[Geek Park IF 2024] Robin Li: What kind of products and developers are needed in the AI-native era?

Original link

Volume 01 Only native AI applications are valuable, and progress in large models is not an opportunity for most people.

• Many highs and disappointments: Although artificial intelligence technology in the past has made progress in specific application areas (such as playing Go, face recognition), it is difficult to form a universal product due to its limitations and dispersion.
• Turning point: The emergence of large models has changed this situation, with unprecedented versatility and learning capabilities, the so-called "intelligent emergence." This feature enables large models to be quickly applied and generate value in various scenarios.
• Excitement about large models: generality and the potential to create valuable solutions across domains.

The main opportunity in the future is actually the application of AI on top of the model.

• On top of the basic model, there must be tens of thousands or even millions of AI native applications before the value of this large model can be reflected.

In the past year, the main excitement among the media, society, and the public has been on the basic model, and has not transferred to the native application of AI. I am more or less anxious.

By analogy, for example, in the mobile Internet era, there are Android and iOS. In fact, there are only two. Today, I think WeChat and TikTok are no less valuable than iOS or Android. Why don't people take it seriously and spend energy on making native applications, and pay attention to the progress of the model every day. In fact, the progress of large models is not an opportunity for most people. Only a very small number of technicians who are very related to this technology study and track these things. I think it is valuable.

• You should roll it, but you have to roll it in the right place . It is normal for people to roll up. When any new thing or big opportunity comes, everyone will definitely see it step by step, and they will rush to it, and then there will be a process of sweeping away the sand. This is very normal.

• Evaluate progress on large models

  • We should pay attention to its performance in actual application scenarios, and technology must still serve the application. And not just scores and chart rankings. You really have to create value for this society. This is really not something you can just publish a few papers or run an evaluation.
  • Beating the rankings may be beneficial to financing, but financing is also indirect. Investors are willing to give you money because they think you can eventually make money through these things.

• Models should be goal-oriented and application-oriented.

• The evaluation of the quality of a model is the core competitiveness of a modeling company.

• The advancement of model technology depends on its effectiveness and commercial value in specific application scenarios.

• The main line of large model competition is the optimization of inference cost and effect, rather than just relying on great efforts to produce miracles.

  • The saying that great efforts can produce miracles refers more to the exploration process of a large model from 0 to 1. OpenAI used enough computing power and done enough data training when others did not know that this road could work. Finally ran out of this road. Everyone said that they actually did not invent a new algorithm, but used Transformer. In the end, they achieved very good results because they used enough cards.
  • But I feel that going forward, this game will not be played like this, it will not be a miracle produced by great efforts, but more towards the opposite of it. This is the same as the law of all business competitions. Whoever is more efficient will win. You can raise funds, and I can raise funds. In the end, I used 10 yuan to achieve 100-point results, and you spent 10 yuan to achieve 120-point results. Over time, you will win. In other words, in order to achieve a 100-point effect, I spend 100 yuan and you spend 80 yuan, and you win. In the end, the value reflected in the application is the price.

• Enterprises should focus on end-to-end optimization, improve efficiency, and reduce costs to win in the competition.

  • After training, you have to apply it. The application is reasoning. Is your reasoning cost lower than others for the same effect, or is your effect better than others for the same cost? This is the main line of future competition.

02 Large companies represent backward productivity. Don’t look at what big companies are doing.

1. Influence of policy environment and public opinion: The country’s industrial policy and public opinion environment are crucial to the success of large model applications. Policy support and media attention can drive development in this area.
2. Integration of existing enterprises and large models: Focus on how existing enterprises can better utilize large models, especially the positive impact on key business indicators. This aspect has received relatively little attention, but it is an important point in the development of large-scale model applications.
3. Challenges to the response speed of large companies: Large companies often respond slowly, which may represent lagging productivity, so we cannot just look at the practices of large companies. For innovation in large model applications, more participation and exploration are crucial.
4. Embrace the new era and break muscle memory: In order to cope with the challenges of the new era, it is necessary to abandon past practices, give up muscle memory, and constantly reconstruct and redo existing business, especially in the application of large models.

The large model is a foundation. If there are valuable applications on it, even if the industry is successful, it will become bigger and bigger.

How can existing enterprises and existing businesses make better use of the model? What is important is whether our existing enterprise, no matter what it originally does, can have a positive effect on its key business indicators after using a large model. This area has received relatively low attention. I think if this level of attention can be raised, it will also be a very important point for making a large model and making it bigger.

Even companies that have raised funds through large models are increasingly talking about developing applications. I think this is gradually embarking on a healthier track. Whether it is your own model or someone else's model, its value will ultimately be reflected through application.

If it can find a super app on its own, that's great. If you can't find it, I think it can empower others to develop successful native applications based on its model. I do feel that hundreds of basic models are a huge waste of social resources, and more resources should be put into applications in all walks of life, especially when our computing power is still limited. It is hoped that there will be a relatively big change in the understanding of this matter by the whole society.

03 Search itself is also evolving and has the opportunity to become a brand new product

1. Conveying enthusiasm and awareness of AI: In order for the organization to integrate the awareness of AI, it needs to be repeatedly spread in multiple directions, including company meetings, internal live broadcasts, and sermons and discussions with employees. This requires everyone to be engaged and constantly learning, and this iterative process itself is a source of excitement and learning.
2. Transition from "Integration" to "Reconstruction": For the application of large-scale models in business, the key is to reconstruct and redo the business, rather than simply integrating the model into the existing business. This requires that business metrics actually produce positive changes, not just simple access models.
3. The relationship between search and large-scale models: Large-scale models have a direct impact on search. The search function is divided into three parts: "extreme satisfaction", "recommendation stimulation" and "multi-round interaction". This means search will become more intelligent, personalized and interactive. At the same time, large foreign models are also moving closer to search engines, which also makes the future of search uncertain but full of possibilities.
4. Continuous learning and trying: In the application of large-scale AI models, continuous learning and trying are crucial. Every attempt can bring new knowledge to the business, and also help promote the iteration of the model and business, forming a virtuous cycle.

We will hold "thought research meetings" based on the actual business. We will discuss how this technology is related to my business, what muscle memories in the past need to be broken, and what it will look like after being broken... Although I cannot participate in every " Thought Research Conference", but I will read a lot of the summaries they left behind.
I also learned a lot from it, and my own knowledge is constantly iterating. This iterative process is actually very exciting. You always feel that you are learning new things, and you always feel that you understand something that you didn't think of in the past, or didn't think of it this way in the past. Although we have tens of thousands of employees, we all still have something in common in this regard. When you feel that you are constantly learning new things, when you feel that you are constantly seeing new possibilities, everyone becomes more motivated.

04 Using generative AI to transform existing businesses can create greater value

1. The value creation direction of large-scale models: There is no conclusion yet. The value creation of large-scale models may occur in the creation of new "super applications" or the transformation of existing applications. For example, like Microsoft 365 Copilot and Adobe, modifications to existing products may bring huge value.
2. Today, Microsoft 365 Copilot's annual revenue is US$5 billion, which is many times greater than OpenAI's annual revenue. So much new value has been created just by modifying existing products. Everyone still needs to pay more attention to the combination of large models and their existing businesses. Maybe engineers will tell you at the beginning that the effect is not good and that this thing is of no value to us. In fact, no, you have to look carefully and put requirements on the model. Finally, after several rounds of iterations, the effect will come out.
3. Companies need to clarify their strategic boundaries: In the field of large models, companies need to clarify their strategic boundaries and trade-offs, that is, decide what to do and what not to do. This may involve stopping something that has been done in the past, which is a difficult but necessary thing for CEOs to face. Generative AI is such a huge opportunity that it may reshape the entire society, so no one company can seize all the opportunities.
4. Market competition and survival: A company should take advantage of market competition. If it does the best in the market, it will have a reason to survive.

What I always say is: What is strategy? Strategy is about making trade-offs, deciding what to do and what not to do. It's easier to decide what to do than it is to decide what not to do, especially if it's something you've already done in the past. You decided to say that my resources should not be invested in this area anymore. This is a bit invasive, and everyone will feel pain and have emotional problems.
But as a CEO, you have to make these decisions. There must be something to do and something not to do, and we must cut off the things that should not be done.
For outside entrepreneurs and partners, you are actually competing with the entire market. If you are the best in the market, then you have a reason to survive. If you are defeated by any player in the market, you will not survive.

05 Startup companies can create three to five Super Apps and thousands of vertical applications

1. For entrepreneurs, if new technologies easily replace their products, it means that what they are doing lacks unique value and they need to adjust their direction.
2. Big companies will take away most of the dividends in the AI ​​field, but this does not mean that startups have no opportunities. Startups can still develop valuable applications.
3. The definition of AI-Native (AI native) is still being iterated, and it can currently be understood as applications with natural language interaction characteristics. But this is still an open concept that requires long-term exploration.
4. In the AI ​​era, the core of developing AI-Native applications is still solving problems and creating value. However, the requirements and approaches will differ for product managers, R&D personnel, and company organizational capabilities. For example, the ratio of PM to R&D may change, and PM may be more independent in the early testing phase, with less need for R&D involvement.

We also found that the threshold for a pure natural language interactive interface is not low. In the past, everyone was used to swiping up to see new content. Today, you ask them to enter a paragraph before seeing the content. This is a high threshold for many users.
On the one hand, the user needs we want to solve are more complex needs. Not only does it require you to type a paragraph, it even requires multiple rounds of interaction to figure it out. But we can also lower the user threshold by combining natural language interfaces with graphical user interfaces.
Furthermore, we still have to do things that past technologies couldn’t do. In the past, it was called recognition needs, such as face recognition. In fact, most of the requirements for generating classes cannot be achieved by past technologies. It still depends on where it can create greater value. Then the applications it generates are likely to be AI native applications.

06 For successful product managers in the AI ​​era, learning ability is the most important

Product manager traits suitable for the AGI era

• The successful product managers of the future may not be one type of person, but a combination of all types of people. These product managers need to have basic qualities, strong learning ability, a feel for the product and market, and not be afraid of technology. They need to focus on business needs and drive technology development, rather than just relying on existing technology.
  • Learning ability is the most important thing. He is not afraid of these technologies, he will learn them, and he is not satisfied with the utilization of existing technologies, but also requires technology. Although your technology is not able to do this now, I ask you to do it for me, so that I can put my products are made.
  • Technology does iterate too fast. Our best technology is also very immature technology. You must tell me what my business needs are, and I will force these engineers to meet my needs. You haven't made it now, OK, but how long will it take you to make it? A PM who can make such requirements is a qualified PM.

Advice for innovators

• It is recommended that entrepreneurs and developers try new technologies as early as possible, constantly iterate their knowledge, and find a path that suits their own development. This is a long-term opportunity, but early opportunities need to be seized to avoid falling behind the competition. The best technology is the technology that suits you and matches your scenario. Even people who have not done AI may have greater opportunities.
  • Trial and Practice: A lot of trial and practice is important. The current understanding of the application of large models is still very limited, and continuous attempts are needed to explore possible applications and innovations. Most of the possibilities today have not been tried yet. Entrepreneurs and developers must try them. Regardless of whether this road succeeds or fails, it is a valuable experience and lesson. Even if it doesn't work out, you know it doesn't work out. If it works out, it's a big opportunity.
  • Time scale: It is crucial to grasp the time scale in entrepreneurship. Opportunities are long-term, but seizing them early can keep you ahead of the competition.
  • Match technology selection with scenarios: Technology selection should match your own scenarios and needs. You don’t have to use the so-called most advanced technology, but you should choose the technology that suits your scenario.
  • Value Orientation: The value and potential of technology, the key is to have a positive impact on the key indicators of the core business, which is the most important.
  • Development threshold: The development threshold for entering the AI ​​era is not high, and even non-AI practitioners may have great opportunities.

Summary: Try and seize opportunities for native applications

With the advent of the big model era, the real value lies in native applications, and native applications are great opportunities for large manufacturers, small and medium-sized enterprises, and entrepreneurs. I hope everyone will grasp it as soon as possible and try as many as possible. I think you will be able to find a path that suits your own development.