Prediction is not intelligence

It's fascinating how we define intelligence.

Mostly when it comes to evaluating AI for intelligence we evaluate against math. Moravec's paradox suggests it's easier to evaluate computers against hard logical problems than actions a one year old takes.

Things like picking up a cup, learning to walk or to balance is wildly harder to pin down and measure, yet it's what biological intelligence evolved to do.

We humans or animals in general have acquired intelligence over hundreds of thousands of years of evolution. It took a lot of trial and error to get to the level we are now.

How one big questions is how does a baby learn things?

We have a notion of supervised learning where we learn by imitation. Does this work at all?

But when we zoom out to the open, interactive world, imitation is only part of the story. Babies learn by acting, by getting feedback that isn't a label but a cascade of physical consequences: the cup falls, someone laughs, a hand recoils. That stream of outcomes supplies goals → implicit rewards → that the child discovers and optimizes for.

Richard Sutton (one of the founding fathers of reinforcement learning), argues that intelligence is fundamentally about achieving goals through interaction and learning from experience.

LLMs are breathtaking at predicting and mimicking human text. But prediction of tokens is not the same as learning from the consequences of actions in the world. A prediction task doesn't necessarily give the model a goal that changes the external environment and provides feedback for continual learning.

That's also one of the reasons that post training RL fine tuning is becoming a crucial step in LLM training.