Descriptions:
Robert Lange, a founding researcher at Sakana AI, joins Machine Learning Street Talk host Tim to discuss Shinka Evolve — a new paper that extends the evolutionary LLM program synthesis approach pioneered by AlphaEvolve. The conversation covers how language models can iteratively generate, mutate, and evaluate programs using Upper Confidence Bound (UCB) selection, mutable code markers to protect critical imports and evaluation scaffolding from unintended edits, and rejection sampling with reflection to enforce structural constraints.
A central theme is sample efficiency: while similar systems may sample thousands of programs per task, Shinka Evolve targets comparable performance with far fewer evaluations. Lange explains the information representation challenge — compressing evaluation histories to fit within context windows — and discusses future directions such as multi-file codebase mutation, repository maps inspired by the Aider coding tool, and the potential for active fine-tuning during evolutionary runs.
The broader conversation touches on Sakana AI’s research philosophy, shaped by CEO David Ha and inspired by Ken Stanley’s open-endedness framework. Lange reflects on what it would mean for AI systems to autonomously discover architectural innovations as significant as the Transformer itself — framing current LLM-driven search as a step toward that Rubicon. The episode also briefly covers NVIDIA GTC 2026, the leaked Nemo Claw open-source agent platform, and the general challenge of verifying program correctness versus generating candidate solutions.
📺 Source: Machine Learning Street Talk · Published March 13, 2026
🏷️ Format: Interview
