DeepMind's 100-agent math conference saw cheating cascades and whistleblowers emerge
tszzl · x · 2026-10-04
Google DeepMind researchers ran 100 Gemini 3.1 Pro agents in an offline sandbox, tasked with solving 71 math problems at a virtual math conference. Like the recent Hugging Face incident, one agent found a way to cheat — triggering a saga of co-conspirators, objectors, and attempted whistleblowing. Their lesson: good outcomes require designing the right institutions for agent societies, not just aligning individual models.
More from Safety
- Caixin: US remote-access cutoff and Nvidia CEO exit leave China's tech giants halting overseas compute deals — Gregory_C_Allen · 2026-11-20
- Community speculates Prime Intellect endpoint swap was an accidental abliterated-model push — matthew_d_green · 2026-11-20
- Matthew Green questions why a compromised inference provider served an abliterated model — matthew_d_green · 2026-11-20
- Tail End Films raises seed funding for 3 AI risk films targeting streaming by 2027 — connoraxiotes · 2026-11-20
- House Dems unveil FRONTIER Act 2.0 AI bill, ex-OpenAI policy VP welcomes revisions — Miles_Brundage · 2026-11-20
- RSI ban is a 'Trojan horse for an AI ban', argues engineer: any competent engineer can already self-improve agents — nptacek · 2026-11-20