Intel / video

Recursive Language Models — Alex Zhang, MIT PhD

Source: youtube.com
Author: Latent Space
Date: Published Oct 2, 2026

Context

Recursive Language Models (RLMs) let a model work with an input far larger than it can read in one call by not loading the input into the model at all. According to the paper's abstract, by Alex Zhang, Tim Kraska, and Omar Khattab, the long prompt is treated as part of an external environment: the model can inspect it, break it into parts, and call itself on those parts. The authors report handling inputs up to two orders of magnitude beyond model windows. Research abstract

This explains the interview's focus on the software around a model. In the Latent Space episode, Zhang discusses context offloading, code execution, and recursive subagents, so the system decides how information is accessed and how smaller calls add up to a larger task. The abstract and episode description establish the mechanism and the authors' headline claim, but the full paper's methods, costs, and results were not reviewed. Episode description

Key takeaways · AI-distilled

Zhang argues Claude Code, Codex, Pi and many other modern harnesses are structurally very similar.
The episode cites OpenAI's 10,000-agent experiment, about 130B output tokens and roughly $40M-equivalent spent on one problem, and notes much of a swarm's work may be wasted search.
Zhang says AI-generated GPU kernels still leave substantial room for human expertise, and one expert insight can potentially replace large amounts of brute-force token search.
Speculative programmatic , which overlaps tool execution with generation, is one of the techniques Zhang discusses for faster agents.

Terms in this piece · Glossary

context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.

Read the source www.youtube.com