[D] If reasoning requires optimization rather than generation, what does that mean for the scaling paradigm?

About this article

Been digging into the architectural differences between autoregressive LLMs and Energy-Based Models (EBMs) for reasoning tasks, especially with LeCun's recent push towards optimization-based architectures. The premise is that true reasoning should be an optimization problem (finding a state that minimizes an energy function satisfying constraints), rather than next-token prediction. If reasoning inherently requires this optimization loop, does brute-force scaling of autoregressive models hit ...