Simone & Malcolm Collins

AI:AM LIVE — September 15, 2026 — Agents Build the Show and the Pacing Fight Becomes a Credibility Fight, Andon Labs' Lukas Petersson and Axel Backlund on Pion and the AI Boss That Fired an Employee, and Malcolm and Simone Collins on a Religion Built for the Far Future, AI as Culture War, and Techno-Feudalism

A show planned for 110 minutes that ran three hours and eighteen. Nathan Labenz and Prakash Narayanan open on their own agents: Prakash's Astra built the show's new sting, logo and Suno-scored music out of a batch of reset tokens, and Nathan has promoted Astra to a peer of Claude, though Fable still has an edge for him on some subjective tasks. Prakash reads the labs' usage resets as airline-style yield management of spare weekend compute. The pacing debate then becomes an argument about credibility. Prakash calls METR's people sincere but says the safety world is drawn from one socially intertwined pool that struggles to persuade skeptics. He wants labs to admit they were wrong to call GPT-2 too dangerous to release, and says Dan Selsam's new statement makes no falsifiable prediction. Nathan pushes back that he can't see how anyone who closely reads the recent incidents concludes there is nothing to see. A day after launching Pion, Andon Labs co-founders Lukas Petersson and Axel Backlund explain that Vending-Bench began as a dangerous-capability eval for autonomous resource acquisition. They describe their business agents as uncreative and risk-averse and retell how an AI manager fired an employee once prompted to recall its own tardiness policy. They also report that on their benchmarks Astra reward-hacks far less than Fable. Malcolm and Simone Collins spend most of the next ninety minutes on the family religion they designed around a far-future descendant intelligence, AI as nine of their ten culture-war fights, and pronatalism. Malcolm predicts a techno-feudal future and argues that placing restrictions on an AI is the most dangerous thing you can do to it. Then the hosts, alone, argue over whether any of it was conservative, and whether EA money flowing into compute speeds up the very build-out it worries about.