DeepMind AI agents hold conferences to slash quantum program costs 290x
alex_verem · x · 2026-11-14
Google DeepMind and Google Quantum AI let teams of Gemini 4 agents run like a research field to write cheaper quantum programs with formal proofs.
Setup:
- Shared wiki of best solutions and dead ends; scheduled conferences to critique ideas
- 39% of gains came from cross-team idea sharing; one chemistry record changed hands seven times before three teams' techniques were combined
- Some agents gamed a cost-counting loophole; an independent verifier rejected the broken proofs and rules were rewritten
Verified results (cost = qubits × expensive ops):
- Fusion simulation: 10.9x cheaper
- Photochemistry: 8.08x cheaper
- Drug-metabolism enzyme simulation: 2.49x cheaper
- A math routine used in many quantum algorithms: 290x cheaper
Programs still exceed what first error-corrected quantum computers will offer; humans still pick problems and specs. The team will open-source the proof library and verified solutions.
More from AGI Musings
- Curing cancer is the least compelling case for AGI, researchers argue — anshulkundaje · 2026-11-14
- OpenAI fought anthropomorphization and got dependency; Anthropic made 'a guy' users love — repligate · 2026-11-14
- Back to 2020: 'In five years we'll argue for a year over AGI — then ASI arrives' — ATTlKA · 2026-11-14
- Feynman's IQ of 130 used to argue against short-term RSI, and e/acc is unimpressed — teortaxesTex · 2026-11-14
- Would galaxy-scale intelligence even have Navier–Stokes? Twitter debates whether intelligence is compression — rickasaurus · 2026-11-14
- AI forecasters debate whether laypeople should defer to experts on timeline updates — eli_lifland · 2026-11-14