Think of a river - the water carves the banks, the banks channel the water. Which is the real river?

A model is like the river banks, it is not intelligent in itself. Neither is activity by itself. Intelligence comes from their co-constitutive relation.

You can see the structure+flow coupling in many places. The model weights (banks) vs the training set (water). The contextual tokens (banks) vs the mass of probability for the next token (water). Agent environment (banks) vs agent activity (water).

The trick here is that neither is capable of explaining the process, or more fundamental than the other. It is a recursive process rewriting its own rules.

And we know from Godel, Turing and Chaitin that recursive processes exhibit incompleteness, undecidability and incompressibility. There is no way to predict them from outside, they can only be known only through unrolling, you have to be inside, or you have to be it to know it.

I suspect that the amount of freely available text, video, and audio will decrease once entities like Disney start winning lawsuits. Alternatively, anyone who doesn’t want their work to be crawled and trained from will keep their work behind login/paywalls moving forward.

I think it’s entirely fair that there should be a robots.txt type of mechanism for sites to opt out of being crawled for training, and we might see places like the EU come up with something of that sort.

what pattern are human color/light sensors arranged in? maybe we should replicate that pattern? an organic arrangement discovered via simulated annealing?

all this bitmap x/y display stuff is very pre-AI-ish. old tech. victorian era clockwork mechanism. built to make it easy to reason about for humans, before the advent of neuron-soup. maybe we can do better?