“To understand a sentence is to know the conditions under which it is true.”
Came across this when reading the 1990s book 'Modern Philosophy' by Roger Scruton, p259, summarizing the philosopher Donald Davidson's 1960s essay 'Truth and Meaning'.
Implications to current AI/AGI debate? Do LLMs "understand" the sentences they generate?
9 comments
[ 3.2 ms ] story [ 25.1 ms ] threadhttps://champollion.com/research/
But the tl;dr is that event semantics is a way of parsing sentences to logical assertions about whether specific events or classes of events have/had/will occur, and that this is a general way of understanding nearly all forms of spoken or written communication (the "nearly" being only due to relatively uninteresting things like interjections and fillers). It wasn't always the case that Davidsonian event semantics was considered complete as a formal framework, but more recent work by Champollion and others have helped fill in the gaps. I'm thinking of this paper in particular:
https://champollion.com/wp-content/uploads/2018/06/2015-inte...
So a Davidsonian way of understanding a sentence is to turn it into a formal logic statement about the past/present/future or even counter-factual state of the world. And the truth of a statement lies in how well that formal logic matches up with actual reality. To understand a sentence in this framework is to have worked through the logical consequences of that interpretation and know what conditions would make it true and what conditions would make it false.
On the other hand if I tell you "The king of Thailand is riding on an elephant." there is a sense in which you intuitively understand this sentence as soon as you hear it without having worked through all the various edge cases in which it might be true or false. (Am I talking about a painting of the king riding an elephant? Or a movie on YouTube? Or is the king actually riding the elephant right now, at this very moment? etc.)
The problem with your original question is that you are conflating these two meanings of "to understand." When you hear a spoken utterance or read a written sentence, you form a temporary and contextual model in your mind linking its constituent concepts together. A LLM does exactly the same thing. Davidson's event semantics provides a formal framework for clarifying, refining, and extending that understanding with rigorous logical analysis, but that's not something you, I, or a LLM normally do. It is still a valuable thing to do, but shouldn't be conflated with the natural everyday sense of "understanding."
But I do understand your subsequent explanation, the distinction between event semantics and everyday understanding, at least to some extent. Thank you. It reminds me of Frege's distinction between sense and reference, which I've read about second-hand.
https://gist.github.com/partiallytyped/deb53061e26f7a77ebde0...
Given the self-referential nature of this response, one might even express that to some degree they do.
AFAICT, the only difference between human brains understanding relationships and LLMs is that the former is grounded on the real world and the other is not. That doesn't imply that the LLMs have no understanding assuming understanding is creating associations at different degrees of abstraction. Only that their understanding is not grounded in things we perceive as true through our interactions.
The claim however assumes that there is an objective reality that is true which is a whole different can of worms that needs to be opened before we can even argue whether there is sentience or understanding in the first place.
Furthermore, humans themselves are not grounded in an objective reality either. While LLMs might hallucinate reality in one way, we hallucinate in others (cf religions), we hold irrational opinions, we perform irrational actions that are not beneficial to us, our brains become narrowminded in the physical sense, and so on.
Re: "objective reality" your argument can be made even stronger using Kant's transcendental philosophy which says that we are limited at apprehending reality by our fundamental conceptions e.g. of space and time ("categories"). Similarly, perhaps, LLMs are limited by other fundamentals. So we don't even have to bring irrationality into the argument. We both have incomplete grasp on "objective" reality, even if we are being fully rational.
I see it the other way around: propositional logic is a very limited sub-domain of natural language. It may be the case that you can model human brains with propositional logic, but the kinds of sentences you create are very, very different. They're more along the lines of "Neuron 1 has a .25 weight connected to Neuron 2" than "A BLT is a sandwich made out of bacon, lettuce, tomato, and bread".
Focusing on sentences is designed to highlight this mapping of language to propositional logic, while missing everything else about discourse. A sentence like "I would like a cheeseburger" is not understood by "the conditions under which I'm hungry". That sentence is a request from one person to some other, who is not visible in the sentence at all but is part of the conditions in which that sentence is uttered.
As it comes to LLMs, I think that is the opposite of helpful. LLMs are not expert systems, which do come closer to trying to establish truth conditions. Those were abandoned because they were poor models of human cognition.
Whether LLMs do human cognition either remains an open question, but I don't think this direction is a very helpful one for understanding it.