The absolutely relentless deluge of breaking news from the AI world over the last few months has been pretty much impossible to adequately keep up with,1 but the following is possibly the most important diagram I have encountered so far in 2025.
It is from a report released yesterday by the AI evals organization METR. In short: Due to the highly multi-dimensional nature of intelligence, we have so far been struggling to find a measure of AI capabilities that is appropriate for predicting when to expect AI with transformative societal effects. Since AI capabilities seem bottlenecked by the their ability to devise and execute plans in many steps, METR's proposal to measure capabilities in terms of how humanly long tasks AIs succeed at seems very promising. Their main result is that the length of tasks that AIs can handle seems to be growing exponentially, with an astonishingly short doubling time of around 7 months. Obviously the trend is in no way guaranteed to continue, but... extrapolating it just a few years into the future is vertigo-inducing.
Here is METR's report, and here is their summary blog post about the work.
Footnote
1) Anyone wishing to nevertheless try to do so is strongly advised to follow Zvi Mowshowitz' newsletter Don't Worry About the Vase.
Inga kommentarer:
Skicka en kommentar