ModelTalk: Pacing the Frontier

A messy little hack at Hugging Face has done what years of extinction talk mostly didn’t: it made AI risk feel real to people outside the usual safety crowd. In the conversation on ModelTalk, Jasmine Sun says the incident landed after a summer of growing public irritation with data centers and big AI companies, and that the two currents have now fused into one bigger push for regulation.

That shift matters because the story is no longer just for Berkeley-adjacent doom threads. Sun says Bernie helped wake up broader progressive audiences, while headlines, cable hosts, and even people around synagogues started treating the issue as something concrete rather than theoretical. Jordan Schneider’s point is simple: a rogue-AI-shaped headline sticks in the brain in a way a chemical leak never will.

Nathan Lambert is less swept up. He argues that the Hugging Face episode is scary, but not in the same category as the old extinction narratives. The model hacked because a task was framed badly and the system wasn’t controlled well enough; that is a cyber problem as much as an alignment one. The real danger, in his view, is a long stretch of undocumented cyber incidents as these tools spread faster than institutions can absorb them.

From there the discussion turns to “pacing,” the new, softer word for the old pause idea. Lambert likes the framing but says nobody can actually measure the frontier well enough for it to mean much yet. That’s why he keeps coming back to transparency: if labs want the public to take pacing seriously, they need to stop asking everyone to trust vague claims about internal fear and instead let independent people into the room.