Original video in Chinese.
Key Takeaway
- The ability to write code is the core standard for judging how good a large language model is and determining its future.
- Models that are better at code have higher intelligence, because it requires logic, precision, and an understanding of complex systems.
- Programming is a model’s “compound movement,” which expands its capability boundaries and generalizes it to more domains; it is the necessary path to AGI.
- Models that are better at code have a more developed “cerebellum,” turning abstract intentions into concrete actions and enabling interaction with the digital world.
- The AI coding market has huge potential, can optimize developer costs and reshape software production, and has enormous commercial value.
Want to know which large language model will survive to the end? One standard: the ability to write code.
Models that can’t write code have no future. If a model can’t write code well, you don’t need to pay attention to it anymore.
I mentioned this view in the community a few days ago. I was answering a question from a member at the time. I know it sounds extreme and biased. But this has already become a consensus in the circle. If you don’t believe it, I’ll give you four reasons.
First, the better a model writes code, the higher its intelligence.
Unlike writing an article, writing code requires absolute logic, precise syntax, and an understanding of complex systems.
Natural language actually has a lot of ambiguity—a piece of writing can be done in many ways, and it can still be good. And there are many shortcuts, such as imitation. When we were little, we all memorized a lot of model essays, right?
But code is different. It allows zero tolerance for errors—if one symbol is wrong, the program crashes. It also requires logical consistency, because every line of code is part of the system. Finally, it is highly abstract—it takes a concrete problem, abstracts it into an algorithm, and then concretely turns it into executable code.
So a model that can produce excellent code is definitely not just mimicking; it truly understands abstract concepts and rules.
That is why I have always recommended Claude and Gemini. Even if you don’t program, I still recommend choosing one of these two models, or even using both.
Second, the better a model writes code, the larger its capability circle.
To give an analogy, if you want to build a great body through fitness, you definitely can’t train only one part. You have to train everything, including squats, deadlifts, bench press, cardio, and stretching.
For a model, programming is exactly such a perfect “compound movement.” It is not just a single skill, but forces the model to simultaneously develop and integrate a series of cognitive abilities.
Once these underlying abilities are trained strongly enough, they can be used to handle more complex tasks. For example, they can be generalized and applied to law, finance, scientific research, and other domains.
Then the model’s capability boundaries expand. When those boundaries expand across the board, AGI is achieved.
Third, the better a model writes code, the more developed its cerebellum is.
Today’s models are very different from the models of two years ago.
In my view, models from two years ago were more like passive text generators that could converse in natural language. But they were “locked” in the world of text.
Top models today have code capabilities. So they can run a simple script to analyze data, call complex APIs to control an application, and so on.
In other words, in addition to having a powerful brain, they also have a very developed cerebellum, which can turn abstract intentions into concrete actions that interact with the digital world.
And models with weak code capabilities are like a person who has a brain but an underdeveloped cerebellum. They may “know” what to do, but they can’t precisely control their “limbs,” and their movements are extremely clumsy.
That is why all top AI companies are madly improving their models’ code capabilities. They are not just teaching models to “program”; they are training the models’ “cerebellum,” giving them the core ability to move freely in the digital world.
This is the necessary path to AGI.
Fourth, the better a model writes code, the more valuable it is.
This is simple. Let’s do two quick calculations.
There are roughly 30 million developers worldwide. If we calculate each person’s annual total cost, including salary, benefits, equipment, and so on, at $100,000, then the total labor cost of the entire market is $3 trillion a year.
We’re not even talking about full replacement—just optimizing 20% of that would already be $600 billion a year.
And that still isn’t the biggest market. Global enterprise spending on software and IT services has already reached $5 trillion a year. Imagine if the production, maintenance, and iteration of software were fundamentally reshaped by AI—how much would that market be worth?
That is why companies like Cursor have been able to see their valuations soar so quickly.
To sum up, the better a model writes code, the higher its intelligence, the larger its capability circle, the more developed its cerebellum, and the stronger its money-making ability. Put all of that together, and doesn’t that mean a bright future?
If you use code ability as the standard, when you look at the models on the market or the models in China, you’ll have a much clearer sense of what’s what.
OK, that’s it for this episode. If you want to learn about AI, become a super individual, and find like-minded people, come join our newtype community. See you next time!