"AI" is a Terrible Name

Finkle-McGraw brightened a bit. "You know, when I was a lad they called it A.I. Artificial intelligence."

Hackworth allowed himself a tight, narrow, and brief smile. "Well, there's something to be said for cheekiness, I suppose."

-- Neal Stephenson, The Diamond Age, 1995

Naming things is hard¹. 😂

Sadly, it's also important. Giving something a name is, at minimum, giving everyone who will ever use it a hint about how you want them to think about it. There are a lot of different ways you can approach this, but whatever you're naming, it's worth considering how you want people to think about it, and about how you should reflect that in its name.

Any such consideration will reveal that "artificial intelligence" is a terrible name for our current crop of tools, because they are not intelligent (even though they're obviously artificial).

Our current tools are Large Language Models (LLMs), which I've taken to calling "shoggoths". Glossing over a huge amount of very important detail, they're all basically glorified Markov chains: they take a godawful amount of training data, slice it up into features represented by tensors of absurdly high dimensionality, and generate output by looking at what sequences of these tensors happen in the training data.

Shoggoths can obviously do some really impressive things this way. A properly-trained shoggoth is far, far better than I am at writing CSS and JavaScript, for example², and while shoggoths may or may not be better than I am at Go, they're certainly far faster. This opens some really cool doors, but it's more about mechanical brute force than intelligence.

We can draw an analogy to calculators here. A calculator is far better than I am at doing arithmetic, but everyone these days understands that a calculator is just a tool that uses a kind of silicon-based brute force to do things inhumanly fast. No one today would claim that calculator - or a calculator program running on a computer, or the computer itself - is intelligent.

Back in the early days, though, people would talk about how the calculator or computer "could think"³. It took awhile for the collective "us" to be educated enough to get past that.

I think that's pretty much where we are with LLMs right now. Lots of hype, lots of misunderstanding, lots of misuse because the name so encourages us buy into the hype. (And sure, there are folks who argue that the glorified Markov chain is all human intelligence is, but I don't buy that -- good topic for a future article, maybe.)

What would be better? Stephenson in The Diamond Age actually coined a name I like far far better: "pseudo-intelligence" or "PI". It's a throwaway in the book, but it fits ever so much better: what we have right now are a crop of tools that mimic intelligence even though they aren't intelligent.

I'd forgotten about the term "PI" until I reread The Diamond Age while recovering from surgery (🤦‍♂️). But I've wanted something better than "AI" for a long time, and who knows, maybe it's not too late.

A constant reminder that these things are only mimicking intelligence would be a good thing.