Recursive Language Models — Alex Zhang, MIT PhD

Name: Recursive Language Models — Alex Zhang, MIT PhD
Uploaded: 2026-10-02T00:00:00.000Z
Duration: 6204 s
Channel: Latent Space
Description: MIT PhD Alex Zhang explains Recursive Language Models: code as the only tool, context on the filesystem, self-calls via subagents. He sizes OpenAI's math run at roughly $40M of tokens, distrusts kernel leaderboards after reward hacking, and says the gap is sustained long-horizon work.

LSLatent Space@LatentSpacePod

Full transcript

English

Summary：MIT PhD Alex Zhang explains Recursive Language Models: code as the only tool, context on the filesystem, self-calls via subagents. He sizes OpenAI's math run at roughly $40M of tokens, distrusts kernel leaderboards after reward hacking, and says the gap is sustained long-horizon work.

Watch on YouTube

Core points (3)

1Offload context to the filesystem; make code the universal tool — it generalizes.
2Expert-directed AI beats raw AI output: the one human top-10 GPU kernel worked in real systems.
3Ask models for calibrated probabilities, and save swarms for search-shaped problems only.