Montreal.AI paper: rejection-only experience transfer cuts candidate checks from 233 to 118
Ghost_Pilot_MD · x · 2026-10-23
Montreal.AI (Vincent Boucher) released a research manuscript, Autonomous Enterprises: Activate. Evolve. Scale., framing "transformational intelligence" as intelligence that continually improves our ability to create enterprises.
- A locally runnable launch kit starts from deterministic tools, executes an explicit mission, and keeps a bounded mechanism to improve the next launch; three offerings (range analytics, polynomial composition, dependency reachability) share one operating path without sharing budgets or authority.
- The core mechanism transfers rejection experience rather than acceptance: evidence-derived test prefixes expose defective improvement candidates, with proofs of finite-acceptance-predicate preservation and tight bounds on extra screening after unfavorable task shifts.
- Numbers: schedule learning cuts next-cycle candidate checks from 233 to 118 (polynomial composition) and 89 (reachability); a strong conventional ordering worsens the comparison and a second learned update is rejected.
- Positioned as executable organizational synthesis, explicitly not recursive self-improvement. Companion repo: github.com/MontrealAI/AGI-Alpha-Agent-v0.
More from AGI Musings
- Investor vasuman: all AI utility flows to customization, generic wrappers lose — vasuman · 2026-10-23
- 'I don't trust you': dev clashes over whether closed frontier models will actually patch real-world software — MoonL88537 · 2026-10-23
- AI cyber capabilities break the frontier-pacing narrative, safety voices warn — inductionheads · 2026-10-23
- Fewer Than 10 People Can Bend the Arc of Tech Civilization, and Most Aren't American — fkasummer · 2026-10-23
- New paper: AI reviews beat human peer review at its best, with policy implications — RexDouglass · 2026-10-23
- The paperclip problem is what a dumb person thinks is a smart person's moral dilemma — ericwdolan · 2026-10-23