Will AI Kill Our Creativity: The Lock Is One You Put on Yourself

The main Attention operator in DeepSeek v4.1 was built by an engineer named Liu Shengyu. He pushed its performance past Flash Attention — hardcore low-level optimization.

He published an essay titled “I Have No Choice but to Bury My Talent in Yesterday,” saying AI can now think three hundred words per second, finish a block of code in twenty seconds, and within six months to a year, write operators as good as his. After the essay went viral, he clarified: it’s not job loss, it’s a career shift — he’s losing the way he loved to work. His analogy: after the knitting machine was invented, someone who’d knitted for twenty years would still operate it better than most, but what they used to enjoy was the knitting itself. Now the job is managing the machine. AI isn’t taking away your livelihood — it’s taking away the passion.

But the most critical line in his essay flew past most people: he worries that students now rely entirely on AI for homework and experiments, that engineering skills are degrading, producing mountains of low-quality code.

Think three years ahead: everyone in your company will use AI, but no one will be able to say the plan is wrong, and no one will think of a path AI can’t. Your company’s ceiling becomes the model’s ceiling — and your competitor bought the same model.

In The Three-Body Problem, the Trisolarans sent a Sophon to turn experimental results into noise — you can’t measure accurately, so theory can’t advance. But today you don’t need a Sophon. The Sophon stops you from measuring; AI stops you from wanting to measure. Nobody’s forcing you. This lock, we put it on ourselves.

This isn’t about using AI less — I use it the hardest. Catching AI’s mistakes just holds the line; moving forward means thinking of what it can’t. Tim Cook said at Duke’s commencement: No big challenge has ever been solved, and no lasting improvement has ever been achieved, unless people dare to try something different. Yet AI always gives the same answer — the best path computed from everything the world has already done. Your competitor asks the same question and gets the same answer.

Thinking differently comes down to two things. First, when you hit a problem, spend five minutes thinking on your own before asking AI — once you see its answer, you’ll never think of anything else. Second, and more importantly: reason from first principles. Don’t ask how others do it; ask why it has to be done this way.

When I was building a voice input method, the hotword dictionary exceeded a hundred words and introduced more noise than it corrected. The conventional path was iterating the model and optimizing algorithms. But I stopped to think — the real issue wasn’t model weakness; the dictionary simply shouldn’t be static. It should change with context. The final solution was dynamic context-based injection. Not a single line of the model was changed.

Liu Shengyu supports open-source, accessible AI and opposes frontier technology being monopolized by a few companies. I’d go one layer deeper: technology can be open-sourced, answers can be open-sourced, but questions can’t. Today, nobody’s stopping you — you just stopped wanting to. The part of you that’s still thinking for yourself is the part that can still move forward.

EP0099 · 2026-09-17 · 04:38. The picture uses a digital human, and the voice uses my cloned voice. This is an English translation of the approved Chinese script used to create the narration.

The main Attention operator in DeepSeek’s v4.1 was written by an engineer named Liu Shengyu. An operator is basically the layer of code that determines how fast a model runs. He pushed its performance past Flash Attention — the kind of person who squeezes every bit of speed out of low-level, hardcore engineering.

This same person published an essay a couple of days ago, titled “I Have No Choice but to Bury My Talent in Yesterday.” He said AI can now think three hundred words per second, finish writing an entire block of code in twenty seconds, and within six months to a year, the operators it writes will be as good as his. So he has to go from being someone who writes operators himself to someone who directs AI to write them. He called this being a “mech pilot” — a few more gears in your hands, a little less rhythm in your heart.

The essay went viral instantly. Everyone said this was the first cry of AI unemployment anxiety. Right after, he published a follow-up to clarify: no, I’m not losing my job — I need to change careers. What I was talking about is losing the way I love to work.

He made a really good analogy. After the knitting machine was invented, someone who’d been knitting for twenty years would still be better than average at operating the machine. But what you used to enjoy was the knitting itself. From now on, your job is managing the knitting machine.

This analogy nails one thing: AI isn’t taking away your livelihood — it’s taking away the passion you had for the work.

But honestly, the most critical sentence in his essay flew right past most people. He said he worries that students now rely entirely on AI for homework and experiments, that their engineering abilities are degrading, and that they’ll end up producing mountains of low-quality code.

Think three years ahead. Translated into boss-speak: three years from now, everyone in your company will know how to use AI, but not a single person will be able to stand up and say this plan is wrong. Worse still, not a single person will be able to come up with a path that AI can’t. So where’s your company’s ceiling? It’s the ceiling of whichever model you bought. But your competitor bought the same model — everyone’s stuck at the same height, and nobody gets past it.

So these past couple of days, I’ve been thinking about that thing in The Three-Body Problem. How did the Trisolarans lock down human science? They sent a Sophon that turned all your experimental results into noise — you can’t measure accurately, so theory can’t advance, right? But look at what’s happening now. You don’t even need a Sophon. The Sophon stops you from measuring; AI stops you from wanting to measure. Nobody’s forcing you, there are no aliens. This lock, we put it on ourselves.

So what do we do? I’m not going to tell you to use AI less — that’s empty talk. I use it the hardest. Being able to tell whether AI is right or wrong — that just means you can hold the line. But for a company to move forward, you need to think of what AI can’t.

Tim Cook said something at Duke University’s commencement that really stuck with me. His exact words: No big challenge has ever been solved, and no lasting improvement has ever been achieved, unless people dare to try something different. That means no great problem was ever truly solved, and no real progress ever lasted, unless someone dared to do something different.

But have you noticed? The answer AI gives you is always the same. It learned from everything the world has already done and computed the best path above the average line. Your competitor asks the same question and gets the same answer.

So how do you come up with something different? I think it comes down to two things.

The first is to stop and think. Right now, everyone’s habit is to ask AI the moment they hit a problem, right? But once you see its answer, you’ll never think of anything else. So flip the order — spend five minutes thinking on your own, then ask it.

The second is more important: think about the essence of the problem from first principles. Don’t ask how others do it — ask why it has to be done this way. Once you see the essence clearly, the path you find is often more direct than the one everyone else is walking.

When I was building a voice input method, I ran into this: once the hotword dictionary exceeded a hundred words, the noise it introduced was worse than what it corrected. The conventional approach? Iterate the model, optimize the algorithm — everyone takes this path. But I stopped and thought: the real issue isn’t that the model isn’t strong enough. The dictionary simply shouldn’t be static — it should change with your current context. So the final solution was dynamic context-based injection. I didn’t swap out a single line of the model.

At the end of his essay, Liu Shengyu said he supports open-source, accessible AI and opposes cutting-edge technology being monopolized by a handful of companies. I completely agree, and I want to go one layer deeper: technology can be open-sourced, answers can be open-sourced, but questions can’t. The Trisolarans locked down science by stopping you from experimenting. Today, nobody’s stopping you — you just stopped wanting to. The part of you that’s still thinking for yourself is the part that can still move forward.

That’s all for today. See you tomorrow — bye.