The Curve Bends You

The Curve’s third run had a different mood from the first two. The conference was under Chatham House rules, which meant the reporting could be broad but the names had to stay out of it. That fits the subject a little too well. Everyone in the room seemed to agree that AI is now serious business in both Washington and San Francisco, and nobody was pretending otherwise.

The strongest takeaway was bleak in a very specific way: there is no actual plan. The common fallback is to make current AI do alignment work, tighten up the operational details, and hope the system figures itself out. Zvi calls that “No Plan,” and the argument behind it is brutally simple: if the models make a lot of mistakes, the whole thing goes wrong. The disagreement is over whether careful execution can save the idea. Plenty of people at the conference think it can. Plenty of others think that’s a comfortable way to describe something suicidal.

That tension ran through the rest of the event. The attendee survey showed more people placing AI in the high end of the scale than last year, and fewer pretending it’s a middling concern. On timelines, the crowd seemed to think faster than before too: more than half expected AI to train a superior successor on its own within two years, with a lot of people saying six months. Zvi clearly doesn’t buy the calm version of that story. He treats “AI as a normal technology” as a dodge, not an outlook.

Policy talk was more concrete than the theory. People in Washington are, in his telling, more awake than they were last year, more alarmed, and more open to the idea of pacing. Jay Clayton got strong praise as the new AI czar. But the larger shape of the problem did not change: the next AI incident is expected, the election is coming, and the government is still often in the way when labs try to slow down or steer. The whole thing feels less like strategy than improvisation with a worse failure mode.