MIT's Alex Zhang: Academic freedom requires taking bold steps in AI research

Summary

In a Latent Space podcast episode featuring MIT PhD student Alex Zhang, the discussion covers his research on Recursive Language Models (RLM), GPU core generation, and multi-agent swarms. Zhang also argues that 2026 will be defined by models viewing their own prompts not as fixed inputs, but as mutable objects.

Why it matters

Zhang's RLM framework proposes that a model can iteratively rewrite its own context and delegate subtasks to sub-agents. This approach represents a concrete architectural shift from the current prompt-response paradigm. Early results like Prime Agent, which claims to have solved ARC-AGI-3 before OpenAI's Astra, are drawing attention. If the thesis that 'the scaffold is a compositional generalizer' holds true, it could transform how research labs configure agent scaffolds far more deeply than simply scaling model weights.

Highlights

• RLMs enable iterative self-improvement within a single inference loop by allowing a model to view its own prompt as an external, mutable object.
• Prime Agent, a self-modifying RLM scaffold, became the first system to approximately solve ARC-AGI-3, outperforming OpenAI's Astra.
• OpenAI's 10,000-agent experiment generated approximately 130 billion output tokens, with an estimated cost equivalent to $40 million. However, a significant portion of the swarm's search process was wasted, and convergence remains difficult.

Critical take

This content is merely a set of notes from a 101-minute podcast episode presented as analysis. The provided 'transcript' cuts off after two minutes; consequently, readers find no verifiable technical depth beyond promotional phrases like 'clearly rising superstar PhD student' and a jargon-heavy topic list. Claims such as '2026 will be the year of RLMs' are the speculation of a single researcher, not a proven trend. Furthermore, the ARC-AGI-3 result is cited without sharing methodological or reproducibility details.