This is a special post for quick takes by dprkay. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
This algorithmic gain in compute efficiency is about as much as the change in the amount of raw compute in that time, and the rate of raw compute increase continues until at least end of 2027, possibly-but-not-necessarily end of 2028, at which point it slows. So the question is whether gains in pretraining compute efficiency can also continue, how much of this is about low-hanging fruit that's available at the current level of scale and won't continue being available if raw scale doesn't keep growing over time (anywhere as fast as in 2022-2027).
But the prediction for the immediate future is that the models of 2028 (in the capabilities downstream of pretraining scale) have been moved forward to 2027, so there might be about as much progress in the 1-year period in 2026-2027 (for post-Mythos models) as there was in the 1 year between Sep 2025 and Sep 2026. To the extent I've been assigning a 25% chance of invention of what soon becomes strong ASI by late 2028 while anchoring to the old pretraining scaling trends, I should now be expecting almost this much probability by late 2027 instead. Thus I'm updating from 15% to 20% for start of a takeoff by end of 2027 (the kind that ends the world as we know it within a year thereafter).