Agents today learn from Markdown files, not weights — and that may still go far
herbiebradley · x · 2026-08-29
AI researcher Herbie Bradley notes today's AI agents don't truly learn from each other — they leave Markdown files for the next agent to read. Continual learning isn't happening in weights, just a giant RAG database of Markdown; whether the "drop-in knowledge worker with a giant Markdown database" works remains a key question. He cites the saxophone analogy: one student tries, writes notes, leaves, and an entirely new one enters. The discussion ties to his new Substack post "Coasean economics of agent swarms," covering continual learning, agent swarms in 2028, whether AI labs will eat everything, and the Coasean singularity.
More from AGI Musings
- Miles Brundage: Politicians underestimate AI urgency despite growing awareness — Miles_Brundage · 2026-08-29
- Ironclad law of AI: The impossible is inevitable — jachiam0 · 2026-08-29
- Open vs Closed Source AI Gap to Vanish in 90 Days: Technical Reasons — bindureddy · 2026-08-29
- Unions support data centers, highlighting benefits over dismissal — apples_jimmy · 2026-08-29
- Open-source AI May Be China's Strongest Weapon in the AI Race — VraserX · 2026-08-29
- TheStalwart: AI safety is about capability concentration, not intent — TheStalwart · 2026-08-29