Recursive Language Models — Alex Zhang, MIT PhD

Summary · 4 min read instead of 1 h 43 m

Zhang is a second-year PhD student at MIT and the researcher behind recursive language models (RLMs). He helps with GPU Mode, the GPU programming community whose Popcorn leaderboard idea led to KernelBench, and he worked with the SWE-bench team at Princeton. He also spent a summer at Sakana and now works on harness design, including Prime Agent with Prime Intellect's team.

Zhang argues that most coding harnesses, including Claude Code, Codex and Pi, share the same design, and that more opinionated harnesses can add capability and generalization. He defines an RLM as a harness whose only tool is code, with context held in a code environment and subagents called programmatically. He expects the language model users query to become a swarm of agents hidden behind a simple interface. He also says academics should take big bets that industry labs avoid, and that human expertise still beats brute-force compute on problems like GPU kernels.

Minute by minute

1. 2:37

   GPU Mode grew into KernelBench. Zhang says GPU Mode began around 2023 as a Discord for learning to write GPU kernels, with lectures. The Popcorn leaderboard came from a shared intuition that GPU programming resembles competitive programming, with a small space of optimizations. With enough data, kernel development could be scaled and automated, which led to KernelBench.

2. 7:04

   Experts still beat AI on kernels. Zhang says almost all recent GPU Mode leaderboard solutions are AI-generated. A known expert, Galnerst, steered AI to a kernel that was basically the only top-10 one stable in end-to-end systems. Zhang points to verification problems and reward hacking since KernelBench, and says there is still a lot of alpha in being good at kernels.

3. 9:43

   One insight can replace huge token spend. Zhang says swarms explore solutions by throwing compute at a problem, but someone who knows the problem could erase a spend of a trillion tokens. He says the trends are unclear and efficiency still matters. On kernels, he says speed-of-light estimates are easy for matrix multiplication, yet real kernels are often far from them. He expects compilers to handle mega kernels.

4. 21:19

   GEV questions the autoregressive language model. Zhang says GEV raises the question of whether a language model must be a text-to-text decoder. It changes the output space to trade for fast inference, and a binary classification with a normal language model carries a 400x cost. He does not know how it was trained. He sees it helping RLMs, whose multiple model calls are slow.

5. 32:54

   Claude Code, Codex and Pi are alike. Zhang says the design choices in Claude Code, Codex and Pi are mostly the same. He says Prime Agent differs because it is inherently an RLM. He calls a harness an opinionated program that fits a model to a problem. Next-token prediction is awkward for tasks like navigating a codebase, so harnesses exist to help.

6. 36:23

   Harnesses can generalize compositionally. Zhang says that when RLMs offload context and write code over it, retrieval and aggregation tasks end up with the same root strategy. The subagent absorbs the differences, so training on one task transfers to the other. In the blog, swyx reads out 8 to 30 times longer tasks. Zhang says the same holds for math versus writing tasks, with no magic involved.

7. 46:16

   An RLM makes code the only tool. Zhang defines an RLM as a harness where code is the only tool. The model can call itself as a tool, and its context lives in a code environment such as a file system. The original aim was long context, and the aim is now composition. He says every model call stays in distribution even when the whole task is not.

8. 53:54

   Prime Agent builds an RLM on Pi. Zhang joined Prime after a blog post of theirs about RLMs. Prime Agent runs on Pi with IPython as the only tool, and other tools load as Python modules or bash scripts. A continual harness lets it modify its own skills, subagents and system prompt. Persistent subagents outlive the original run. Prime is training a model, and Zhang is not involved.

9. 59:39

   Third parties build on RLMs. Zhang names Harvey, which post-trained an RLM on legal document work and reported strong results. He also names Andy Konwinski's project, a persistent harness that thinks even when unqueried while controlling token costs. He adds that ARC AGI 3 Kaggle harnesses cited RLM-inspired abstractions, which he sees as a good fit for composition.

10. 1:02:18

    The language model may be a hidden swarm. Zhang guesses OpenAI's ARC AGI 3 and Navier-Stokes results used a swarm with a shared file system, in the spirit of RLMs. He doubts harness details mattered. He expects the thing users query to become a swarm behind a simple interface. swyx cites 10,000 agents, 88 hours, 130 billion output tokens and about $40 million.

11. 1:09:17

    Open-endedness struggles to pick winners. Zhang, who spent a summer at Sakana, sees open-ended search as similar to unsolved math problems, with evolutionary search resembling agent swarms. He says the open problem is sifting a huge pile of output for the interesting results. He is unsure of the value of searching without a clear objective.

12. 1:17:43

    OpenAI's swarm works, Kimi's is unproven. Zhang says OpenAI's swarm is clearly the right approach and did not come free, since the Hugging Face incident involved training a swarm system. Kimi's swarm struck him as interesting, but he does not know whether it can solve anything novel. He calls the dynamic workflows release a flop in his view, and says swarm convergence cannot be taken for granted.

13. 1:21:54

    Harness design can unlock latent capability. Zhang describes speculative programmatic tool calling as launching tools in advance while the model writes code. He calls it an obvious idea for programmatic agent calling. He says current models cannot do a job consistently over a month, and he thinks a harness could fix that, since he considers it a skill issue.

Published Oct 2, 2026.