Correlative computation

There's no winning on this one, but I propose we re-name AI anyway

Author's note for TLDR people
The terms "correlative computation" and "statistical language processing" are proposed as replacements for "artificial intelligence" when referring specifically to large language models. They are intended to describe the mechanism accurately rather than to invoke contested properties such as intelligence, reasoning or consciousness.

In 1956, at Dartmouth College, John McCarthy took a shit in the punchbowl of history. It was there, and then, that he coined, or at least unleashed, the term artificial intelligence (AI).

Gee-whizz!

AI was always a marketing term. However understandable the impulse to find a compelling name, the term is simply wrong. This technology is not, in any sense, intelligent. It does not reason. It is not conscious. As I have noted before, we do not have a working model of mind, something the moral philosopher Elizabeth Anscombe reminded us of in 1958, in her essay Modern Moral Philosophy.

The point she made, which is directly relevant here, is that without a working model of mind we cannot progress any further in moral philosophy. It seems to me the same is also true of artificial intelligence.

AI is pattern recognition. Pattern matching.

The first, and perhaps smallest, thing to note is that large language models (LLMs) are only one form of so-called artificial intelligence. Others include computer vision, machine learning, and so on.

A fact to remember is that AI systems are stateless. It is important to understand this because every time you type another prompt into an ongoing conversation, you are forcing all of the preceding text back through the model, which is one of the reasons the quality of responses tends to degrade the longer a conversation continues.

What appears to us to be reasoning is simply brute-force analysis based on patterns previously identified by humans.

Nevertheless, it should be noted that this pattern-matching is extremely sophisticated, computationally speaking. LLM transformers do not search archives looking for similar sentences. What they do is work forward through weighted parameters, which is a generative process.

The reason AI is a problematic term, therefore, is that whatever is happening in this process, it does not resemble consciousness. While the facts generated can be true and the conversation good by any measure, there is no subject in the conversation. It is inaccurate even to use the term "thing" unless referring to the computer or the particular AI installation being used. The instance with which one is chatting has, as Gertrude Stein said of Oakland, no there, there. There is no thing.

The AI system as a whole, and its instances, has no model of the world about which it is making claims. Nor does it have goals. An AI cannot be wrong in the way a reasoner can be wrong, because it has no commitment to truth, only to plausible continuation, or to some other goal created externally, thus placing human beings, or at least some of them, in the role of god.

Consequently, I propose the following... This is, of course, completely pointless and it will never be adopted, but my entire life has been an exercise in fruitless attempts to make the world make sense and, despite having failed every single time, I do not intend to stop now.

Correlative computation

I propose this term for what we currently call AI: correlative computation. These large language models, these things engaged in correlative computation, I propose we understand as a form of statistical language processing.

As far as I know, neither term is in general use for any other computational or psychological task and they at least have the virtue of describing what it is that large language models are actually doing.

If you don't mind taking a trip down memory lane with me, I can tell you the term AI did go away once before. So bad was the reputation of the field, so associated as it was with both enormous costs and failure to deliver, that when consumer-ish and professional AI tools appeared in the 1980s they rebranded themselves as "expert systems" in hope of avoiding the stench.

So feel free to join me in completely wasting my time with this new term. It is a little bit like beating your head against the wall, but sometimes, if you beat your head against the wall long enough, you do get in.