So, how should we be using these artificial intelligences? The grand name for what are, at present, really just Large Language Models (LLMs). The obvious answer is “as much as they're useful for” but that doesn't, not particularly, advance our understanding all that far.
As ever, economics has some insight here. Not the full answer, but some insight. The distinction is made between a “substitute” and a “complement” (no, not compliment, this is not about how that dress looks nice). A substitute is when we use something instead of some other way of performing that task. A complement is something that aids us, makes us better, at performing that task.
This then goes on to have significant implications for what happens next: A spade is a substitute for having to use our hands when digging up potatoes. Now that we have plentiful spades, no one does use their hands to dig up the potatoes (and, at scale, a tractor etc is a substitute for a spade and so on).
A complement is, say, a book. We could learn everything by heart (which is why a lot of old stories are in poetry, rhyme and metre make memorization easier, very useful before books) but once we mastered writing and then printing we can write it all down. Have libraries and things. We can even say that written prose is a substitute for memorized poetry. But the books are a complement to our task of preserving knowledge. Education then becomes, really, learning how to reason and how to look things up. Instead of having to learn, by rote, all the facts. Calculators are a complement to our ability to do addition and long division and so on, even as they're a substitute for slide rules (and it's wholly possible that more than 50% of readers here won't even know what one of those is).
With this basic distinction in mind we can then think about AI and LLMs. What seems to be happening is that sometimes they are substitutes and others complements. I've seen Paul Graham (who founded Y Combinator, the Silicon Valley startup accelerator) say that AI is better than bad programmers in themselves, but allow good programmers to be even better and more productive.
So, AI is a substitute for bad programming, a complement to good.
Which is how we've probably got to view AI -- as we've got to view all tools in fact. They replace the need to do some tasks. Or rather, they replace some people doing some tasks. They also make some other people simply better at doing those tasks. Which is why there is the replacement of some of those people of course. We end up getting the things we need done by the better people at those tasks -- plus the tool -- and those not good at the task get freed up to go and do something else.
This is true of tractors and spades and it's apparently true already of ChatGPT.
The discussion in this newspaper that sparked this was the question of, well, how do we deal with AI in education? The answer is the same as the one in that discussion about using hands, spades or tractors in farming. Education is going to be, as it already is in agricultural colleges, the teaching of which tool to use at which time as a complement or a substitute.
True, true, that doesn't give us a specific answer but then good economics doesn't. It gives us a logical structure around which we can think about the process. Economics very rarely says “this is the way to do it!” and most claims of that certainty aren't good economics. But economics is very good at describing the system we should use to find out which is the way to do it. You know, freedom, markets, try lots of things and do more of what works, less of what doesn't.
We still use human hands to pick strawberries even though we've tractors to plough the rice paddy. AI's not going to be any different.
Tim Worstall is a senior fellow at the Adam Smith Institute in London.
