[karpierz]

Yes, that's true. But ChatGPT isn't trained to play any language game other than: "make a plausible sounding sentence".

The issue it has with facts is the same issue it'll have with any language game, in that it only understands how to mimic playing the game, without understanding the underlying rules of the game.

[williamcotton]

Nit: Tractatus is early Wittgenstein. Language games come from Philosophical Investigations-era Wittgenstein which is almost a complete rejection of his approach in Tractatus.

[karpierz]

Ah yeah, my mistake. I'd assumed that the GP was referring to the later work and echoing its core premise (language has no intrinsic meaning; or rather takes on meaning from context and intent).

As an aside to anyone reading, would highly recommend internalizing the Tractatus. It really helps side step a lot of "high minded" debates (ex: is X conscious, is it moral to do Y) by actually making you ground what you actually mean when you say "conscious", "moral", etc.

[williamcotton]

In many ways LLMs are more in support of the Philosophical Investigations era understanding of language and less in support of the logical positivist understanding of language put forth my Frege...

Like, the Frege approach is like the symbolic AI approach... define the rules, derive the meaning from the rules.

The PI approach is like the LLM approach... derive the rules and meaning from "experiencing" the use of the language.

Eg, we don't need to define what a "game" is. We know when something is a game or not even if we can't come up with an explicit set of rules for defining all instances of what we would otherwise "feel" was a game.

I'm running low on blood sugar...

[karpierz]

I think the PI approach still requires that the agents in the game have a shared internal model of the game.

You can make Searle's Chinese Room argument, but I always find "assume there exists a book that contains every conversation ever" as a flawed premise.

[mLuby]

Well, if there've been about 120B humans ever, and we speak fewer than 1B words per lifetime, and the average word takes 1 byte to store, that's about a fifth of all data stored in AWS (according to Wolfram Alpha). It's undoubtedly a lot, and yet clearly within human capability. And of course that ignores optimizations that'd certainly drop that high estimate by many orders of magnitude.

[karpierz]

I think you're misunderstanding Searle's Chinese Room. It has a response for every sequence of conversation, ever. It doesn't store every conversation that has happened; it stores every possible conversation that's possible, whether it'll ever happen or not.

It would be able to handle the following exchange:

Person: "Here's a cool question, ready?" Room: "Ready." Person: "What was the last message I sent to you?"

It can respond appropriately to the following sentence:

Person: "Hey, I'm gonna say something. Here is a sentence. Can you repeat the previous sentence back to me?"

Otherwise, why bother with all of this AI stuff? Just build Searle's Chinese Room as an index and you have a perfect chatbot.