22 Insights on AI in 2025

1

After the official release of GPT - 4, it caused a global sensation. For me, it was the first time I realized that large models had the ability to surpass human rational thinking. Immediately afterwards, I published several review articles in one go, launched new courses, and developed new products.

Two years later, however, Altman led the way in writing a letter to the US government, calling for restrictions on DeepSeek. Since the palace coup, the OpenAI team has hardly launched any remarkable major achievements. Currently, the open - source community generally believes that open - source large models have substantially surpassed closed - source large models in multiple dimensions.

The Way of Heaven is subtle, while human desires are perilous. The more successful one is, the more one should be vigilant against the hidden risks in the "human nature system" of the management team. The development of the OpenAI team can truly be described as "succeeding because of Altman and failing because of Altman."

2

Sometimes I really admire engineers.

For cognitive scientists, models are more like abstract architectures. Whether it's Simon or Hinton, the models they proposed are mostly simple programs, not even real programs. Sometimes, the models constructed by cognitive scientists are just abstract diagrams.

However, engineers have put these models into practice and developed a unique ecosystem and culture.

In my speech "Standing at the Starting Point of 3000 Years of Prosperity" in 2024, I mentioned that the future era will belong to modelers.

The work results of scientists are experiments and papers, while the work results of modelers are models and code.

In the future, organizations that do not have the ability to produce models and code will gradually fall behind and eventually be eliminated.

Similarly, individuals who cannot understand models and code will be like those who couldn't understand scientific experiments and papers in the 20th century, gradually becoming marginalized in research work.

3

From the perspective of this major AI revolution, the most prominent change is that models have become a production factor.

In my New Year's Eve speech "The Tao and Tools of Life Development" in 2023, there was a golden line: Conversation is production. This is what makes the new era different from the information age: the information age emphasized "the medium is the message," while the new era is about "conversation is production."

And today I want to say that the model is the production factor. Its format, architecture, encapsulation, modification, generation, and application all together constitute a crucial proposition for the future.

The most special thing about "the model is the production factor" for human society is that what we think, do, and even our past creations can be solidified into models and called and used by more people.

The example Hinton gave in an interview was very wonderful: It's difficult for us to understand a person's internal biases, but we can relatively easily quantify the biases of a model and make adjustments accordingly.

This is the uniqueness of this major AI revolution.

4

With the advent of the AI era, there are three professions that are still difficult to be replaced: teachers, consultants, and doctors, physiotherapists, and masseurs.

These three professions are interesting because they stem from the inherent limitations of humans.

Teachers: Since human behavior is reflexive, teachers are needed to create an atmosphere and stimulate the internal driving force of learners.

Consultants: The human inner world is extremely complex, with a large amount of ineffable and even embarrassing hidden content. This content is often difficult to express completely in language and is more conveyed through body language and overall feelings.

Doctors, physiotherapists, and masseurs: This stems from the handling of the human body. The human body is a complex system, and the tolerance for treatment errors is extremely low. Therefore, experienced professionals are particularly important.

5

Before the AI era, things might have been like this:

If you couldn't design, you'd hire a designer; if you couldn't code, you'd hire three engineers: front - end, back - end, and algorithm. To manage everyone, you'd hire a product manager. To sell the product, you'd hire three operators, responsible for content, community, and channels respectively, and you'd also need to hire an operations manager to manage them.

But now, the trend has changed and it's becoming like this:

Can't do this? Then try if the first large AI model can handle it; can't do that? Then try if the second large AI model can do it. To further improve efficiency, you can let a third large AI model write an agent program to coordinate their operations. If none of the large models on the market can meet your specific needs, then you can do it yourself and use large AI models to train a new AI model, and even part of the training data can be generated by AI.

In the next three years, all kinds of AI - based software and scientific research results will experience an explosive growth.

This will have a huge impact on the social form and also bring some dividends.

6

Knowledge can be divided into two categories: before 2023 and after 2023.

Knowledge before 2023 is knowledge that has not been contaminated by AI.

Knowledge after 2023 has almost all been contaminated by AI. Today, I saw a news article that made me laugh and cry: A professional media outlet reporting on AI wrote an article about an investor who graduated from Zhejiang University to follow the trend, but they wrote that Zhang Lei graduated from Zhejiang University as an undergraduate... As we all know, he graduated from Renmin University of China. Why did the journalist make such a low - level error? Because that list was obviously generated by AI.

For anyone who collaborates with me, I use one indicator to judge whether they are reliable: Was the first version of the manuscript directly generated by AI?

If they did so, I hardly collaborate with them on writing anymore.

Giving up writing the first version of the manuscript by oneself actually means giving up the ability of independent thinking.

To borrow my previous analogy: The difficulty of writing lies in: first "getting pregnant", then waiting for the child to be born and grow up, and only then can we talk about taking photos of it and using beauty filters. Many people mistakenly think that AI can not only complete the "getting pregnant" but also help the child grow up. This idea is really absurd.

This is tantamount to giving up the possibility of continuously improving one's own abilities.

7

For those outside the AI field, the most difficult part of understanding large models is actually a shift in thinking mode, from thought experiments and experimental science to symbolic thinking and computational simulation.

8

After two years of AI development, my previous predictions were basically correct: Great progress has been made in the fields of AI Coding and AI Science, creating several new trillion - dollar markets.

But to be honest, although I had predicted the emergence of some new trends and technologies, some achievements still greatly exceeded my expectations and gave me a sense of surprise and amazement.

AI Coding is more like a meta - method and meta - skill in the digital world. Similar to the human world, we all have to master listening, speaking, reading, and writing to communicate well.

Similarly, in the digital world, we all have to master the ability to CRUD entities. In the era of the Internet + mobile Internet, a programmer could CRUD at most 30 entities a day, which was the limit, an operation on objects in the digital world. And this ability was limited to programmers.

The transformation brought about by AI Coding is that any ordinary person with a higher education, and even any literate person in the future, can use AI tools to CRUD hundreds or thousands of entities and present the relationships between entities to form the final software.

In the area of AI Science, on the one hand, it benefits more from the progress of AI Coding; on the other hand, it actually makes up for the loopholes in human rational thinking and the lack of imagination.

9

Before the AI era, the half - life of certain technical knowledge was generally about 3 - 10 years, but now it has generally become 18 months to 3 years.

The requirements for a person's learning ability have become higher, and the requirements for a person's cognitive ability have also become higher.

Correspondingly, the time spent on handling technical trivialities has decreased, and the requirement for rote - learning ability has become lower.

10

This evening, when I went for a run in the park, I had a strong feeling that in the AI era, psychology may become more important. But the psychology I'm talking about may not be the current outdated set of psychology. It's more like these:

1. Cognition: How can humans better constrain the instructions they give to AI to reduce cognitive complexity?

2. Emotion: How can humans better help AI understand the context or situation humans are in?

3. Motivation: How can humans fine - tune the weights of AI to enhance love or efficiency?

4. Action: How can humans better cooperate with AI's actions?

5. Personality: How can humans better understand the personalities of different large AI models?

6. Society: How can humans better understand the new social community composed of AI agents and large AI models?

Traditional psychology is centered on humans, with other organisms as supplements. Since other organisms can hardly understand human intentionality, traditional philosophy says that humans are lonely creatures in the universe.

But now it's very different. Through engineering means, such as SFT and pre - processing of large - scale corpora, in fact, large AI models have built - in some human intentionality and can at least have extremely smooth conversations. In a sense, it is already an agent.

In this way, the new - generation psychology will gradually transition from being human - centered to including not only the part of psychology centered on humans, but also the part centered on large AI models and the part of human - AI collaboration.

Before the era of large AI models, psychology was actually a rather abstract discipline. Improving human cognition and emotions brought limited benefits to individuals and society. Even the various methods that psychologists boasted about for successful improvement were often not that successful and were often just empty talk.

But in the era of large AI models, it's very different. Cognition is reality.

Improving the cognition and emotions of a large AI model will immediately lead to a huge leap in human productivity.

Fast - growing AI companies like Anthropic and Notion actually hire many psychology doctors.

11

Anthropic actually defined three concepts strongly related to the "human knowledge structure": Project, MCP, Skills.

This sentence is extremely crucial. A project actually provides the context and constrains the input - output boundaries; an agent actually provides the acting subject and constrains the action boundaries; skills provide the verification criteria or examples and constrain the ability boundaries; MCP provides the communication method between the model and external things and constrains the model's throughput.

12

The times are actually undergoing a huge change. Knowledge work can be accelerated by 10 times, 30 times, or 100 times.

It's just that many people can't feel it. In the field of AI Coding, there is a general consensus: at least 10 times the speed.

Why exactly, precisely, and first, does AI speed - up occur in the field of [AI Coding]? In fact, it's the result of multiple forces:

1. Too much code, data, and implicit knowledge have been accumulated on Github.

2. The mainstream development frameworks in different fields are relatively easy to sort out.

3. The results of program development are easy to verify.

4. Large language models are inherently good at processing complex texts, and code is a typical complex text.

5. With nearly 100 billion US dollars in investment, a large amount of manual annotation, and a large amount of synthetic data, [AI Coding] has been continuously pushed to new milestones.

6. Programmers are the group that accepts new things the fastest and is most willing to invest in productivity tools.

13

The word "Agents" is actually an abstract and evolving entity (object), not a static and specific instance.

For example, in 2023, the industry generally understood Agents as writing different prompts, like writing something like: Assume you are a writer now.

In 2024, the industry began to add some features to Agents, and networked retrieval was generally supported. RAG began to be widely popularized. Popular features like having a conversation with a PDF, networked search, and knowledge - base answering became the Agents of 2024.

However, in 2025, people found that Agents also needed to include MCP, Project, Skills, and more scripts, global rules, knowledge bases, memory banks, etc.

So, what will be added to Agents in 2026?

So, don't make Agents seem too mysterious. Strictly speaking, it is restricted by four things:

1. Optimization on the LLM side: The multi - hop reasoning support ability of the large model itself. The more hops, the stronger the supported Agents ability.

2. Optimization on the Prompt side: Prompt engineering, or context engineering, actually means the same thing.

3. Optimization on the MCP side: All kinds of MCPs can be better understood by the large model, and more sufficient error - reporting and feedback information can be provided.

4. Optimization on the knowledge - base side: The knowledge base is actually far more than just a document - based knowledge base. It also includes various global rules, project context information. Some call it Project, some call it Skills, and various memory banks. Some call it MemoBank, but generally speaking, it provides more information.

Finally, it's the turn of optimization on the Agents side: That is, when designing multiple Agents, how to make them compete, cooperate with each other, and how to schedule them?

14

Now, a new era of work has really arrived. There used to be a term: digital literacy, corresponding to the digital divide. Now, there is a new divide: the cognitive divide, which is magnified by AI. With the support of AI, there is a significant gap in work efficiency.

In the future, in countless fields, 10 - times work efficiency will no longer be far away, just like the current situation in the field of AI Coding.

15

After two years, the substantial productivity leap brought about by large AI models first occurred in the field of AI Coding. It really gives a feeling of being within reason but not generally predicted in advance. Previously, people mainly talked about the subversion of AI in fields like graphic design, audio - video, and personal creation. But in these fields, because of the great personal emotional value, there can't be any subversion in the short term.

However, the leap in the field of AI Coding is real, and productivity has leaped.

The three key time nodes of this leap are May, August, and October 2025. Now, it's time for large - scale follow - up in China.

What I'm more looking forward to is what unexpected things might happen when [cognitive ability] is combined with [AI ability]. After all, [cognition is reality]. After all, my [cognitive ability] and the research of the [Open Mind Ecosystem] on how to improve a person's [cognitive ability] are, I think, very advanced.

16

This wave of AI productivity revolution is extremely huge and fundamental. Everyone must pay attention to it. Currently, I have clearly predicted