Posted on Friday 5 Dec 2025. 1,960 words, 10 links. By Matt Webb.
I think we often just draw a straight arrow from “collect training data,” like ingest pages from Wikipedia or see what people are saying to the chatbot, to “now the AI model is better and therefore it wins.”
But I think it’s worth thinking about what that arrow actually means. Like, what is the mechanism here?
Now all of this is just my mental model for what’s going on.
With that caveat:
To my mind, the era-defining AI company is the one that is the first to close two self-accelerating loops.
Both are to do with training data. The first is the general theory; the second is specific.