a powerful enabler for enterprise services, though we ordinary people may not feel it.
Zhu Xiaohu: Right, because AI-to-C scenarios are not that smart yet. To be honest, most of the traffic generated on the C-end today is somewhat marginal and somewhat non-compliant. On the To B side, however, I see real explosive growth.
Phoenix Finance: But OpenAI's position in the U.S. doesn't seem so solid. Why can the U.S. fund so many LLMs on such a large scale? Given that LLMs have high demands for data, scenarios, and algorithms, where do these funds and data come from?
Zhu Xiaohu: They don't really have that many. There are probably no more than 10 true underlying models. China has over 200, so there is a big difference.
But this really needs long-term investment. Especially today, reaching GPT-3.5 or GPT-4 might require tens to hundreds of millions of dollars. Moving forward, it will require a financial scale of up to one billion dollars.
If GPT-5 is ten times smarter than GPT-4 and requires one billion dollars in funding, not everyone can afford it. So next year, out of these over 200 models in China, only about 10-20 LLMs might make it to the semifinals.
Phoenix Finance: Are the semifinals only next year? Some have already concluded that only BAT (Baidu, Alibaba, Tencent), ByteDance, and Huawei have a chance to stand out.
Zhu Xiaohu: No, these are just predictions. Relatively speaking, indeed only these large companies, with their financial scale, can afford to invest, for example, one billion dollars, like in the 486 or 586 (central processing units in computers). If a startup company's later rounds of financing can't keep up, then they definitely wouldn't dare to invest in 486 or 586.
Phoenix Finance: So, you're saying startups still have a chance?
Zhu Xiaohu: Right, they must find the right vertical landing scenarios. In any vertical scenario, we see that the cost of training a vertical model is not high.
This year, we have a company that can make tens of millions in revenue, and training a vertical model might only require around 100 GPU cards. This cost is relatively controllable.
Phoenix Finance: But you actually invested in quite a few AIGC projects this year.
Zhu Xiaohu: Yes, we invest in vertical application scenarios.
Phoenix Finance: Very specific verticals.
Zhu Xiaohu: Right, all vertically grounded. They must immediately meet with customers, be validated by customers, and be willing to sign contracts and pay.
OpenAI’s potential for creating economic value for humanity may be 10 times greater than Google's.
Phoenix Finance: OpenAI’s founders are now seeking hundreds of billions of dollars in financing from the Middle East. Do they really need that much money?
Zhu Xiaohu: So it all depends on the story they tell next. If they talk about GPT-5, with its intelligence magnified tenfold, increasing programming efficiency by over 90%, that would be a disruptive change. Because as I said before, GPT-4 cannot achieve commercial quality in many scenarios.
For example, in coding and programming, if it only achieves a 50% increase in efficiency, that saves the work of junior programmers, but it requires a lot more senior programmers to check for errors. Many companies wouldn’t want to use it because senior programmers are harder to find.
Phoenix Finance: And more expensive.
Zhu Xiaohu: Exactly. But if efficiency increases by 90%, companies will be willing to use it. It could eliminate most junior programmers, requiring only a few senior programmers to check for errors. Many companies have done the math. The market would open up suddenly.
Phoenix Finance: OpenAI is telling a story of superintelligence. Isn’t this story a bit premature?
Zhu Xiaohu: That’s why we have to look at GPT-5. Today’s GPT-4 certainly can’t do it. If GPT-5 can indeed magnify intelligence tenfold and work wonders, especially in high-difficulty, high-intelligence tasks like coding, and increase efficiency by over 90%, that would be very impressive.
Phoenix Finance: This year, ChatGPT has indeed kept many big tech leaders awake at night, with people like Zhang Yiming reading papers in the wee hours. Is this a significant platform opportunity?
Zhu Xiaohu: Everyone can now see the possibility of a path forward. Previously, because of multimodal capabilities and the emergence of big models like Transformers (deep learning models), everyone only theoretically felt the possibility but didn’t know if investing billions of dollars would yield results, so many were hesitant to invest.
But OpenAI and Microsoft, through their investments, have seen the effects, and now everyone believes in it. If miracles can be achieved by throwing money at it, then many big companies are willing to do it.
Phoenix Finance: In fact, American companies have already blazed a trail for us.
Zhu Xiaohu: Right.
Phoenix Finance: Lu Qi’s widely circulated article on LLMs this year mentioned that OpenAI’s future scale could be ten times that of Google. Do you think this statement is exaggerated?
Zhu Xiaohu: It depends on how you define it. It’s still not very certain whether OpenAI will become an independent company, but if it empowers big players by improving programming efficiency by over 90%, then it could be a massive boost for Microsoft and GitHub.
It could indeed increase the human economy by more than ten times, but this economy might be reflected in NVIDIA, Microsoft, and many of the existing big companies. Whether a new startup can emerge that’s ten times bigger than Google, we feel that’s still quite challenging. At least for now, we don’t see that possibility.
Phoenix Finance: There is still skepticism about OpenAI, as many don’t know where its profit points are. But actually, it could easily make money from advertising.
Zhu Xiaohu: No, it doesn’t need to sell ads. It makes money from usage, like electricity and water.
Phoenix Finance: Infrastructure.
Zhu Xiaohu: Yes, like infrastructure.
Phoenix Finance: So profit is not actually at the core of the skepticism.
Zhu Xiaohu: Right, profit is not the issue. The core is whether it can penetrate more scenarios, achieve commercial quality, and land in more scenarios.
Phoenix Finance: Then, retain more users.
Zhu Xiaohu: Right, as long as it reaches that commercial quality, users will definitely come.
Zhu Xiaohu: Today is just the era of AIGC’s iPhone 1 or iPhone 2.
Phoenix Finance: The unicorns you invested in at that time, looking back now, are actually platform opportunities. But they didn’t start with the intention of being platforms. Is there a kind of randomness to this? Is it particularly hard to find platform opportunities now?
Zhu Xiaohu: Yes, it is indeed more difficult to find platform opportunities today. That’s because there are no new traffic dividends. The internet overall still needs a traffic dividend, which made it particularly easy to rise in the early days, and more scalable. So that was a dividend of an era, including Didi, Eleme, and Xiaohongshu, which all benefited from the massive era dividend of mobile internet.
Phoenix Finance: In 2016, Wang Xing made a judgment about the internet market, saying that it had entered the second half of the game. But now the second half of the internet has been so prolonged. It’s been over 7 years, and it seems like there are no new companies emerging.
Zhu Xiaohu: So why is everyone so interested in AIGC? It’s because they are hoping AI can create a new platform opportunity.
Today we only see that in B2B applications and in enterprise services, we might find some landing scenarios earlier. For B2C, we feel we might need to wait a bit longer. Today is just the era of AIGC’s iPhone 1 or iPhone 2, and it might take another year or two to reach the era of iPhone 3. If the intelligence of GPT-5 can be magnified by ten times, then it might usher in a new B2C opportunity.
Phoenix Finance: What variables have the emergence of AI brought to the internet market? When we look at a company now, will we add some new core points to it, saying this thing is more important?
Zhu Xiaohu: The core is still the data loop. Only with a data loop can you build your own core barriers and continuously improve your performance.
Phoenix Finance: Is the data loop referring to a scenario loop?
Zhu Xiaohu: It’s about the usage and feedback from your users in this scenario. You need to be able to capture user feedback. For example, Midjourney can capture user feedback. It can generate four images and see which one the user will choose, so that it can continuously improve its model.