When AI starts to attempt reverse fine-tuning of humans, how should we harness this new intelligence?

"Everyone is asking me if I should quickly study lobsters. Actually, I think most people are using it wrongly."

"Product managers in the AI field shouldn't study lobsters. They should study the entire system of Anthropic. I think this is a compulsory course for every product manager."

"If you directly ask AI whether the product you've made is good and what suggestions there are for improvement in the document, it can only give you a bunch of reasonable but useless nonsense. How can you instantly call on the world's top business minds, like Steve Jobs and Elon Musk?"

"In the middle of the night, I was so scared that I almost jumped out of bed because AI almost predicted every question I was going to ask."

...

In the 9th episode of "Shan You's Exploration Stream", Professor Li Shanyou invited Zhang Fan, the founder of Yuanli Intelligence. As the former COO of Zhipu AI, he witnessed the battle of hundreds of models and has in - depth and essential thinking about the commercial implementation of large models. This conversation rips apart all the appearances and puts forward a highly impactful judgment: Most people's current way of using AI essentially still stays at the level of "watching fireworks".

The problem never lies in the tools, but in cognition. The real dividing line is never "whether you can use AI", but rather: Can you recognize the essence of the model, define the boundaries of the problem, and force out the "optimal solution" that belongs to you from the infinite solution space?

If you're anxious about the uncertainties brought by AI, or you've vaguely realized that "just learning the tools is not enough", then this conversation might help you achieve a cognitive leap.

The following are the selected contents of the podcast.

True experts are "controlling" the models

Li Shanyou: Now many people are struggling with whether to install "Lobster". Actually, it's just like in 2023 when everyone rushed to learn to write prompts, use Sora, learn about Kouzi, and learn about Manus... Looking back now, these are all just appearances and are changing too fast.

Zhang Fan: The current market changes are generally in line with our judgment from last March and April, but now the situation is a bit like "fireworks". So the next proposition will be the controllability of the "firework" ability of the models. Those who really understand models well, like Anthropic, almost aim to ensure that intelligence doesn't get out of control. It's not about watching it do flips, but about really helping us with complex business. We've vaguely found some means to leverage control.

Li Shanyou: The point you just made is very interesting. Can you elaborate?

Zhang Fan: Take writing prompts as an example. Beginners like to write very long prompts, thinking that the more information, the better. Actually, this undermines the model's generalization ability. If you impose too many restrictions, it won't perform well; if you give too little, it won't achieve the goal. The key is to find that boundary.

The output of a model on a problem is essentially a normal distribution. The closer to the center, the higher the probability and the more general it is, but the result is often mediocre. For example, if you ask "What tastes the best?", it will give you the most widely - recognized answer in the world, but it definitely won't be outstanding. The truly good solutions must be at the edges. The essence of writing a prompt is to change the shape of this distribution, push the edges to the center, and thus obtain a better solution. So how can we leverage it?

For example, suppose you want to create a product plan. If you directly ask "Please help me evaluate whether this is a good product and what suggestions there are for improvement in the document?", I can clearly say that it can only give you a mediocre solution, all reasonable but useless nonsense.

So how can we activate the edge solutions?

I'll first ask "Please tell me who the world's top product managers are?" It will mention Steve Jobs, Elon Musk, etc.

Then I'll ask, "Assume you are Steve Jobs. Summarize the 5 core methods of his problem - thinking." At this step, the model will move "Steve Jobs" from the edge to the center and give several solutions from Steve Jobs' perspective.

Finally, I'll say, "Please judge my plan from Steve Jobs' perspective using these methods."

Then I'll do the same thing with Elon Musk. One emphasizes the intersection of humanities and technology, and the other emphasizes the first - principles thinking. These are all edge solutions. When we summarize different edge solutions and extreme suggestions from different perspectives, we'll get a more effective judgment.

This is to tap into the model's original ability. In this era, thinking itself has become cheap. How to control the direction of thinking is what everyone should do today. If we can standardize and productize this, it will be completely different.

Going "vertical" is the moat

Zhang Fan: In the future, people won't think about how to write specifications, but about how to guide the model to self - compete, self - learn, self - verify, and self - compete. Then knowledge will emerge continuously. Our company is working on this, trying to turn it into a standardized product, allowing people to control the model's self - evolution with a lower threshold.

Li Shanyou: To make the model self - evolve, will you use other people's models or your own?

Zhang Fan: I'll use other people's models. You can think of the base model as human DNA, which is the optimal solution obtained through countless generations of natural selection. To create civilization and explore knowledge, you don't need to start from DNA. You just need to build a university.

Li Shanyou: How do you make it grow on its own? Can only model companies adjust the weights?

Zhang Fan: I think adjusting weights will become a very low - threshold thing in the future. In the past, adjusting a general model required extremely high computing power and a large amount of data, and only model companies could do it. But today, we don't need a general model. We can completely add our own preferences in the vertical field. It's like recruiting employees. You don't need to create a person from cells.

Li Shanyou: Do you mean that vertical models can adjust weights?

Zhang Fan: Yes. That's how humans are. Genes determine a general weight, so they have common sense. But whether they are born in China or the United States, and whether they study engineering or medicine, they are constantly changing their own weights.

The optimal solution can't be guessed through a general model. It must be constructed in our own environment. I think this is the key in this era, which is different from the previous logic.

Today, all AI base models are doing entropy reduction and control. For example, Claude's MCP, skill, cowork, etc., are all redefining the vertical scope. I think we will definitely enter the era of weight engineering in the future. Everyone can define their own environment at a very low cost and train the optimal solution in their own environment on the base model.

I think the core competitiveness of an enterprise today is not to train a model, but to define its own vertical knowledge, MCP, skills, etc. Many enterprises are wrong in spending a lot of energy on infrastructure now.

Li Shanyou: Your insight is very good. Where do you place your core competitiveness?

Zhang Fan: The base model solves the general ability, but there is still a long way to go from the general ability to a truly efficient "employee" in an enterprise.

I think what's missing here is a "university", that is, how to "specialize" this "employee". For example, some study business, some study code. Different enterprises have different edge directions, but 80% or even 90% of their edge directions may be the same. I can put these 80% of the common parts together, and enterprises don't need to start from building a "university".

What we provide is the best industry practice - for example, what 20 strategies a salesperson has and how to evaluate them. At the enterprise level, whether you sell hats or pants, online or offline, through live - streaming or on the shelves, that's your own preference.

Li Shanyou: How do you set the fees?

Zhang Fan: There are usually three charging models in the market. The first one is charging by seat, following the software logic. I don't think it works. It's still the thinking of the previous era. Some people say that if you spend 10,000 yuan to hire a person, and I charge you 2,000 yuan to create an intelligent agent for you, it's cheaper, right? I don't think this logic works either. The competition in this era is between intelligent agents. Pricing an excavator based on how many people it can replace is unreasonable.

The second one is paying by results. For example, you pay me for each item I help you sell. The result is likely that you blame me for not selling well, and I blame you for having bad goods. Finally, we'll suspect each other.

I think the most reasonable logic is to pay by energy. Making intelligence as metered as electricity and water, charging by unit, is the future.

But it will make people think that intelligence has become extremely cheap. What enterprises need to do is to adapt it well, exert your creativity on the basis of its cheapness, and connect it to the physical world to bring benefits. From our perspective, we just want to accelerate the adaptation process and make it connect to the physical world faster.

The three major schools of current AI applications

Zhang Fan: Now intelligent agents have been divided into schools. There are currently three major schools of AI product - based applications, corresponding to three different philosophies.

The first school is OpenAI, which is called the romanticism of AI. All its products have extremely high imagination thresholds. When you look at each one, you feel that it can outperform everything. For example, Sora is supposed to replace Douyin, and Atlas is supposed to replace Chrome. They are all like huge fireworks, but the retention rate is extremely low.

OpenAI's philosophy is to sell the imagination of AI. It doesn't let you use the products, but makes you feel that AI is infinite, thus promoting the "Stargate" and greater investment. Sam Altman essentially has an investor background, so he sees the state of the entire market ten years later.

Another product philosophy is Google's practical school of AI. The products even seem ordinary, but the retention rate keeps rising. The most typical one is Notebook LM.

At first, I thought it was just a knowledge base. This kind of thing has been over - used in China. But the more I use it, the more I find that it's not essentially a knowledge base, but a management pipeline for model knowledge - quickly supplementing information through Deep Search, processing it into podcasts, documents, PPTs, etc. You'll find that it's a pipeline assembly line for knowledge production. You'll find that the more you use it, the better it gets, and you'll become more and more dependent on it. So I think Google's underlying principle is productivity.

There is an even more powerful player, which is Anthropic, called the geekism or the originalism of AI.

Regardless of the first two types, they are all making products for humans, while Anthropic is making products for agents. I remember Elon Musk said something very shocking some time ago. He said that within the next six months, there won't be today's C - language in AI coding. It should be written directly from binary.

This logic is very first - principles. C - language is essentially for humans to read. Why does it need to be for humans to read? Because machines couldn't do it before, so engineers needed to make fine - tuning in a language that humans could understand. Now, with direct binary, there is no loss from the compiler layer, and the efficiency is higher.

Li Shanyou: Then a large number of programmers will really lose their jobs.

Zhang Fan: Definitely. We're not judging whether it's good or bad. We're just saying that this will definitely happen. When Sora first appeared, we knew that Seedance would follow. Don't expect any turns in this era. There are no more turns.

Why is AI more like "electricity"?

Li Shanyou: Regarding the basic models, do you think the iteration speed of these basic models is accelerating or not?

Zhang Fan: It's accelerating.

Li Shanyou: But why doesn't it give people a sense of amazement anymore?

Zhang Fan: There is a logic here. The greater the contrast between a thing and human cognition, the more likely people are to be amazed. But when people are amazed, this thing is often useless. When ChatGPT - 3 first came out, everyone was amazed. You never thought that a model could talk to you, but at that time, it couldn't really do anything. It was just for novelty.

When ChatGPT - 5.2 came out, most people were disappointed, but I still think there are significant changes. When we no longer feel amazed by a technology, it might just mean that it has entered the deep - water area of application.

AI is more like electricity. Now no one mentions "using electricity" anymore, but electricity has reshaped all businesses. The same is true for AI. AI can't exert its value independently. It must be embedded in the physical world and reshape the links of the physical world to exert its value. So I think when AI seems to disappear, that's when its real value appears.

Li Shanyou: Is there a big change in the application level? For example, what do you think of the emergence of Lobster?

Zhang Fan: Lobster is a bit like when DeepSeek was released during the Chinese New Year. All enterprises were crazy about buying all - in - one machines. But as for what to do next, no one knew.

People using Lobster today may feel that they haven't been left behind by the times. But specifically, has it helped you solve an important current problem? Is there any difference between using it to generate daily reports or analyze stocks and directly giving the requirements to the model? If there is no difference, it's just a form of show - off.

Li Shanyou: So do you think Lobster is still a "firework"?

Zhang Fan: Lobster has a good architectural design and a good product feel, but it's still a long way from being practical. For ordinary users, repeatedly adjusting Lobster won't bring any substantial help in mastering the model's ability. Just like learning prompts and playing with Kouzi in the early days, looking back today, it was still useless, and the more you learned, the more anxious you became. So I think we still need to learn first - principles.

A compulsory course for every AI product manager

Li Shanyou: Do you think the gap between Chinese and American models is widening or narrowing?

Zhang Fan: Personally, I think the gap is not widening. It seems that the United States can't build a high - enough barrier in the field of models, so China's trend of catching up is quite certain. And from the application perspective, if you create a sales intelligent agent, the effect may only differ by 5% whether you use a domestic model or an overseas model, but the cost may differ by 10 times. To some extent, domestic models are also doing quite well.

To some extent, I think the model business doesn't have network effects. So you can see that OpenAI can't hold on to its barrier, and Google has caught up.

Li Shanyou: Why can Google catch up?

Zhang Fan: Google has no weaknesses in the entire chain from chips to clusters, to base models, to applications, and to the ecosystem. These chains have network effects, but the model itself doesn't. So I think Google will definitely be one of the greatest companies in the future. It has transformed the AI war from a model war to an ecosystem war, and the ecosystem war affects your personal private weights and context precipitation, which is almost an insoluble problem.

Li Shanyou: How did Google wake up?

Zhang Fan: Actually, we don't have a clear answer. But I think the first reason is the return of the founder. In the AI era, you'll find that professional managers can't do high - risk, high - return things. Only the founder can withstand anti - consensus and bear great risks in the short term.

Second, OpenAI has used up its old advantages.

Third, there is no non - competition clause in Silicon Valley, and knowledge flows at a high speed, so no one can be three months ahead of others. What remains to compete is cash flow, resources, cards.