I have a question about Tarski's T-schemas. Would the categorization of various T-schemas under the guise of conceptual analysis, lead to a more truthful semantic model than what we see today with large-language-training-models? — Shawn

Um, I can almost understand this. Can you develop it more?

Um, I can almost understand this. Can you develop it more? — tim wood

Yes, well the rationale is that given that languages utilize extensively the use of concepts to talk about various issues, like space or time or physics, then I surmise that by appealing to T-schemas, that a user of language would be better able to understand concepts with regards to what can be rendered or said truthfully about a concept in a language (atomic sentences).

Tarski's T-Schemas (TTS), near as I can tell, rely on very plain and simple language, so-called atomic sentences. - I may be getting this wrong. - If your question is, whether simple language as an alternative to difficult language is an aid to understanding, it seems to me the answer is obvious.

I see reference to propositions of the form "P(x)" and "'P(x)' is true." And it is noted that the first is a statement about x, and the second about the sentence P(x). Since the truth of the first and the claim of the second both depend on a rigorous definition of truth, and there ain't one - well, there is but it's not helpful here - then I remain to be educated on what the exercise is about, if anything.

Well, the aim here is to have a rendering or model of language that aligns with truth of concepts and what I imagine "archetypes" in understanding concepts as truthful. This whole use as meaning from Wittgenstein's family resemblance and language games is kinda something I wanted to see disambiguated from a Tarskian view on semantics...

I'll watch. But I cannot quite understand, or understand at all, what you're writing. Why not try expressing it in some atomic sentences? It seems to me also that truthful is not truth - two different words meaning two different things.

Do you mean you find Wittgenstein ambiguous and maybe TTSs could disambiguate it?

As far as I can tell, the OP is asking if Tarski's T-Schemas can be used to develop better LLMs, ones that do not come out with false statements, since they are constrained to produce statements that are actually true.

How does that work?

Loooks to be two very different things. To get anywhere would have to show how T-sentences could be used here. And T-sentences do not make use of atomic sentences, but of translated sentences. The p in <"p" is true iff p> does not need to be atomic.

Think it's all too vague.

ones that do not come out with false statements, since they are constrained to produce statements that are actually true. — Jamal

My bias is that truth has a heart, is a function of human being and communication. LLMs aren't human and don't have hearts, therefore cannot be truth tellers. Facts are a different matter. Given (or understood) that all facts are historical facts, enunciating them simply a matter of reproducing records - which to be sure can be done inaccurately.

No doubt we'll all come to regard machines as truth tellers - and in many cases have already. But (imho) that is real tragedy in the making.

LLMs have no contact with reality, but are themselves textual artefacts. They never touch the ground but are free-floating in the sea of language. There's nothing like banging your head against a brick wall for convincing you of its solidity; likewise spades hitting 'bedrock'. They know not whereof they speak, having as yet no senses of anything that is not language and image. Therefore their talk is all talk and no trousers.

Like intellectuals, you cannot trust them further than they can throw you.