Google's Multi-Agent AI Proves Long-Standing Math Conjectures
Google released Antigravity’s /teamwork, a multi-agent AI system pairing Gemini models with automated verification tools to help researchers solve long-standing research problems. The platform addresses two core AI limitations: single model attempts frequently fail to find correct solutions, and models lose focus during extended tasks as their context windows expand. /teamwork deploys parallel agent exploration paired with automated verification (Lean proof systems, test suites, hardware simulators) while partitioning each agent’s context to only the information it needs. Independent auditor agents verify results, preventing reasoning from earlier phases from contaminating subsequent work.
Research collaborations between Google and academic teams have produced multiple significant breakthroughs. The system proved a 2013 information theory conjecture, generating a formal Lean 4 proof requiring 50 million lines of code. Additional achievements include solving open problems in quantum computing and optimization, discovering new algorithmic improvements for matrix operations, and producing a polynomial-time solution to a problem previously requiring exponential time. Google is making /teamwork available externally for researchers to apply to their own projects.
Read the full article
Continue reading at Hacker News →This is an AI-generated summary. Read the original for the full story.