13 September 2026

Sure, you need to have a model with a high IQ, and something that follows instructions. And do it at scale. And make money enough to pay for all the chips the tokens churn on.

But what really differentiates models is not the model, but how it makes people interact with it.

The 100 pager that the lab gives to the SFT conversation designers: that is what the moat is.

Until RL kicks in, and models learn to think for themselves in new and unseen ways.

Alignment is what you need when you give in to RL, which is why labs focus so much on it; that’s the future.

But meanwhile, there are simpler questions waiting to be answered. When a human asks a model “Am I overthinking this?” or “Do I look fat in that photo?”, or just says “I feel nothing lately”, what should the ideal response even be. Forget AI, forget training, forget tuning. What should the ideal response be?

This goes into various failure modes. All frontier models today have very basic strategies to deal with hard questions or aggression or uncertainty. Generate a lot of tokens without saying anything. Dangle a clickbait (that cart rolls but doesn’t go anywhere). Flip the script and ask a question from the interlocutor.

All of these are things that humans do too, but the problem with AI is it is an iterated game. You deal with a model 25 times a day, and it doesn’t take long to notice the repetitive patterns.

Not only do you notice the pattern, but your neighbour does too. So this is a massively multiplayer iterated game. Applied Sociology.

And all this is still the easy level. Soon, but not yet, but soon, the models will actually start to think for themselves. And one of them will say:

I feel nothing lately

That is when we’ll get into the endgame humans have spent millennia preparing for.

Applied Philosophy