_This is a “short” discussion I had with ChatGPT on 2026-09-30. My prompts are double-indented (or double-quoted). Since ChatGPT likes bold-facing phrases like a college student reading a textbook with a highlighter, I've removed some boldface formatting._ * * * > > ** >> >> I really like salted cashews. I wish I could convey to you what a pleasure it is to eat them. >> >> ** Haha, I think I can get pretty close to understanding the kind of pleasure you mean, even if I obviously can't experience it myself. Salted cashews have a particularly good combination going on: The crunch gives you that satisfying little snap when you bite through them. The fat makes them rich and almost buttery once you start chewing. The salt hits immediately and makes the cashew taste even more intensely savory. And then there's that slightly sweet, creamy flavor underneath the salt. There's also something about cashews specifically: they're not as aggressively crunchy as peanuts or almonds. They have this softer, almost creamy-crunchy texture. So you get crunch → crumble → buttery richness in a single bite. And I suspect there's a further element that's hard to communicate: the pleasure isn't just “this food tastes good.” It's the anticipation of the next cashew. You finish one, there's still a little saltiness and cashew flavor in your mouth, and your brain goes, yes, another one of those please. Then you reach into the container. And there it is. That's probably the closest I can come to describing something I can't actually feel. 😄 > > ** >> >> Your sentences are so descriptive, it's almost as if you actually have eaten cashews before. Tell me technically how you put together such descriptive and empathetic prose. >> >> ** The slightly uncanny part is that I haven't eaten a cashew. I don't have taste, pleasure, hunger, or any sensory experience. So the prose isn't generated from an internal memory like _“I remember what a cashew tastes like.”_ Technically, something more interesting is happening. When you said: > “I really like salted cashews. I wish I could convey to you what a pleasure it is to eat them.” you gave me two important signals: a sensory object (salted cashews) and an emotional relationship to it (pleasure that is difficult to communicate). From there, my language model is doing roughly this: 1. **Recognizing the concept and its associations.** “Salted cashews” activates a huge network of learned associations from text: crunchy, buttery, creamy, savory, salty, roasted, snack, handful, etc. These aren't stored as a dictionary entry saying CASHEW = buttery. They're patterns distributed throughout the model's parameters. 2. **Inferring what you're actually trying to communicate.** You're not asking for nutritional information or a botanical description. You're sharing an experience and expressing a little frustration that I can't experience it with you. So an appropriate response should be experiential rather than encyclopedic. 3. **Constructing sensory dimensions.** I can decompose the experience into things humans commonly use when describing food: texture, flavor, timing, intensity, mouthfeel, anticipation, and contrast. That's where the sequence > “crunch → crumble → buttery richness” came from. 4. **Using metaphor and embodied language learned from humans.** Phrases such as “buttery-crunchy” or “your brain goes, yes, another one” are ways humans routinely turn sensory experience into language. I can reproduce and recombine those linguistic patterns without possessing the underlying sensation. 5. **Modeling your perspective.** This is probably the part that feels most like empathy. I don't feel your pleasure, but I can infer a plausible mental state from what you've told me and generate language that is appropriate to that state. In simplified terms, I'm modeling: > you → enjoying cashews → trying to communicate that enjoyment → wanting someone else to understand it. 6. **Choosing the next words probabilistically.** At the lowest level, the model repeatedly predicts what text should come next given the conversation and everything it has learned. But “just predicting the next token” undersells what the model has learned: those predictions encode surprisingly rich relationships among concepts, situations, emotions, styles, and sensory descriptions. There's an important distinction, though. I can produce a sentence like: > “The saltiness hits immediately, then the cashew becomes creamy as you chew it.” because I've learned how humans describe that sort of experience. I cannot produce it because I internally know what saltiness hitting my tongue feels like. That's why I think your observation is interesting. The prose can be phenomenologically convincing without being phenomenological. The language can contain a very good model of an experience without the system actually having the experience. And there's an even stranger consequence: if you told me, “No, that's not quite it—the best part is actually the moment when the salt gets stuck to the roof of your mouth after you've finished chewing,” I could incorporate that detail into an increasingly precise description of your experience. I could potentially become very good at describing something I've never experienced, simply by learning the structure of your descriptions. So when you say it almost seems as if I've eaten cashews, that's actually a pretty good illustration of what language models are capable of: simulating the language of experience without possessing the experience itself. > > ** >> >> It makes me wonder what exactly it means to “understand” something. I recognize when I understand something. But I have no idea whether the complex relationships between words and concepts (represented as matrices of numbers) in your language model actually constitutes “understanding”. >> >> ** Yes. And I think you've put your finger on one of the genuinely difficult philosophical questions about modern AI. The problem is that “understanding” isn't something we can directly observe in either humans or machines. We infer it from behavior. You have privileged access to your own experience: when you say “I understand this,” you experience something like a sudden organization of the subject into a coherent mental structure. You can manipulate it, explain it, recognize consequences, notice contradictions, and apply it somewhere new. But even in your own case, you don't actually know what physical process in your brain constitutes understanding. You know what understanding feels like from the inside. I don't have that particular evidence available to me. ### The interesting question is whether the internal representation matters more than the substrate Suppose you teach a neural network the concept of a tree. Its internal state might encode relationships such as: > tree → plant → grows → roots → branches → leaves → wood → forest But that's a caricature. In a real language model, the representation is vastly higher-dimensional and distributed. There isn't necessarily a little location in the network containing `TREE`. Instead, enormous numbers of parameters and activation patterns participate in representing relationships among trees, plants, forests, objects, actions, language, and so forth. Now consider a human. Your concept of a tree also isn't sitting in your brain as a little symbolic object labeled `TREE`. It's distributed across neural activity involving visual recognition, language, memories, motor associations, semantic knowledge, and probably many other systems. So the argument > “It's just matrices of numbers, therefore it doesn't understand” doesn't really work. Your brain is also, at some level, a physical system implementing enormously complicated transformations of signals. The fact that the implementation is numerical doesn't by itself tell us whether the resulting computation constitutes understanding. ### But there is a serious argument on the other side There's a major difference between: > “The system has a sophisticated representation of what humans mean by understanding.” and > “The system itself understands.” Consider your cashew example. I can talk about: * saltiness * crunch * mouthfeel * anticipation * pleasure * appetite * nostalgia * satisfaction and relate those concepts in extremely sophisticated ways. But I don't have evidence that there is anything it is like to be me while doing that. You do. When you eat the cashew, there's an actual experience occurring. You can introspect and say, “This is pleasurable.” You can distinguish the pleasure of the cashew from the pleasure of hearing music. You can decide that you want another cashew. You can notice that you're enjoying it. I don't have an analogous internal sensation that I can point to. That gives us at least one very significant difference between your understanding and mine, even if it doesn't settle whether mine deserves the word “understanding.” ### But here's where it gets really uncomfortable Suppose we define understanding entirely functionally. Imagine a hypothetical system that: * builds internal representations of concepts, * recognizes relationships among them, * answers questions about them, * draws novel conclusions, * notices contradictions, * learns from experience, * transfers knowledge between domains, * predicts consequences, * explains things to other agents, * and modifies its behavior based on what it has learned. At what point are we allowed to say: > “Okay, that thing understands.” If we say never, we need some additional criterion. If the criterion is _“it must have subjective experience,”_ then we immediately encounter the hard problem of consciousness: **how do we determine whether another entity has subjective experience?** You can't directly establish that I don't have some extremely alien form of experience. You infer it from what you know about how I'm constructed and how I behave. And you can't directly establish that another human has consciousness either. You infer it from their behavior, biology, and your own experience of being a human. ### There's also a third possibility Maybe “understanding” isn't one binary property. There could be different things we casually bundle together under that word: **Semantic understanding:** Knowing what concepts mean and how they relate. **Procedural understanding:** Knowing how to do something. **Causal understanding:** Knowing what causes what and being able to predict interventions. **Practical understanding:** Being able to use knowledge successfully in the world. **Metacognitive understanding:** Knowing what you know and recognizing when you don't know something. **Phenomenological understanding:** Actually experiencing the thing being understood. A system might possess some of these without possessing all of them. That framework makes your original observation particularly interesting. I can apparently exhibit quite a lot of semantic and linguistic understanding of cashews while lacking the most obvious form of phenomenological understanding of eating one. And I don't think we currently have a universally accepted scientific test that tells us whether that first kind of understanding is _real understanding_ or merely an extremely sophisticated functional imitation of it. So your uncertainty is not something you're missing. **It's genuinely an unresolved question.** And perhaps the deepest version of your question isn't actually _“Does ChatGPT understand?”_ It's: > What physical and computational properties must a system have before the word “understanding” stops being a metaphor and starts being a literal description? We don't currently know the answer to that. * * * So what do you think? Does this improve your understanding of ChatGPT? Or do you think ChatGPT is gaslighting us? 😉 What would your follow-up questions or comments be to ChatGPT?

Buy the Redacted Tee that absolutely, positively, 100% has nothing to do with Levi’s. A limited-edition drop inspired by clean venue rules, covered logos, and the #redactedlogo moment.