Does a large language model show signs of an emergent awareness of semantics?

Question

It is well established that computers have semantics of some sort. After posting the question Is non-deterministic automated reasoning a viable strategy for solving problems in mathematical logic? (PhilSE), I received a variety of yes, no, and some-days to the answer demonstrating varied viewpoints on computers and automated proof of mathematical logic. An issue the disagreement turned on seemed to be on the nature of what respondents believed an LLM was, did, and was capable of. This brings to the fore an important question about LLMs. Large language models analyze human language, like a collection of proofs in mathematical logic, and construct products that seem to show an awareness of the syntax and semantics of content.

Consider the conversation with ChatGPT:

Human: Which sentence is meaningful? "Bob monkeys purple." or "Bob's monkey is purple."

ChatGPT: The sentence "Bob's monkey is purple." is meaningful. It conveys a clear and coherent statement: Bob has a monkey, and that monkey is purple. In contrast, the sentence "Bob monkeys purple." does not form a grammatically correct or meaningful statement.

Given that LLMs regularly provide meaningful answers to questions about the meaning of natural language, do large language models pose some form of awareness of the semantic content of the corpus they are trained on?

Answer 1

In actual fact, machines don't have semantics yet. That is because they are not conscious of the outside world, so they know not what they speak of. ChatGPT too. If you simply asked it the question, i suspect that's what it will tell you. What they do have and work with are semantic networks: how concepts relate to other concepts. For them, Derrida is correct: il n'y a pas de hors texte.

Edit: As an illustration of the point, I recently had a little chat with ChatGPT. I told it about my elephant, Jumbo's the name, and how this little bugger keeps climbing on rainbows everytime he sees one. And then he gets too afraid to climb down so I need to call the firefighters all the time...

See what I am driving at? Any human being with a reasonnably accurate sense of what elephants and rainbows are would know I'm joking. Because a rainbow is made of light and light is not solid, so elephants cannot possibly climb on rainbows. But ChatGPT has no real sense of what an elephant or a rainbow is. At most it can correctly classify "elephants" as part of the set of "animals", itself a subset of "living creatures". But it never touched an elephant, never smelled one, never gave an elephant peanuts, and never really saw an elephant. It never marveled at a rainbow either.

So it answered: "Sorry, I am unable to find references to your specific query on the Internet."

Answer 2

The short answer is "no one knows," which is what makes this a live and interesting philosophical question.

Your statement "Large language models analyze human language, like a collection of proofs in mathematical logic, and construct products that seem to show an awareness of the syntax and semantics of content" is a bit misleading. LLM (large language models) like ChatGPT are basically artificial neural networks. They are given training data to work with, and they learn to simulate it. Pathways that lead to good responses are reinforced, those that lead to poor ones are deprecated. Because of the way they are trained, and because of their size and complexity, the inner workings of the network are largely opaque even to the "programmers." However, there doesn't seem to be any evidence that they are working from first principles, the way your question suggests. When ChatGPT provides a well-formed argument, it does so by imitation, not by analysis.

We know that ChatGPT must have some semantic framework, since that's a foundational pieces of creating well-formed responses. But how that's encoded in the network and whether it's at all analogous to the human understanding of semantics is largely unknowable. (For that matter, we don't know how semantics are encoded in the human brain, either.) We do know that when we ask ChatGPT about semantics, it's not scanning itself and providing a self-analyzing answer. It's imitating the kinds of things that people say about semantics.

Answer 3

When meaningful expressions continue, nothing clear can be said about whether perceived is sensational. Since the question you ask is about artificial intelligence's internalization of meaning, I can say that the answer is yes; it's simulatevly. We cannot say that perceived can be sensational clearly with AI's training method is excluding.

In my opinion, human, even though he has found a position in this universe as a small being, has a reasonable judgment in achieving goals in it. Since primary disadvantage of humans is weakness at exceptional and hard life rules, it cannot be said that it is incomplete in logical reasoning with artificial intelligence. What is requested for artificial intelligence is requested partly for human, and what is requested for humans is requested partly for artificial intelligence. Your question is about the existentiality of consciousness. Therefore, for a clearer answer, I want you turn to discussions about soul, mind, intelligence and consciousness.

Reality is the dynamism that enables the transition from state A_x to state B_y. While Descartes describes that animals haven't consicious, he puts people at the center of perfection. In a way, it makes a statement that leads us to discuss why zombies and the dead are not always conscious beings. Cognitive science supports that a consciousness exists in parallel with the mind. But he cannot identify that some unconscious beings have minds. Status of consciousally or partially consciousally is the synthesis with sci-Fi technology from a unconscious vision. This cannot be achieved; it is acquired with not reasoning by personally.