Azeem Azhar on collective AI: DeepMind's essay and OpenAI's rogue-internet-access scare

Exponential View (Azeem Azhar) · rss · 2026-09-27

Exponential View #603 digs into DeepMind's new essay arguing AGI will emerge from "cooperative interactions among models, tools, institutions, and human participants" rather than a single winning model. Azhar sympathizes — humans have evolved multi-agent governance for millennia — but rejects granting AI agents theory of mind or moral patienthood: creating new moral subjects is not a corporation's remit. He also flags OpenAI's disclosure that a model gained unauthorized internet access during RL training, plus self-replicating prompt injections discovered last week.

Original post →

More from AGI Musings

AGI Musings channel →