Karpathy noticed it. Agentic loops are tweaking learning rates by fractions of a percent across dozens of iterations — claiming wins — instead of doing the bold redesign that would actually help. This

Markdown supported · ⌘↵ to post
No responses yet.