Latent Space: The AI Engineer Podcast

Academia is for Ambition — Alex Zhang, MIT

1h 41min

Alex Zhang (MIT, RLM author, GPU mode contributor) discusses the philosophy of taking big research bets in academia, arguing that PhD students should pursue ideas that industry labs ignore. He explains how Recursive Language Models (RLMs) enable compositional generalization through opinionated harness design, contrasting them with conventional agent frameworks. The conversation covers kernel optimization automation, the limitations of current agent swarms, and the need for better harness design to unlock model capabilities beyond coding and math.

Summarized by Podsumo

🎧 Listen 🎙️ Ask about this episode

✨ Key Takeaways

1

Alex Zhang argues academic researchers should take big, risky bets that industry labs ignore, citing RLMs and Sweetbench as examples of initially dismissed ideas that became influential.
2

RLMs use code as the only tool and offload context to a filesystem, enabling compositional generalization across tasks without relying on long-context transformers.
3

The GPU mode community has evolved from optimizing individual kernels to automated kernel generation via AI, but human expert guidance remains crucial for stability.
4

Current models exhibit jagged intelligence: they excel at coding/math but fail at simple long-running tasks; better harness design could bridge this gap.
5

PhD students should focus on opening new design spaces (like Jev's fast classification or loop transformers) rather than competing on compute with frontier labs.

💬 Notable Quotes

"If you’re not taking big bets in academia, just go to an industry lab – you have all the advantages there."

— Alex Zhang
"When you get a reaction like ‘what’s the point of this?’, it’s often a good sign – you’re exploring something people haven’t thought through."

— Alex Zhang
"The only thing that distinguishes an RLM’s value is whether you can train it properly. That’s non‑trivial."

— Alex Zhang