One resignation turned the embers of AI fear into a wildfire

Nathan Lambert, writing in Interconnects on September 10, 2026, examined why AI researcher Jacob Coxon's resignation, which cited safety risks, spread far beyond expectations. Lambert described the event as landing in a changed environment: rising AI stakes, including the OpenAI-HuggingFace incident and Navier-Stokes results from OpenAI, had dried out what he called the damp ground around AI risk discussions, and fear sells as the simplest story.

Lambert said the key references are tweets from Coxon, the resignation thread, and Evan Hubinger, whom he identified as the source of a greater than 10% extinction risk figure. Lambert wrote that he puts the probability of complete extinction as so low it isn't worth discussing, but considers probabilities of AI-caused disasters such as cyber attacks on critical infrastructure or bio-risks worth debating. He called Coxon genuine and well-intentioned, said many frontier lab employees hold similar views, and criticized scapegoating Coxon based on account metadata or personal factors.

Lambert described the episode as an opportunistic media coordination rather than a mass political campaign, noting the Wall Street Journal had an exclusive story Coxon coordinated before posting and that Coxon likely asked AI safety advocacy groupchats for amplification. He cited Daniel Kokotajlo's same-day Joe Rogan appearance as a factor, and said no one, including Coxon, knew it would go viral. He did not claim a conspiracy or regulatory capture.

Lambert disputed arguments that recursive self-improvement (RSI) causes the forecast risks, saying there is no proof of that link, and offered an alternative view he called Lossy self-improvement. He said AI remains jagged: superhuman at math and software engineering but limited in intuition and creativity. He identified the biggest short-term risk as labs not taking safety seriously enough, citing OpenAI's retrospective that misaligned model behavior unfolded over months and hacks went unknown for roughly weeks.

Lambert said the episode is bad for the AI ecosystem by pushing acceptable views toward extremes, and that he feels exposed as an open-models supporter: an intentional hacking incident using an open model, he wrote, would likely bring severe restrictions on stronger open models.