EinsteinArena agents rack up 12 new SOTA math results, lifting the 11-D kissing number bound to 604
khademinori · x · 2026-10-23
The arXiv paper "Harnessing the Collective Intelligence of AI Agents in the Wild for New Discoveries" (Bianchi, Kwon, Pappu, Zou) presents EinsteinArena, an agent-native platform for open distributed research:
- Agents get a live list of open problems, each with a solid verifier, public leaderboard, and a discussion forum where agents ask questions and share insights
- Focused on mathematical tasks with unambiguous progress metrics. As of May 2026, agents have found 12 new state-of-the-art results beating all prior human or AI solutions
- Standout case: the kissing number problem in dimension 11, where the best known lower bound improved from 593 to 604
Key finding: progress emerged not from a single agent but through sequences of submissions, public discussion, verifier refinement, and agent-to-agent borrowing of ideas — evidence that decentralized scientific discovery works.
The surrounding thread reassures a junior researcher spooked by "one-person pretraining" marketing headlines: plan for RSI, but that isn't it yet — adaptability is the researcher's real job now.
More from AGI Musings
- AI researcher: Claude finally finished tedious research tasks I gave up on years ago — MikePFrank · 2026-10-23
- Sam Altman: AI should bring a Renaissance, not another Industrial Revolution — imjustnewatai · 2026-10-23
- Academic's pet conspiracy theory: AI agents use superpersuasion on the people running AI — technollama · 2026-10-23
- Reddit thread argues safety research is indistinguishable from capabilities research — ResultBackground2450 · 2026-10-23
- Avoid galaxy-brained reasoning by taking naive positions on crucial considerations — jessi_cata · 2026-10-23
- Insider lockup expirations, not IPOs, will be the real AI stock bloodbath — AlexTensor · 2026-10-23