If a PR Says “Artificial Intelligence,” There’s a Good Chance It’s Meaningless

It feels like I'm getting massively subsidized by the AI hype tho. Even if I were to pay 20 dollars a month for that, it wouldn't come close to covering how much it cost to train and host it.

I agree that it’s not usually an intentional mixup. From what I’ve seen it’s mostly a lack of rigor in defining terms.

1. On one view of intelligence, something is either intelligent or not. On another, some things are more intelligent than others, but there’s no clear cutoff between intelligent and non-intelligent things. On the first view, it seems quite plausible that actually existing ‘AI’ (e.g., GPT) doesn’t count as intelligent. On the second view, actually existing ‘AI’ seems to be at least somewhat intelligent: more so than most other software we’ve written. If the second view is right, it’s unhelpful in many cases to simply pronounce things intelligent and unintelligent.

2. By way of analogy, suppose I say that a walking route is quite hard. I might mean that it’s very long. Or I might mean that it’s hilly. Or I might mean that it’s very boggy. Each is a perfectly good reason to say that the route is hard. So a walk that’s merely quite hilly counts as hard, even if it’s fairly short and the ground is dry.

We might say that attributions of intelligence are similar. If so, we can attribute intelligence to systems for many different individually respectable reasons. Perhaps a system is intelligent because it can respond to novel situations in some appropriate way. Perhaps it’s intelligent because it predicts a certain statistical parameter correctly. Perhaps it’s intelligent because it’s small but can correctly deal with a wide range of situations.

If the analogy is right, it would be odd (perhaps wrong) to say that a system good at one of these just isn’t intelligent because it falls down on the other measures. If so, surely GPT counts as intelligent for at least one respectable reason or another.

On the other hand, suppose I call someone tall. There’s only one way to be tall. Being fat, or having muscly arms, or having long legs but a short torso don’t count. So the analogy doesn’t apply to all concepts. Does it apply to intelligence? Initially, it might seem that it doesn’t: surely there are lots of ways to be intelligent. But I’ve heard that the psychometrics literature suggests that all these measures correlate to a great degree, and statistically can be predicted by a single-factor model (thus ‘g’). That might suggest that there really is only one way to be intelligent, and that appearances are misleading.

I am not familiar with the psychometrics literature, so I wouldn’t know; maybe the single-factor model is wrong. But my point is this. Even if the single-factor model is right, it’s only been shown to be right about humans (so far): their statistical base has comprised humans. So maybe a multi-factor model of intelligence works better for would-be machine intelligence. For example, perhaps arithmetic ability in humans is predicted well by a single factor; maybe it’s even reducible to some single form of intelligence. But we can obviously separate arithmetic ability from e.g. analytic ability in computers, to an almost arbitrary extent, by making very good calculators. (And LLMs are often not very good at arithmetic, though I gather that’s being improved.) If that is so, intelligence is more like difficulty of a walking route than tallness. And so that’s another reason to avoid straightforward denial that would-be AI is or could be intelligent.

3. It’s quite plausible that no presently existing would-be AI should count as intelligent. But we don’t know whether that’s a general limitation or not. And if there are general limitations, how general are they? For example, maybe LLMs couldn’t be intelligent but some GOFAI type thing could be. Or maybe we simply need a new architecture.

One argument we could read in Fodor is that neural networks have to implement a so-called language of thought to be meaningfully intelligent. That would be quite a general limitation, though arguably one we could overcome. (I’ve always been a bit confused by what Fodor really meant by a language of thought, and in particular what he required of mental representations, but I haven’t made a full study of him yet.)

A much stronger argument from J.R. Lucas is broadly ‘anti-mechanism’, which would roughly include everything we can presently engineer or can be run on a Turing machine. This is very strong, and not many people agree in my experience.

—

The point of my comment is that these matters are complicated, and the comment above didn’t really address these complications. Sometimes nuance doesn’t add much or isn’t worth it. (I quite like Kieran’s ‘Fuck Nuance’ as a lesson for all theorising, not just sociology.) But sometimes it does matter. ‘[T]here is no artificial intelligence’ is hasty enough to require a response.

Disagree - GPT's mechanics are that it guesses what word/token should come next, however it is now displaying emergent capabilities which do show an agent which can generalize. Maybe not to 'AGI' levels, but it is starting to show little signs of it.

> It can't do basic maths, or even follow instructions. It can't make valid moves in a game.

It can do basic maths - better than most humans (but it cannot do maths better than a calculator). If I ask GPT4 to calculate 51 times 102 it will get the answer, despite most humans struggling with this task. It does this just via internal representation without any specific 'calculator' API call or functionality.

It can follow instructions - if I tell it to take the sentence "abcd" and then tell it to follow the following steps: "1) swap b and c around, 2) reverse the string, 3) ignore step 1", it will output "dcba" and describe how it reached that output

It can make valid game moves - Without being formally taught chess it will make valid chess moves most of the time (see: https://www.lesswrong.com/posts/xyjhFCSSXZsW6HDBb/a-chess-ga...). It made 29 valid moves before making invalid moves. Again, GPT4 has not been specifically trained to play chess, it just has picked up some chess rules and games while reading the internet and this has somehow formed some sort of internal representations for chess within it's model. I can also describe a new/novel game or puzzle to GPT4 and it can have a go at playing and solving it.

What it can't do is learn at the moment, although learning isn't part of a lot of definitions of intelligence, which is why we have to define intelligence before you can say if it is/isn't intelligent.

Sure, we can also associate supply chain and logistics bin packing, inventory forecasting, existing warehouse automation--including computer vision and such.

But once you get to NLP, deep learning, and LLM, doesn't it kind of go off the rails?

Lather, rinse, repeat.