Marr versus Friston is a fight neuroscience never settled

David Marr argued in 1976 that intelligence has three independent levels: computational, algorithmic, implementational. Karl Friston argued in 2010 that intelligence has one principle: free-energy minimization, instantiated everywhere from cells to societies.

Both are influential. Both have produced productive research programs. They are not compatible.

Marr's framework treats intelligence as a layered system where understanding any level does not require understanding the others. Friston's framework treats intelligence as a single principle expressed at multiple scales, where the same equation describes neuron firing, perception, action, and high-level cognition. The first says decomposition first, then unification. The second says unification first, decomposition second.

Neuroscience has not chosen. The field has produced strong work in both traditions and has mostly avoided the question of which framing is correct, on the assumption that both can be useful. AI is now inheriting the same unresolved choice. Researchers committed to mechanism design, including transformer architectures, MoE routing, and reasoning loops, implicitly work in Marr's tradition. Researchers committed to general principles, including predictive coding, world-model frameworks, and active inference, implicitly work in Friston's.

The two communities mostly do not talk to each other. They will eventually have to, because the question of whether intelligence has unifying principles or only useful decompositions matters for what AI alignment can hope to achieve. A unifying-principle world is more controllable. A useful-decompositions world is not.