Hurtling through 2026

In mid-January, I made predictions about AI capabilities for the year 2026. By early March, my forecast of the 50% time horizon on software tasks already seemed clearly too conservative. And then a couple weeks ago, amid the flurry of news about dozens of esoteric theorems falling to AI, I decided that my prediction that AIs wouldn’t be able to get a paper into a top math journal had now clearly fallen. As it turns out, this was a bit more ambiguous than I assumed (as I’ll discuss below), but it prompted me to revisit all my qualitative predictions from that post. I posted a quick thread to X about this a few days ago, and I’ll go through the analysis in a bit more detail in this post. I made predictions in five categories: game play, logistics, video, game design, and math. For each category, I made a low-end forecast (a task I thought AIs had an ~80% chance of doing by the end of the year) and a high-end forecast (a harder task I thought AIs had only a ~20% chance of doing by the end of the year). I scored these predictions with a lot of help…