When AI Discovers the Next Transformer — Robert Lange
Machine Learning Street Talk · 1:18:07 · 4 months ago
Shinka Evolve applies evolutionary algorithms to LLMs to make scientific discovery more efficient and autonomous, though humans remain necessary to steer the research process and provide deep understanding.
-
Better efficiency — The framework cuts down computational costs by using evolutionary search, allowing for progress with far fewer evaluations than typical AI methods .
-
Parallel exploration — Programs are organized into separate islands, enabling the system to evolve multiple candidate solutions simultaneously and share insights across the database .
-
Smart model selection — A strategy called UCB bandits automatically tracks which model performs best for a task, adjusting usage between frontier models like Gemini or GPT-5 in real-time .
-
The problem gap — Current systems optimize solutions for pre-defined tasks well but struggle to invent new, useful problems; real progress requires co-evolving both the problem and the solution .
-
Proven performance — The system successfully optimized circle packing beyond standard results and achieved second place in a competitive programming contest .
-
Human guidance — Autonomous systems currently act as amplifiers for human creativity rather than replacements, as they often get stuck without human steering or deep understanding .
-
Future research vision — The ideal workflow involves an agentic system that runs experiments overnight, letting the scientist act as a shepherd for the work rather than manually executing tests .
-
How does the UCB bandit strategy allocate different LLMs during the program evolution process?