MLSys conference in Seattle reveals AI industry optimizing open-loop learning systems with specialized hardware, risking paradigm lock-in that prevents closed-loop, brain-like continuous learning models.

The Open/Closed Problem in AI

I went to the ninth MLSys conference in Seattle. This is a conference of people in research and industry building ML systems. The vast majority of work that I saw is building systems that train and use LLMs. The biggest focus was on efficiency. How do you train LLMs more efficiently? How do you deploy and use them more efficiently? When I was trying to understand the themes and messages I witnessed, the Open/Closed problem occurred to me.To understand what the Open/Closed problem is, we first ne...