I think most people have the relationship between intelligence and prediction backwards. They treat intelligence as the core mechanism, and prediction as just something it produces. But I'm more and more convinced prediction is the fundamental process, and intelligence is what emerges from it once the system doing the predicting becomes complex enough.

The common criticism of LLMs is that they're "just predicting the next token." This is lazy. You could say mathematics is "just manipulating symbols" technically true, but it ignores that those symbols consistently correspond to real structure. Maxwell's equations described electromagnetic waves before anyone observed them. Einstein's math described gravitational effects we couldn't measure yet. The patterns in the symbols lead to something real, not just more symbols.

Language has the same property. It's a compressed record of essentially every human observation, error, and discovery accumulated over thousands of years. The structure of reality shaped that record. When a system becomes good enough at predicting the statistical patterns in language, it's not just learning word associations. It's recovering the structures that generated those patterns in the first place. Reconstructing a model of the world from the data humans left behind.

I think, intelligence was never a special capacity that allows prediction. Prediction is the more fundamental thing, and intelligence is just what happens when a predictive system reaches sufficient complexity. Doesn't matter if it's a brain, natural selection, or a neural network.

So, conscious thought as we experience it isn't really the source of intelligence. It's just what the prediction process feels like from the inside.

What's your take on this?