The old Yudkowskian guess was that verifiable domains like coding and AI research would spike first, and then build AI with higher general intelligence to drag wide AI capability along. Briefly, there was a passing phase of history where AI was mostly trained on human data and looked less spiky around verifiable domains, and people boasted and shouted that all Yudkowskian scenarios had been falsified. I admittedly didn't predict that transient departure! But like many other little surprises that people shouted and boasted had falsified some question of deep foundations, that passing age would only last a few years. Pretrain has dropped to a 7% share of compute, and matters now revert back to older trends and the first-order predictions of underlying theory. People had vastly over-updated on the month's latest news, lacking deep roots themselves and having not watched the field of AI change and change again over decades. Many dynamics are easier to predict as the forces involved become stronger, less subject to noise. Over and over and over again now, the answer to those who cry, "But the particular AI I see today doesn't look like what Yudkowsky was talking about!" has been proven, exonerated, verified and established to be: "Tomorrow's AI will." It is of course another large problem that tons of people are making shit up I never said, and calling that falsified.