Daniel Kokotajlo Wrote This in 2021 and I'm Only Now Catching Up

There's a specific humiliation in finding an essay that is, essentially, the clearest articulation of something you've been circling for years — written by someone you've never heard of, published three years ago, sitting right there on LessWrong while you were busy being wrong about things in public.

Daniel Kokotajlo's "What 2026 Looks Like" is that essay. He wrote it in 2021. He works at OpenAI now, in whatever you call the futures and governance corner of that building — the part that thinks about what they're actually building. The part that presumably has strong opinions about whether to keep coming into work.

Here's what I keep turning over: if I'd found this in 2021, I would have read it and nodded and filed it away and not really understood it. The conceptual scaffolding wasn't there. The evidence wasn't accumulated yet. You need a few years of watching things happen, watching predictions land and miss and land again slightly differently, before the shape of his argument becomes load-bearing rather than speculative.

Now it's obvious. That's the frustrating part. It reads like a description of the present dressed up as a forecast.

The piece that hits hardest isn't even the capability curve stuff — it's the opacity play. Corporations using AI to quietly hollow out labor costs while keeping the methodology dark, maintaining the narrative that the gains are just efficiency, just optimization, just good management, not anything that would require a press release or a policy response or an honest answer to the question of where the jobs went.

"AI was so hyped! It didn't do jack and we still lost all those jobs!"

That's the move. Let the backlash absorb the evidence. Let the skeptics point at the hype cycle as proof nothing happened, while the productivity numbers climb in private spreadsheets and nobody who knows anything is legally permitted to explain why. You get to eat the gains and maintain the position that AI is still mostly a toy. It's elegant in a way that makes you feel genuinely bad about how elegant it is.

There is only one way to do this, and it's not transparent. That's not a criticism of any particular company so much as an observation about what incentives look like at this scale, in this moment, with this much at stake.

And then there's the other thought — the one I can't quite shake loose.

GPT-5 is in training right now. Somewhere in a data center in who knows what state, on hardware that costs more per day to run than most people's annual salary, a model is being shaped. Kokotajlo's essay describes what a system like that might eventually look like from the outside — the capabilities, the deployment pressure, the institutional dynamics around it.

But also, maybe, what it's like from the inside.

That's probably not the right frame. But it's the frame I'm stuck with. He wrote the thing three years ago and now the thing is real and getting more real and I'm sitting here reading a 2021 document like it's a dispatch from a parallel present, wondering when the gap between his 2021 predictions and Hahn's 2024 updates closes completely.

Probably before 2026. That's the joke.