AI Notes — April 16

Energy-Based Models (EBM)

EBM is not new. Energy-Based Models have been around in machine learning for decades:

• Hopfield Network (1980s) — is EBM. John Hopfield won the 2024 Nobel Prize in Physics for this.
• Boltzmann Machines — Hinton's work (one of the deep learning godfathers), also an EBM variant.
• Diffusion Models — the technology behind Stable Diffusion and Midjourney, highly related to EBM.

This isn't fringe science — it has a solid mathematical foundation and is a mainstream research direction.

Who Is Saying "EBM Will Lead the Next Revolution"?

The loudest advocate is Yann LeCun (Meta AI Chief Scientist, one of the deep learning godfathers). Since 2022 he's been saying in various venues: "Autoregressive LLMs have fundamental flaws; the future belongs to JEPA (Joint Embedding Predictive Architecture)" — which is essentially a form of EBM.

This isn't a fringe claim — it's a research bet by one of the top figures in the field.