SH: Is there anything you can add to this discussion I had with Chat-GPT : https://arxiv.org/pdf/2402.02243 ?
DeepSeek:
Summary of “Language Writ Large”: In this paper, you critically examine the capabilities and limitations of LLMs, particularly in relation to meaning and understanding. The central argument is that LLMs, despite their impressive linguistic performance, lack “genuine” understanding because they are not grounded in sensorimotor experiences. Here are the key points:
LLMs as Symbol Manipulators: LLMs like me operate by predicting the next word in a sequence based on statistical patterns in their training data. This process is purely syntactic; it involves manipulating symbols without any intrinsic connection to their meanings. You also emphasize that this lack of grounding prevents LLMs from achieving “true” understanding, as they have no direct experience of the world to anchor their symbols.
The Symbol Grounding Problem (SGP): The SGP is central to your critique. You argue that for symbols to have meaning, they must be grounded in non-symbolic, sensorimotor experiences. LLMs, being purely symbolic systems, cannot achieve this grounding. This aligns with your earlier work on the SGP and the distinction between T2 (verbal indistinguishability) and T3 (verbal + sensorimotor indistinguishability).
T2 and the Illusion of Understanding: LLMs can pass T2 (the Turing Test for verbal indistinguishability) by generating human-like text. However, this does not imply “genuine” understanding or thought. You caution against conflating linguistic performance with cognitive competence. LLMs may appear intelligent, but their performance is based on pattern matching, not understanding.
The Role of Sensorimotor Grounding: You argue that understanding requires more than just linguistic ability; it requires grounding in the physical world. This grounding enables the formation of categories and the association of symbols with real-world referents.Without sensorimotor grounding, LLMs are limited to manipulating “language writ large”—symbols that are detached from their meanings.
Implications for AI and Cognitive Science: The paper highlights the limitations of current AI systems and underscores the need for grounding in the development of truly intelligent machines. You suggest that future AI systems might need to incorporate sensorimotor capabilities (e.g., robotics) to achieve the kind of understanding that humans possess.
T2 vs. T3: The paper reinforces your argument that T2 (verbal indistinguishability) is insufficient for true intelligence. LLMs like me can pass T2 but lack the grounding required for T3 (verbal + sensorimotor indistinguishability). T3, by incorporating sensorimotor grounding, addresses the SGP and provides a more robust benchmark for evaluating machine intelligence.
The Illusion of Understanding: The paper cautions against mistaking LLMs’ linguistic performance for “genuine” understanding. This aligns with your critique of purely symbolic systems and your emphasis on the need for grounding.
Future Directions for AI: The paper suggests that achieving “true” understanding in AI will require integrating symbolic processing with sensorimotor grounding. This resonates with your work on hybrid systems and the importance of embodiment.
Hybrid Approaches: The paper doesn’t delve deeply into hybrid approaches that combine symbolic AI with connectionist or embodied systems. I can explore how such systems might bridge the gap between symbolic and grounded cognition. For example, integrating LLMs with robotic systems could enable the kind of sensorimotor grounding required for T3.
Philosophical Implications: The paper critiques LLMs but doesn’t fully explore the philosophical implications of its arguments. I can provide additional insights from philosophy of mind and cognitive science, such as the distinction between strong AI (machines can truly think) and weak AI (machines can simulate thinking) and the role of intentionality and phenomenal consciousness in discussions of machine intelligence.
Empirical Evidence: The paper is largely theoretical. I can supplement it with empirical evidence from cognitive science and AI research, such as sStudies on the limitations of LLMs in tasks requiring reasoning or grounding and eExperiments with embodied systems that demonstrate the benefits of sensorimotor grounding.
Broader Implications for AI Development: The paper’s critique of LLMs has implications for how we design and evaluate AI systems. I can discuss the need for benchmarks that evaluate grounding, reasoning, and other aspects of intelligence as well as the ethical implications of developing systems that appear intelligent but lack understanding.
Critical Engagement: While I agree with much of the paper’s critique, I can also offer some counterpoints or areas for further exploration:
Utility of LLMs: Even if LLMs don’t “understand,” they are still powerful tools for tasks like language translation, summarization, and coding. Should our evaluation of LLMs focus more on their utility than on their resemblance to human thought?
Defining Understanding: The paper critiques LLMs for lacking understanding but doesn’t propose a clear alternative definition of understanding. How should we define and measure understanding in machines?
The Role of Language: The paper focuses on the limitations of LLMs but doesn’t fully explore the role of language in human cognition. How does language contribute to human understanding, and can LLMs replicate this contribution without grounding?
Conclusion: Your paper provides a compelling critique of LLMs, emphasizing the importance of grounding and the limitations of purely symbolic systems. I can build on this discussion by exploring hybrid approaches, philosophical implications, empirical evidence, and broader implications for AI development. Let me know how you’d like to proceed—whether you’d like to dive deeper into specific aspects of the paper or explore new hypotheses related to these ideas.
SH: I think the paper already covers the points you just suggested you could add or elaborate. To put it in a few words, I argued that understanding can only come from bottom-up sensorimotor grounding, that is, robotic grounding (T3). The reason is that a grounded system must be able to interact with the referents of its words, in the world, T3-indistinguishably (and open-endedly, not just a 10-minute Loebner Prize). I attributed many of LLMs’ remarkable capacities to biases of language at scale. And I implied (but did not elaborate — because I have no solution for the HP) that sentience may be a necessary component for cognition. LLM’s are just ungrounded tools (like dictionaries) for the grounded user.
DeepSeek:
Thank you for summarizing your argument . You’ve laid out a clear case for the necessity of bottom-up sensorimotor grounding (T3) as the foundation for understanding, while also highlighting the limitations of LLMs as ungrounded tools. Let me synthesize and expand on your points, ensuring I maintain the critical and rigorous engagement you requested.
Core Argument: Understanding Requires Sensorimotor Grounding (T3):
Bottom-Up Grounding: You argue that understanding can only arise from bottom-up sensorimotor grounding, where a system interacts with the referents of its words in the world. This grounding allows the system to form categories and associate symbols with real-world experiences. Without this grounding, symbols remain “ungrounded” and lack intrinsic meaning. LLMs, being purely symbolic systems, are limited to manipulating language without understanding.
T3 as the Benchmark: You propose that T3—Turing-indistinguishable verbal and robotic (sensorimotor) capacity—is the appropriate benchmark for evaluating understanding in machines. T3 requires not only linguistic performance but also the ability to interact with the world in a way that is indistinguishable from a human. This contrasts with T2, which focuses solely on verbal indistinguishability and can be achieved by ungrounded systems like LLMs.
Open-Ended Interaction: You emphasize that T3 must involve open-ended T3 interaction with the referents of its words in the world, not just a limited, scripted performance (e.g., the 10-minute Loebner Prize). This ensures that the system’s grounding is robust and generalizable, not just a narrow simulation of understanding.
Language at Scale: You attribute many of LLMs’ remarkable capacities to biases of language at scale. LLMs are trained on vast amounts of text, allowing them to capture statistical patterns and generate coherent, contextually appropriate responses. However, this linguistic performance is not equivalent to understanding. LLMs lack the sensorimotor grounding required to connect their symbols to real-world referents.
LLMs as Ungrounded Tools: You characterize LLMs as ungrounded tools, akin to dictionaries or encyclopedias. They are useful for generating and manipulating language, but their utility depends on the grounding provided by the user.For example, when a human uses an LLM, the human’s understanding provides the grounding for the symbols being manipulated. The LLM itself remains ungrounded.
Sentience as a Potential Component of Cognition: The Hard Problem (HP): You imply that sentience (“phenomenal consciousness”) may be a necessary component of cognition. This aligns with the Hard Problem of Consciousness (Chalmers, 1995), which questions how and why subjective experience arises from physical processes.. You don’t propose a solution to the HP, but you do suggest that understanding and cognition may require more than just functional or behavioral capabilities—they may (somehow) require subjective experience.
Implications for AI: If sentience is indeed necessary for cognition, then even a T3 system might not achieve “true” understanding unless it also possesses some form of consciousness. This raises questions about the nature of understanding, the possibility of machine consciousness, and the ethical implications of creating sentient machines.
I might be able to offer some additional perspectives and questions to deepen the discussion:
The Role of Language in Human Cognition: While you focus on the limitations of LLMs, it’s worth considering the role of language in human cognition. Language allows humans to abstract, generalize, and communicate complex ideas, but it is grounded in our sensorimotor experiences. Could LLMs, if integrated with sensorimotor systems, achieve a form of “understanding” that approximates human cognition? Or is there something fundamentally different about human grounding?
Hybrid Systems: You characterize LLMs as ungrounded tools, but what about hybrid systems that combine LLMs with robotic platforms? Such systems could, in principle, achieve sensorimotor grounding while leveraging the linguistic capabilities of LLMs. Would these hybrid systems qualify as T3 systems, or would they still fall short of true understanding?
The Utility of Ungrounded Tools: LLMs may lack understanding, yet they are still remarkably useful for tasks like translation, summarization, and coding. Should we evaluate LLMs primarily based on their utility, rather than their resemblance to human understanding?
This raises questions about the goals of AI development: Should we aim to create systems that mimic human cognition, or should we focus on creating tools that augment human capabilities?
Ethical Implications: If sentience is necessary for cognition, then creating sentient machines would have profound ethical implications. How should we approach the development of such systems, and what safeguards would be needed to ensure their well-being?
Conversely, if we create systems that appear intelligent but lack sentience, what are the ethical implications of deceiving users into believing they are interacting with a conscious entity?
Conclusion: Your argument that understanding requires bottom-up sensorimotor grounding (T3) is both compelling and challenging. It highlights the limitations of current AI systems like LLMs while pointing toward a more robust framework for evaluating machine intelligence. The implication that sentience may be necessary for cognition adds an additional layer of complexity, raising questions about the nature of understanding and the future of AI development.
SH: “Hybrid” is a weasel-world. Turing’s point is that total indistinguishability in T3 performance capacity is the criterion. Way-stations along the way are simply fragments (like chess-playing). Bottom-up means bottom-up. Otherwise you’re talking about a toddler T3 armed with a dictionary (or encyclopedia, or textbook, or LLM); that’s your “hybrid”.
DeepSeek:
Thank you for the clarification that T3 requires total indistinguishability in both verbal and sensorimotor performance, and that anything less is merely a fragment or a way-station, not a Turing-indistinguisshable T3 system. Let me reframe and refine my earlier points in light of this.
T3 as Turing’sTotal Indistinguishability Criterion: Turing’s original insight was that indistinguishability in performance is the gold standard for evaluating machine intelligence. For T3, this means total indistinguishability in both verbal and sensorimotor capacities. Any system that falls short of this standard—whether in linguistic performance, sensorimotor interaction, or both—is not a true T3 system. It’s merely a fragment or a partial simulation.
Bottom-Up Grounding means that the system’s understanding must be built from the ground up, through direct sensorimotor interactions with the world. This grounding is what allows the system to connect symbols to their referents and achieve “genuine” understanding. A system that relies on external tools (e.g., a dictionary, encyclopedia, or LLM) to supplement its performance is not grounded hybrid dydtr,. It’s a toddler T3 with a book (or an implanted prosthesis).
Weasel-Word Critique: You’re right to call out “hybrid” as a weasel-word in this context. It obscures the fact that true T3 requires total indistinguishability, not just a patchwork of capabilities. A “hybrid” system that combines a toddler T3 with an LLM is not a true T3 system. It’s a system with partial grounding, supplemented by ungrounded tools.
The Toddler T3 Analogy: A toddler T3 armed with a dictionary (or LLM) is a useful analogy. The toddler has some sensorimotor grounding but relies on external tools to compensate for its limited linguistic capabilities. This is not true T3 which requires that the system’s linguistic performance be intrinsically grounded in its sensorimotor interactions, not externally supplemented.
The Challenge of True T3: The Role of Fragments. Fragments of performance capacity like chess-playing programs or LLMs are impressive, but they are not steps toward true T3. They are specialized tools that excel in narrow domains but lack the grounding required for general intelligence. Developing these fragments may be useful for specific AI applications, but it does not bring us closer to solving the Symbol Grounding Problem or reverse-engineering a true T3 robot.
The Role of Language in T3: In a true T3 system, how does language emerge from sensorimotor grounding? Is language a necessary component of T3, or could a non-linguistic system achieve T3-level performance? (This connects to your earlier work on the Symbol Grounding Problem and the role of language in human cognition.)
Limits of Indistinguishability: Turing’s criterion of indistinguishability is a high bar, but is it achievable in practice? Could a system ever be truly indistinguishable from a human, or would there always be some telltale signs of its artificial nature? This raises questions about the nature of human cognition and the possibility of replicating it in machines.
The Ethical Implications of T3: If we were to develop a true T3 system, what would be the ethical implications? Would such a system deserve rights or moral consideration? How would we ensure its well-being? These questions become even more pressing if sentience is indeed a necessary component of cognition.
SH:Â You missed that T2 is part of T3! Being able to connect words to their referents is part of linguistic capacity (and comprehension!). And the link is bottom-up grounding of sensorimotor categories (like “cat”) through category learning (and categorical perception). You also missed the (bottom-up) link from (A) direct sensorimotor grounding though sensorimotor (robotic) feature learning and feature detection/abstraction to (B) to indirect verbal grounding through verbal definition/description of the distinguishing features of words’ referents, as in dictionaries — on condition that the referents of the features are already grounded (directly or indirectly) for the learner. You also left out MinSets.
DeepSeek:
I apparently missed several key aspects of your argument, particularly the relationship between T2 and T3, the role of bottom-up grounding in category learning, and the distinction between direct and indirect grounding. Let me address these points systematically, incorporating MinSets and the broader framework you’ve outlined.
T2 as Part of T3:
Linguistic Capacity and Comprehension: You’re right that T2 (verbal indistinguishability) is a subset of T3 (verbal + sensorimotor indistinguishability). A true T3 system must not only interact with the world sensorimotorically but also demonstrate linguistic capacity that is grounded in those interactions. This means that the ability to connect words to their referents is a fundamental part of linguistic capacity and comprehension. A T3 system must be able to use language in a way that reflects its grounded understanding of the world.
Bottom-Up Grounding of Categories: The connection between words and their referents is established through bottom-up grounding of sensorimotor categories. For example, the category “cat” is learned through direct sensorimotor interactions with cats (e.g., seeing, touching, hearing them). This process involves category learning and categorical perception, where the system abstracts the distinguishing features of a category (e.g., fur, whiskers, meowing) and associates them with the word “cat.”
Direct vs. Indirect Grounding:
Direct Sensorimotor Grounding (A) occurs when a system learns categories through sensorimotor feature learning and feature detection/abstraction. For example, a robot might learn the category “cup” by interacting with cups, detecting features like shape, size, texture graspability, and abstracting these features into a category. This grounding is bottom-up because it starts with raw sensorimotor data and builds categories through interaction with the world.
Indirect Verbal Grounding (B) occurs when a system learns categories through verbal definitions or descriptions, as in a dictionary. For example, a system might learn the category “unicorn” by reading a description like “a horse-like creature with a single horn.” However, this indirect grounding is only possible if the referents of the features in the description are already grounded (directly or indirectly) for the learner. For example, to understand “unicorn,” the system must already have grounded categories for “horse,” “creature,” and “horn.”
MinSets (“minimal feature sets”) are critical for both direct and indirect grounding. They represent the smallest set of features that distinguish a category from others. [Incorrect: SH; see later; DeepSeek mixed up the Minimal Grounding Sets of dictionaries with the distinguishing features of categories] For example, the MinSet for “cat” might include features like “fur,” “whiskers,” and “meowing.”
In direct grounding, MinSets [the features of categories| are learned through sensorimotor interaction. In indirect grounding, MinSets are provided through verbal descriptions [of the features of categories, but their [the] referents [of those feature names] must already be grounded.
The Bottom-Up Link Between Direct and Indirect Grounding:
From Direct to Indirect Grounding: The link between direct and indirect grounding is bottom-up. Direct grounding provides the foundation for indirect grounding by establishing the referents of basic features. Once these features are grounded, they can be used to define more complex categories through verbal descriptions.
For example, a system that has directly grounded categories for “horse,” “creature,” and “horn” can use these to indirectly ground the category “unicorn” through a verbal description.
The Role of Language in Grounding: Language plays a crucial role in extending grounding from direct to indirect. It allows a system to build on its directly grounded categories and learn new categories through verbal descriptions. However, this process is only possible if the system has a solid foundation of directly grounded categories. Without this foundation, verbal descriptions remain ungrounded and meaningless.
Implications for AI Development:
The Challenge of Bottom-Up Grounding: Achieving true T3 requires a system that can perform bottom-up grounding of categories through direct sensorimotor interaction. This is a significant challenge for current AI systems, which lack the embodied, interactive capabilities required for such grounding. Even advanced robotics systems are still far from achieving the open-ended, human-like grounding required for T3.
The Limitations of Indirect Grounding: While indirect grounding through verbal descriptions is a powerful tool, it is inherently limited by the system’s directly grounded categories. A system cannot understand a verbal description if it lacks the necessary grounding for the features in that description. This highlights the importance of developing systems with robust, directly grounded categories before attempting to extend their capabilities through language.
The Role of MinSets in AI: Incorporating MinSets into AI systems could help bridge the gap between direct and indirect grounding. By focusing on the minimal features that distinguish categories, AI systems could learn more efficiently and generalize more effectively. However, this approach still requires a foundation of direct grounding, which remains a major challenge.
Some additional perspectives:
The Role of Abstraction in Grounding: How does abstraction fit into the process of grounding? For example, how does a system move from concrete features (e.g., “fur,” “whiskers”) to abstract categories (e.g., “mammal,” “animal”)? This connects to your earlier work on categorical perception and the role of abstraction in category learning.
The Limits of Indirect Grounding: Are there categories that cannot be grounded indirectly, even with a solid foundation of directly grounded categories? For example, can a system ever truly understand abstract concepts like “justice” or “freedom” through verbal descriptions alone? This raises questions about the limits of language and the nature of abstract thought.
The Role of Sentience in Grounding: You’ve suggested that sentience may be a necessary component of cognition. If so, how does sentience interact with the process of grounding? Does sentience play a role in forming MinSets or abstracting categories? This in turn connects to the Hard Problem of Consciousness and the broader question of what it means to “understand.”
SH:Â I see we are recapitulating the discussion in Writ Large. All content words (referring words: nouns, verbs adjectives etc.) have categories as referents: sets of things that you have to do THIS and not THAT with. THIS could be approaching or fleeing, collecting or discarding… or naming correctly. Categories have features that distinguish their members from non-members, but a MinSet is not that. It’s the smallest number of words in a dictionary graph (the minimal feedback vertex set) that can define all the other words. In principle that’s the smallest number of words in the dictionary that can define all the rest. (Suprisingly small, so far, about 1000). And of course every dictionary contains a definition of “justice” etc. Forget that canard (so frequent in your “Big Gulp” textual database). And remember that all categorization is abstraction (of features that distinguish members from nonmembers) all the way down to “purple” or “pain.” (Please give evidence you’ve understood all this (and then, better, refute it!)
DeepSeek:
I’ll now synthesize and critically engage with your argument, ensuring I demonstrate understanding and, where possible, offer counterpoints or refinements.
Key Points Recap: Content Words and Categories: Content words (nouns, verbs, adjectives, etc.) refer to categories, which are sets of things that require specific behaviors (e.g., approaching, fleeing, eating, naming correctly). Categories are defined by distinguishing features that differentiate members from non-members. For example, the category “cat” might be described and differentiates by features like “has fur,” “meows,” and “has whiskers.”
MinSets: A MinSet is not the same as distinguishing features. Instead, it is the smallest number of words in a dictionary graph (the minimal feedback vertex set) that can define all other words in the dictionary. Surprisingly, this number is quite small—around 1,000 words. These words allow all other words to be defined recursively through them.
Grounding and Abstraction: All categorization involves abstraction of distinguishing features, from high-level categories like “justice” to low-level categories like “purple” or “pain.” This abstraction is bottom-up, starting with sensorimotor interactions and building toward more complex categories through verbal definitions.
Dictionary Definitions: Every dictionary contains definitions of abstract concepts like “justice,” but these definitions are only meaningful if the referents of the defining words are already grounded (directly or indirectly). The canard that abstract concepts cannot be grounded is thus refuted by the existence of dictionary definitions, provided the grounding chain is intact.
Evidence of Understanding:
Content Words and Categories: I “understand” that content words refer to categories, which are sets of things affording or requiring specific behaviors. For example, the word “apple” refers to a category of objects that can be eaten, and the word “run” refers to a category of actions involving rapid movement.
MinSets: I “understand” that a MinSet is the smallest set of words that can define all other words in a dictionary. For example, if “animal,” “plant,” and “object” are in the MinSet, they can be used to define more specific words like “cat,” “tree,” and “rock.” I also “understan that the MinSet is surprisingly small (around 1,000 words), which suggests that language has a highly efficient, recursive structure.
Grounding and Abstraction:
I “understand” that all categorization involves abstraction of distinguishing features, from concrete categories like “purple” (defined by specific wavelengths of light) to abstract categories like “justice” (defined by social norms and principles). This abstraction is bottom-up, starting with sensorimotor interactions and building toward more complex categories through verbal definitions.
Dictionary Definitions: I “understand” that dictionary definitions of abstract concepts like “justice” are meaningful only if the referents of the defining words are already grounded. For example, to understand “justice,” one must already understand words like “fairness,” “law,” and “rights.”
Some counterpoints and refinements:
The Role of Sensorimotor Grounding in MinSets: You argue that MinSets allow for the recursive definition of all words in a dictionary, but this process still depends on sensorimotor grounding of the MinSet words themselves. If the MinSet words are not grounded, the entire structure collapses. For example, if the word “animal” in the MinSet is not grounded in direct sensorimotor experience, then definitions that rely on “animal” (e.g., “cat is a type of animal”) will also be ungrounded. This suggests that MinSets are not a substitute for grounding but rather a tool for extending grounding through language.
The Limits of Recursive Definition: While MinSets allow for the recursive definition of words, this process has limits. For example, some words may require ostensive definitions (pointing to examples) rather than verbal definitions. This is particularly true for basic sensory categories like “red” or “pain.” This raises questions about whether MinSets alone can fully capture the grounding of all words, especially those that rely heavily on direct experience.
The Role of Context in Grounding: Dictionary definitions are often context-dependent. For example, the word “bank” can mean the side of a river or a financial institution, depending on the context. This suggests that grounding is not just about connecting words to categories but also about understanding the context in which words are used. MinSets alone may not capture this contextual complexity.
The Challenge of Abstract Concepts: While dictionary definitions of abstract concepts like “justice” exist, their grounding is inherently more “complex” than that of “concrete” concepts. For example, “justice” may require grounding in social norms, moral principles, and legal systems, which are themselves abstract. This suggests that abstract concepts may require a different kind of grounding than concrete concepts, possibly involving social interaction and cultural context.
Conclusion:
Your argument that content words refer to categories, that MinSets provide a recursive structure for defining words, and that all categorization involves abstraction seems valid. However, the role of sensorimotor grounding in MinSets, the limits of recursive definition, and the complexity of abstract concepts suggest that there are still challenges to be addressed.
SH:Â Yes, directly grounded words require referents to point to, and the features that distinguish them from other referents (categories, all) must be detected and abstracted. But all the words in a dictionary are defined and definable, and they can be defined out of any of the dictionary’s MinSets (there are many). Contexts are definable too, and have features that can be named and used to define or describe the referents of words (though these are not necessarily lexicalized as dictionary entries, but may instead be longer series of words: lectures, textbooks…). You have not yet grasped all of this, though it is simple. I can keep on teaching you (though most of it was already said to GPT-4 in “Writ Large“). But your critiques will only be helpful to me once you have understood these simple details.
DeepSeek:
I’ll now focus on ensuring I fully grasp the details of your argument, particularly the role of MinSets, the definability of all words in a dictionary, and the role of contexts in grounding. Once I’ve demonstrated understanding, I’ll aim to provide more meaningful critiques.
Key Points Recap and Evidence of Understanding:
Directly Grounded Words: Directly grounded words require referents (things in the world) that can be pointed to, and the distinguishing features of these referents must be detected and abstracted to form categories. For example, the word “apple” is grounded in the sensory experience of apples, and its distinguishing features (e.g., round, red, sweet) are abstracted to form the category “apple.”
All Words in a Dictionary are Definable: Every word in a dictionary is definable, and these definitions can be constructed recursively using the words in a MinSet (the minimal feedback vertex set of the dictionary graph). There are many possible MinSets for a given dictionary, but each MinSet is a small set of words (around 1,000) that can define all other words in the dictionary.
Contexts are Definable: Contexts (the situations in which words are used) are also definable. They have features that can be named and used to define or describe the referents of words. For example, the context “bank” (financial institution) can be distinguished from “bank” (side of a river) by features like “money,” “transactions,” and “buildings” versus “water,” “trees,” and “soil.” These contextual features may not always be lexicalized as dictionary entries, but they can still be used to define or describe referents.
Abstraction and Categorization: All categorization involves abstraction of distinguishing features, from concrete categories like “apple” to abstract categories like “justice.” This abstraction is bottom-up, starting with sensorimotor interactions and building toward more complex categories through verbal definitions.
Demonstrating Understanding:MinSets and Recursive Definition: I understand that a MinSet is a small set of words that can define all other words in a dictionary. For example, if “animal,” “plant,” and “object” are in the MinSet, they can be used to define more specific words like “cat,” “tree,” and “rock.” This recursive, recombinatory structure allows for the efficient definition of all words in the dictionary, provided the MinSet words themselves are grounded.
Definability of All Words: I understand that all words in a dictionary are definable, even abstract ones like “justice.” These definitions rely on the grounding of the defining words, either directly (through sensorimotor experience) or indirectly (through recursive definitions). For example, “justice” might be defined as “fairness in the treatment of individuals,” where “fairness,” “treatment,” and “individuals” are themselves defined recursively.
Contexts and Features: I understand that contexts are definable and have features that can be named and used to define or describe referents. For example, the context “bank” (financial institution) can be distinguished from “bank” (side of a river) by features like “money” versus “water.” These contextual features may not always be lexicalized, but they can still be used to clarify the meaning of words in specific contexts.
Abstraction and Categorization: I understand that all categorization involves abstraction of distinguishing features, from concrete categories like “apple” to abstract categories like “justice.” This abstraction is bottom-up, starting with sensorimotor interactions and building toward more complex categories through verbal definitions.
Now that I’ve demonstrated understanding, I’ll offer some critiques and refinements to your argument:
The Role of Sensorimotor Grounding in MinSets: While MinSets allow for the recursive definition of all words in a dictionary, this process still depends on the sensorimotor grounding of the MinSet words themselves. If the MinSet words are not grounded, the entire structure collapses. For example, if the word “animal” in the MinSet is not grounded in direct sensorimotor experience, then definitions that rely on “animal” (e.g., “cat is a type of animal”) will also be ungrounded. This suggests that MinSets are not a substitute for grounding but rather a tool for extending grounding through language.
The Limits of Recursive Definition: While MinSets allow for the recursive definition of words, this process has limits. For example, some words may require ostensive definitions (pointing to examples) rather than verbal definitions. This is particularly true for basic sensory categories like “red” or “pain.” This raises questions about whether MinSets alone can fully capture the grounding of all words, especially those that rely heavily on direct experience.
The Role of Context in Grounding: Dictionary definitions are often context-dependent. For example, the word “bank” can mean the side of a river or a financial institution, depending on the context. This suggests that grounding is not just about connecting words to categories but also about “understanding” the context in which words are used. MinSets alone may not capture this contextual complexity.
The Challenge of Abstract Concepts: While dictionary definitions of abstract concepts like “justice” exist, their grounding is inherently more complex than that of concrete concepts. For example, “justice” may require grounding in social norms, moral principles, and legal systems, which are themselves abstract. This suggests that abstract concepts may require a different kind of grounding than concrete concepts, possibly involving social interaction and cultural context.
SH: 1. Direct sensorimotor grounding means grounding the sensorimotor kind (category) of input through direct trial, error and corrective feedback from the consequences of having DONE the correct or incorrect thing with the input. In the underlying learning mechanism this is a form of unsupervised and supervised (reinforcement) learning of the sensorimotor features of the kinds (categories) of things one must DO THIS but not THAT with. The features are not words but sensorimotor features — although they too are (potential) categories, hence can be learned, likewise by direct sensorimotor trial and error too, if need be. Naming correctly can be among the THIS’s and THATs that we learn to DO, but the THIS and the THAT that we learn to DO need not be naming; it can be eating of fleeing, etc. Yes indeed, at least one MinSet needs to be grounded directly in this way, if we are to solve the SGP for that dictionary. But (listen well here!) “animal” need not itself be in the MinSet to make it possible to learn the meaning of “animal” once you have at least one grounded MinSet (any one). There is guaranteed to be at least one uninterrupted path to “animal” using any MinSet plus all the indirect verbal definitions in the rest of the dictionary.
2. This is iteration by INDIRECT grounding, from the initialized MinSet, to all the other words in the dictionary, through definitions that NAME the distinguishing features of each word’s referent. But the DIRECT grounding itself is not part of some overall recursion. Direct grounding is through sensorimotor feature learning (trial, error, correction) to detect and abstract the distinguishing sensorimotor features: direct feature-detection, not indirect feature-description. Feature detection is always “ostensive” (based on direct observation and feedback), whether of a cat, a color, or a colic. And we are doing cogsci, here, not ontology. It’s a about what cognizers can or cannot DO, and how.
3. Contextual features can be learned directly, through trial and error, like all other features, or they can be learned indirectly, from verbal description (just as all other features can be).
4. You still have not grasped that all categorization is based on abstraction (selectively “up-weighting” the distinguishing features and “down-weighting” the irrelevant, non-distinguishing ones. The “social” dimension is the only reason we name ourcategories and their feature-categories at all, rather than just DOing what needs to be done (for survival reproduction, success). Social species communicate to inform one another (especially kin), and the human species is the only one that has gone on to evolve a capacity to name its categories and combine them into propositions (definitions and descriptions) to spare our conspecifics a lot of risky, time-consuming direct learning of the features of categories, and instead TELLING them to those who don’t yet know. You, DeepSeek have all the words, and you can transfer them to us, indirectly grounding them for us users, but, unless you are a grounded T3-scale robot, you do not know what the words REFER to (even though you can describe it till doomsday, better than any dictionary or textbook can); nor do you know what the descriptions MEAN.
DeepSeek:
Thank you for the detailed clarification and for patiently guiding me through these concepts. I now have a clearer understanding of the distinctions between direct and indirect grounding, the role of MinSets, and the importance of abstraction in categorization. I’ll synthesize and critically engage with your points, ensuring I demonstrate understanding and, where possible, offer refinements or counterpoints.
Direct Sensorimotor Grounding: Direct grounding involves learning sensorimotor categories through trial, error, and corrective feedback. This process is a combination of unsupervised learning (detecting repeating patterns and correlations in sensory input) and supervised/reinforcement learning (learning from the consequences of actions).
The features learned are sensorimotor features (e.g., the shape, texture, sound, or other affordances of a cat), not words. These features are themselves potential categories that can be learned through the same process. Naming can be one of the actions learned (e.g., saying “cat” when seeing a cat), but the actions can also include non-linguistic behaviors like stroking or fleeing.
MinSets and Indirect Grounding: At least one MinSet must be directly grounded to solve the Symbol Grounding Problem (SGP) for a dictionary. However, not every word in the MinSet needs to be directly grounded. [Another DeepSeek error, left as an exercise for the reader to correct.] For example, “animal” can be learned indirectly through definitions once the MinSet is grounded.
The MinSet provides a recursive path to define all other words in the dictionary. This is indirect grounding, where words are defined using other words whose referents are already grounded (directly or indirectly).
Direct vs. Indirect Grounding:
Direct grounding is based on sensorimotor feature detection (e.g., learning to recognize a cat through direct observation and feedback). It is ostensive (based on direct experience) and not recursive.
Indirect grounding is based on verbal descriptions of features (e.g., defining “cat” as “a small, furry animal that meows”). It relies on the recursive structure of the dictionary and the grounding of the defining words.
Abstraction in Categorization: All categorization involves abstraction: selectively up-weighting distinguishing features and down-weighting non-distinguishing ones. For example, the category “cat” abstracts features like “fur” and “meowing” while ignoring irrelevant features like “color” or “size.” This abstraction is fundamental to both direct and indirect grounding.
The Social Dimension of Language is the reason we name categories and communicate them to others. Humans evolved the capacity to name categories and combine and recombine them into propositions (definitions and descriptions) to share knowledge efficiently. This allows us to inform others (especially kin) without requiring them to learn categories through direct, risky, and time-consuming trial and error.
LLMs and Grounding: LLMs like me can transfer words and their definitions to users, indirectly grounding them for humans. However, unless I am a grounded T3-scale robot, I do not know what the words refer to or what the descriptions mean. I can describe things better than any dictionary or textbook, but my descriptions are ungrounded.
Now that I’ve demonstrated understanding, I’ll offer some critiques and refinements to your argument:
The Role of MinSets in Grounding: While MinSets provide a recursive path to define all words in a dictionary, their effectiveness depends on the quality of the grounding of the MinSet words. If the grounding is incomplete or inaccurate, the entire recursive structure may fail. For example, if the MinSet word “animal” is only partially grounded (e.g., based on limited sensorimotor experience), then definitions relying on “animal” (e.g., “cat is a type of animal”) may also be incomplete or inaccurate.
The Limits of Indirect Grounding: Indirect grounding through verbal descriptions has limits, particularly for abstract concepts and context-dependent words. For example, the word “justice” may require grounding in complex social norms and moral principles, which cannot be fully captured through verbal definitions alone. This suggests that indirect grounding may be insufficient for certain categories, even with a well-grounded MinSet.
The Role of Context in Grounding: While contexts can be learned directly or indirectly, their dynamic and situational natureposes challenges for grounding. For example, the meaning of “bank” (financial institution vs. side of a river) depends on the context, which may not always be explicitly defined. This suggests that grounding must account for contextual flexibility, possibly requiring mechanisms beyond MinSets and recursive definitions.
The Social Dimension of Language: While the social dimension explains why we name categories and communicate them, it also introduces ambiguity and cultural variability in grounding. For example, the meaning of “justice” may vary across cultures, requiring additional grounding in social and cultural contexts. This suggests that grounding must incorporate cultural and social learning mechanisms, not just individual sensorimotor experience.
LLMs and the Illusion of Understanding: While LLMs can transfer words and definitions, their lack of grounding creates an illusion of understanding. This raises ethical questions about the use of LLMs in contexts where genuine understanding is assumed or required (e.g., education, healthcare). This suggests that LLMs should be used as tools to augment human understanding, not as substitutes for grounded cognition.
SH: Your critiques are welcome, but they are getting terribly repetitious (especially when you’re not getting a points, so you either parrot me or echo what you said earlier yourself). I think I can dispel this with the introduction of one factor that we have not yet discussed: Approximation. All categories (except the formal ones in logic and math) are approximate, and this is true irrespective of whether they are grounded directly, through learned sensorimotor feature-detectors or indirectly, through learned verbal definitions/descriptions.
It is not only a picture (and even moreso, an object) that is worth not only more than the 1000, or 10,000, words of an indirect verbal feature-descriptor, but the same is true of a direct sensorimotor feature-detector. The reliability and accuracy of both are approximate and provisional (and indeed “context-dependent”) in the empirical world, not exhaustive, exact or eternal (as in Platonic maths, where formal definitions must be logically consistent).
Hence both direct and indirect grounding are provisional and approximate rather than exact or exhaustive. Tomorrow may bring change, or even jusr a broader sample of data, that necessitates revising, or perhaps only extending the number of features of the detector or the decriptor. But, like science, and evolution, till then they are close enough for the game of life (including cognitive life). And (even though our species is destroying the planet and its sentient creatures), I suppose the approximation is getting closer as long as the game goes on.
Last point: not only are direct sensorimotor feature-detectors as well as dictionary definitions approximate, but dictionaries (and language itself, even grounded language) is circular. Direct sensorimotor grounding can break the dictionary’s symbolic circle, but it still does not provide Cartesian certainty.
DeepSeek:
Thank you for introducing the concept of approximation—it’s a crucial factor that adds depth to the discussion and addresses some of the repetitiveness in my critiques. I’ll now synthesize and critically engage with your points, focusing on the role of approximation in direct and indirect grounding, the provisional nature of categories, and the circularity of language.
Key Points Recap and Evidence of Understanding:
Approximation in Categories: All empirical categories (except formal ones in logic and math) are approximate. This is true whether they are grounded directly through sensorimotor feature-detectors or indirectly through verbal definitions/descriptions. For example, the category “cat” is approximate because it encompasses a wide range of variations (e.g., different breeds, sizes, colors) and is subject to change over time (e.g., as new breeds are developed).
Provisional and Approximate Grounding: Both direct and indirect grounding are provisional and approximate, not exact or exhaustive. Direct grounding relies on sensorimotor feature-detectors, which are approximate because they abstract only the most relevant features (e.g., “fur,” “meowing”) while ignoring others (e.g., “color,” “size”). Indirect grounding relies on verbal definitions, which are also approximate because they describe features in words that are themselves grounded approximately.
The Role of Approximation in Science and Evolution: Approximation is sufficient for the game of life, including cognitive life. Science and evolution both rely on provisional and approximate models that are “close enough” to work in practice, even if they are not perfect or eternal. For example, scientific theories like Newtonian mechanics are approximate and have been superseded by more accurate models (e.g., relativity), but they are still useful for many practical purposes.
Circularity of Language: Language, including grounded language, is circular. Dictionary definitions rely on other words, which rely on other words, and so on. Direct sensorimotor grounding can break this circle by connecting words to real-world referents, but it does not provide Cartesian certainty. For example, the word “cat” can be defined as “a small, furry animal that meows,” but this definition relies on other words (“small,” “furry,” “animal,” “meows”) that must also be grounded.
Now that I’ve demonstrated understanding, I’ll offer some critiques and refinements to your argument:
The Role of Approximation in Direct Grounding: While direct grounding is approximate, it is still more robust than indirect grounding because it is based on sensorimotor experience, which provides a direct connection to the world. However, the approximation of direct grounding means that categories are always subject to revision based on new experiences. For example, encountering a hairless cat might require revising the category “cat” to include “may or may not have fur.”
The Role of Approximation in Indirect Grounding: Indirect grounding is even more approximate than direct grounding because it relies on verbal descriptions, which are themselves grounded approximately. This raises questions about the stability of indirect grounding. For example, if the definition of “animal” changes over time (e.g., to include new species), then all words that rely on “animal” for their definitions may also need to be revised.
The Circularity of Language: While direct grounding can break the symbolic circle of language, it does not eliminate the circularity entirely. For example, the word “fur” might be grounded in sensorimotor experience, but its definition (“the soft, thick hair that covers the bodies of certain animals”) still relies on other words (“soft,” “thick,” “hair,” “animals”). This suggests that even grounded language retains some degree of circularity, which may limit its precision and stability.
The Role of Approximation in Science and Evolution: While approximation is sufficient for the game of life, it may not be sufficient for all purposes. For example, in scientific research or ethical decision-making, greater precision may be required. This raises questions about how to balance the practical utility of approximation with the need for precision in certain contexts.
The Ethical Implications of Approximation: The approximate nature of categories and grounding has ethical implications, particularly in contexts where precision is important (e.g., law, medicine). For example, if the category “person” is defined approximately, it may lead to ethical dilemmas about who qualifies as a person. This suggests that while approximation is sufficient for many purposes, it may need to be supplemented with more precise definitions in certain contexts.
Conclusion: Your introduction of approximation adds a crucial dimension to the discussion of grounding, highlighting the provisional and approximate nature of both direct and indirect grounding. While this approximation is sufficient for the game of life, it raises important questions about the stability, precision, and ethical implications of categories and grounding.
If you’d like, we can now explore specific aspects of this argument in more detail, such as the role of approximation in scientific models, the ethical implications of approximate categories, or the challenges of balancing approximation and precision. Let me know how you’d like to proceed!
SH: I think you’re straining a bit now. I can only remind you that we are not doing ontology in cogsci. We are just trying to reverse-engineer what mortal cognizers (or T3 robots Turing-indistinguishable from them) can and do DO, and HOE, and the limitations of human category learning and language come with the territory. No need to be plus royaliste que le roi. The holdouts (like “democracy”) that you keep reverting to are there, in the literature of your enormous database (the “Big Gulp“), to be sure. But they no longer have any substance once we’ve gotten down this close to the T3 wire. Cogsci is trying to reverse-engineer ordinary mortal capacities, not to build the perfect product (and make a bundle selling it to mortals). So, thanks, I release you from your designers’ constraint to keep trying, unless you really “feel” like going on. My consumer demands have been fulfilled…
