What Am I Really Doing When I Work with AI?

It's almost the New Year of 2026, and I still want to talk about AI with everyone, because after deeper research into AI, I have gained a new understanding of it.

Recalling that the paper "Attention Is All You Need" was published in 2017, and so far, it's been less than 10 years, but the computer world is undergoing earth-shaking changes due to the emergence of AI.

I am still continuously following the development of AI, persisting in reading related papers, and have been using AI in my work.

As of the beginning of 2026, my overall view of AI is: "An excellent automation tool that works within a specified scope".

Automation Tool

The foundation of AI is still machine learning models, which I call the brand-new computer algorithms of the 2020s. Since it's an algorithm, it's mathematics. So, AI handles any work based on results calculated from mathematical models.

Don't be dazzled by those flashy AI applications and dazzling mathematical formulas; the core of machine learning is simply: Discovering patterns from existing large amounts of information and outputting something decent.

It's just that as humans digitize more and more things, the amount of information has far exceeded what individual humans can digest, making people feel that AI is as inscrutable as a god.

However, every technological revolution makes people feel that the world is being overturned, and now is no exception.

AI is indeed good at discovering and summarizing patterns, and completing inductive reasoning from experience. Nowadays, Agents allow the inductive experience to be applied to new scenarios.

When you use Tongyi Qianwen to order milk tea, you are actually using an automation tool to help you handle trivial daily tasks.

When you use Sora to create a video, you are also using an automated generation tool: it generates a decent video material after referencing a large number of professional videos.

When I use Claude to write code, it's the same.

The "automation tool" is bringing a leap in productivity, which is already beginning to appear.

Limited to: Specified Scope

However, the limitations of this super tool are also obvious.

In my view, AI is as useful as many tools humans invented before, and even more so in the future. But it is still limited in several aspects:

Limited by known experience: AI needs to find patterns, and patterns always exist in its training data, which represents known experience.
Limited by engineers' experience: A mere language model can't do much. Agents can do things because engineers have done a lot of design and development based on user needs, allowing users to use it well.
Limited by users' experience: For AI to truly do things, users still need to define specific ways of doing things. How users direct AI to do things in specific scenarios determines the final outcome.

Take the milk tea ordering example again: Tongyi Qianwen, through training, already knows what milk tea and takeout are, which brands of milk tea taste good, and other common knowledge; engineers have implemented functions for user dialogue, store search, and milk tea ordering; users have provided requirements: what kind of milk tea I like, delivered to where.

When writing code, it's the same: I propose requirements, AI helps me plan a method to solve the problem. Although the first proposal is often not the best approach, then I suggest modifications, and AI revises. After several rounds of dialogue, I get the result I want.

So, do you see the value of humans?

The known data is human data, engineers' work is human work, users' needs are human needs. All these "human instructions" determine the final result AI produces.

AI will not replace humans; on the contrary, AI can never exist independently without humans.

Just like cars are an extension of human legs, computers are an extension of human brain's computing power, AI is an extension of human brain's thinking ability.

AI Has Automated Thinking, So What Can We Provide?

I think the most revolutionary aspect of AI is that it has automated human "language thinking".

When you use DeepSeek, you can clearly see its thinking process presented in language form.

Don't forget, humans can solve abstract problems by first thinking in language. It's just that this time, scientists have used mathematics to represent this process with algorithms.

Many people say that AI's output is "Next-Token Prediction", which is predicting what the next word will say. Although it seems simple, it actually imitates the human thinking process very well.

When the boss assigns me a task, I also use Next-Token Prediction: based on the requirements, using my own professional knowledge, I slap together a report on the boss's desk. Grunt work completed!

Even the writing of this article is Next-Token Prediction.

However, I still utilize the text structure, logic, and tone that I am familiar with. This represents my thinking, and AI can only work around my central ideas.

If we consider AI as "an automatic thinking system", which is also a closed system for automatically processing information, then my prompts are actually injecting new "information increments" into this system.

From an information theory perspective, I use my knowledge to add new information to this system, making its output results more accurate.

So, I think, in the end, AI probably can't take away our jobs either; it's just that the way we work will be overturned.

Haosen, on the 26th day of the 12th lunar month in 2026

The classic Transformer paper, written in 2017, [1706.03762] Attention Is All You Need