Recursive Language Models: Claude Code, Agent Swarms, Big Research Bets, & the $40M AI Problem Solver MIT’s alex zhang explains why Claude Code, Codex, and Pi are basically the same, how RLMs use code, context offloading, and recursive subagents to generalize across tasks, why one expert can sometimes replace... a trillion-token brute-force search, what OpenAI’s 10,000-agent, 130B-output-token experiment reveals about the future of language models, and why academia’s biggest advantage is the freedom to take weird, ambitious research bets.show more

@a1zhang I watched this and I will watch it again. My Favorite Latent Space session yet! Thank you @a1zhang you are my hero! :)

@a1zhang treating Claude Code, Codex and Pi as the same shape clicked for me the harness stays, the model under it keeps rotating

@a1zhang @a1zhang just shout out @dosco as this one guy who uses RLM in Ax lol

@a1zhang Treating Claude Code, Codex, and Pi as the same shape is the useful bit. The harness and review loop matter more than which model sits underneath this week.

@a1zhang Yes. Deterministic gates first. A second model is still correlated; the compiler is not.

@a1zhang If Claude Code, Codex, and Pi are the same shape, the moat is the harness and the eval loop. How much of that $40M problem solver is recursion vs better context offload?

@a1zhang oracle: recursive calls help only when each step leaves a traceable state

@a1zhang HI, I’ve been using AINews to keep up with AI, but I noticed the website hasn’t been updated since September 9. 😅 Is AINews coming back? Would love to see the daily updates return!

@a1zhang working on new one

@a1zhang How much of that gain remains when cost and recovery from failed runs are part of the test? MIT CSAIL reports a median 26% gain over GPT-5 compaction on long-context tasks, but operational value will depend on orchestration overhead.

@a1zhang RLMs make context routing explicit, but recursive calls need a budget and stopping rule. Otherwise context offloading just hides runaway coordination cost.

@a1zhang The leverage in AI often comes from changing the search strategy, not just scaling the search. A good expert can turn brute force into a much smaller, structured problem.

@a1zhang 代理群集听着猛，我卡在凭证隔离——集群里每个 agent 是各自单独授权工具，还是共享同一套会话密钥？一个节点挂了，别的节点会不会捡到半截旧 token

@a1zhang The context offloading angle is the key. Recursive agents are less about adding more calls and more about giving each subproblem a bounded working set.

@a1zhang Recursive loops make the setup part of capability. As the system rewrites its path, an external judge matters more.

@a1zhang Recursive subagents are a compute bill with a principal problem attached. Who authorized the leaf when the parent only had a vague task?

@a1zhang Who or what reviewed the output of the 10,000-agent run? Review is the step I see fail first when subagents multiply, since a model checking its own diff costs more compute than writing it.

@a1zhang RLM 的实用价值在于把“上下文长度”问题改写成“可编排的计算”问题：代码负责筛选与卸载，子代理负责局部推理，主代理只汇总结果。真正落地时建议同时报告递归深度、工具调用次数和失败回退率，否则很难和普通长上下文方案公平比较。