AI is not just AI. Or, why you should learn the difference

Software
,
1

This is not about the ethics of AI and machine learning (ML) as such, but about understanding how different types of AI and ML work, so you can make informed ethical decisions for yourself and not just add more noise to interminable discussions. Keep all replies technical.

It’s clear from the various discussions, both here and elsewhere, that most people simply don’t have a clue about what AI/machine learning (ML) actually is and how it works. Or, perhaps more accurately, conflates the various types of AI/ML, often seemingly believing that they all work the same way as LLMs (ChatGPT and chatty friends) or diffusion models (Dall-E, Stable Diffusion, Nano Banana and so on). Reality is that, apart from the foundational mathematics, they often have very little in common - whether it’s the training data needed or what they actually do when used. And some of what is branded as AI may not even have anything to do with machine learning, but simply got the label because some marketing department thought it would sell.

So before you throw yet another baby out with the bathwater, because “AI”, do yourself - and others - a favour by learning why masking, denoising or conjuring up non-existent cats really should not be treated the same way. Whether you end up being for or against AI is not the point (and I honestly don’t care either way), just that you cease being an ignoramus lighting yet another discussion on fire through your lack of knowledge.

If you’re in a hurry, Andy Hutchinson has a really good explanation (modulo a few technical inaccuracies).

If you want to go a little deeper, How AI Works by Ronald T. Kneusel is a great book. It’s a fairly short and easy read that explains the mathematics, training and functioning of all the various types of ML and AI, while only requiring basic arithmetic to understand. It actually goes in enough detail that you could probably build your own model from scratch, if you really wanted to. Seriously, read it. Yes. Now.

3Blue1Brown has a series of videos explaining various aspects of ML/AI. While it can be watched on its own, I would say it’s even better as a companion to How AI Works.
3Blue1Brown: Neural networks

11 Likes
2

Makes me think software should endeavour to use more correct terms - ML or AI - where applicable, so as not to add to the confusion.

3

ML is a field of AI.

4

These labels “AI” and “ML” are thrown about with very little rigor. I am particular irked by that “artificial intelligence” moniker, as it intentionally harkens back to that science fiction idea of a true intelligence. Yet what we actually built under that name is “just” a very fancy autocorrect.

What we call “artificial intelligence” is a particular family of transformer models (a subtype of neural networks) with extremely broad training data, essentially the biggest datasets ever assembled. They are trained –at staggering cost– on the sum total of recorded human artifacts with the implicit hope that out of so much humanity, a human spark must be conferred.

It is nothing short if astounding that such an architecture can even approach “intelligent” behaviour, but calling it thus is just a marketing ploy called a whishful mnemonic.

What is often called “machine learning”, in contrast, is smaller neural networks of a very similar architecture, but trained on narrow datasets. These are thus only useful for narrow tasks, such as denoising or image segmentation.

But as you might imagine, the borders between the two is very fluent. Scale one up and add some broader data, and you get broader applicability. Scale one down and restricted the data and you get a more targeted model. I find it more useful to think of them as “narrow”, or “broad” instead of more or less “intelligent”.

In reality, I think “AI” is a marketing term for “the thing AI companies sell”. Thoroughly tautologic of course, but that’s the entire point. They sell the dream, not the product. The product, on the other hand, is very fancy autocorrect. Which is tremendously useful in many regards, but “intelligent”? Is it even “artificial”, given that it is trained on organic human artifacts?

Sorry, I’ll get off my soap box. I just hate these terms so much.

4 Likes
5

Yes perhaps artificial generation would be more suitable than artificial intelligence.

6

On the contrary, “AI” has a very accurate, specific meaning.

It means “we are on the ascending part of the hype cycle, give us money.”

This usage has been consistent for five decades now. Once the current wonder technology reaches diminishing returns, it will revert to a more specific term.

4 Likes
7

I had a text book at uni in the 90’s labelled AI which included things like expert systems and trivially simple neural nets. AI has been a broad catch all term for a very long time and I find it odd that people are recently trying to restrict it to specific technologies.

1 Like
8

I think the issue is that older systems showed hints of intelligence, so artificial intelligence seemed sensible. No one was seriously claming the systems were human equivalent. Now that the hype is going all in for AGI, the term is being viewed from a different angle.

9

In my country, there is a chunk of metal which is commonly referred to as an Assault Rifle. There is a technical definition for Assualt Rifle, and that chunk of metal does not satisfy the definition. That chunk is fundamentally no different than a chunk of metal called a Hunting Rifle. But, even after years of ‘education’, people’s perception of that chunk of metal has not really changed.

The technical distinction is, for better or worse, a distinction without a difference for many people.