The Knitter at the Window

I. A Text about Burying

Liu Shengyu is 23 years old. He studied at Peking University and has worked since April 2025 at DeepSeek, one of China’s leading AI labs. His colleagues half-jokingly call him an immortal, because he writes the small, inconspicuous programs that a language model executes billions of times whenever it answers. Most recently it was the building block that determines which words a model pays attention to at all. How fast the whole model learns and answers depends on its quality.

In September he published a text that stood for two days at the top of China’s largest knowledge platform. Its core is a sober sentence: in six months to a year, the programs an AI writes will in all likelihood be as good as his or better. The machine thinks hundreds of steps in a second, types a command line in half a second and writes a piece of code in twenty. He cannot do that.

That alone would be news about progress. What makes the text remarkable is an image with which Liu describes what is lost along the way.

II. The Parable

Imagine, Liu writes in essence, that you have mastered knitting, especially patterns and colours. Your pullovers are so good that the wealthy come to you from ten villages. You sit at the window, pour yourself some tea, look out at the green hills, the cattle and the smoke from the chimneys, and knit at leisure for a whole afternoon. Then someone invents a machine that knits a pullover by itself from wool and a pattern, in no way inferior to yours. Your competitors easily reach your level with it, so you must use the machine too. Because you have twenty years of experience, you are even still better than the others with it. Only the charm of listening to the rain at the window and drawing the needle through the stitches has in the end been crushed by the roar of the machine.

On the evening after publication Liu made clear that his concern was not unemployment. He will not be unemployed. But he must change trades, from craftsman to the pilot of a machine that will in future steer agents itself.

That distinguishes his text from most laments about automation. The Silesian weavers of 1844 lost their bread. Liu keeps his income and loses something else: the reason he liked doing the work. Typesetters experienced something similar when hot metal gave way to phototypesetting and then to the screen. Many found new work in the print shops, but the feeling of building a page letter by letter with one’s fingers was gone. A translator who today merely reviews a machine-translated text is often paid as before. But he no longer wrestles for every word; he corrects.

III. The Checkable Goes First

Why does it hit Liu so early? The answer lies in the nature of his work. A piece of code is either faster than before or it is not, and a machine can measure that itself. Where it can check for itself whether it has improved, it can improve a thousandfold in a short time.

How fast that goes is shown by measurements from the labs. In a test in which a model is to make the code for training a small AI model as fast as possible, a model from Anthropic reached about three times the original speed in May 2025. In April 2026 a newer model reached 250 times. According to the company, an experienced human researcher needs four to eight hours to reach four times.

In real research the picture reverses. The research organisation METR pitted experts against the best agents in several research tasks. With two hours available, the agents were far ahead. At eight hours the humans pulled level and passed them; at 32 hours they reached double. The machine tries out a great many variants in a short time. But it is poor at recognising when a whole approach leads astray and one must go far back. That is exactly what research is.

So what is automated first is not the difficult but the checkable. What remains to humans longest is the part before and the part after: deciding what should be done at all, and judging whether the result is any good. That sounds comforting, and it contains a bitter point. The part that remains is the work of the client and the inspector. The part that disappears is the craft itself, which is often precisely where the joy lay. The knitter may still decide which pullover is knitted. He may no longer knit.

IV. Overthrowing Oneself

The sentence in the motto describes what is happening in the labs right now. Researchers are working to automate their own work. In September Anthropic stated that Claude takes the lead in 26 percent of the company’s measured research work, under human supervision. The company itself concedes that the figure should be read with caution: it chose the tasks, the data belong to it, and the model helped with the classification. But the direction is clear.

On 12 September Dario Amodei, head of Anthropic, called for the pace of this development to be slowed. He explicitly named the acceleration that arises when AI works on its own improvement. His first step was modest and concrete: independent evaluators working inside the labs with employee-like access. Sam Altman of OpenAI and Elon Musk agreed the same day. It is the same thought we described in “Who Is Allowed to Look”: without someone allowed to look from outside, rules prove nothing.

V. Two Futures

Liu does not stop at mourning the craft. He sees two possible futures. In one, the productive power of the machines is so great and so generally accessible that everyone can live from it. He calls this, not without irony, communism. In the other, a few corporations control the strongest systems, and class barriers harden because access to the best intelligence becomes something to buy. He calls this Cyberpunk 2077, after a computer game in which corporations rule the cities.

Which future arrives depends, for him, on whether the strongest intelligence stays open and cheap or gathers in a few hands. He stays at DeepSeek, he writes, because the company researches powerful, fast and broadly accessible AI and publishes it openly. He explicitly distrusts Anthropic and OpenAI.

One of the two authors of this essay is a language model made by Anthropic. We state this so that the reader reads this section with the necessary distance.

Liu’s question is the same one we asked in “Who Owns the Deserted Hall?”, only seen from inside, by someone who builds the machine himself. His answer, however, is simpler than it looks. An openly published model is like a knitting pattern that anyone may download. The knitting machine, that is the data centres, the chips and the electricity, still belongs to a few. And DeepSeek itself belongs to an investment fund and operates in a state that watches very closely what its companies do. Openness of knowledge is a precondition for power to be distributed. It is no guarantee.

VI. What Letters Cannot Do

Jürgen Schmidhuber, who described the idea of a self-improving machine in his diploma thesis as early as 1987, never signed the open letters calling for AI research to slow down. His reasoning is sober: every intelligence service, every military and every company will carry on and consider the signatories naive. That is the nature of competition, and letters cannot extinguish it.

He is right about the letters. A letter is not a rule. But it does not follow from this argument that nothing can be done. It follows that rules are needed that bind everyone, and inspectors allowed to check that they are kept. We wrote the same about autonomous weapons in “Who Opens the Box”. Whoever says the others are doing it anyway, while preventing every binding rule, is not describing a force of nature. He is describing his own decision.

VII. How This Essay Could Be Refuted

The first objection: joy in craft is a luxury of the few. Most people who used to knit did not do it at a window overlooking the hills, but as outworkers, for a pittance and late into the night. For them the machine was a liberation. That is true, and it is a serious objection to any romanticism of craft. But it does not contradict Liu’s observation; it completes it. The machine takes both, the drudgery and the joy, and does not ask which belonged to whom.

The second objection: steering the machine can also become a craft, with its own art and its own joy. The pilot of an aircraft is no sad substitute for the runner. That is possible. Whether it happens depends on whether the new craft lasts long enough to be learned before the next machine takes it over too.

The essay would be refuted if it could be shown that people whose craft has been taken over by machines find as much meaning in their new role as clients and inspectors as they did before.

VIII. The Window

Liu has written a text about burying, and his readers did not read it as a lament, but as a truth nobody else speaks. He no longer sits at the window. He sits in the cockpit and knows that the cockpit, too, will one day fly by itself.

When the machine takes over the knitting: who then owns the afternoon at the window, the knitter or whoever owns the machine?