Ulysses, List, Thank you, Ulysses, for your thoughtful response. Just to make sure I've got your argumentation right, I believe you are saying that LLMs function primarily as *models, *so akin to *diagrams, *thus a type of *icon*. You argue that LLMs encode relationships between lexical tokens rather than operating with symbols *per se*.
Further, you describe the 'attention mechanism' in LLMs as a *diagram* of indexical relationships between *tokens*, such that each token in a sequence 'points to' possible 'next' tokens, thus aligning with Peirce’s notion of *rhematic indexical *signs as *those which *suggest possible outcomes without guaranteeing the truth of these 'outcomes'. You give the example of LMM 'hallucinations' which you see as equivalent to Peircean *abductions, * that these are plausible but uncertain inferences. This seems to me questionable: for are all (or even many) of these AI 'hallucinations' truly 'plausible'? Be that as it may, if these hallucinations are quasi-abductions then this might further support your idea that LLMs generate responses that are *iconic* (as *possible* representations of meaning) rather than *symbolic* (that is, representations fully grounded in rule-based -- lawful -- interpretation). But, again, I'm not at all certain about this. You further remark that, while LLMs lack direct real-world referentiality, they do establish indexical relationships within language itself, and that through external applications that they can be coupled to real-world effects. This too seems questionable to me, at least in the sense that an actual human 'effector' (metaphorically speaking) would seem to be required for this to happen in fact. You conclude that LLMs do *not* fully internalize *law* or *habitual reasoning*, these being key aspects of 3ns. Thus, so-called reasoning models are better seen as *lexical simulations* (i.e.,icons) of reasoning rather than true symbolic reasoning systems since they lack true 3ns'. And so, while LLMs do indeed excel at manipulating *tokens *of symbols, they do not function *as* *symbols* themselves: their operations remain fundamentally *iconic*. In a word, you argue that LLMs do *not* engage in *symbolic* reasoning but instead operate through *icons of symbols; *that genAI is powerful at token manipulation while not being fully engaged in symbolic thought or 'habitual' reasoning in the Peircean sense. But then why write: "I don’t say this to dispute Leggs’ account. I think we agree more than we disagree. I just want to point out that because ‘[tokens of] symbols are LLMs strong suit’ it is easy to forget that LLMs are not, at their core, symbolic"? (I won't get into here Peirce's argument that there *are no pure symbols*, that they *only* manifest in concrete reality* through* tokens; that is, that a symbol as a general class or law has its reality independent of any particular instance, that it *must be* 'instantiated' in some form to *function* as a symbol at all). Best, Gary R On Sat, Feb 22, 2025 at 10:32 PM Ulysses <[email protected]> wrote: > 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. > > 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). > > Current LLMs lack robust ‘thirdness’ — they do not fully learn/habitualize > law. 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 don’t say this to dispute Leggs’ account. I think we agree more than we > disagree. I just want to point out that because ‘[tokens of] symbols are > LLMs strong suit’ it is easy to forget that LLMs are not, at their core, > symbolic. > > > On Sun, 23 Feb 2025 at 6:08 am, Gary Richmond <[email protected]> > wrote: > >> List, Jon, >> >> Jon, thanks for posting the abstract of Cathy Legg's interesting, timely >> and, in my view, important chapter, "Peirce and Generative AI," to appear >> in the forthcoming, *Pragmatism Revisited. See: * >> https://www.academia.edu/127744327/Peirce_and_Generative_AI. I read it >> yesterday and found it to be solid in its research methodology, >> argumentation, and conclusions. >> >> The importance of her chapter for Peirce scholars -- and for AI engineers >> who might hope to address AI's present weaknesses -- lies in the possible >> consequences of applying Peirce’s semiotics to the challenges and >> limitations of genAI. Legg clearly sees the promise, challenges, and >> dangers inherent in genAI yet, since it appears that it will likely be here >> to stay in one form or another, that it ought to be as logically -- >> *meaningfully* -- sound as is possible. She argues that Peirce's >> "pragmatist epistemology" provides a framework for both understanding and >> improving AI’s role as a cognitive tool. >> >> Comparing the elements of signification of AI to Peirce’s triadic model >> of signs involving not only symbols, but also indices, and icons, one sees >> the limitations of AI’s essentially *symbolic *processing and its >> deficiencies in indexical and iconic signification, all three of which Legg >> views as essential for genuine meaning-making in the quest for real >> knowledge and truth. >> >> She begins her argumentation with a discussion of meaning, contrasting >> Peirce’s public and dynamic model of sign interpretation with the >> Cartesian private, static approach, which early AI engineers adopted. While >> LLMs (Large Language Models) learn meanings through statistical >> associations, they lack connections to real-world objects (i.e., they lack >> real indices) and logical (iconic) structures, thus limiting their ability >> to perform true 'artificial sign interpretation'. Without these AI >> generated text remains but a mirror of human discourse rather than an >> active participant in authentic meaning-making. >> >> Legg quotes Peirce that “a sign is not a sign unless it translates itself >> into another sign in which it is more fully developed” and that “no present >> actual thought [. . .] has any meaning, any intellectual value; for this >> lies not in what is actually thought, but in what this thought may be >> connected with in representation by subsequent thoughts; so that the >> meaning of a thought is altogether something virtual.” She argues that >> Peirce’s epistemology further challenges mainstream "representationalist >> realism" which treats truth as but a correspondence between propositions >> and discrete facts, contrasting it to Peirce’s "rich relational realism*." >> *She quotes S. Vallor that “today’s most advanced AI systems are >> constructed as immense mirrors of human intelligence. They do not think for >> themselves; instead, they generate complex reflections cast by our recorded >> thoughts, judgments, desires, needs, perceptions, expectations, and >> imaginings.” >> >> So Legg concludes that while AI might assist human inquiry, it cannot -- >> at least not yet -- function as an *autonomous inquirer* because it >> lacks the necessary semiotic relations to reality, viz. true indexes and >> icons. >> >> Peirce’s vision of knowledge as a living, growing body of truth suggests >> that to be 'genuinely epistemic' AI must engage in dynamic, interactive >> reasoning beyond mere 'pattern recognition' (as Gary Furhman has also >> recently suggested). This insight challenges AI engineers to design systems >> that do more than produce 'plausible' text; they must integrate signs >> (including icons and indices) in ways that truly reflect the structure of >> reality. >> >> I hope this short summary helps to suggest the depth and richness of >> Cathy's argumentation which includes a brief history of the development of >> AI, LLMs, etc., how, for example, early AI research relied on explicit >> representations of facts and rules, followed by a shift to deep learning >> and neural networks, thus leading to LLMs capable of generating >> intelligible text. >> >> I encourage interested List members to read her chapter which, as Jon >> noted, can be found (along with many other of Legg's excellent papers) at >> Academia (see the link in the first paragraph above). >> >> Best, >> >> Gary R >> >> On Wed, Feb 19, 2025 at 9:10 AM Jon Alan Schmidt < >> [email protected]> wrote: >> >>> List: >>> >>> Since (so-called) "artificial intelligence" was a topic of discussion >>> here recently, I thought that it might be of interest that Catherine Legg >>> wrote a chapter on "Peirce and Generative AI" for a forthcoming book that >>> Robert Lane is editing, *Pragmatism Revisited*, and has posted the >>> prepublication version at >>> https://www.academia.edu/127744327/Peirce_and_Generative_AI. Here is >>> the abstract. >>> >>> Early artificial intelligence research was dominated by intellectualist >>> assumptions, producing explicit representation of facts and rules in “good >>> old-fashioned AI”. After this approach foundered, emphasis shifted to deep >>> learning in neural networks, leading to the creation of Large Language >>> Models which have shown remarkable capacity to automatically generate >>> intelligible texts. This new phase of AI is already producing profound >>> social consequences which invite philosophical reflection. This paper >>> argues that Charles Peirce’s philosophy throws valuable light on genAI’s >>> capabilities first with regard to meaning, then knowledge and truth. >>> Firstly, I explore how Peirce’s icon/index/symbol distinction illuminates >>> the functioning of genAI. I argue that genAI’s engineers have skilfully >>> captured a form of symbolicity, but no other sign-kind. In lacking >>> indexical signs, LLMs lack connection with, and accountability to, >>> particular worldly objects. In lacking iconic signs, LLMs are >>> insufficiently disciplined by structural – most notably logical – >>> relationships. Then I argue that GenAI’s astounding stream of articulate, >>> truth-semblant, yet worthless texts issues a timely reckoning to modern >>> philosophy’s representational realism. By contrast, Peirce’s pragmatism >>> scaffolds a rich relational realism (Gili and Maddalena 2022), which shows >>> how meaningful concepts, and a grasp of truth, can only occur across >>> multiple cognitive systems who are simultaneously richly related with one >>> another, and a shared environment in which they continually act and receive >>> feedback, within a logical space of reasons. As Peirce himself noted, “Mere >>> knowledge, though it be systematized, may be a dead memory; while by >>> science we all habitually mean a living and growing body of truth”. >>> >>> >>> Regards, >>> >>> Jon Alan Schmidt - Olathe, Kansas, USA >>> Structural Engineer, Synechist Philosopher, Lutheran Christian >>> www.LinkedIn.com/in/JonAlanSchmidt / twitter.com/JonAlanSchmidt >>> _ _ _ _ _ _ _ _ _ _ >>> 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. >> >> _ _ _ _ _ _ _ _ _ _ >> 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. > >
_ _ _ _ _ _ _ _ _ _ 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.
