Li Kaifu: For AI 2.0 startups, don't compete with big companies in burning money. Vertically integrating and deeply developing applications is the best solution | WISE 2024 King of Business.

The environment is constantly changing, and the times are always evolving. The "Kings of Business" follow the waves of the times, insist on creating, and seek new driving forces. Based on the current major transformation of the Chinese economy, the WISE2024 Kings of Business Conference aims to discover the truly resilient "Kings of Business" and explore the "right things" in the Chinese business wave.

From November 28 to 29, the two-day 36Kr WISE2024 Kings of Business Conference was grandly held in Beijing. As an all-star event in the Chinese business field, the WISE Conference is already in its twelfth year, witnessing the resilience and potential of Chinese business in an ever-changing era.

This is a year that is somewhat ambiguous and has more changes than stability. Compared to the past decade, everyone's pace is slowing down, and development is more rational. 2024 is also a year to seek new economic momentum, and the new industrial changes have put forward higher requirements for the adaptability of each subject. This year's WISE Conference takes Hard But Right Thing as the theme. In 2024, what is the right thing has become a topic we want to discuss more.

What is the "right thing"? In today's AI industry, disassembling such a question has become particularly complicated.

From 2023 to 2024, we have witnessed the great changes in the AI 2.0 large model track. The Scaling Law, which was once a consensus, has begun to be challenged, and it is still undecided whether the killer application will first appear in the C-end or the B-end. What was right yesterday may have become a misjudgment at present.

Facing such a rapidly changing industry, Kai-Fu Lee may be the AI practitioner who has the most say in "how to do the right thing".

He has many labels: He once served as a global executive of Apple, Microsoft, and Google, an investor who has invested in more than ten Chinese AI unicorns, and now he is also an entrepreneur of the AI large model who has personally entered the game - the founder of the AI 2.0 unicorn 01Wanwu.

Right: CEO of 01Wanwu and Chairman of Innovation Works, Kai-Fu Lee, Left: CEO of 36Kr, Feng Dagang

The following is the conversation between Feng Dagang, CEO of 36Kr, and Kai-Fu Lee, CEO of 01Wanwu and Chairman of Innovation Works:

The right thing is not calculated meticulously, but felt with the heart.

Feng Dagang: When it comes to "the right thing", Teacher Kai-Fu, what kind of right things do you think you have done in the past two years, and what is the logic of doing the right thing?

Kai-Fu Lee: I think the most important thing I have done is to establish 01Wanwu. I dare not say that this unicorn company has been completely successful, but it represents a way to make the right choice.

I firmly believe what Jobs said that life cannot make a permanent plan for the future. But if you look back at everything you have done in the past, they can be strung into a line. If every decision you make follows your heart, then in the end you will find that they are all right and are paving the way for the future.

When I was 20 years old, I chose to devote my life to AI. After that, I built various products based on AI: I returned to China, established the Microsoft Research Asia, and later founded Innovation Works. I have seen the glorious moment of AI 1.0 and also the challenges it faces. So I think everything I am doing at the moment is something I must do.

When making choices for the future, another suggestion is: Don't wait until everything is certain to make a decision. By the time you start, what you see, others will also see.

So, how to make a choice when things are still uncertain? I think history will give us a lot of inspiration. We can see that some rules of the mobile Internet era and the AI 1.0 era will definitely be repeated in today's AI 2.0 large model era. But at the same time, we must also see that the AI 2.0 large model era has its own characteristics, and we cannot stick to the old rules.

We can see how the past PC applications and mobile Internet applications rose, what opportunities and challenges toC and toB respectively face, what kind of path AI companies have gone through, what caused the initial wave, and what caused the challenges later, and whether it will eventually become a great company...

These lessons and learnings are very difficult. I suggest that making the right decision does not actually guarantee that it will be ultimately correct, but it guarantees that you have done your homework, know what you want to do in your heart, and then pursue it without hesitation and never give up.

Feng Dagang: Will you be a bit hesitant in your heart when making a decision? Because I have experienced many people starting their own businesses. Because the success or failure of starting a business is uncertain, this decision is not easy. But at the same time, when you feel that a certain point in the past is strung together and becomes something that has to be done and must be done today, it is actually a huge excitement and joy.

Kai-Fu Lee: It can be understood in this way. When you need to determine whether something is something you must do, you can ask yourself such a question: "If you fail after trying" and "If you don't do it today", which one will make you more regretful?

Several important decisions in my life are made based on this principle, and this time is no exception. My career in the past is not all successful cases, but what makes me very proud is that at every juncture, I have gathered a group of very excellent talents. Later, some of them have made many great companies, such as Kuaishou, and some have become very excellent VCs and senior executives of large companies, such as Jiang Fan. They were all cultivated in that era.

So sometimes success is not necessarily that you have made a company with a market value of hundreds of billions or trillions, but it may be that history gives you such an opportunity, that is, to cultivate such a group of people to make the entire environment better. Of course, there are many other successful works, but I think we must be brave to keep trying, understand our own hearts, and don't regret when one thing is over, and then look for the next opportunity.

Feng Dagang: I personally summarize, not representing Teacher Lee's meaning: The right thing is not calculated meticulously, the right thing is felt with the heart.

I know that Teacher Kai-Fu Lee has gone to Southeast Asia, the Middle East, and many other places in recent months. I want to know what your feelings are. Do people in other countries view AI from a different perspective than us?

Kai-Fu Lee: I think Silicon Valley is still a very important learning goal for us. They are already very certain to be a global leader and a company that everyone pays attention to. Several things I saw this time:

One is that OpenAI will continue to research and develop, but the progress of model iteration will not be as fast as before. Everyone is greatly shocked. It is not that OpenAI has changed many things today, but that there is not only a Scaling law in the pre-training stage in the world, but also a Scaling law in the reasoning stage. Then I also saw that the entire AI 2.0 ecosystem in the United States is developing very fast now. To a certain extent, it is because ChatGPT originated from the United States and swept the world, and at that time only OpenAI had it. Therefore, the entire United States quickly completed the process of integrating large models into various applications. Users have been educated and have established relevant cognitions: Chatbots can not only chat, but also help you do something.

In AI 1.0, China not only actively embraced new technologies, but also led the world in landing applications. In the AI 2.0 era, ChatGPT is outstanding but not open to China. At present, several large model companies in China are doing very well, but China has not yet ushered in its own ChatGPT moment. Many users still do not understand, have not been exposed to, and do not have the habit of using it. This is something we must change in the next year.

Regarding the situation in various parts of the world, it is actually roughly in line with people's expectations. In terms of technical awareness, the United States is in the leading position, followed by China, while the Middle East and Southeast Asian markets still face greater challenges in understanding the overall market.

The arrival of this large model era is different from the development of previous operating systems. The operating system is mainly a technology that has nothing to do with language, culture, law, values, and religion. However, today, large models like ChatGPT have encountered dissatisfaction from users in many non-European and American countries. Because in the training process of these models, although they do not intentionally add specific values or tendencies, but if the data mainly comes from a certain country, such as the United States, then in the eyes of other countries, the answers provided by the model may appear biased, wrong, or discriminatory. For example, asking ChatGPT whether Palestinians should be treated fairly, the answer may not meet the expectations of most people, because too much US data has caused the model to learn a biased view.

Therefore, I think the future trend will be the era of "one country, one model". Each country and region will have its own unique large model, which will reflect the local market characteristics, user values, people's value orientations, and different laws, regulations, and religious beliefs. This is an inevitable trend. Unlike the situation where operating systems can be used universally in the past, it is difficult to achieve global unity of large models.

Feng Dagang: Teacher Lee's insights have greatly inspired me. I previously thought that every country has its own large model, which may be caused by trade disputes. But from a fundamental level, it is difficult to build a model that can unify all cultures because there are significant differences between countries.

Kai-Fu Lee: Yes, China and the United States have different market leaders. Maybe everyone thinks that the Chinese market is large, the technology is strong, and the entrepreneurs are strong. But in fact, I think every rich country, such as Saudi Arabia, and every country with a particularly large user base, such as India and Indonesia, will eventually have its own large model.

This model may be developed by companies from other countries, or it may be fine-tuned based on an open source model. But ChatGPT, especially considering the way American big companies do things, must first satisfy the United States, then Europe, and the others later. "Later" brings the market, allowing entrepreneurs in these countries or entrepreneurs outside the United States to seize such a time window to develop their own large models.

Continuously simply pursuing "bigger models" will only make Nvidia's stock continue to soar.

Feng Dagang: Last year's Scaling law, we called it "Great efforts lead to miracles", and the effect was really good. This point is no longer a consensus this year. It seems to have reached its limit. Do we have a new consensus? What is the new consensus?

Kai-Fu Lee: I think the Scaling law still holds. Now spending 10 times the money can only get some improvements, but the improvements still exist.

The Scaling law has not stopped, but investing money in it is no longer a good business. For an investor or a large company, for example, GPT-4 is trained with 100 million US dollars, GPT-5 is trained with 1 billion US dollars, and GPT-6 may be 10 billion US dollars, and GPT-7 may be 100 billion US dollars.

These are incredible numbers. I think most companies will not take this path. OpenAI will, but I think this is not a path suitable for China. On the one hand, spending a large amount of money to do uncertain things is not the behavior of Chinese large companies in the past, nor is it a path that a startup company can take.

Another point worth noting is that continuing to desperately improve the model parameters to make a bigger model will only continue to drive Nvidia's stock to soar. American large companies are still frantically buying GPUs, but this is not a good thing for the entire ecosystem.

In fact, what we need to think about is how to enable every entrepreneur here to build an AI Native application. To achieve this, the biggest bottleneck today is that the cost is too high. We may have forgotten that when GPT-4 came out more than a year ago, its price was 75 US dollars per million tokens. And today, the Yi-Lightning "Lightning" model made by 01Wanwu is 0.99 RMB per million tokens. And the performance of the Yi-Lightning model is definitely stronger than that of GPT-4 at that time. We even surpassed GPT-4o in May this year, and our price is 1/500 of GPT-4 last year and 1/30 of GPT-4o.

We can be sure that the cost of the entire industry is decreasing. If you are doing AI entrepreneurship today, if you think the AI model is still not good enough or still expensive, I can tell you for sure that you can make a prediction for the future by referring to the changes from last year to this year: Within one and a half years, the price has dropped by 500 times, and at the same time, the model's ability has also been greatly improved. If you still think it is expensive today, there is a 99% chance that it will not be expensive next year. After another year, it may not only not be expensive, but also be able to support the applications you want to do.

So the thinking mode now should no longer be the same as last year. Including me, last year everyone considered based on the Scaling law, to get more GPUs and make bigger models. Now it is clear that the application era has arrived. What should be emphasized now is how to make a very good model very fast and very cheap, so that everyone can use it, so that the era of universal benefit can arrive, and so that the prosperous entrepreneurial ecosystem can be repeated in the AI 2.0 era.

Feng Dagang: The number 500 times is extremely persuasive, almost free.

Kai-Fu Lee: Yes. But we will also make better models, and they will be cheaper next year than this year. Can you imagine that when mobile phones and PCs first came out, we were all very excited. This year, there are better PCs, and we are very happy if they are 30% cheaper than last year. The same is true for mobile phones. But what was just mentioned is a 500-fold difference. If a car, a house, and a mobile phone can be reduced in price like this, what kind of economic benefits do you think it will be.

Feng Dagang: We can also see the strategy of 01Wanwu, which has not changed much. We adhere to the "Trinity" strategy. Do you think the "Trinity" strategy is the right thing in the AI industry?

Kai-Fu Lee: Yes, if we do them separately, one group of people does applications, one group of people does models, and one group of people builds AI Infra, then the final result will not be the most efficient.

If you believe what we just said, our goal is to greatly reduce the cost in order to allow applications to bloom. What we hope to do is the Trinity.

Of course, it is difficult enough for a startup company to do the Trinity. How to do three things at once? How to do it? Our method is vertical integration. How to do vertical integration? When Infra and the model are not yet solidified, optimize both at the same time to get the fastest, cheapest, and most efficient.

We hope to make the large model cheaper. The first question is why is it expensive? Because GPUs are expensive. Can GPU be used less? Then use memory instead. We make Memory cache, remember everything that has been calculated and may be reused, and directly load this part of the cache when necessary to maximize the reuse of previously performed calculations. At the same time, when making the model, we need to consider what architecture, how many GPUs, how much memory, and what kind of Memory cache we are using, so that we can design a reasoning-friendly model based on this.

<