What's this about?
Alex Zhang connects GPU-kernel evaluation, agent harness design, Recursive Language Models and multi-agent swarms. He argues that models often have more latent capability than a simple chat loop exposes.
What exactly is a Recursive Language Model?
An RLM treats a large prompt as external data. It can write code to inspect or split that data and call language models recursively on selected pieces instead of forcing one pass over everything.
Why is this discussion important?
full 90-minute timestamped transcript. The practical lesson is to pair generation with strong verification and expert priors: brute-force search can top a benchmark yet produce fragile kernels. Key sections begin at 00:07:38, 00:29:03, 00:44:24, 01:00:30 and 01:20:06.