DeepMind paper: Gemini-based agent discovers and Lean-verifies new rules for 11 Busy Beaver machines
rohanpaul_ai · x · 2026-10-22
A Google DeepMind and Cambridge paper, "At the Limits of Computability: Steps Towards Autonomous AI Discovery on the Busy Beaver Frontier" (arXiv:2610.22603), shows AI moving from answering human-posed math problems to open-ended discovery.
Key points:
- Why Busy Beaver: it's an uncomputable frontier — no algorithm can compute BB(n) for every n — making it a natural setting for sustained open-ended conjecture and proof beyond static benchmarks.
- Method: a Gemini-based agentic pipeline integrating pattern recognition, automated conjecturing, informal mathematical reasoning, and formal theorem proving in Lean, discovering Turing machine "accelerations" — compact rules summarizing many execution steps.
- Results: the system autonomously discovered and formally verified novel Collatz-like accelerations for eleven BB(6) machines with no prior documented analysis, testing edge cases in a simulator; classifications are provisional pending community acceptance.
- Authors include Marco Dos Santos and Tom Zahavy.
More from AGI Musings
- ARK's Sam Korus: Tesla Optimus academy and David Silver's super learner converge — skorusARK · 2026-10-22
- VraserX mocks Dario Amodei: those warning about AI want to decide who gets access — VraserX · 2026-10-22
- We're close to losing the ability to see what AIs are actually thinking — Puzzleheaded-King584 · 2026-10-22
- Wikipedia has lost 25% of traffic—1B+ monthly visitors—as AI intercepts readers — _akpiper · 2026-10-22
- François Fleuret's definition of understanding: no unconscious consistency checks ring an alarm — francoisfleuret · 2026-10-22
- Axios scoop: US banks detect unauthorized AI agents inside customer accounts — daniel_mac8 · 2026-10-22