Adding Value in the Age of AI

If you wrote out all possible combinations of, say, 1000 letters, the vast number of them would be nonsense. And the ones that do make sense are likely mostly mundane - really interesting ideas are super rare (this is pretty much the idea of Borges’ Library of Babel). This means that if you pick some random 1000 letter string, the odds if it being useful and new are essentially 0.

So what do you do? Well, you typically stay “in the distribution” - you say and do and think things that other people have tried before and found useful. You can get out of that distribution, but the farther you go, the less likely you are to have something that makes sense (this is true of genes too: you have to stay near the set of genes that is know to be viable, long mutations usually die).

This is fine, it’s what we’ve been doing for millennia: slow and gradual mutation on the existing distribution, finding nearby new ideas. Sometimes people make large jumps and get lucky, sometimes we get stuck on a local maxima and can’t get off without them, but mostly, we stay near the known and safe. Most people don’t think many thoughts that haven’t been thought before.

But…AI is the distribution. LLMs are now trained on most of the language that we have found to be useful - the ‘valid’ collection of points in the space. That means it’s hard to create new value if you’re inside the distribution, because by definition, the LLM can do that too, and likely cheaper and faster.

So you have to get out of the distribution! Creativity and invention! Which is cool, except its dark out there, baby! The farther you go, the less likely it is to be valuable. But hurry, because the LLM is right behind you, doing it’s own incremental expansion of the distribution.

I don’t know that there is an answer or even advice, just an observation. There has always been some pressure to explore the margins outside, but most people haven’t felt much of it in their lives. That pressure is now increasing and broadening - and the challenge is harder. There is a narrower band between the next incremental idea and the next non-viable one. And that’s likely to be true in all fields, increasingly.