[{"data":1,"prerenderedAt":123},["ShallowReactive",2],{"post-contra-stochastic-parrots-the-role-of-meaning-in-assessing-ai-intelligence":3},["Reactive",4],{"id":5,"date":6,"date_gmt":7,"guid":8,"modified":10,"modified_gmt":11,"slug":12,"status":13,"type":14,"link":15,"title":16,"content":18,"excerpt":21,"author":23,"featured_media":24,"comment_status":25,"ping_status":25,"sticky":20,"template":26,"format":27,"meta":28,"categories":29,"tags":31,"class_list":32,"_links":39,"_embedded":81,"tagsResolved":122},965,"08. 08. 2026","2026-08-08T21:45:10",{"rendered":9},"https://cms.local.test/?p=965","2026-08-09T00:05:59","2026-08-08T22:05:59","contra-stochastic-parrots-the-role-of-meaning-in-assessing-ai-intelligence","publish","post","https://cms.local.test/contra-stochastic-parrots-the-role-of-meaning-in-assessing-ai-intelligence/",{"rendered":17},"Contra Stochastic Parrots: the role of meaning in assessing AI intelligence",{"rendered":19,"protected":20},"\u003Cp>In this essay, I argue that LLMs are not stochastic parrots by evaluating the claims by (Bender et al., 2021) who coined the term “stochastic parrot” to express the view that the output of LLMs is 1) haphazard, 2) incoherent, 3) meaningless, because, following (Harnad, 1990), it produces symbols that are not grounded in the human phenomenal existence inside a physical world. I present a concept of “meaning networks”, a network-based account of the meaning of terms that consists in weighted, multimodal, socially stabilized associations that constitute what a term means for a language user. The main argument of the essay is to show that LLMs produce outputs that inherit grounding indirectly from human produced language, which is itself grounded. The use of linguistic data as an interface into human existence in the physical world suffers from a lossy compression: any interface that translates the human existence, which is a “continuous signal” (following computer science terminology), into discrete units, such as linguistic symbols, loses the richness of the detail. The grounding is therefore both indirect and lossy. Compared to human-level intelligence and meaning-use, LLMs occupy a never-before seen niche of systems that produce a human-level linguistic output whose meaning has a purely symbol-based modality. The meaning networks generated by LLMs can be understood as an extreme example of language-laden human intelligence.\u003C/p>\n\u003Ch2>\u003Ca id=\"post-965-_geao7idfhv2\">\u003C/a>Intro\u003C/h2>\n\u003Cp>Since at least the second half of the 20th century, our society has been labeled the information society, network society, post-industrial society, etc.\u003Csup>\u003Csup>\u003Ca id=\"post-965-footnote-ref-0\" href=\"#post-965-footnote-0\">[1]\u003C/a>\u003C/sup>\u003C/sup> All of the terms reflect the dominant mode of how society works. Information, networks, abstractions, and symbolic knowledge are the currencies of the day. Highly educated individuals with an aptitude for analytical, abstract thought and processing information are the sought-after professionals. The advent of cognitivism in philosophy of mind and cognitive sciences marked the beginning of the era when the metaphor of humans as information processors started to be taken seriously even at the frontier of our search for understanding who we are and how our minds work. Simon and Newell’s work on Logic Theorist and General Problem Solver software systems, Weizenbaum’s ELIZA, Winograd’s SHRDLU, and even Chomsky’s Syntactic structures and later the biological turn in linguistics were a few of many theories and projects to explain human cognition by appeal to logic and language as the essential ingredients of human intelligence. It is true that since then, the cognitive sciences and AI fields underwent major changes. The connectionist and 4E paradigm shifts emphasized the use of neural nets and the role of the embedded body in situated and broader cultural-societal contexts for creating a more faithful model of how brain processes and human cognition work. But the focus on understanding \u003Cem>intelligence\u003C/em>, both human and artificial, stays primary.\u003C/p>\n\u003Cp>Yet, the introduction of ChatGPT, Bard, Claude, and other chatbots based on the large language model (LLM) architecture forces us to re-consider who we are. LLMs show that the use of language, being the most distinct human feature compared to \u003Cem>all \u003C/em>other animal species, can be automatized to produce human-like responses, passing a Turing test with ease, surpassing the average human at handling factual data, logical reasoning or language abilities. If LLMs beat humans in the tasks that we consider uniquely human, is a LLM-based artificial intelligence becoming more like us than we are? It seems preposterous to claim that the LLM, a formal symbol-string prediction machine, can attain a status of being human. Unless we are hypocrites, we cannot stop praising human intelligence and language competency as the highest \u003Cem>human\u003C/em> cognitive values the moment another entity becomes superior to us. To avoid an ontological conundrum of what it means to be human, I will ask differently: can LLMs even in theory become more intelligent than humans? To again avoid defining what intelligence is, I will narrow down my inquiry further still: Whatever being intelligent means, I presuppose that the human-like intelligence requires understanding; and understanding requires the grasp of what words, sentences, or concepts mean. Without an understanding of the meaning, there cannot be intelligence.\u003C/p>\n\u003Ch2>\u003Ca id=\"post-965-_hy70hds010iq\">\u003C/a>Meaning networks: an intersection of phenomenology and linguistic meanings\u003C/h2>\n\u003Cp>It is important to clarify what I mean by the meaning. I will not present here an overview of theories of meaning in analytical philosophy. What I \u003Cem>can\u003C/em> outline here is a minimal theory of meaning that clashes both with \u003Cem>referential theory of meaning\u003C/em> and \u003Cem>truth-conditional\u003C/em> \u003Cem>theory of meaning\u003C/em>.\u003C/p>\n\u003Cp>Imagine that we want to talk about the meaning of the word “mother”. My mother passed away a few months ago. What my mind generates when thinking about the meaning of the word is as follows: I remember a visual image of a sullen face of my mother and the coldness of her forehead while I kissed her when she lay in the coffin at the church. I remember the smell of the incense, while hearing the notes of Dancing Queen by ABBA played by the string quartet on a viola, violins, and cello. I have tactile memories of hugging my ill mother a few days before her passing. I feel non-objectifiable feelings of sadness, remorse, and happiness when thinking about the word. Also, my mind gravitates towards the topic of what it means to be a mother I discussed with my pregnant fiancée. My mind is also aware of purely linguistic links the word “mother” has with other words like “father”, “pregnancy”, “female”, “children”, etc. and linguistic relationships between the word “mother” and other features of the language (“my” pronoun, “mothers” plural, “not breathing” gerund, etc.). I am also aware of connotations of the word “mother” that I assimilated into my understanding of the word when reading newspapers, literature or learning intersubjectively what mothers can do or are like when talking to other people.\u003C/p>\n\u003Cp>The meaning of the word “mother” is for me both private and public. The private meaning consists of elements of various perceptual modalities and linguistic symbols. The public, external meaning consists of what I heard and assimilated from my community and society at large usually in the form of linguistic symbols (i.e. words and sentences), which may have affected my existing non-linguistic, perception-derived sub-meanings via a feedback loop (e.g. when I heard how badly someone’s mother behaved, I started to appreciate my mother more and feel more love for her, modifying my prior feeling-derived sub-meanings).\u003C/p>\n\u003Cp>If we borrow further from the network theory the basic terms “node” and “edges”, their equivalent would be\u003C/p>\n\u003Cp>node = words / concepts\u003C/p>\n\u003Cp>edge = meaning element that can be of many modalities: perceptual or symbolic representations, feelings, other nodes mediated by the symbolic representations\u003C/p>\n\u003Cp>As we can see, the meaning of a word is primarily a meaning for me, existing in my mind. It is a multi-modal “meaning network” of both perceptual and linguistic elements. This “meaning network” is not static, but dynamically reacts to my situated existence in the world, any new perceptual or linguistic elements can be added or possibly removed (or rather overridden) at any time.\u003C/p>\n\u003Cp>The term “mother” cannot refer to just one existing “object” in the world. I set aside the Fregean distinction of “sense” and “reference”, which presupposes stable referents and sense. Instead, I want to propose a multi-modal, network-based semantics where the meaning of terms equals the meaning networks themselves. With that, I can integrate sense data, feelings, visual imaginary, false beliefs, words, concepts, memories of many modalities for which I have distinct and clear sense that they constitute what I mean when I use the term “mother”, or terms like “my mother” or “Helena Ferencová”.\u003C/p>\n\u003Cp>The meaning network is amenable to the influence of the public and shared meanings of the word “mother” and it helps to defend the theory of meaning proposed here against the accusation of the private meaning, which as argued by Wittgenstein is impossible\u003Csup>\u003Csup>\u003Ca id=\"post-965-footnote-ref-1\" href=\"#post-965-footnote-1\">[2]\u003C/a>\u003C/sup>\u003C/sup>. The community I am a member of does participate in the constitution of what “mother” means for me and I’m aware that our personal experiences and thinking about mothers co-constitute each other’s meaning networks and the shared culture influences our meaning networks to have considerable overlaps. That implies there must be stable elements that are fixed or their stability is dynamically and continually enforced by the linguistic or perceptual interaction with others. In other words, the meaning networks can be dynamic, yet their stability is constantly negotiated via interaction with others. I cannot change the whole meaning network of what I mean by “mother” because I would be talking about someone or something else. If the language is to keep a communicative function, I need to keep the contents of my meaning networks (in other words, the meaning of terms I use) rather stable so that I am reasonably understood by most people. If what I mean differs radically from anyone else, I need to explain myself to others and by that update their own meaning networks, provided they are open to cognitively integrate new information.\u003C/p>\n\u003Cp>As far as truth-conditions go, the truth may be an element in my meaning network, but does not exhaust the entire concept of meaning. My subjective feeling of sorrow is right now more meaning-bearing than any conception of truth. It is not to say that a truth-condition does not participate in the meaning of terms, but is moved down, as it were, to the level of other meaning elements.\u003C/p>\n\u003Cp>It seems plausible that the dynamic aspect of meaning networks is partially realised by the external world influencing my meaning elements stored in my memory. But my meaning elements can change due to my memory going wrong, when I forget about words or experiences that are related to my mother. Also, due to the effects of time, it is likely that some meaning elements will partially fade away, whereas others will remain strong or even stronger. If we borrow terminology from neural networks that LLMs themselves are based on, we can assign each meaning element a \u003Cem>weight\u003C/em> that signals how strongly or weakly a given meaning element participates in constituting the meaning. If I burn an incense stick at home tomorrow, it may increase the weight of the meaning elements related to smell, church, coffin, funeral in the meaning of the word “mother”.\u003C/p>\n\u003Cp>The individual meaning elements may also be false. I may have started associating black velvet jackets with the meaning of “mother” due to my belief that she was clothed in this dress at the funeral. In fact, it may have been a dark navy velvet dress, but due to poor lighting conditions, the color appeared to me black. Some meaning elements thus may be added to my meaning network on the basis of a false belief. But what is indubitably true is that this false meaning element participates in the meaning of “mother”. I think the same can be said of meaning elements derived from references to hallucinations or things that have no proven existence, yet they are part of my meaning networks, such as the concept of “heaven” or “God”.\u003C/p>\n\u003Cp>Importantly, there is still a place for purely symbolic or linguistic meaning in the meaning network. Other words and linguistic features that can be purely formal (based on the position of words in a sentence, as required by the syntactical rules of a language) do help co-constitute the overall meaning network. But again, the linguistic meaning is only a part of the meaning network. It does not exhaust the full meaning of the word “mother” either.\u003C/p>\n\u003Cp>I remain agnostic as to what percentage of the overall meaning networks is constituted by a given modality. Some people may be more visually inclined, others rely on linguistic elements in their meaning networks. Not all modalities described here will be present in others‘ meaning networks. For example, a person suffering from a condition called aphantasia is unable to generate mental imagery. We can expect that their meaning network will not rely on any visual sense-data, or the influence of visual sense-data will be greatly diminished.\u003C/p>\n\u003Cp>The theory of meaning vis-à-vis the concept of meaning network relies on a few minimal requirements: it is not just formal-symbolic or purely linguistic, it is both internal/private and external/community-shared, it is constituted by several modalities, among them, a linguistic symbols and relations are just one of many meaning-constituting elements. Meaning elements that constitute the meaning have weights that signal how strongly or weakly the given meaning element participates in the overall meaning.\u003C/p>\n\u003Ch2>\u003Ca id=\"post-965-_my37juc1t5pp\">\u003C/a>Chinese room and Stochastic parrots\u003C/h2>\n\u003Cp>Not all meaning networks are created equally. There are people with deeper understanding of concepts, things, bodily actions etc. than others. What is going on exactly when we say that someone’s understanding is deep or shallow?\u003C/p>\n\u003Cp>In the famous Chinese room experiment\u003Csup>\u003Csup>\u003Ca id=\"post-965-footnote-ref-2\" href=\"#post-965-footnote-2\">[3]\u003C/a>\u003C/sup>\u003C/sup> Searle shows us what it means to have exactly zero understanding of the meaning of words and sentences. Searle, who does not understand any Chinese, sits in a sealed room. From one side he is given a set of Chinese symbols that represent a written input by a native Chinese speaker. He is also given a set of Chinese symbols for answers and the instructions written in English that he understands how to correlate the input symbols with the output symbols, pretending to have a natural conversation in Chinese with a person outside of the room. Searle has no understanding of the meaning of the Chinese symbols; he correlates them purely formally based on their shape (i.e. form). Searle argues that such manipulation of formal symbols based on a fixed set of rules is analogous to how computers work. Because computers also work with formal symbols only, they have no understanding of the meaning of the symbols. Searle, like computers, does not understand Chinese because they work on the formal and syntactical level that lacks understanding of the meaning of words, or as Searle would put it: rule-based manipulation of formal symbols lacks mental states, which are preconditions for an intrinsic intentionality\u003Csup>\u003Csup>\u003Ca id=\"post-965-footnote-ref-3\" href=\"#post-965-footnote-3\">[4]\u003C/a>\u003C/sup>\u003C/sup>. Intentionality is a property of mental states, pointing out that each mental state has a content; they are about something.\u003C/p>\n\u003Cp>In a way, we could answer the question if LLM-based AI is intelligent right here by saying that LLMs have no mental states, therefore whatever they do lacks intentionality, aboutness or mental content. And if their states are not about anything, they cannot understand the meaning.\u003C/p>\n\u003Cp>I will not argue here if LLMs have mental states or intentionality. Let’s say it is a question of definition or empirical question. I want to imagine what kind of mental states, or rather what kind of meaning networks do LLMs work with, whether they have mental states or not.\u003C/p>\n\u003Cp>LLMs work by stitching together strings of words based on the probability of such a combination derived from high-dimensional representations latent in the large corpus of linguistic data. (Bender et al., 2021) coined the term “stochastic parrot” to describe that the LLM architecture produces seemingly coherent strings of words, but without any reference to the meaning of those strings of data, just like the parrot animal, which has no linguistic competence and just repeats the phonetic sounds it heard back to humans who interpret the mechanically repeated sounds as distinct phoneme units of distinguishable word(s). (Bender et al., 2021) argue that LLMs generate those strings stitched together “haphazardly” and that their coherence is purely “in the eyes of the beholder”, suggesting the coherence is not inherent in the output of LLMs itself, but rather externally given by the human interpreters of LLM-produced outputs. (Bender & Koller, 2020) then elucidate their understanding of the meaning by subscribing to the referential theory of meaning where they take the meaning to be “the relation between the form and something external to language…” (p. 5187). In other words, for a human or computer to be a competent speaker and understand the meaning of words used, they have to “ground” their understanding of words in the real world. (Bender & Koller, 2020) cite (Harnad, 1990) and his “Symbol grounding problem” to argue that, without the grounding of meanings in the physical world, the system (human, computer) cannot understand the language; it would be analogous to learning a language just by studying a dictionary. Meaningful responses must make connections between the words and the physical world.\u003C/p>\n\u003Cp>As I attempt to explain below, I disagree with Bender et al. in three specific cases:\u003C/p>\n\u003Col>\n\u003Cli>meaning must refer to something outside of language\u003C/li>\n\u003Cli>LLMs produce outputs haphazardly\u003C/li>\n\u003Cli>coherence is in the eyes of beholder\u003C/li>\n\u003C/ol>\n\u003Ch2>\u003Ca id=\"post-965-_gknh07wai9q7\">\u003C/a>Grounded and Ungrounded Meanings\u003C/h2>\n\u003Cp>(Harnad, 1990) in the era of the symbolic AI paradigm and in line with Searle’s Chinese room argument writes contra the idea that the purely formal manipulation of symbols can lead to understanding. The main thrust of his argumentation is that for symbolic systems to understand a language, they must ground the bottom level meaning of words in non-symbolic representations coming from sense-data of our perception of and interaction with the physical world.\u003C/p>\n\u003Cp>To “ground” their meanings, the systems must feature equivalents of two features that Harnad calls 1) Iconic representations and 2) Categorical representations and only then it is possible to give a linguistic label on top of them through 3) Symbolic representation.\u003C/p>\n\u003Cp>\u003Cem>1 Iconic representations\u003C/em>\u003C/p>\n\u003Cp>Iconic representations are raw perceptual data of some specific objects, which we yet cannot identify or name. In the case of vision, iconic representations would be the shapes that objects, such as horses, following Harnad’s own example, cast on our retinas. The mind would keep iconic representations of horses as analogs of their real shapes. At this level, the only functions that are needed according to Harnad are the ones to discriminate stored shapes of horses based on similarity/difference. The important fact is that there is a non-arbitrary, isomorphic relationship between the icon and what it refers to, because the iconic representation has a structural relationship to its origin.\u003C/p>\n\u003Cp>\u003Cem>2 Categorical representation\u003C/em>\u003C/p>\n\u003Cp>According to Harnad, the icons are not enough for meaning. The world is too complex and we need faster identification of what shapes and sense data from our sensory organs are of a specific category. For that, we need to abstract invariant features from the iconic representations and create categories from these invariants, so that in future when we encounter a horse-like shaped object, we can identify faster if this object belongs to the category that has horse-like attributes. At this level, we still do not use the word “horse” to describe the categorical representation, but we identify horses by \u003Cem>the shape of their heads, being four-legged, having a tail\u003C/em>, etc.\u003C/p>\n\u003Cp>\u003Cem>3 Symbolic representation\u003C/em>\u003C/p>\n\u003Cp>Only at this level does the language come into a play. Once we have the “horse” as a set of abstracted features of what resembles a real-life horse, we associate the symbol “horse” in our mind with our perceptually grounded and experiential understanding of a horse. Despite the fact that symbols at this level are completely arbitrary – the shape of the symbol/word “horse” bears no resemblance to the actual horse – they are meaningful because their semantic content (meaning) is inherited from the iconic and categorical representations. On this view, what makes symbols meaningful is indeed their perceptually-derived references to the external world.\u003C/p>\n\u003Cp>Harnad interestingly allows for the possibility that we can create meaningful symbols that have no direct iconic and categorical representations. Harnad gives an example of a word \u003Cem>“zebra”\u003C/em>:\u003C/p>\n\u003Cp>\u003Cem>“zebra”\u003C/em> = “horse” + “stripes”.\u003C/p>\n\u003Cp>If the system has iconic and categorical representations of the word “stripes”, then Harnad admits that the symbolic representation will \u003Cem>inherit \u003C/em>its grounding based on the grounded symbolic representations of “horse” and “stripes”. What does it mean? The system will have a meaningful word “zebra” despite not having any direct perceptual reference of a zebra in the external world. The meaning is inherited, or derivative. Its only direct meanings are references to the linguistic symbols “horse” and “stripes”.\u003C/p>\n\u003Cp>It seems to me that (Bender & Koller, 2020), who rely on (Harnad, 1990), incorrectly assume that Harnad disallows meanings other than those that are grounded. But based on my reading of (Harnad, 1990), he allows meanings of symbols/words not having a direct reference to the physical world, provided that its real references are symbols/words whose meanings are grounded in the physical world. Exactly like the example of zebra = “horse” + “stripes”.\u003C/p>\n\u003Cp>Once we break apart the necessity of a direct connection between the meaning of a word/symbol and its relationship with the physical world, there appears nothing that prevents us from injecting a number of intermediate symbols as the only real meanings of an original symbol, provided some symbol, a last one, is grounded in the physical world.\u003C/p>\n\u003Cp>I think we have undermined the first claim \u003Cem>1) meaning must refer to something outside of language\u003C/em>. We just saw that a meaning \u003Cem>can refer to something inside of language\u003C/em>, but to be in line with (Harnad, 1990), must \u003Cem>eventually\u003C/em> jump out of the linguistic space into a physical space of perceptions and iconic and categorical representations to be grounded, which for (Harnad, 1990) is a necessary a way out of an infinite regress of linguistic symbols referring \u003Cem>only \u003C/em>to other linguistic symbols, never outside of the language system.\u003C/p>\n\u003Ch2>\u003Ca id=\"post-965-_yni9wq403zva\">\u003C/a>Sparse and dense meaning networks\u003C/h2>\n\u003Cp>Returning to our concept of meaning networks, how can we improve our understanding of which meanings are \u003Cem>deep\u003C/em> or \u003Cem>dense\u003C/em>, and which ones are \u003Cem>shallow or sparse\u003C/em>?\u003C/p>\n\u003Cp>Using the symbol grounding theory by (Harnad, 1990), I think we can claim that my descriptions of the meaning network representing the meaning of the word “mother”, as used by me, are a case of a \u003Cem>grounded meaning par excellence\u003C/em>.\u003C/p>\n\u003Cp>Now imagine that I never had a mother. In fact, I have never read anything about mothers, except for the information that a mother is a human female that has at least one child. Imagine further that I have a grounded understanding of symbols “human”, “female” and “child”. How does such a shallow understanding of “mother” differ compared to my personal and deeper understanding?\u003C/p>\n\u003Cp>In the first case, the meaning network would be rather sparse. It would consist only of potentially three linguistic symbols. In the second example, the graphical schema of the meaning of “mother” as used by myself would be a dense network of perceptually derived data, lots of linguistic symbols that are grounded, and perhaps also ungrounded symbols such as “heaven” or “God” with which, for the sake of the argument, even in theory I cannot have any experiential relationship and all direct meanings tied to these two words are linguistic symbols. Some of them may be grounded at lower levels.\u003C/p>\n\u003Cp>As we said, we can have \u003Cem>weighted\u003C/em> edges (meaning elements) of the meaning network connected to a node (word/symbol), as the standard graph theory in mathematics allows. The density and sparseness of meaning networks would not be calculated merely by the number of meaning elements. But the weight of participation of individual meaning elements in the constitution of the overall meaning must be taken into account.\u003C/p>\n\u003Cp>The modelling of meaning as networks also entails that the meaning is not a binary value. Towards the left side of the spectrum, the meaning is increasingly sparse. Towards the other side, the meaning is getting higher in its density.\u003C/p>\n\u003Ch2>\u003Ca id=\"post-965-_ovq933hufu4w\">\u003C/a>Coherence and haphazardness of LLMs\u003C/h2>\n\u003Cp>Previously, we said that (Bender et al., 2021) claim that\u003C/p>\n\u003Cp>2) LLMs produce outputs haphazardly\u003C/p>\n\u003Cp>3) coherence is in the eyes of beholder\u003C/p>\n\u003Cp>The claims 2) and 3) are in fact logically inter-dependent. If LLMs produce outputs haphazardly, any coherence and meaning that there is must be either by sheer chance, or reconstructed by the external system interpreting the output of LLMs. Such system can be, for example, humans.\u003C/p>\n\u003Cp>LLMs are symbol prediction machines. And their success in statistical prediction and pattern matching is dependent on the size of the corpus of data, supplemented by humans. The use of statistics and prediction does not make LLMs haphazard\u003Csup>\u003Csup>\u003Ca id=\"post-965-footnote-ref-4\" href=\"#post-965-footnote-4\">[5]\u003C/a>\u003C/sup>\u003C/sup>.\u003C/p>\n\u003Cp>Unless we are concerned with the fact that the data themselves are haphazard and incoherent. But that is not the case that (Bender et al., 2021) argue for. Whatever we can say about the quality of human discourse, especially the one coming from the online sources, if the data used by LLMs are derived from discourse of native speakers, the coherence is not in the eyes of the beholder, it is objectively embedded within the corpus of the data that LLMs work with, exactly because the coherence is initially created by humans themselves. What LLMs do, it appears to me, is that they apply statistics and prediction to the corpus of data that are linguistic compressions of all kinds of sparse and dense meaning networks created by humans themselves. LLMs can tap into the meaning networks through the exploitation of a fact that \u003Cem>all grounded meanings \u003C/em>or \u003Cem>all meaning networks\u003C/em> can be made explicit by a transformation into pure symbolic forms, i.e. into a language. The language is a window, or an interface into human meaning networks and all meanings grounded in an external physical world.\u003C/p>\n\u003Ch2>\u003Ca id=\"post-965-_yg0v14k93zld\">\u003C/a>Language as an interface into human meaning networks\u003C/h2>\n\u003Cp>We do know that LLMs work by statistically predicting the best next word that fits previously generated strings of texts and their context. If we mean by “stochastic” in the accusation of “stochastic parrot” to merely say that LLMs use statistics to generate their responses, it cannot be considered an argument against LLMs on its own, especially when human brains do certainly employ their own statistical prediction processes.\u003Csup>\u003Csup>\u003Ca id=\"post-965-footnote-ref-5\" href=\"#post-965-footnote-5\">[6]\u003C/a>\u003C/sup>\u003C/sup> Then, what is left is to critique the idea of LLMs being mere “parrots”? Those parrots that can talk imitate human speech sounds rather faithfully but cannot go beyond the data they heard and also, the sounds have no meaning to the parrots themselves. LLMs clearly go beyond the data, generating novel sentences. We can dismiss the first part of the accusations of being a parrot. What about the meaning of the accusation?\u003C/p>\n\u003Cp>LLMs have no bodies, so far. The meaning of their words and utterances cannot be \u003Cem>grounded\u003C/em> in the physical world. This in itself, as we have argued previously, does not constitute a problem leading to meaninglessness, if the words that LLMs output are at some level grounded in the physical world occupied and experienced by humans. Are they? I want to argue that indeed the words produced by LLMs are grounded, albeit by proxy of symbolic mediators that function as an interface into human meaning networks.\u003C/p>\n\u003Cp>Imagine that each human has a meaning network for each word. We admitted that the meaning networks are not fully private, but partially shared within the language community. When we talk and write and want to “get the meaning across”, we compress our complex and messy meaning networks into a language. By its design, the language compresses and makes complex things discrete and separated into graspable units. I can keep in my mind an overwhelming experience of a funeral. Yet, when I wrote into my diary about the funeral, I was forced to strip down the experience into words and sentences. In fact, all sources of writing, from newspapers to poetry and novels are artifacts in which we use the cognitive technology of writing to make seemingly indescribable amenable to linguistic description.\u003C/p>\n\u003Cp>After the compression that comes with using a language, we transform all the meaning elements that are not linguistic into some form of linguistic strings. These strings will be grounded because their meaning eventually refers back to other symbolic representations that are grounded or refer to iconic and categorical representations directly. The more meaning networks overlap for some words, the faster the invariant features of the meaning of the word will get into LLMs, because they are such powerful statistical machines that capture regularities and patterns in the data that even humans do not notice.\u003C/p>\n\u003Cp>After the grounded meanings from human physical experience in the world are compressed into the linguistic symbols that may or may not constitute meaning networks of other words, LLMs can tap into this linguistically compressed groundedness of the words.\u003C/p>\n\u003Cp>By proxy and being parasitic on human groundedness of the meaning, the LLMs use a language as an \u003Cem>interface\u003C/em> \u003Cem>to\u003C/em> \u003Cem>tap\u003C/em> into our meaning networks mediated by all the language output that we as humanity produced and used as input data for the LLMs to learn from. But using an interface comes with a price we have to pay. This act of using an interface – I previously called it \u003Cem>interfaciality\u003Csup>\u003Csup>\u003Ca id=\"post-965-footnote-ref-6\" href=\"#post-965-footnote-6\">[7]\u003C/a>\u003C/sup>\u003C/sup> –\u003C/em> is a lossy compression because the interface inevitably simplifies and reduces. Let me give an example. My feelings and memories and ideas constituting the meaning of the word “mother” I keep in my mind have no specific discrete forms. They are, to use a computer science term, a continuous signal exactly because my experiential (phenomenal) reality that co-constitutes the memory network (along with symbolic representations) is not discrete either. But by using a language as an interface into the meaning network of the word “mother”, I am reducing the continuous signals into discrete forms of words to create discrete, symbolic representations. Inevitably, a lot of detail is lost due to the fact that to make a linguistic representation of a phenomenal reality is a lossy process.\u003C/p>\n\u003Cp>How lossy is it? Depends on the density of a linguistic representation and how deep it taps into the entirety of the meaning networks for each word, and higher order elements like a sentence or paragraph. What are the meanings of these higher order elements? As I see it, the meaning is constituted by a process of merging together meaning networks of each word and \u003Cem>traversing\u003C/em> through the merged meaning networks where the valid rules for such traversing are governed by the syntactic rules of a given language. The lossiness of a linguistic representation would be different between an average person and, say, a Nobel laureate in literature. Whatever the ontological status of experiential reality is, a masterful writer or speaker can use their skill of using the language to bring closer the continuous and irreducible phenomenal reality into something that is graspable and meaningful for us.\u003C/p>\n\u003Ch2>\u003Ca id=\"post-965-_osctxq6frn4d\">\u003C/a>LLMs have dense, modally superficial meaning networks\u003C/h2>\n\u003Cp>For us humans it is normal to enrich our own meaning networks by a cross-pollution that happens when we interact only with linguistic representations of others, such as the literature, that makes the meaning networks denser, despite the fact we have no direct experiential (iconic, categorical representations) references with whatever we read in great novels or poetry or philosophy. Most people who read about neuroscience have no iconic or categorical representations of the terms like “frontal lobes”, “predictive coding”, or “glia cells”. Yet, we are happy to talk about it. A random person who talks about neuroscience has a very sparse meaning network regarding the topic, most likely having few if any symbolic representations that are grounded in an experiential reality. Presumably, most if not all meaning elements in her meaning networks related to the topic would be ungrounded symbolic/linguistic representations, with a few examples of films, photos or interviews with brain patients co-constituting what the topic means to her. If this neuroscience student told us anything about the brains, she would lack the meaning grounded in experiential reality of seeing and operating on the brain, lacking emotional connections to the patients etc, but would be able to recite latest theories, concepts, whole books on the topic. Our neuroscientist would be an expert theoretician on the topic: would analyze concepts, find similarities and differences, could offer criticism based on the coherence to the existing body of knowledge, and may even predict new theories based on seeing patterns and connections in the body of knowledge which no one else could see. Our neuroscience theoretician could even pass a medical school, but we would think that something is missing in her understanding of the topic.\u003C/p>\n\u003Cp>The LLMs are an extreme case of our neuroscience theoretician as far as meaning networks are concerned. Out of the coherent and structured linguistic data that we humans historically produced, LLMs reconstruct patterns that exist in our meaning networks and easily find new ones that no human could possibly do. Through the language, LLMs have an accessible \u003Cem>interface\u003C/em> into our discrete symbolic representation of continuous and irreducible phenomenal reality, which is inevitably a lossy process, but again the sheer amount of the linguistic data may shed more light on the existential reality than any person could possibly hope for. Any modification of meaning networks of LLMs consists of adding, removing, or modifying symbolic representations only. LLMs thus have extremely dense meaning networks with an unimaginable amount of weighted meaning elements. But what they lack, similarly to our theoretical neuroscientist, is a diversity in modality of those meaning elements. Most meaning elements in the meaning networks of the theoretical neuroscientist are symbolic representations. Whereas LLMs are the extreme limit of such “modally superficial” meaning networks where only one of the modalities – symbolic representation – can be used for coding the meaning elements.\u003C/p>\n\u003Ch2>\u003Ca id=\"post-965-_xtrzflbprc93\">\u003C/a>Discussion as summary\u003C/h2>\n\u003Cp>We never encountered an entity that would readily cite, analyse and synthetise the history of human experiential reality mediated through the lossy compression of linguistic representations. To answer whether LLMs can achieve a human-like intelligence depends, of course, on what we mean by “human”. Because of our societal tendency to privilege cognitive and intellectual processes, with the language endowment distinctly a human ability, the language abilities are at a top position on the list of what separates us from the rest of the animal kingdom. In that regard, LLMs have surpassed us. No human will ever be able to hold, analyse, synthesize, and traverse across the meaning networks built from the entire corpus of human knowledge and experience. Despite the fact that LLMs exhibit quantitatively dense but modally superficial meaning networks, the meaning of the output that LLMs generate is not stochastic, nor repetitive. For that reason, I suggest we should reject the current claim of LMMs being stochastic parrots. Instead, the meaning networks latent in the linguistic corpus and embedded in the higher dimensional networks of LLMs has a solid pedigree based on the grounded symbols referencing the human experiential reality. So what, if LLMs themselves did not experience it? We humans, with theoreticians and philosophers at the forefront, readily use words/symbolic representations whose meaning network consists only of other words and concepts represented linguistically, because we have no other experience with it except for other words and concepts.\u003C/p>\n\u003Cp>The density of meaning networks embedded in LLMs is far superior to what our brains can hold. So what, if the meaning networks are implemented in silicon rather than carbon?\u003C/p>\n\u003Ch2>\u003Ca id=\"post-965-_za3x0518451d\">\u003C/a>References\u003C/h2>\n\u003Cp>Bell, D. (1999). \u003Cem>The coming of post-industrial society: a venture in social forecasting\u003C/em> (Special anniversary ed. /). Basic Books.\u003C/p>\n\u003Cp>Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data. In \u003Cem>Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics\u003C/em>. Association for Computational Linguistics. https://doi.org/10.18653/v1/2020.acl-main.463\u003C/p>\n\u003Cp>Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots, 610-623. https://doi.org/10.1145/3442188.3445922\u003C/p>\n\u003Cp>Castells, M., & Cardoso, G. (2006). \u003Cem>The network society: from knowledge to policy\u003C/em>. Johns Hopkins Center for Transatlantic Relations.\u003C/p>\n\u003Cp>Clark, A. (2013a). Are we predictive engines? Perils, prospects, and the puzzle of the porous perceiver. \u003Cem>Behavioral and Brain Sciences\u003C/em>, \u003Cem>36\u003C/em>(3), 233-253. https://doi.org/10.1017/s0140525x12002440\u003C/p>\n\u003Cp>Clark, A. (2013b). Whatever next? Predictive brains, situated agents, and the future of cognitive science. \u003Cem>Behavioral and Brain Sciences\u003C/em>, \u003Cem>36\u003C/em>(3), 181-204. https://doi.org/10.1017/s0140525x12000477\u003C/p>\n\u003Cp>Dijk, J. van. (2006). \u003Cem>The network society: social aspects of new media\u003C/em> (2nd ed.). Sage Publications.\u003C/p>\n\u003Cp>Ferenc, J. (2018). \u003Cem>Postkognitivistické HCI: Vidět interface jako sociotechnický vztah\u003C/em> [Diplomová práce]. Univerzita Karlova.\u003C/p>\n\u003Cp>Harnad, S. (1990). The symbol grounding problem. \u003Cem>Physica D: Nonlinear Phenomena\u003C/em>, \u003Cem>42\u003C/em>(1-3), 335-346. \u003Ca href=\"https://doi.org/10.1016/0167-2789(90)90087-6\">https://doi.org/10.1016/0167-2789(90)90087-6\u003C/a>\u003C/p>\n\u003Cp>Searle, J. R. (1980). Minds, brains, and programs. \u003Cem>Behavioral and Brain Sciences\u003C/em>, \u003Cem>3\u003C/em>(3), 417-424. https://doi.org/10.1017/S0140525X00005756\u003C/p>\n\u003Cp>Webster, F. (2006). \u003Cem>Theories of the information society\u003C/em> (3rd ed.). Routledge.\u003C/p>\n\u003Cp>Wittgenstein, L. (2009). \u003Cem>Philosophical investigations\u003C/em> (Rev. 4th ed.). Wiley-Blackwell.\u003C/p>\n\u003Col>\n\u003Cli id=\"post-965-footnote-0\">Bell, 1999; Castells, 2006; Dijk, 2006; Webster, 2006 \u003Ca href=\"#post-965-footnote-ref-0\">↑\u003C/a>\u003C/li>\n\u003Cli id=\"post-965-footnote-1\">Wittgenstein, 2009 \u003Ca href=\"#post-965-footnote-ref-1\">↑\u003C/a>\u003C/li>\n\u003Cli id=\"post-965-footnote-2\">Searle, 1980 \u003Ca href=\"#post-965-footnote-ref-2\">↑\u003C/a>\u003C/li>\n\u003Cli id=\"post-965-footnote-3\">ibid, Searle, 1980 \u003Ca href=\"#post-965-footnote-ref-3\">↑\u003C/a>\u003C/li>\n\u003Cli id=\"post-965-footnote-4\">If I know that you are a female from the US and your name starts with the letters “JAN”. There is a good chance that the fourth letter in your name is “E”. I cannot be sure entirely, but the answer stems from a statistical analysis of the set of all US female names and the frequency of the letters in the names. This is not a haphazard process to get to the answer. Similarly, LLMs use statistics to get to their answer. I am not sure why the use of statistical methods would constitute an haphazard process of working with data. \u003Ca href=\"#post-965-footnote-ref-4\">↑\u003C/a>\u003C/li>\n\u003Cli id=\"post-965-footnote-5\">Clark, 2013a; Clark, 2013b \u003Ca href=\"#post-965-footnote-ref-5\">↑\u003C/a>\u003C/li>\n\u003Cli id=\"post-965-footnote-6\">Ferenc, 2018 \u003Ca href=\"#post-965-footnote-ref-6\">↑\u003C/a>\u003C/li>\n\u003C/ol>\n\u003Cp> \u003C/p>\n",false,{"rendered":22,"protected":20},"\u003Cp>In this essay, I argue that LLMs are not stochastic parrots by evaluating the claims by (Bender et al., 2021) who coined the term “stochastic parrot” to express the view that the output of LLMs is 1) haphazard, 2) incoherent, 3) meaningless, because, following (Harnad, 1990), it produces symbols that are not grounded in the […]\u003C/p>\n",1,0,"open","","standard",{"footnotes":26},[30],4,[],[33,14,34,35,36,37,38],"post-965","type-post","status-publish","format-standard","hentry","category-article",{"self":40,"collection":46,"about":49,"author":52,"replies":56,"version-history":59,"predecessor-version":63,"wp:attachment":67,"wp:term":70,"curies":77},[41],{"href":42,"targetHints":43},"https://cms.local.test/wp-json/wp/v2/posts/965",{"allow":44},[45],"GET",[47],{"href":48},"https://cms.local.test/wp-json/wp/v2/posts",[50],{"href":51},"https://cms.local.test/wp-json/wp/v2/types/post",[53],{"embeddable":54,"href":55},true,"https://cms.local.test/wp-json/wp/v2/users/1",[57],{"embeddable":54,"href":58},"https://cms.local.test/wp-json/wp/v2/comments?post=965",[60],{"count":61,"href":62},3,"https://cms.local.test/wp-json/wp/v2/posts/965/revisions",[64],{"id":65,"href":66},971,"https://cms.local.test/wp-json/wp/v2/posts/965/revisions/971",[68],{"href":69},"https://cms.local.test/wp-json/wp/v2/media?parent=965",[71,74],{"taxonomy":72,"embeddable":54,"href":73},"category","https://cms.local.test/wp-json/wp/v2/categories?post=965",{"taxonomy":75,"embeddable":54,"href":76},"post_tag","https://cms.local.test/wp-json/wp/v2/tags?post=965",[78],{"name":79,"href":80,"templated":54},"wp","https://api.w.org/{rel}",{"author":82,"wp:term":99},[83],{"id":23,"name":84,"url":85,"description":26,"link":86,"slug":84,"avatar_urls":87,"_links":91},"jakubferenc","https://cms.local.test/wordpress","https://cms.local.test/author/jakubferenc/",{"24":88,"48":89,"96":90},"https://secure.gravatar.com/avatar/be0a3bc58c835417264df524152162a21507cbbd47b541b516705d259a516c1f?s=24&d=mm&r=g","https://secure.gravatar.com/avatar/be0a3bc58c835417264df524152162a21507cbbd47b541b516705d259a516c1f?s=48&d=mm&r=g","https://secure.gravatar.com/avatar/be0a3bc58c835417264df524152162a21507cbbd47b541b516705d259a516c1f?s=96&d=mm&r=g",{"self":92,"collection":96},[93],{"href":55,"targetHints":94},{"allow":95},[45],[97],{"href":98},"https://cms.local.test/wp-json/wp/v2/users",[100,121],[101],{"id":30,"link":102,"name":103,"slug":103,"taxonomy":72,"_links":104},"https://cms.local.test/category/article/","article",{"self":105,"collection":110,"about":113,"wp:post_type":116,"curies":119},[106],{"href":107,"targetHints":108},"https://cms.local.test/wp-json/wp/v2/categories/4",{"allow":109},[45],[111],{"href":112},"https://cms.local.test/wp-json/wp/v2/categories",[114],{"href":115},"https://cms.local.test/wp-json/wp/v2/taxonomies/category",[117],{"href":118},"https://cms.local.test/wp-json/wp/v2/posts?categories=4",[120],{"name":79,"href":80,"templated":54},[],[],1786226783336]