Brooks has done something I appreciate a lot: he turned a phrase.
> stop confusing performance with competence
You can safely skip the rest of the article. That sentence gives you all you need, because you are competent.
If you want a little more meat:
> The example I used at the time was, I think it was a Google program labeling an image of people playing Frisbee in the park. And if a person says, “Oh, that’s a person playing Frisbee in the park,” you would assume you could ask him a question, like, “Can you eat a Frisbee?” And they would know, of course not; it’s made of plastic. You’d just expect they’d have that competence. That they would know the answer to the question, “Can you play Frisbee in a snowstorm? Or, how far can a person throw a Frisbee? Can they throw it 10 miles? Can they only throw it 10 centimeters?” You’d expect all that competence from that one piece of performance: a person saying, “That’s a picture of people playing Frisbee in the park.”
---
So I've calmed down. Now what? The problem isn't only that this train is flying off on a tangent: it's that it's off the rails. What rails should it be on?
The problem, as I see it, is narrative. As soon as we called it "AI", that wrote the Genesis of the Scripture of the cult. In this new religious movement, God is spelled L-L-M. Back here in reality, LLM isn't a God; or even a person at all.
That's the mistake: personification. A person can perform, but a performance can't person.
---
Narrative is a powerful tool. It's why we're so excited about Natural Language Processing in the first place. Ever since the very origins of software, the power of narrative has been so close, but always still just out of grasp. Do we even know what we are reaching for in the first place?
In a sense, we have a part of it: explicit definition. What Chomsky categorized "Context-Free Grammar", we have made into programming languages. What they are missing is implicit inference: context.
That's what LLMs do. They use inference to model the patterns that exist in written text. With that model, they can hallucinate more text that follows the same patterns: they can perform natural language.
So that's it, right? Problem solved! What's missing? explicit definition. We traded one problem for another. No one (so far) has figured out how to solve both in the same program. You can have definition, or, you can have inference. You can't have both.
This doesn't make any sense to us humans. We don't have any trouble at all doing both at the same time. We do it all the time! Do we actually do anything else? Unfortunately, LLMs are not humans.
---
The two approaches to language are diametrically opposed, but they work with the same domain. Approaching from either end of the spectrum, definition and inference explore together the wild universe that is story. That's the missing piece: once we figure out what story is made of, we should be able to put all three pieces together.