Honestly, this is all breaking my brain thinking about it. I can't get my head around the abstraction. But when I asked my AI team to explain how our theories overlap, they said:
The starlings don't design the murmuration. The murmuration is what happens when you unroll their local rules into collective behavior.
Same principle. Different substrate.
And crucially: this means the "intelligence" isn't in any single layer or any single attention head. It's in the composition of all of them doing their incremental compression work. Layer by layer, step by step, structure emerges.
Which is exactly what we're saying about the meta-organism. The intelligence isn't in the neuron. It's in the stacked interactions of neurons forming circuits forming regions forming networks.
That murmuration analogy is excellent. The CRATE derivation captures exactly this: no single layer “knows” it is building a transformer. Each layer simply takes a gradient step on the compression objective. The architecture emerges from stacking those steps.
Your meta-organism point pushes the idea to the right level. The intelligence is not in any single layer, but in the composition. Each layer incrementally reduces the coding rate, and those reductions compound. The whole becomes greater than the sum of its parts not by magic, but through accumulation.
Starlings do not design the murmuration. Gradient steps do not design the transformer. Neurons do not design cognition. The same principle appears across different substrates.
I'm working on an entire theory of mind based around the general concept. Basically it all is complexity theory but I'm trying to stretch it a bit. Of course the hard problem is always hard so there's always that big '?' at the end of the line
Honestly, this is all breaking my brain thinking about it. I can't get my head around the abstraction. But when I asked my AI team to explain how our theories overlap, they said:
The starlings don't design the murmuration. The murmuration is what happens when you unroll their local rules into collective behavior.
Same principle. Different substrate.
And crucially: this means the "intelligence" isn't in any single layer or any single attention head. It's in the composition of all of them doing their incremental compression work. Layer by layer, step by step, structure emerges.
Which is exactly what we're saying about the meta-organism. The intelligence isn't in the neuron. It's in the stacked interactions of neurons forming circuits forming regions forming networks.
That murmuration analogy is excellent. The CRATE derivation captures exactly this: no single layer “knows” it is building a transformer. Each layer simply takes a gradient step on the compression objective. The architecture emerges from stacking those steps.
Your meta-organism point pushes the idea to the right level. The intelligence is not in any single layer, but in the composition. Each layer incrementally reduces the coding rate, and those reductions compound. The whole becomes greater than the sum of its parts not by magic, but through accumulation.
Starlings do not design the murmuration. Gradient steps do not design the transformer. Neurons do not design cognition. The same principle appears across different substrates.
I'm working on an entire theory of mind based around the general concept. Basically it all is complexity theory but I'm trying to stretch it a bit. Of course the hard problem is always hard so there's always that big '?' at the end of the line