Dialogue with investor Chen Yu: Before the large-scale model base capacity is improved, it will be difficult for the application to change qualitatively
2024-08-22
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
The development of big models is still in the process of “crossing the river by feeling the stones”. Regardless of whether they choose the B-end (enterprise-oriented) market or the C-end (consumer-oriented) market, big model companies face similar difficulties and do not have a clear and established business model.
In August, Big Model startups reported financing again. Dark Side of the Moon and Zero One Everything reported completing a new round of financing. Dark Side of the Moon completed the latest round of financing of over US$300 million, with the latest valuation reaching US$3.3 billion. Although capital once again showed its favor for AI unicorn companies, in the eyes of investors, Big Model is still "crossing the river by feeling the stones" in its development.
"Everyone knows that big AI models are good, but how can we deploy them internally to maximize their effectiveness?" "Whether the big model chooses the B-end market or the C-end market, there is currently no clear and feasible business model." Chen Yu, a partner of Yunqi Capital, who has been investing in cutting-edge technology, smart driving and other fields for ten years, said bluntly in an interview with The Paper (www.thepapaper.cn).
Chen Yu, Partner at Yunqi Capital
Before joining Yunqi, Chen Yu worked as an engineer at Google and CTO (Chief Technology Officer) of a domestic listed company. Chen Yu had entered the big model industry as early as 2021. That year, he participated in the investment of the big model startup MiniMax and became the latter's angel investor. MiniMax is also one of the current big model star startups. Chen Yu believes that "every company will have its own big model application in the future, but there is no good business profit model at present."
The Paper: We have recently been conducting research on some large-scale startups and have discovered that there are many uncertainties in how startups can achieve sustainable development. On the one hand, the investment is extremely large, and on the other hand, financing is not easy. As an investor, what do you think of these issues?
Chen Yu:Indeed, this year, big model companies seem to have raised money, and the valuations of star startups have reached 20 to 30 billion US dollars. However, the venture capital environment for technology companies is actually very cruel. Capital cannot support unprofitable companies indefinitely. Startups must eventually form their own business models and generate scaled revenue. But the reality is that it is already very difficult for big model startups to achieve scaled revenue, let alone profitability. Regardless of whether they choose the B-end (enterprise-oriented) market or the C-end (consumer-oriented) market, big model companies face similar difficulties and do not have a clear and established business model.
The B-end market actually helps corporate customers deploy private models. Currently, the profit margins of the B-end market are shrinking rapidly. A large model project that could sell for tens of millions last year may only sell for 1 million this year. There are too many open source large models in the market that can be used as shells, and the competition is very fierce. The B-end business involves pre-sales services, contract signing, project implementation (such as model supervision and fine-tuning), and after-sales maintenance, all of which require manpower. Originally, large model companies may have gross profit margins if they sold for tens of millions, but now they don’t even have gross profit margins.
The computing power consumed by large-model companies to train base models is usually in the hundreds of millions, and currently large-model companies have no way to share the training costs through revenue from B-side business.
The Paper: At the moment, for client side, is it not cost-effective to deploy large private models?
Chen Yu:Many companies are reluctant to miss out on the AI wave, but have yet to find a clear path to integrate AI with their own business. The same is true for Party A, whose biggest problem is that they know they need to use big models, but they don’t know how to implement and deploy big models within the company, and how to maximize the effectiveness of big models. In fact, this requires a lot of time to communicate with big model companies.
When enterprises deploy large models privately, data may become a real obstacle. What kind of data is suitable for fine-tuning, how to clean the data, and how to label useful data all require manpower and material resources. The selection of training data and model alignment will greatly affect the effectiveness of the model, so there is no large model that can be applied to all scenarios.
The Paper: The B-end market is not easy to do business in, so what about the C-end market?
Chen Yu:The product applications of large models for the C-end mass market can be roughly divided into two categories: one is emotional companionship conversation robots, such as Xingye under MiniMax and the well-known C.ai; the second category is productivity tools. In fact, all large models are making these products. The problem is that as long as large companies do not charge for a day, it is difficult for other companies to charge, because it essentially has no user stickiness. For large companies with good cash flow, the best strategy is to outlast their competitors, so it will be difficult for domestic startups to charge through productivity tools for the C-end. In the end, there will definitely be companies that cannot raise money and are eliminated.
The Paper: Many large model companies are now developing overseas businesses. What do you think of this phenomenon?
Chen Yu:After a large model reaches a certain valuation, it must answer a question, that is, how to achieve scale and generate revenue. The overseas market is larger, and overseas customers are more willing to pay than domestic ones. Therefore, the companies I am currently focusing on are also those with overseas businesses. The profit model of large models going overseas is not much different from that in the domestic market, but it will face competition from foreign large model products, which is extremely challenging.
The Paper: From last year to this year, everyone’s focus on big models has shifted from technology to applications, but there has been no explosion in applications. What’s the reason?
Chen Yu:From last year to this year, the inference cost of large models has been greatly reduced, but the reason why there are no good application scenarios is still limited by the capabilities of the large models themselves, such as reasoning ability and hallucination problems.
Now everyone feels that big models can be used in many places, but they can't be used anywhere. This is because big models are still subject to hallucinations and other problems before they are adjusted. This is also the reason why it is difficult for us to invest. Although everyone is looking at the direction of AI applications, the problem is that the foundation capabilities of big models are progressing slowly. Before the big model capabilities are broken through, it is difficult for the lower-level applications to have substantial changes.
So don't have too high expectations at the moment, but be optimistic about the future. Technology iteration is very fast. If the level of the big model last year was the level of a junior high school student, it may be the level of a sophomore in high school now. It will continue to learn and improve just like people.
The biggest change in the field of large models this year is cost, which has dropped by at least two orders of magnitude. The cost reduction is the result of market pressure on one hand, and also benefits from technological advancement on the other. Innovation in underlying algorithms has indeed always existed, leading to cost reduction.
We must be patient with the development of technology. The experience of big model companies is only one or two years, and now everyone is just crossing the river by feeling the stones. However, the development trend of technology is ultimately the democratization of technology, which makes the threshold for everyone to use it lower. So wait a little longer, I think every enterprise will have its own big model application in the future. Today, it may cost tens of millions for an enterprise to deploy its private model, but in another ten years, it will no longer need to spend such a high cost. I think that in the future, the private big model deployment of enterprises may be more tool-based and product-based.
As technology matures, it will become more standardized and the cost will be lower. All technological developments follow this path. The high cost is only a short-term phenomenon. When I shared with our LPs (investors) in November last year, I said that the cost of large models would drop by two orders of magnitude in the next three years. I didn’t expect that this would be achieved six months later. In fact, the world is changing faster than any of us can imagine.
The Paper reporter Yu Yan
(This article is from The Paper. For more original information, please download the "The Paper" APP)
