[StrictDabbler]

Current AI doesn't give us what Hofstadter has said he wants from AI.

If you try ChatGPT on any of his creative analogy problems it fails completely. Given abc->abd, what does xyz go to? ChatGPT says xyd until you further prompt it.

It is a little disappointing that useful language is translation is possible with that little understanding of content. Most sentences a person might choose to say don't mean very much.

I'm depressed by that too. There's something about the facility and the emptiness of these models that suggests our ability to think is a lame trick.

I think "competent" is the wrong adjective, though. "Capable" is more accurate and the distinction there is the issue.

[gwern]

> Given abc->abd, what does xyz go to? ChatGPT says xyd until you further prompt it.

Because you chose to phrase it in a way predictably broken by BPEs (https://gwern.net/gpt-3#bpes), like so many character-based puzzles people 'disprove' GPT with, as Melanie Mitchell already demonstrated with your exact example 3 years ago: https://twitter.com/MelMitchell1/status/1285270704313610241

[StrictDabbler]

I'm sure she did use this exact example because Hofstadter and others have been using it since the mid 80s, yes. I don't think he invented the test. I have seen the chatbot fail at it and I've seen it succeed, more or less at random.

You are, however, missing the point.

If you hold a small dog over water it will make swimming motions. It didn't learn those motions. They're encoded into the genes that make its neural matrix and they go off even if it isn't actually in the water.

The ability of ChatGPT to perform (what we consider to be) extremely sophisticated linguistic operations while also being dumb as a sack of hammers about simple questions is upsetting.

A system that sometimes fails the abc->abd test but will easily rewrite the lyrics to "Baby Got Back" in the style of Shakespeare and then Bukowski and then translate it into passable Japanese suggests that many of the activities that we consider creative and useful are actually shallow, stupid, and easy imitations of things we've heard from other people.

I have no urge to disprove ChatGPT. It's extraordinary and powerful. I had just hoped that doing interesting things linguistically was a good measure of internal understanding.

With the way LLMs work we know there is no stateful model in there, no cognitive loop, no understanding, no eventual emergent consciousness. The weights are precalculated and it does not reconsider them. No aspect of the system is based on any possible notion of an "I" that is feeling and looking at itself. There is no memory space set up to store that selfhood and it forgets everything between conversations. The slate is wiped clean.

Yet here it is playing word games beautifully. Failing very basic mathematical questions and writing kick-ass rock lyrics.

As a proponent of emergent strong AI I'd assumed that these abilities depended on self awareness and that any AI capable of doing these neat tricks would be conscious. Instead it appears that self-awareness is an optional curse, an evolutionary side-effect that allows us to suffer and be disappointed in ourselves but is not necessary for creating poetry or visual art. The parts of T.S. Eliot we care about were just a dog wiggling his legs over something he thought was a pond.

There's a sadness there. The point is definitely not "look how dumb the robot is, hurr-de-durr."

[gwern]

> You are, however, missing the point.

No, you are. You are making a lot of hay out of a simple, dumb, trivial technical flaw, which is easily fixed, and reading deep lessons and 'sadness' into it. It's not interesting at all (which is part of why I'm angry I have to waste so much of my life explaining to people why their character-based gotcha is actually uninformative and irrelevant).

It no more tells us anything about deep learning or human intelligence than would waving a piece of paper 'ABC is to XYZ as etc' in front of a blind person and them promptly failing the test, even if they are a skilled song writer. Duh! Of course they fail the test! They're blind. Their eyes don't let them see the individual characters, anymore than a BPE tokenization lets a LLM see individual characters. You would be a fool to conclude that 'blind humans are not actually intelligent, and yet, he wrote such good lyrics, isn't it sad that the parts of T.S. Eliot we cared about were just some sort of wiggling dog leg' on the basis of failure due to such an obvious mechanical perceptual flaw. Maybe blind people are less intelligent on average due to their handicap, sure, maybe human intelligence is in fact a sad dog leg wiggling around - both of those are entirely possible - but you need a different measurement and it should have been obvious to you and everyone else before you wasted your time waving the paper. The waving paper told you nothing of interest whatsoever.