List, I had forwarded Ulysses' response to my summary of Cathy Legg's paper on Peirce and Generative AI to her, and she responded off List to Ulysses, Ccing me. As her comments are of interest in this discussion, I asked her if I might forward them to Peirce-L and she gave me permission to do so.
After an introductory note to Ulysses (I took the liberty of omitting a short postscript to it), Cathy's comments are interleaved in boldface between his original comments (which can be found in full in this thread). Best, Gary R Hi Ulysses, Thank you for engaging in such depth with my paper on the Peirce-L. Gary forwarded your post to me, and yours is the first response informed by sufficient knowledge of Peirce's semiotics to actually challenge my argument, which I really appreciate! I've taken the liberty of writing some responses below. Cheers, Cathy ---------- Forwarded message --------- From: *Ulysses* <[email protected]> Date: Sat, Feb 22, 2025 at 10:32 PM Subject: Re: [PEIRCE-L] Peirce and Generative AI To: <[email protected]>, Gary Richmond <[email protected]> Cc: Jon Alan Schmidt <[email protected]> For what it is worth, I tend to understand LLM operations as NOT symbolic (in the peircean sense). Large Language Models are first and foremost *models* ie diagrams ie icons of language. Just as peirce argued that algebraic formula are diagrams, one can see LLMs as massive intricate algebraic expressions that encode positional relationships between words. The attention mechanism is, from a peircean perspective, a diagram of the indexical (spatiotemporal) relationships between lexical tokens. Every token sequence “points to” an array of possible next tokens. (Think rhematic indexical). This view helps explain phenomena like ‘hallucinations’ which, like abductions, are iconic of possible responses to a query and, like abductions, are not guaranteed to be factual or accurate—they are only possibly true. *>> These are good points, but there is an aspect of iconicity that I think LLM 'icons of language' miss, and that is its 'structural hardness' - the fact that certain aspects of a given structure force certain other aspects of that structure to be a certain way. This enables what Wittgenstein called 'the hardness of the logical must'. (And I have argued, in previous publications, it's the only thing that can enable it - see e.g. my 2012 piece "The Hardness of the Iconic Must..."). By contrast, an LLM doesn't see p as irrevocably sundered from ~p, just rarely found alongside it (but that might change).* *>> Also, insofar as the patterns of association between terms that LLMs trace are indefinitely generalizable, I think they are symbolic rather than iconic. * While I agree that LLMs lack indexical relationships to many real world dynamical objects they nevertheless do encode indexical relationships to other lexical tokens. This enables LLMs to be in causal and dynamical relationships with the world through application interfaces that are dynamically coupled to other objects in the world. Consider coding agents that predict code which actually compiles and affects changes in the world. I tended to think of LLM outputs as austinian “performances” / promises whose felicity conditions are checked in the future (ie at run time for code, or by some other social convention for language). *>> I'd say that's somewhat true, but it depends what you mean by 'dynamically coupled'. I see genuine dynamic coupling as an ongoing continuous feedback loop, so, for instance, if I'm perceiving a yellow chair before me, and it suddenly transforms into a yellow snake, I will react and try to update my beliefs. If an LLM writes a piece of code and only in the future is it tested...not entirely dynamic. * Current LLMs lack robust ‘thirdness’ — they do not fully learn/habitualize law. *>> Yes, and this is really important.* At best they parrot (iconize) reasoning. Even so-called “reasoning models” are better understood as lexical simulations (icons) of reasoning. This may change with new architectures that incorporate test-time learning, multi-modal models, and recurrent reasoning models. The fact that LLMs are so adept at manipulating tokens of symbols without being fully symbolic is quite fascinating. *>> I know! I agree. This is why Gary Marcus says that only through 'NeuroSymbolic' systems will we make ongoing progress with AI. I would be interested to know your thoughts on that. *
_ _ _ _ _ _ _ _ _ _ ARISBE: THE PEIRCE GATEWAY is now at https://cspeirce.com and, just as well, at https://www.cspeirce.com . It'll take a while to repair / update all the links! ► PEIRCE-L subscribers: Click on "Reply List" or "Reply All" to REPLY ON PEIRCE-L to this message. PEIRCE-L posts should go to [email protected] . ► To UNSUBSCRIBE, send a message NOT to PEIRCE-L but to [email protected] with UNSUBSCRIBE PEIRCE-L in the SUBJECT LINE of the message and nothing in the body. More at https://list.iu.edu/sympa/help/user-signoff.html . ► PEIRCE-L is owned by THE PEIRCE GROUP; moderated by Gary Richmond; and co-managed by him and Ben Udell.
