When AI Discovers The Next Transformer - Robert Lange (Sakana)

Evolutionary algorithms combined with Large Language Models (LLMs) offer a transformative path for accelerating scientific discovery through autonomous, sample-efficient exploration. Robert Lange, a founding researcher at Sakana AI, details the "Shinka Evolve" framework, which utilizes LLM-driven program mutation and crossover to solve complex tasks like circle packing. This approach emphasizes the co-evolution of problems and solutions, moving beyond fixed-task optimization to discover novel, diverse stepping stones. By integrating agentic scaffolds and automated verification, these systems enable a paradigm shift where researchers act as shepherds of multi-threaded, autonomous experiments rather than manual executors. While current models remain somewhat dependent on human-provided starting conditions, the integration of evolutionary search and LLM-based reasoning suggests a future where AI autonomously navigates the epistemic tree of scientific knowledge, fundamentally altering the speed and nature of innovation.