A useful metric for AI’s ability is the duration of a task it can complete (e.g., METR measures AI success on lengthy, expert-level tasks). State-of-the-art models can now complete tasks that take an expert almost a day, a horizon that doubles about every seven months. Human abilities grow along the same dimension from childhood to college: from coloring a page of line art to reading a chapter book, writing a five-paragraph essay, and completing a senior project. There is something unsettling about the juxtaposition of sixteen years and seven months, though.

The increasing time over which AI can work autonomously (from /goal to delivered project) means we must rethink education. To prevent AI’s enfeebling effects, I suspect we must divorce education from AI altogether in our early years, but then, at some point, compress the timeline for higher-education with a focus on application: using AI in a way that exceeds the value AI itself can provide. Of course, AI’s capability is a fast-moving target, about which it is hard to think.

What I do know is that traditional higher education scaffolds assignments across a semester (reading, proposals, outlines, drafts, presentations, etc.) and curriculum across a major (foundations, methods, writing intensives, capstones, etc.). This process has been valuable (to society and to me personally), but will it continue to be so? Increasing our ability to take on more complex tasks over longer periods is itself lengthy (many years) and expensive (tens to hundreds of thousands of dollars). A year or so ago, AI could do any one of the course assignments well; now it can do the whole semester’s sequence within an hour.

Extending a person’s horizon takes sixteen years and a small fortune, while AI’s is doubling within a single year. Yes, it cost a large fortune to train these increasingly capable AIs, but the steepness of AI’s capability curves is unsettling.