2

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?

17
  • 1
    @gs The thrust of this question is in the direction of clarifying whether given some canonical notion of awareness, that the performance of an LLM rises to the criteria.
    – J D
    Oct 6, 2023 at 16:27
  • 1
    Not yet, see SEP:"GPT-3 is not close to passing the Turing Test: GPT-3 neither perceives nor acts, and it is, at best, highly contentious whether it is a site of understanding." Current AI cannot even imitate awareness well enough, let alone suggest a genuine article. Chalmers does call it "a potential mindless path to artificial general intelligence". However, we would have to see AI handle tasks outside a predefined context they were programmed/"trained" for to talk about even genuine intelligence.
    – Conifold
    Oct 6, 2023 at 22:42
  • 1
    "The force that through the green fuse drives the flower."
    – Scott Rowe
    Oct 7, 2023 at 3:24
  • 1
    @ScottRowe That poem was emotionally emancipatory. Thank you.
    – J D
    Oct 7, 2023 at 19:35
  • 1
    @user4894 The claim 'Computers are bit-twiddlers' is a slogan, and one that doesn't comport. And at the minimum, a clothes dryer with a moisture sensor is aware of whether or not my clothes are dry. It seems odd to argue otherwise for what use would such a sensor be if not help the computer be aware of the water content? I would encourage you to move beyond a non-cognitive boo and share whatever depth, if any, lies under the slogan.
    – J D
    Oct 7, 2023 at 19:42

3 Answers 3

1

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.

6
  • 2
    As always, your arguments are succinct, persuasive, and show an unusually uncommon regard for agnosticism which is always refreshing to me as a fallibilist. The transformer model does indeed have a ANN core, and ANN seem to have captured the capacity for the ostensible emergence that we observe in neural computation. The argument that we proceed from first principles with our semantic awareness and grounding is for the linguistically ignorant. Language acquisition can happen entirely without them. Brains barrel towards grammar despite grammarians. Humans learn by imitation, not analysis...
    – J D
    Oct 12, 2023 at 21:40
  • 1
    and Tomasello puts forth a strong philosophical framework. Langacker, Lakoff, and Jackendoff have some strong theoretical frameworks that point in the direction towards neural computation as an ontology between mind and body, and experimental philosophy and NLP continues to make gains. My company is working on a NLP intermediary that uses LLMs in conjunction with formal systems, and a lot of the market space is engaged in similar enterprises. Thanks for a level-headed analysis.
    – J D
    Oct 12, 2023 at 21:45
  • If a language model were trained to translate its arguments into mathematical logic, the correctness of those arguments could be verified by an automated proof checker. Feb 25 at 20:01
  • @AndersonGreen - That SOUNDS possible, but unless you can cite a LLM that can do this (accurately and consistently) we don't know if it's actually possible. There's a lot of things that we think LLMs should be able to do that we haven't been able to accomplish.// However, even if they definitely can provide that for us, that doesn't mean that's what they are doing deep down. Feb 26 at 14:01
  • 1
    @ChrisSunami I believe that this is possible because there have been several successful implementations of theorem provers using large language models. Feb 26 at 20:36
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."

1
0

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 Ax to state By. 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.

19
  • 1
    I would say that the question provokes discussion on consciousness by asking if computer technology manifests some small measure of it. If we can conclude that consciouness and mind is not a crisp category of meaning, and that as a fuzzy logic might apply, that consciousness is a measure of behavior on a scale from 0 to 1 of the reals. I will upvote for the novel experience of providing an answer in Turkish since it demonstrates that translation systems clearly are becoming more proficient in the meaning of languages.
    – J D
    Oct 6, 2023 at 19:51
  • Also, +1 for raising the language of simulation which begs the question, what is real? This is also a received position, that somehow consciousness either is or isn't, and Descartes maintained that animals were not conscious, but mere automata. Cognitive science supports that consciousness occurs in degrees.
    – J D
    Oct 6, 2023 at 19:54
  • 1
    Gerçek, A_{x} durumundan B_{y} durumuna geçişi sağlayıcı dinamizmdir. Descartes hayvanların bilinçli olmadığını ifade ettiğinde insanı benmerkeze koyuyordu. Bir yönden de zombilerin ve ölülerin neden daima bilinçli varlıklar olmadığını tartışmamıza yol açıcı bir ifadede bulunuyor. Bilişsel bilim, bir bilincin zihin ile paralel bir şekilde varolduğunu destekliyor. Fakat bazı bilinçsiz varlıkların zihne sahip olduğunu tanımlayamıyor. Kısmen bilinçli veya bilinçli olma durumu bilinçsiz varlık için bilimkurgu teknolojisinin sentezidir. Bu, elde edilemez; kişisel açıdan nedensizce edinilir.
    – fkybrd
    Oct 6, 2023 at 21:16
  • "Bei mir bist du schoen"
    – Scott Rowe
    Oct 7, 2023 at 3:34
  • 1
    @ScottRowe Bella, bella. Wunderbar! ; )
    – J D
    Oct 7, 2023 at 19:52

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .