The Next Human Bottleneck

AI has made intelligence cheap, and I mean really quite cheap. Anyone can now write software, analyse a bit of data or launch a business in an afternoon, which is rather remarkable when you stop to think about it. So the bottleneck, I'd suggest, has moved from the technology to us. Our working memory is fairly small, and if you're running five AI agents at once you end up with five lots of output to check, make sense of and act on. AI takes the work out of each task, which is lovely, but it rather piles it back on when you try to manage them all.

That does, I think, change which skills are in demand. The first is handling cognitive load, and since none of us can really grow our working memory, that mostly comes down to experience and good systems. The people who get the most done probably won't be the ones juggling the most. They'll be the ones who've set things up so they only get called when something goes wrong, and are otherwise left in peace. The second is judgement. A hundred agents can launch a dozen businesses, which is all very well, but someone still has to know which one is any good.

The slightly awkward bit is that both skills come from doing hard things yourself. In one study cited by the OECD, maths students who used AI freely scored 48% higher in practice, but once it was taken away they did 17% worse than students who'd never had it at all. A version that gave hints rather than answers avoided most of that drop, which is rather encouraging. So I'd gently suggest that schools set harder problems, test children both with and without AI, and mark how they got to the answer as well as the answer itself. At the moment we mostly test what children remember. It might be rather more useful to test what they can do.