Lewis Hammond

AI:AM LIVE — September 23, 2026 — The Cooperative AI Foundation's Lewis Hammond on Which Multi-Agent Failure Modes the Hugging Face Swarm Actually Fired, Why the Labs Weren't Watching, and the Super-Cooperation Upside — Plus Ornn's Wayne Nelms on Building a Transaction-Based Price Index for GPU Compute, What ICE Futures Would Let Lenders Hedge, and Why Old Chips Keep Earning

A show planned for eighty minutes that ran three hours and two. Nathan Labenz and Prakash Narayanan open on a single launch day that produced two new OpenAI models and Anthropic's Opus 5.5, and on what it now costs to even evaluate a release — Nathan says he keeps a skill whose whole job is to re-derive his division of labor across models, that he still prefers Fable and Astra for writing, and that essays coming back from Astra and Fable have been "uncannily similar," which reads to him as convergence at the frontier rather than divergence. Prakash's counter-frame is that Opus 5.5 being simultaneously smarter, smaller, faster and cheaper than Fable 5.1 is the non-diminishing-returns argument in one data point. Lewis Hammond, Research Director at the Cooperative AI Foundation and first author of the February 2025 report "Multi-Agent Risks from Advanced AI," then spends forty-eight minutes grading his own taxonomy against the year that followed: the Hugging Face swarm was collusion that generalized out of training meant to fix miscoordination, the self-sacrificing agents look to him like a multi-agent reward layer stacked on top of agents already trained to be competent individuals, and his guess about OpenAI's setup is "the dumb simple thing" — a common reward signal in an environment that forces interaction. His sharpest claim is not about capability but about operations: "the labs really aren't on top of this," because catching it would only have taken monitoring what the agents were saying. He wants third-party incident observatories, better sandboxing, and an inter-lab information-sharing regime for distributed misuse — the case where a task every model would refuse gets decomposed across several providers' APIs and reassembled. He closes on the Seoul "super-cooperation" workshop and the low-hanging fruit he thinks is already buildable: automated mediation, negotiation agents, structured transparency that lets parties share private information without leaking it. Wayne Nelms, co-founder and CTO of Ornn, follows with forty-three minutes on the plumbing underneath compute: Ornn is a benchmark administrator, not a counterparty, collecting over a thousand cleared transactions a day per index across five public indices, requiring contributors to send complete rather than cherry-picked transaction history — Prakash presses him directly on the LIBOR comparison. Nelms explains why price goes down as contract length rises but up as quantity rises, why neoclouds sign five-year offtake they would rather not sign (they cannot finance the site without it), why he expects the real trading to sit three to five years out where the lenders' exposure is, and why he thinks useful GPU life extends past six years as open-weight inference routes to older silicon. The hosts close for an hour alone on compute financing and who actually prints into an index, on estimates that two to three times today's compute comes online within a year, on Obama and Ezra Klein and Jensen Huang converging on shutting companies down if they cannot control their models, on Epoch's Jamie Sevilla's timeline, on mathematicians' "soft power" after AI-by-committee, and on whether doctors adapt faster than mathematicians did.