A few days ago, DeepSeek v4.1 was released, pushing the ceiling of small-model capabilities up another notch.
The pace of AI development has far exceeded everyone's expectations. From the earliest ChatGPT, which could only babble in conversation and had a context length of just a few thousand tokens, to reasoning-capable OpenAI o1, DeepSeek R1, and Kimi K1.5 Thinking, only two short years passed; from reasoning models to today's agents that can smoothly execute commands in various harness tools and complete complex tasks, only a year and a half has passed. It is hard to imagine what AI will look like, how powerful it will be, whether it will have acquired the ability to self-evolve, and whether it will have deeply penetrated fields such as embodied intelligence if we wait another year, two years, or three.
AI Is Getting Better at Writing Operators
AI has also advanced rapidly in the field I work in: operator design and implementation. In just one year, it has transformed from a little assistant that could only help me look up documentation, read code, and find bugs into a master operator writer that can independently read CUDA, PTX, and SASS code, use professional tools to analyze the stall time of each instruction, and then independently optimize operators. I believe that in the near future, it will also be able to independently design operator schedules, evaluate the performance of different scheduling schemes, and implement and optimize them.
Of course I am proud of DeepSeek v4.1's success—after all, I wrote its main Attention operator [1], and its excellence is an affirmation of my operator work. But the wheels of the times roll on, and no one can stop technological progress. I am well aware that in another six months or a year, operators written by AI will most likely be just as good as mine, or even surpass mine. AI can think 300 tokens per second, type a line of command in half a second, and write a piece of code in twenty seconds; I cannot. AI can keep improving in model depth, reasoning intensity, tool-call volume (the frequency of interaction with the environment), and even parallelism; I cannot.
Humanity has never been hesitant about destroying itself. Why, then, knowing full well that 'the better I write operators, the faster our new models will train and infer, the faster model capabilities will advance, and the sooner I will be replaced,' do I still choose to optimize operators as much as I can? For one thing, it is because writing operators is like playing a game for me; it gives me enormous pleasure. When I invent a new technique or see my operator's performance rise, the excitement I feel is no less than that of a speedrunner breaking their own past record. At the same time, when I see my operator's performance far exceed the vendor's official operator, I also feel a great sense of pride. But beyond that, a more important reason is that even if I just 'gave up' or deliberately sabotaged model training, other companies' models would continue to develop as usual and eventually kill me off anyway. 'Of course I hope I won't be revolutionized, but if I must be, I hope the one who revolutionizes me is myself.' When everyone is so determined to destroy themselves, I have no choice but to join this cruel arms race.
What About Me?
When the day comes that AI's operator-writing ability truly surpasses mine, what will happen to me?
My judgment is this: I won't exactly become 'unemployed,' but I will have to 'change careers.' I may still keep my rice bowl, but it could mean I never again get to do the work I once loved.
I once made a judgment about the changes of the times and my own future position: because the times are changing so fast (the AI development above is a good example), I cannot predict at all what will happen in five or ten years. But no matter what, I believe that with my vision, judgment, initiative, and intelligence, I can stay at the table of the times and once again stand at the forefront. However, this judgment can only guarantee that I won't be 'unemployed'; it cannot guarantee that I won't need to 'change careers.' Rather, it encourages me to avoid unemployment by changing careers.
So what does changing careers mean? It means I need to give up the field of operator design, implementation, and optimization that I have cultivated for so long and that I love, and instead become an Agent's 'mecha pilot.' Before, my interests, what I was good at, and what industry needed were basically aligned; now, AI has turned what I am good at into something it is even better at, and has shifted industry demand from 'people who can write high-performance operators' to 'people who can use AI to produce high-performance operators faster.' To adapt to industry demand, I will inevitably have to abandon the direction I once loved and turn to an unknown new direction. I believe that with my understanding of engineering, upper-layer model requirements, and lower-layer hardware, I can continue to produce operators with high quality and efficiency. I also know that I might love this new direction (or might not), but the feeling of having one's passion taken away is truly not pleasant. That quiet joy of sitting at my workstation and calmly writing operators for an entire afternoon may become a swan song this summer. I have no choice but to bury my talent in yesterday and become a mecha pilot. There are more gears in my hands, but fewer beats in my heart.
Here's a vivid analogy: you are a master of knitting sweaters, especially skilled at weaving various patterns and matching different colors. The sweaters you knit are sturdy and beautifully patterned. Rich people from miles around come to ask you to knit for them, and you have made quite a bit of money from it. At the same time, you thoroughly enjoy the feeling of sitting by the window, brewing a pot of clear tea, gazing at the green mountains, blue waters, cattle, sheep, and cooking smoke outside, and quietly knitting for an entire afternoon. But one day, someone invents a miraculous machine that, given only yarn and a pattern, can automatically knit sweaters whose quality and texture are no worse than those you make by hand, and far faster than you. You know very well that your peers can easily reach your former level with this machine, so you have no choice but to use it too. You also know that, thanks to the knitting skills you have accumulated over the past twenty years, even if everyone has a machine, your speed and quality can still surpass your peers. But that delight of listening to the rain by the window, threading the needle, and slowly passing the time is ultimately crushed by the roar of the machine.
I know this is helpless, but there is no way around it. I can keep my job, but the loves of old will most likely have to be abandoned. I am someone who keeps reason and emotion fairly separate; when a problem needs to be handled rationally, I can be very rational, but sometimes I also show an emotional side. I remember that when I moved out of the rented apartment I had lived in for a year, I cried hard, unwilling to part with the memories of the past. Saying goodbye today to the era of hand-written operators and human-brain optimization is undoubtedly crueler than that.
I don't know if any readers feel the same way, but I suppose that's just how it has to be.
What About Everyone Else?
As AI keeps advancing, I am also worried about some questions:
• Will today's students be much more inclined to use AI to complete homework, especially practice-oriented Labs? Imagine two choices: one is to slog through eight hours to finish a Lab, and perhaps still not get full marks; the other is to fire up an AI model and, for a few cents and a few minutes, have AI directly write full-mark code. Which would most students choose?
• The previous point will lead to a severe shortage of engineering ability among many students, including the ability to organize code, build systems, think about potential future requirements and address them in design in advance, abstract, and so on. So, against the backdrop of ever-stronger AI, is this kind of 'engineering ability' still necessary? Will these engineering abilities gradually be abandoned by the times like the old ability to 'skillfully write x86 assembly,' or will they remain valuable forever like the ability to 'understand the entire computer system from software to systems to hardware'? If it is the latter, then it is dangerous—a person with very poor engineering ability, paired with AI, can produce shit mountains at several times the previous efficiency, burying all kinds of disasters in systems and making the world even more half-assed.
• In future society, will power be more important than technology or IQ?
These questions, perhaps, will need the times themselves to answer.
Conclusion
Along with AI's development, future society may tend toward two extremes: communism and Cyberpunk 2077. In the former, productive forces are greatly liberated, and people's living standards rise markedly (I'll just write this much, otherwise I'm afraid it won't pass censorship); in the latter, a few tech companies control most resources, only a tiny number of people can use the most advanced AI and various technologies and achieve an effect close to 'mechanical ascension,' while most people can only use very feeble AI. Moving up social classes will become increasingly difficult: you must first have the strongest AI in order to cross class boundaries, creating a dead loop.
Guess: if Anthropic were to forever hold the world's most advanced AI, would future society become communism or 2077? Take a guess.
Therefore, I still believe that the most cutting-edge intelligence should be supplied to everyone in an open and inexpensive way. I do not trust Anthropic or OpenAI to do this, and I especially do not want Anthropic to hold the most advanced AI or AGI; to put it exaggeratedly, the seriousness is no less than letting Hitler acquire atomic bomb technology before the Allies. This is also why I chose to stay at DeepSeek and have kept staying: we research powerful, fast, and inclusive AI and open-source it, which may pull the world back a little from the 2077 end.
May all be well with the world in the future. May all the beauty be blessed.
[1] 'Main Attention' here only includes MQA attention with head dim = 512, and does not include the indexer used to select the top-k important tokens; that part was written by other (also very strong) colleagues (and their AI Agents).