Machine Learning and Friends Lunch: Louis Kirsch, Escape Velocity: Training AI Scientists that Train AI
Content
Speaker
Louis Kirsch (Inherent)
Abstract
Recursive self-improvement (RSI) is nearing escape velocity: the point where an AI system sustains its own progress without constant human intervention. The talk traces a path from meta-learned learning algorithms to LLMs that drive science and train the next iteration of models. At Inherent we trained Faraday, a 27B-parameter AI Scientist, with long-horizon RL to replicate figures from 100 research papers. It outperforms frontier models at replication and directs coding agents 100x larger than itself. I close with why replication is a stepping stone to innovation, and how humans could stay in the loop as RSI accelerates.
Speaker Bio
Louis Kirsch is co-founder and Chief Superintelligence Officer of Inherent, the AI lab building recursively self-improving AI that discovers new knowledge. He was previously a Research Scientist at Google DeepMind, where he founded and led the team building LLM-based AI Scientists. He completed his PhD on Automating AI Research with Jürgen Schmidhuber at IDSIA, the Swiss AI Lab, where he pioneered meta-learned learning algorithms that generalize and self-referential systems that improve themselves.