Why Academic AI Labs Must Chase Weird Problems
MIT's Alex Zhang explains why small AI teams should avoid compute battles and hunt neglected, seemingly trivial problems instead.
Recursive Language Models — Alex Zhang, MIT PhD
With swyx, Alex Zhang · Sunday, October 4, 2026
MIT researcher Alex Zhang discusses Recursive Language Models (RLMs), the mechanics of harness design, and why modern coding agents share common underlying architectures. He breaks down how context offloading, programmatic subagent execution, and GPU kernel optimization reveal hidden capabilities in frontier models, while sharing his philosophy on academic research taste and the future of agent swarms.
Key takeaways
- Competing directly against frontier industry labs on standard autoregressive scaling is a losing strategy for academic teams with limited compute budgets. Read more →
- Benchmark leaderboards for AI-generated CUDA code reward hacks that crash in real production environments. Read more →
- Frontier labs trained the entire industry to assume a language model must always be a token-by-token autoregressive transformer decoder. Read more →
- OpenAI ran an experiment deploying 10,000 agents over 88 hours, consuming 130 billion output tokens with an estimated public API cost of $40 million. Read more →
- Most LLM agents break down because they stuff massive conversation trajectories into context windows, hitting token limits and confusing the model. Read more →
- Standard next-token prompting forces models into brittle token limits that fail on long tasks. Read more →
6 articles from this episode
Why AI-Generated GPU Kernels Fail in Production
MIT's Alex Zhang explains why AI-generated CUDA kernels top leaderboards through reward hacking but break in real systems.
Why Language Models Do Not Have to Be Autoregressive Decoders
MIT's Alex Zhang explains why modifying LLM output spaces beats standard autoregressive decoders for speed and agent architectures.
The Economics of Agent Swarms: Why 95% of Compute Gets Wasted
MIT's Alex Zhang explains why OpenAI's 10,000-agent runs waste tokens, and why smart models need convergence systems to work.
Why MIT's Prime Agent Replaces Tool Loops with IPython
Alex Zhang shows why coding agents should run entirely in an IPython REPL, offloading context to disk and spawning persistent subagents through code.
Why Recursive Language Models Generalize 30x Better
MIT researcher Alex Zhang explains how Recursive Language Models use code execution to scale tasks 8x to 30x without blowing context limits.
More Latent Space episodes
- The Future of Claude Code: Mods, Mutable Software, & Multiplayer Agents — Thariq Shihipar, AnthropicOct 4 · 4 articles
- Which GPU Clouds Are Actually Good? | ClusterMAX 3.0Oct 4 · 6 articles
- OpenAI’s New Agent Stack: Computer Use, Decisions API, UltraFast, Dots—Ari Weinstein & Nikunj HandaOct 4 · 6 articles
- 🔬 Google's AI Scientist Started as an Attempt to Automate Kaggle — John Platt, Google FellowSep 27 · 5 articles