MIT researcher Alex Zhang on recursive language models and agent swarms

In this podcast interview, MIT researcher Alex Zhang discusses his work on recursive language models (RLMs), GPU kernel generation, and large-scale multi-agent systems. He argues that many coding agents like Claude Code and Codex share similar underlying structures, and that improving how these systems are built—rather than just the models themselves—could unlock capabilities already present but unused.

The conversation covers topics including context offloading, programmatic subagent calling, and the idea that future AI systems may function as invisible swarms of agents behind a simple interface. Zhang also discusses research from OpenAI, Kimi, and Sakana AI on large-scale agent experiments, and shares his views on what makes for good research taste in PhD-level AI work, including when to pursue ideas that initially seem trivial or strange.