He says he does not trust Amercian companies to do this and compares it, by his own admission as an exaggeration, to Hitler getting the atomic bomb before the Allies (translator's note: take it easy).
“特别是不希望 Anthropic 掌握最先进的人工智能或 AGI，夸张点说其严重性不亚于让希特勒先于盟军掌握原子弹技术。”
To clarify, he said, "he especially don't wish Anthropic to have best AI or AGI first", he did NOT make the Hilter reference for ALL US labs.
This make sense as the extreme hawkishness of Dario's speech had made some Chinese believe Dario is against China as a whole including Chinese people not just the government.
I think it's very important to represent these type of stuff accurately, I would recommend a direct translation for future reference instead of summerisation.
4You are right. To explain that, I am new here in Lesswrong, not knowing if this kind of words would violating the policy.
1These essays never grapple with the fact that you can't open source hardware. To the degree that having more inference compute matters at all to relative power (spoiler alert: it does), open source software alone is going to matter little.
I actually share his general concern about too much concentration of power at the big labs, and in government more broadly, even though I'm perhaps more positively disposed towards Anthropic than the author. But when I read these essays I can't help but feeling like the authors are just narritivizing their pre-chosen set of actions to justify continuing to do what they are already doing.
1I went ahead and translated the essay using Claude Opus 5.5. I think it's worth reading in full (it's not very long).
I Had to Bury My Talent in Yesterday
A few days ago, DeepSeek v4.1 was released, pushing the capabilities of small models up another notch.
AI has developed far faster than anyone expected. It took only two years to go from the first ChatGPT, which could barely babble through a chat and had a context length of just a few thousand tokens, to reasoning models like OpenAI o1, DeepSeek R1, and Kimi K1.5 Thinking. And it took only a year and a half more to go from reasoning models to today's agents, which can smoothly run commands and complete complex tasks across all kinds of harness tools. It's hard to imagine what AI will look like if we wait another one, two, or three years. How powerful will it be? Will it already be able to evolve itself? Will it have deeply penetrated fields like embodied intelligence?
AI keeps getting better at writing kernels
AI has also advanced rapidly in my own field, which is designing and writing kernels. In just one year it went from a little assistant that could only look up documentation, read code, and find bugs for me to a kernel master. It can now independently read CUDA, PTX, and SASS code, use professional tools to analyze the stall time of each instruction, and optimize kernels on its own. I believe that in the near future it will also be able to independently design kernel schedules, evaluate the performance of different scheduling options, and implement and optimize them.
Of course I'm proud of DeepSeek v4.1's success. After all, I wrote its main attention kernel,[1] and its excellence is a validation of my kernels. But the wheels of history keep rolling forward, and no one can stop technological progress. I know perfectly well that in another six months or a year, the kernels AI writes will most likely be as good as mine, or even better. AI can think 300 tokens a second, type a command in half a second, and write a whole piece of code in twenty seconds. I can't. AI can keep scaling up in model depth, thinking intensity, number of tool calls (how often it interacts with its environment), and even parallelism. I can't.
Humanity has never hesitated when it comes to destroying itself. So why do I still do my best to optimize kernels, when I know that the better my kernels are, the faster our new models train and run, the faster model capabilities improve, and the sooner I'll be replaced? Partly it's because writing kernels is like playing a video game for me, and it gives me enormous pleasure. When I invent a new technique, or watch my kernel's performance go up, the thrill is no less than a speedrunner's when they break their own record. And when I see my kernels far outperform the vendor's official ones, I feel a great sense of pride. But there's a more important reason. Even if I just gave up, or deliberately sabotaged model training, other companies' models would keep developing and would end up killing me off all the same. "Of course I'd rather not be revolutionized out of existence, but if it has to happen, I'd rather be the one who does it to myself." When everyone is this determined to destroy themselves, I have no choice but to join this brutal arms race.
What about me?
When the day comes that AI writes kernels better than I do, what will become of me?
My judgment is that I won't be unemployed, but I'll have to change careers. I can keep my job, but I may never again get to do the work I once loved.
I once made a judgment about how the times are changing and where I'll stand in the future. Because things are changing so fast (the AI progress above is a good example), I have no way of predicting what will happen in five or ten years. But whatever happens, I believe that my vision, judgment, initiative, and intelligence will keep me at the table and let me ride the crest of the wave again. However, that judgment only guarantees I won't become unemployed. It doesn't guarantee I won't have to change careers. If anything, it encourages me to change careers precisely so that I don't become unemployed.
So what does changing careers mean? It means giving up kernel design, writing, and optimization, a field I've worked in for a long time and love, and becoming a "mech pilot" for agents instead. Until now, what I was interested in, what I was good at, and what industry needed were basically aligned. Now AI has turned what I'm good at into something it's better at. It has also shifted industry demand from "people who can write high-performance kernels" to "people who can use AI to produce high-performance kernels faster." To keep up with what industry needs, I'll inevitably have to give up the direction I loved and move toward an unknown new one. I believe my understanding of engineering, of what the models above need, and of the hardware below will let me keep producing kernels with high quality and efficiency. I also know I might come to love this new direction (or I might not). But having something you love taken away really doesn't feel good. The quiet joy of sitting at my desk and calmly writing kernels all afternoon may play its final note this summer. I had to bury my talent in yesterday and become a mech pilot. There are more gears in my hands now, but fewer beats in my heart.
Here's an analogy. Say you're an expert at knitting sweaters, especially at weaving all sorts of patterns and matching colors. Your sweaters are well made and beautifully patterned. Wealthy people from miles around come to have you knit for them, and you make a good living from it. You also deeply enjoy sitting by the window with a pot of tea, gazing at the green hills, clear water, cattle and sheep, and chimney smoke outside, quietly knitting through an afternoon. Then one day, someone invents a magical machine. Give it yarn and a pattern, and it knits the sweater automatically, matching the quality and texture of yours, and much faster. You know your competitors can easily reach your old level with this machine, so you have no choice but to use it too. You also know that with the twenty years of knitting skill you've built up, even if everyone has the machine, you can still beat your competitors in speed and quality. But the pleasure of sitting by the window listening to the rain, threading the needle, and letting the hours slowly pass has been crushed under the roar of the machine.
I know there's nothing to be done about it. I can keep my job, but I'll most likely have to give up the thing I used to love. I'm someone whose rational and emotional sides are fairly separate. When a problem calls for reason, I can be very rational, but sometimes my emotional side shows. I remember crying hard when I moved out of an apartment I'd rented for a year, because I didn't want to leave those memories behind. Saying goodbye today to the era of hand-written kernels and human-brain optimization is undoubtedly far more painful than that.
I don't know if any readers feel the same way, but I suppose this is just how it has to be.
What about everyone else?
As AI keeps advancing, I also worry about a few things:
Will today's students be much more likely to use AI to do their homework, especially hands-on labs? Imagine two choices. One is to slog through a lab for eight hours and maybe still not get full marks. The other is to fire up an AI model and have it write full-marks code in a few minutes for a few cents. Which will most students choose?
The above will leave large numbers of students with seriously inadequate engineering skills: the ability to organize code, build systems, anticipate future needs and design for them ahead of time, abstract, and so on. As AI keeps getting stronger, will these engineering skills still be necessary? Will they gradually be discarded like the old skill of "writing x86 assembly fluently," or will they stay valuable forever like the ability to "understand the whole computer system from software to systems to hardware"? If it's the latter, that's dangerous. A person with poor engineering skills, paired with AI, can produce piles of garbage code several times faster than before, planting all sorts of hidden hazards in systems and making the world even more slapdash.
In the society of the future, will power matter more than skill or intelligence?
Perhaps only time itself can answer these questions.
Conclusion
As AI develops, future society may tend toward one of two extremes: communism or Cyberpunk 2077. In the former, productivity is greatly unleashed and people's living standards rise significantly (I'll stop there, or I'm afraid this won't get past the censors). In the latter, a few tech companies control most resources. Only a tiny minority can use the most advanced AI and technology and achieve something close to "mechanical ascension," while most people only get very feeble AI. Crossing class lines becomes harder and harder: you need the strongest AI before you can climb, which creates a vicious cycle.
Take a guess: if Anthropic held the world's most advanced AI forever, would future society become communism or 2077? Go on, guess.
So I still believe that frontier intelligence should be made available to everyone in an open and cheap way. I don't trust Anthropic or OpenAI to do that. In particular, I don't want Anthropic to hold the most advanced AI or AGI. To exaggerate a bit, that would be no less serious than letting Hitler get the atomic bomb before the Allies. This is also why I chose to stay, and keep staying, at DeepSeek: we research powerful, fast, accessible AI and open-source it, which might pull the world back a little from the 2077 end.
May the world of the future be well. May all the beauty be blessed.
- ^
"Main attention" refers only to the MQA attention with head dim = 512. It does not include the indexer that selects the top-k most important tokens; that part was written by other (also very skilled) colleagues and their AI agents.
1
A summary of a Chinese essay, with a few short translated excerpts. All views below are the author's; quotes are my translations.
On 14 September 2026, a DeepSeek engineer writing as intlsy published a WeChat essay titled 我不得不把才华埋葬在昨天 ("I have to bury my talent in yesterday"). He wrote the main attention kernel of DeepSeek v4.1 (the head-dim-512 MQA attention). The essay is a personal reflection on AI taking over his specialty, and it ends with an argument for open-weight frontier AI.
The essay went viral in Chinese Internet: WeChat shows it at "100k+" reads (the platform's display cap), with about 41,000 likes, 84,000 shares, 31,000 recommendations and 1,601 comments as of 7 October 2026. It is probably one of the most widely shared Chinese-language article of what AI automation looks like from inside a frontier lab, and thus a sample of what a large Chinese technical audience is reading and reacting to.
I think it might interest Lesswrong readers.
His argument
1. AI is catching up in his own specialty
He describes AI going very fast, from only an assistant that looked up documentation and found bugs to a system that can read CUDA, PTX and SASS, profile the stall time of each instruction, and optimize kernels on its own, in just a year. He expects that AI-written kernels can match or even beat his own within six months to a year, and expects AI to soon design and evaluate kernel schedules as well. In his opinion, the gap is structural. AI can keep scaling thinking effort, tool calls and parallelism (translator's note: Just like the concept of RSI), while he cannot.
2. Why he keeps speeding up his own replacement
Better kernels make training and inference faster, which speeds up capability gains, which brings his replacement sooner, he says. Anyway, he continues to do it. The first reason is basic. Writing kernels feels like a game to him. The second, which he calls more important, is race logic. If he slacked off, or even sabotaged training, other labs' models would keep advancing and replace him all the same. In his words: "I would rather not be disrupted. But if I must be, I would rather be the one who does it."
3. Not out of work, but forced to change what the work is
He separates losing his job from losing his craft. He expects to stay employed by becoming what he calls a "mech pilot" for agents. Before, his interests, his strengths and industry demand were aligned; AI has split them, shifting demand from people who write fast kernels to people who use AI to produce fast kernels faster. He illustrates this with a master knitter who adopts a knitting machine: still better than competitors who use the same machine, but without the quiet afternoons that made the work worth doing. His summary of the change: "more gears in my hands, but fewer beats in my heart."
4. Worries about everyone else
Three questions:
5. Two futures, and his case for open AI
He frames the future as heading toward one of two extremes: "communism," where productivity gains are broadly shared, or "Cyberpunk 2077," where a few companies control the strongest AI and you need the strongest AI to climb the social ladder, a closed loop (translator's note: typical opinion from a Chinese). His conclusion is that frontier intelligence should be supplied to everyone openly and cheaply. He says he does not trust Amercian companies to do this and compares it, by his own admission as an exaggeration, to Hitler getting the atomic bomb before the Allies (translator's note: take it easy). He gives this as his reason for staying at DeepSeek.