Raffi Krikorian

AI:AM LIVE — September 9, 2026 — A Pretraining Researcher Quits Both Labs, Raffi Krikorian on Project Glasswing and the Agentic Web, Mike Rizkalla's Deliberately Generative-AI-Free Kids' Companion, and Paul Christiano Joins OpenAI's Board

Nathan Labenz and Prakash Narayanan open on Jacob Coxon's resignation — three years of pretraining work across both OpenAI and Anthropic, ended with a public warning that both labs are racing to self-improving superintelligence and gambling with our lives — and split on it: Nathan endorses the core argument, Prakash weighs it against roughly $100 million in lab earnings and calls the 'nobody else is wiser' framing naive, before the two land on what a government could actually impose. Mozilla CTO Raffi Krikorian then takes the practical edge of the same problem, opening with his personal agent man-in-the-middling its own network traffic to log a meal in MyFitnessPal — 'an idiot savant on hacking' — and moving through Project Glasswing, Mozilla's security partnership with Anthropic; the several-hundred-thousand-dollars-a-month cost of continuous AI code scanning; the CQ project for cross-agent knowledge sharing; and the coming split between a human web and an agentic one. Mike Rizkalla, co-founder and CEO of Snorble, makes the opposite bet: a screen-free plush companion for young children built deliberately without generative AI, on a small fixed-intent model with curated scripts, radar instead of a camera, and a safety philosophy written six years before the legislation now catching up to it. The closing returns to lab coordination via John Schulman, argues over an Operation Warp Speed-style antitrust safe harbor, and catches Paul Christiano's appointment to OpenAI's board and Safety and Security Committee as it lands.