Academia is for Ambition — Alex Zhang, MIT

Future "language models" may be invisible agent swarms/harnesses rather than single decoders, and most current harness designs already converge to the same thing.

• RLMs explained: Recursive Language Models restrict a harness to one tool—code execution (e.g. IPython)—letting a model call itself recursively, offload context to disk, and spawn sub-agents, making each sub-call stay 'locally in distribution' even when the overall task is novel.
• Harnesses converge, mostly: Alex Zhang argues Claude Code, Codex, and Pi are structurally nearly identical ('trajectory as a prompt' loops); a cited 'harness tax' paper supports that most harness design choices don't matter much.

Themes, sentiment and reasons are AI-written paraphrases, checked so they never quote an episode. They may contain errors. Not investment advice. Episodes remain the property of their creators.