Book smarts, street smarts, and whatever we’re calling AGI this week

Being told you’re book smart but not street smart is a fairly efficient way of being called clever and an idiot in the same sentence. You know lots of things, you can explain them at length, and apparently you lack the intelligence to recognise that everyone would quite like you to stop. Slightly insulting, obviously, but it does point to something we understand perfectly well about people: intelligence can be a surprisingly uneven business.

Then we get to AI, where solving increasingly difficult maths problems and writing increasingly complicated code gets treated as evidence that general intelligence is nearly here, or already here, depending on who’s on stage. These are impressive capabilities, and being able to solve a new problem goes well beyond spouting back something from a book. But there’s quite a jump from being astonishingly good at these things to being able to make sense of whatever the world throws at you.

Take buying a second-hand car. You can know the specifications, calculate a fair price and arrive armed with an impressive amount of research. Then you have to work out why the seller keeps changing the subject when you ask about the service history, whether the other interested buyer actually exists, and why someone you met ten minutes ago is apparently doing you an enormous personal favour. Knowing about cars gets you so far; understanding what’s happening in that conversation matters quite a lot too.

That’s the street smarts. It includes understanding what people mean, what they want and why those might be different from what they’ve actually said. It’s knowing when to question an instruction because something about the situation doesn’t add up. Much of this comes from being alive and having to deal with the consequences, including seeing the same people tomorrow after being an absolute pain today.

AI can learn about all of that from human accounts of it. Fair enough, people also manage to learn things without personally experiencing them. And expecting a machine to have endured an awkward dinner party before it can understand one would be a fairly odd requirement. But reading an account gives you a version of events that someone has already interpreted. Living through it means working out what’s happening while you’re part of the problem.

Perhaps enough interaction sorts that out, give a system access to the world, let it act and learn from what happens, and maybe it develops something we’d recognise as street smarts. Even then, the feedback needs interpreting though. Negotiate a thousand pounds off that car and you might conclude you’ve done rather well. The seller’s immediate agreement might even feel encouraging. Six weeks later, when the gearbox gives up, the same conversation looks somewhat different.

So there’s a reasonable question here about what our impressive demonstrations actually establish. How much evidence of practical, social judgement should come with a claim of general intelligence? Calling it AGI doesn’t settle that, however good the maths gets. A cafe can put “world’s best coffee” in the window without anyone convening a panel to inspect every other coffee on earth. Being street smart includes recognising that the person selling something may have chosen the definition that suits them.