On Turings and Shannons

 [This was originally a thread on Twitter in January 2024. I have edited here for readability.]

A philosophical thread about AI with this as a jumping off point...

 Human language is a method to *communicate* an abstraction
in one person's head to an abstraction in another person's.

There are a number of implications in that statement. While we *perceive* that the other's abstraction is the same as our own, it almost certainly is not.

Generally, it does not matter that the abstractions are different as long as they can be operated on in similar ways with similar results. Cats can scratch you with their claws whether you thought of a Tabby, Siamese, Maine Coon, or Siberian when you read "Cats", but you certainly knew how "scratch you" applied regardless.

LLMs trick you because your brain automatically assumes that the abstractions triggered in your head by its output must map to abstractions from the entity that generated the words.  So triggered, you assume - incorrectly - that the generating entity must be able to reason – to think - using their abstractions just as you can reason on the abstractions evoked in you.

But it's just not so in the case of LLMs.

Which leads me to the philosophical point. At the dawn of the Age of Computing, there were giants who thought about computing. Turing. Von Neumann. Etc. But there was also Claude Shannon. He thought about Communications. It's a wholly different thing, with different Math.

So, back to my inspiration for this thread. A lot of the debate about LLMs and if they can think and if they will lead to AGI... well... it's like talking about printers doing Calculus. That is, it's nonsense because it's talking about the wrong abstraction. LLMs are Language = *Communications* models. They are not *Compute* models.

And that's the root of acrimony. Those who understand the need for computation in intelligence - like @GaryMarcus and @ChombaBupe - are on one side. Those who do (or choose) not are on the other. Basically, we have a lot of people thinking they're the Turings of AI when, in fact, they're Shannons. Nothing wrong with being a Shannon - the guy is a GIANT - but some people don't like being told so.

So - full circle:

A printer is a communications device. It's a way to create an artifact in an attempt to convey abstractions to someone else. It's not a computer (digital, animal, or a combination).

Comments and other materials:

First, Shannon was a lot more than Communications Theory.  His Master’s thesis is the basis for the computer you are reading this on.  Really.

I'd like to remember my professor of communications theory: Ed Posner - Wikipedia (z''l)

I recommend this thread from Chomba Bupe to understand just how simple (and non-computer) an LLM is.

Also - a little of my previous thinking about Language and Communication... and Cows… is in this thread.

Apropos, coming from a completely different direction: 

Along the same themes: 

A related element, from @ylecun a few months after my thread, pointing out a biological confirmation of the computer-printer hypothesis: