When the barrier to execution falls to zero

When I hear someone say that today's AI is a giant bubble, I usually take it as praise of the highest order. Every genuinely disruptive technology goes through this curve before it spreads everywhere: doubted by people, force-ripened by capital. Once you have personally felt your own efficiency multiplied a hundredfold, you understand that calling this moment an "AI industrial revolution" is no exaggeration, and there is nothing alarmist in it.

The world is now changing fast enough to cause a kind of physical vertigo.

Five years ago today, if I had a deadline that meant working through the night, my task list was usually very clear: write a few backend APIs, run the tests, watch the screen return its green pass signal, and I could close the laptop and go to sleep in peace. Creation back then was typed out line by line, in an intensely linear rhythm, with my energy strictly bound to the passing of physical time.

Back to today, five years later. Same kind of night, same kind of deadline, and my workload has become building a complete official website from zero.

It is far more than a simple prototype, and far beyond the kind of Vibe Coding that never leaves the demo stage. It is a real site that can go live, with complete logic and all kinds of complex material. And in the process, my role changed completely. I shed the identity of the bricklaying programmer at the bottom of the stack and became an exhaustingly attentive foreman and architect.

Most of the actual code and material was handed to AI to generate, but today's intelligence has not yet reached full autonomy. You have to keep watching it, correcting its direction, handling its hallucinations. By the end of that one night, I was more drained than I ever was writing code by hand.

Follow this trend and you can peer at the future. Today a human still has to watch the machine finish a website; five years from now these tools will be stronger and more autonomous, and even the act of "watching" can be dropped. Five years ago I could not have predicted today's way of working. But today, I can already roughly foresee how humans and machines will work together five years from now.

Time, folded flat

This hundredfold gain in efficiency seeped into every corner of life long ago.

I never used to enjoy reading books. Traditional reading demands enormous patience: you follow the author's logic page by page, hunting for the information you need. Today, the friction of acquiring knowledge has dropped close to zero.

Take a physical book at home. My process has become extremely simple. I ask my dad to photograph the contents page and the key chapters and send them to me, and I upload the photos straight into an IMA knowledge base. Within ten minutes I am in a deep conversation with the ideas in that book, extracting exactly the understanding I need.

The same thing happened to my podcast workflow. Editing, cutting, organising the timeline: fiddly tasks that used to consume a great deal of labour now all go to Google AI Studio. Once I have set the framework and the SOP for the process, everything that remains is pure execution. Different AI models have different strengths, some at long-text understanding, some at detailed proofreading. The human job becomes recognising each tool's strengths and snapping them together into a workflow, like building blocks.

The deep end of collaboration, and a container for thinking

People inside this industry feel everything is hurtling forward. But when I turn around and look at the people near me, I see a huge gap in understanding. For 99% of people, using AI still means the first level: one-way "human asks, AI answers." You give an instruction, it returns a passage of text, the conversation ends.

If you treat it as nothing more than a substitute for a search engine, that is an enormous waste of compute. As I see it, collaboration between humans and AI divides cleanly into five levels.

Beyond the most basic Q&A, you can enter the second level and let it help you reflect: before asking, work out what your real question actually is. At the third level you can prompt in reverse, letting the AI question you first, or audit its own output. The fourth level is delegation: like the conductor of an orchestra, you dispatch different AIs to carry out specialist tasks and report back. And at the fifth level, you can set two different AIs debating each other, and in the friction between machine and machine, ideas surface that human effort alone could never reach.

In teaching and sharing, I have tried to distil the intuitions scattered through daily, high-frequency use. What I found is that when people fail to get good output from AI, it is mostly not because they lack technical understanding. It is because they have not worked out what they actually want.

AI never corrects muddled human thinking. It only multiplies the muddle.

To solve this, I distilled a prompting framework called SCOPE. It has completely outgrown the early prompt methods of two or three years ago. It no longer teaches you to feed the machine rigid step-by-step instructions; it teaches you to set up a clear semantic container.

You no longer need to tell the AI "how to think." You only need to define, clearly, the boundaries of success: what your role is, what the full context is, what the final objective is, what form the output should take, and where the lines are that must never be crossed. As AI becomes more and more agentic, clear boundary definition matters far more than fiddly operational steps.

Re-estimating human value in the age of the algorithm

Use AI at its deepest level and you will, unavoidably, meet a final question. It is a question that tormented me for many years.

As a child I studied Go for ten years. In that black-and-white world, as long as you kept training, you could build a solid set of coordinates for good and bad, right and wrong. Then in 2016, AlphaGo defeated humanity outright on the board. What I felt in that moment was not simple shock but a deep weightlessness. The coordinates of value you spent ten years building suddenly lose their anchor.

Later I took a master's degree in AI and joined hackathons and competitions at the frontier. As large language models exploded, I watched AI take over code, painting, video. Every leap of the technology repeated the weightlessness AlphaGo first gave me.

Once a person ties their worth to one concrete executable skill, being replaced is only a matter of time.

If technology has driven the barrier to execution down to zero, what exactly is left for humans? After these years of practice, my conclusion is: Tools change, Creators remain. The tools will keep changing, but the value of the creator is kept forever.

Any software interface you have mastered today, even particular prompting tricks, may be completely obsolete next year. What stays is what machines cannot create from nothing.

Machines are superb at deduction: give one a theory and it can instantly unfold it into countless concrete cases and scenes. What is genuinely scarce in humans is induction: whether, in a plain and noisy physical world, you can notice a phenomenon and shape it into a question.

Can you spot a real pain point in daily life? Can you raise a question that is genuinely valuable and sharp enough? And faced with an ocean of AI-generated options, can you make the judgement that is most human and most on target?

A system that lags, and a steady handing-down

Seeing these trends clearly, I have turned more of my attention to young people and begun exploring education.

In the adult world, many people have already been set hard by the old evaluation system. We are used to hunting for standard answers, used to learning a craft along a fixed path and trading that craft for pay.

But the young people of the next era are born into a world where "if you can think it up, AI can help you build it." In a world like that, the traditional education system looks badly out of date. If schools keep grading them on execution, using a single standard answer to erase their diversity, that is not only a waste; it is stripping away, with our own hands, the cards they will need to survive the future.

Education in China still has a long way to go. Much of the time, the biggest problem with systematic education is not what it teaches but the hidden message it sends: life has standard answers. The real world not only has no standard answers; often it is not even clear what the real question is.

So I do not want to teach them how to use any particular AI tool. Tools are just the process; AI is just the medium. What I want to give them is an environment where they can try, fail, and experiment. In the course of building a real project, they discover the boundaries of the tools for themselves, and feel for themselves what it is like to turn a vague idea into something real.

Good education is never about telling you what to do. It guides you to think from the inside out and to discover for yourself. Teaching a child to ask questions is really teaching them to observe things carefully and to describe what they see.

At its heart, this kind of education is a steady handing-down.

We hand them, early, what we ground out of countless real projects: what we learned about working with people, our instincts about taste, our understanding of how to break out of information cocoons. So that when they one day face, alone and unavoidably, a world completely reshaped by AI, they will have a steadiness inside. They will know that however the tools evolve, as long as they keep the ability to ask questions and make judgements, they will not be easily discarded.

At a stage like this, when the old order keeps dissolving and the new rules are not yet fully written, clear-eyed observation means far more than blind anxiety. Facing algorithms that evolve by the day, I prefer to see them as a mirror: they keep stripping away the mechanical properties we thought were important, and force us to look for the purest value of being human.

1. When machines fully unlock the ceiling on efficiency, what is the floor that decides where things actually go?

2. If the cost of execution approaches free, how should we re-price a singular, good question?

3. Looking back at today from five years in the future: of what we are doing right now, how much is real creation, and how much is just pretending to be busy?