What's this about?
Z.ai reports roughly 3× serving-throughput gains in under two weeks on a 100,000-plus domestic-accelerator cluster; a GLM-powered Infra Agent did much of the work, with human goals, feedback design and high-risk review.
What exactly is dense feedback?
Local, cheap and objectively checkable tests tell an agent why a change helped or regressed. This is more useful than one slow, opaque end-to-end score, and does not establish autonomous successor training.
What's the community saying?
HN 379 points, 264 comments.
- Readers discuss domestic-chip scale and the impact of export restrictions.
- Some question sweeping RSI rhetoric versus the concrete engineering work.
- Users compare serving economics and coding-subscription pricing.