Three Signals
- The improvement of the effect of the next-generation large model and the significant reduction in cost
Frankly speaking, the development of large models in China is still in an early stage. Many brainstormed scenarios are limited by the model's effect when implemented. - Large models are highly stochastic, so we can find some "astonishing" examples in a large number of attempts. This can serve as a product demo, but there is still a long way to go from a demo to a product with good experience.
- This is significantly different from Internet software products. When a software product wants to make the demo effect good, the completion degree of the entire product may be 60%. For an AI product demo, the completion degree may only be 10%, and the remaining 50% depends on luck.
- When the effect of a large model is not good enough, a lot of constraints are needed to make the product experience stable. When the effect of the large model is improved, these constraints become useless and can be removed without any problem.
- Model effect is the most important constraining factor for application development, followed by inference cost.
- Under the current high inference cost of models, the business models of many application scenarios are not viable: the value created for users is not enough to cover the inference cost of the model.
- Although major companies are competing on the price of large model APIs, many are smaller models. Sacrificing effect to reduce cost cannot truly solve the problem. In the future, there may also be a situation where money is burned to subsidize APIs. However, if the business model of an application can only survive with subsidies and suffers a large loss without subsidies, sooner or later, there will be problems with cash flow.
- In the current financing environment in China, the idea of relying on financing and burning money to increase DAU and then exploring business models at the application layer is also not viable.
- Only a real reduction in cost, rather than a reduction in price, can make some application business models viable.
- At the same time, because the core idea of reducing inference cost is to reduce computational volume while ensuring the effect. The reduction in computational volume often leads to an improvement in inference speed. So when the inference cost is significantly reduced, the inference speed often also has a good improvement.
- Although everyone agrees on Moore's Law and that the effect of future models will improve and costs will be significantly reduced, we still have to wait for that moment to arrive.
- So the first signal I observe is:
The cost of the next-generation large model in China should be 80% lower than that of this generation of large models. - Significant revenue growth after existing players add AI
The development of PC and mobile Internet in China has given birth to a large number of extremely good products. These products occupy the usage time of users in various industries and hold extremely high value. - All people, whether the bosses of these products or their customers, users, and partners, have reached a consensus on "this wave of large models will be useful." All products have carried out their own attempts to combine products with large models to some extent, either openly or covertly, last year.
- I had already written a product development idea for Tencent Meeting + AI in February last year. By the end of last year, Tencent Meeting had 陆续 released various functions of the AI assistant.
- These products that have already established themselves in the Internet era possess all the elements. They have market positions, industry know-how, consensus on AI, customer relationships, and sufficient good cash flow, etc. They should be the first to reap the value brought by AI.
- Even if we think that some products may fall behind due to factors such as not recognizing AI, not finding the right scenario, or being too weak in innovation ability, but when we look at the entire market, if there is commercial value in AI, then there must be several products that do a good job and harvest the fruits of AI.
- This is a result-oriented signal. If no product has made money with AI, it may be due to limitations in model effect and cost, or not finding the right combination scenario, or the commercial value being too small to generate revenue, etc.
- In short, the second signal I observe is:
Several mature products in PC and mobile Internet see at least a 30% revenue growth after adding AI. - Consensus application layer entrepreneurs emerge
The primary market quickly reaches a consensus on some entrepreneurs. These entrepreneurs have had very successful experiences in the mobile Internet era. After starting their own businesses, although they haven't determined what they want to do yet, they have quickly obtained a large amount of investment. - They don't lack financing resources, so they don't need PR to obtain resources. On the contrary, they face fierce market competition. As long as they reveal what they are doing, 100 teams will copy their things in the market. When everyone has a consensus on AI but there are not many reliable projects, their best strategy is to keep their heads down and work quietly until they can lead the market by six months to a year. Only then will they consider emerging at an appropriate time.
- And when they emerge, it generally means they have really made good things. When they promote their products on a large scale in the market, a large number of potential AI entrepreneurs will start following their direction. These potential entrepreneurs are not the "fast-paced" group from last year. There will be many thoughtful people, and there will be no lack of some underwater experts.
- The entire market will warm up because of the emergence of these entrepreneurs. In the great waves washing away the sand, investors will make more moves and inspire more entrepreneurs to come out.
- So, the third signal I observe is:
These entrepreneurs who have received several rounds of financing and are recognized by the market emerge and promote their products in the market. - I believe that when all three signals turn green, the great age of exploration in the AI application layer will truly begin. This may take one year, or two years, but at the latest, it should not exceed five years.