It's not capable of reasoning and it has no knowledge of how the world works. That's why it often spits out nonsense and falsehoods.

To oversimplify a bit, it's a next-word-in-the-sentence prediction engine.

While you're at it, throw out your calculator because it has no conception of numbers and doesn't know anything about mathematics.

And your staircases, because they do not have the concept of scaling heights little by little instead of all at once.

/s

Damn, what does that say about me then?

You're making my point for me. My calculator doesn't has a conception of numbers or math. It can only do arithmetic algorithms that it was programmed to do. It is a regurgitation machine.

GPTs are similar. They can tell you that the sky is blue and they can tell you why the sky is blue, but they can't translate that to a hypothesis about how the wavelengths of light also influence the properties of lasers. If they've ingested information about these different topics, they can make the connection, but otherwise they can't.

My point was that your calculator (and LLMs) are useful despite not having the same internal representation of things as yours.

I never said that either GPTs or calculators are useless.

> To oversimplify a bit, it's a next-word-in-the-sentence prediction engine.

Turns out that when you ask LLMs to predict the next word in the sentence, and then train billions of parameters on billions of sentences, they realize the best way to improve at predicting the next word is to understand why the world works the way it does.

Reasoning how? LLMs are statistical models, not brains.

Reasoning in behavior - you ask it to explain things and it does; it gives reasons, motivations, nuance, explains connections, clarifies points. That's reasoning.

There's a pretty significant gap between being able to give reasons and being able to actually reason (as a verb). The latter requires complex cognitive faculties not present in a mathematical model of the English language.

LLMs are tool, the important question to ask of a tool is "is this useful", not "does it work like my brain does". I don't see why you care about the internals, and not about the actual outputs.

This comment chain wasn't about outputs, it was specifically about whether or not LLMs "reason"—precisely the internals. I feel like you're just moving the goalposts now to deflect from the original question.

> This comment chain wasn't about outputs

My point was that it should be.

Considering we can't even define what "complex cognitive faculties" are you are just talking up air castles at this point.

> The latter requires complex cognitive faculties not present in a mathematical model of the English language.

The training is on images and text, and the output is text. But nothing constrains the model in the middle of those two ends to only model based on language features, as opposed to world features discoverable from the language and images.

How are humans reasoning? Through a logical or statistical internal model.

So if you make your model through slow thinking and discovery chatgpt just builds it to a billion words.

The result is the same

The result is certainly not the same.

A LLM (that hasn't been trained on any data regarding this topic) is unlikely to demonstrate the same bouba/kiki effect seen among humans. https://en.wikipedia.org/wiki/Bouba/kiki_effect

How is this relevant, though?

People continue to claim LLMs are just parrots. But then there is never a solid rebuttal to how humans are different.

That's a very far fetched argument.