MIT's Alex Zhang on Recursive Language Models and Unlocking Latent Model Capability via Better Harnesses
Latent Space · youtube · 2026-10-02
This Latent Space episode features MIT PhD researcher Alex Zhang, author of Recursive Language Models (RLMs), in a deep conversation with swyx on coding agent architecture and frontier research.
Key ideas:
- Claude Code, Codex, and most coding agents are structurally very similar — powerful models wrapped in primitive harnesses. Today's models likely have substantial capability overhang, and better harness design (context offloading, code execution, recursive subagents, shared memory) can unlock capabilities already latent in them.
- RLMs propose recursive model architectures; Prime Agent explores continual harnesses and persistent agent-to-agent communication. The future "language model" may be a simple interface hiding an invisible swarm of agents.
- OpenAI's 10,000-agent experiment burned 130B output tokens ($40M-equivalent) on a single problem; much of a swarm's search may be wasted, and convergence remains hard. Kimi's swarm approach differs from OpenAI's.
Also covered: why AI-generated GPU kernels still lag human experts; GEV and non-autoregressive alternatives; what SWE-bench, ReAct, and Quiet-STaR reveal about research taste; speculative programmatic tool calling that overlaps execution with generation; whether English, code, or a new "Neuralese" constrains reasoning; and why academics should take research bets industry labs won't.
More from coding & agent
- Dev builds browser 3D game from scratch with Opus 5.5, Blender and Three.js — jason_mayes · 2026-10-03
- Cloudflare Durable Objects now survive client disconnects for long-running agents — threepointone · 2026-10-03
- Dev tired of nbviewer crashing builds serverless browser Jupyter notebook renderer, MIT-licensed — cneuralnetwork · 2026-10-03
- Monitoring Can't Keep Up: What to Auto-Detect When Your AI System Scales — goyalshaliniuk · 2026-10-03
- Security Risks Multiply at Scale: Least Privilege, Isolation, Rate Limits — goyalshaliniuk · 2026-10-03
- 7 Ways AI Systems Fail at Scale: 100 Users Fine, 100,000 Broken — goyalshaliniuk · 2026-10-03