François CholletVerified account

Central planning is unbeatable as a resource allocation mechanism for small systems. As you attempt to scale it up, it becomes terribly ineffective, to such an extent that decentralized control algorithms wastly outperform it, despite being inherently wasteful

You don't need to look at country-scale economies to observe this effect (obviously free markets >> central planning), it applies even to large companies. Past a certain scale you need teams that compete against each other with overlapping products. But why?

I think it's a technology problem, not an intrinsic issue. Better cybernetics will enable us to scale efficient central control (while taking into account uncertainty and exploration) to increasingly large systems in the future

This is tightly related to AI as well: it's the question of local vs. global optimization in a non-differentiable system, open-ended goal-setting, objective propagation from one module to its neighbors, parent, and internal submodules...

Backprop is a centralized answer, that only applies to differentiable functions where "control" means adjusting some parameters and where the optimization objective is already known. Its range of applicability is minuscule

The objective function isn't always known e.g. GANs. RL problems also involve objectives that can only be sampled from. I get the local vs global argument, but it's unclear what you mean about backprop being limited by objectives.

"known" objective doesn't mean "can be expressed analytically as a func of the data", it means you know what you're optimizing for from the start. Your GAN implementation itself is an expression of a known optimization objective. "Unknown objective" refers to open-ended problems.