Instead of it being an unknown language, its English (a language you know), but every single Noun, Verb, Adjective or Preposition has been changed to Thai (a language you dont know).

The Mæw Nạ̀ng Bn the S̄eụ̄̀x.

If you had sufficient opportunity to study this pile of text, you'd begin to pick out patterns of which words appear together, and what order words often appear in. Suppose you study this giant stack of Thai text for years in isolation. After all this study, you're good enough that given a few written Thai words, you can write sequences of words that are likely to follow, given what you know of these patterns.

Right, and to get good at this task, you'd need to build models in your head. You would think to yourself, right a Mæw tends to nạ̀ng bn a S̄eụ̄̀x, and you would build up a model of the sort of things a Mæw might do, the situations it might be in. In an abstract way. As you absorbed more and more data you would adjust these abstract models to fit the evidence you had.

You dont know what a Mæw is. But if someone asks you about a Mæw, you can talk about how it relates to S̄eụ̄̀x, Plā and H̄nū. You know stuff about Mæw, but its abstract.

And what is “knowing”? If I know that a Mæw tends to nạ̀ng bn a S̄eụ̄̀x, isn’t that the first thing I’ve learned? And couldn’t I continue to learn other properties of Mæws? How many do I need to learn to “know” what a Mæw is?

As for how you would test it, I think one-shot learning would get one closer to proving understanding.

How would you envision a test based on one-shot learning working?

As for one-shot learning, what I was driving at, is that a truly intelligent system should not need to consume millions of documents in order to predict that, say, driving at night puts larger demands on one's vision than driving during the day. Or any other common sense fact. These systems require ingesting the whole frickin' internet in order to maybe kinda sometimes correctly answer some simple questions. Even for questions restricted to the narrow range where the system is indeed grounded: the world of symbols and grammar.

For your example, the concepts of driving, night, vision, all need to be clearly understood, as well as how they relate to each other. The idea of 'common sense' is a good example of something which takes years to develop in humans, and develops to varying extents (although driving at night vs at day is one example, driving while drunk and driving while sober is a different one where humans routinely make poor decisions, or have incorrect beliefs).

It's estimated that humans are exposed to around 11 million bits of information per second.

Assuming humans do not process any data while they sleep (which is almost certainly false): newborns are awake for 8 hours per day, so they 'consume' around 40GB of data per day. This ramps up to around 60GB by the time they're 6 months old. That means that in the first month alone, a newborn has processed 1TB of input.

By the age of six months, they're between 6 and 10TB, and they haven't even said their first word yet. Most babies have experienced more than 20TB of sensory input by the time they say their first word.

Often, children are unable to reason even at a very basic level until they have been exposed to more than 100TB of sensory input. GPT-3, by contrast was trained on a corpus of around 570GB worth of text.

We are simply orders of magnitude away from being able to make a meaningful comparison between GPT-3 and humans and determine conclusively that our 'intelligence' is of a different category to the 'intelligence' displayed by GPT-3.

If you only ever allowed a child to read about adding, without ever being able to physically experiment with putting pieces together and counting them, likely children would not be able to add either.

In fact, many teachers and schools teach children to add using blocks and physical manipulation of objects, not by giving countless examples and documents discussing addition and procedures of addition.

You may feel it's conclusive, and it's your right to think that. I am not sure.

But I think you are right in what you are saying. Basically it not 'seeing' math as a child does, is just another way to say that it doesn't undestand math. It doesn't have a intuitive understanding of numbers. It also can't really experiment. What would experimenting mean in this context? Just more training cycles. This being math, one could have it run random sums and give it the correct answer each time. That's one way to experiment, but that wouldn't solve the issue. At some point it would reach its capacity of absorbing statistical corelations to deal with numbers large enough. It would need more neurons to progress beyond that stage.

Btw. I found this relevant article: https://bdtechtalks.com/2022/06/27/large-language-models-log...

Ultimately, my point is that although the argument is that an LLM doesn’t “know” anything, I am not sure that there is something categorically different in terms of what we “know” vs what an LLM “knows”, we have just had more training on more different types of data (and the ability to experiment for ourselves).

Your perception of what a cat is, however, is most certainly made of nothing but data, encoded as chemical relationships at the neuronal level. And your perception is all there is, as far as you're concerned. The cat is just another shadow on Plato's cave wall.

Arguably you "know" something when you can recognize it outside its usual context, classify it in terms of its relationships with other objects, and anticipate its behavior. To the extent that's true, ML models have been there for quite a while now.

What else besides recognition, classification, and prediction based on either experience or inference is needed for "knowledge?" Doesn't everything human minds can do boil down to pattern recognition and curve fitting at the end of the day?

So back to the questions of "What is knowing?" "Are talking like someone with theory of mind and having a theory of mind the same thing?"

If your argument is that the only way to answer this it to have a first person experience of that consciousness then that's not a scientific question. No one will ever have one for an LLM or any other AI. It's like asking "What's happening right now outside of the observable universe?". If it can't impact us, it's irrelevant to science. If that ever changes it will become relevant, but until then it's not a scientific question. Similarly no person can ever have a first person experience of the consciousness of an LLM, so anything that requires being the LLM isn't relevant.

So that means the only relevant question is what distinction can outside observers make between an agent talking like a theory of mind and having a theory of mind. And given a high enough accuracy / fidelity of responses I think we're only forced to conclude one of two things: 1. Something that is able to simulate having a theory of mind sufficiently well does actually have a theory of mind. OR 2. I am the only person on the planet with a theory of mind, and all of you are all just simulating having but don't actually have one.

It's all "Searle's Chinese room" and "What consciousness is" discussions all over again. And from a scientific point of you either you get into the "it must be implemented identically as me to count" (which is as wrong as saying an object must flap its wings to fly), or you have to conclude the room plus the person combined are knowledgeable and conscious.

But:

- In this context, following on the whole 2nd half of the 20th century where cognitive science and psychology moved past behaviorism and sought explanations of the _mechanisms_ underlying mental phenomena, a scientific discussion doesn't have to restrict itself to only considering what the LLM says. Neither we, nor the LLM are black boxes. Evidence of _how_ we do what we do is part of scientific inquiry.

- But the LLM does _not_ reproduce all the behaviors of an agent with a theory of mind. A two year-old with a developing theory of mind may try to hide food they don't want to eat. A 4-year-old playing hide-and-seek picks locations where they think their play-partner won't look. They take _actions_ which are appropriate for their goals and context which require consideration of the goals of others. The LLM shows elaborate behaviors in one dimension, in which it has been extensively trained. It has no capacity to do anything else, or even receive exposure to non-linguistic contexts.

I am in no way arguing that only meat-based minds can "know". I'm saying that the data, training regime and model structure used for LLMs specifically is extremely impoverished, in that we show it language but no other representation of the things language refers to. Similarly, image-generating AIs know what images look like, but they don't know how bodies or physical objects interact, because they have never been exposed to them. Of _course_ we get LLMs that hallucinate and image-generators that produce messed up bodies.

On the other hand, there are some pretty cool reinforcement-learning results where agents show what looks like cooperation, develop adversarial strategies, etc. There's experiments where software agents collaboratively invent a language to refer to objects in their (virtual) environment to accomplish simple tasks. I think there are a lot of near and medium-term possibilities coming from multi-modal models (i.e. can models trained on related text, images, audio, video) and RL which could yield knowledge of a kind that LLMs simply do not have.

Someone who is blind, doesn't have a lesser concept of knowing even though they are blind. They might not "know" things in the same way a someone who is seeing, but doesn't mean their version of knowing is any less, they just know fewer facts about the world. Specifically the visual facts of what things look like. Their "knowing" functionality is equal to someone who sees.

Similarly, someone who is blind, and deaf also has full ability for "knowing" even if they'll never know things in the visual or auditory spaces.

So my argument is that your premise is wrong, the fact that someone or something has fewer senses doesn't mean it's ability to know is any less.

So back to your LLM the fact it doesn't exists in the real world is not an exclusion from its ability to know. It does not need to have all of those experiences "to know". It will never know the physical meaning of concepts like we do. Just like I'll never know the details of a city block in Jakarta (as I've never been). But not having that experience (or any experiences of multiple senses) doesn't mean I don't know.

LLMs don't need multiple cross connected sensory experiences, nor extensive history with a physical or virtual world to know things.

For an entity "to know" it means it has a model it can use to make predictions.

> For an entity "to know" it means it has a model it can use to make predictions.

Great, every PID controller, every jupyter notebook or excel spreadsheet with a linear regression model, every count-down timer can make predictions and therefore "know" under this definition. But perhaps there's a broader class of things that "make predictions". Down this path lies panpsychism. When I throw a rock, its velocity in the x direction at time t is a great "predictor" of its velocity in the x direction at time t+delta, etc, etc. And maybe there's nothing inconsistent or fundamentally wrong with saying that every part of the physical universe "knows" at least something insofar as it participates in predicting or computing the future. But I think by so over-broadening the concept of knowing, it becomes useless, and impossible to make distinctions that matter.

I still feel this the the point where you're making a difference based on you desired outcome vs the actual system. ChatGPT absolutely does have precepts / a sense. It has a sense of "textual language". It also has a level of sequencing or time w.r.t. word order of that text.

While you're saying experience, it seems like in your definition experience only counts if there is a spatial component to it. Any experience without a physical spatial component to you seems like it's not valid sense or perception.

Again taking this in the specific, imagine someone could only hear via one ear, and that is their only sense. So there is no multi-dimensional positioning of audio, just auditory input. It's clear to me that person can still know things. Now if you also made all audio the same loudness so there is no concept of distance with it, it still would know things. This is now the same a simple audio stream, just like ChatGPT's langauge stream. Spatial existence is not required for knowledge. And from what I'm understanding that is what underpins your definition of a reality/experience (whether physical or virtual).

Or as a final example lets say you are Magnus Carlson. You know a ton about chess, best in the world. You know so much about chess that you can play entire games via chess notation (1. e4, e6 2. d4 e5 ...). So now an alternate world where there is even a version of Magnus that has never sat in front of a chess board and only ever learned chess by people reciting move notation to him. Does the fact that no physical chess boards exist and there is no reality/environment where chess exists mean he doesn't know chess? Even if chess were nothing but streams of move notations it still would be the same game, and someone could still be an expert at it knowing more than anyone else.

I feel your intuition is leading your logic astray here. There is no need for a physical or virtual environment/reality for something to know.

- I disagree that inputs to an LLM as a sequence of encoded tokens constitute a "a sense" or "percepts". If inputs are not related to any external reality, I don't consider those to be perception, any more than any numpy array I feed to any function is a "percept".

- I think you're begging the question by trying to start with a person and strip down their perceptual universe. I think that comes with a bunch of unstated structural assumptions which just aren't true for LLMs. I think space/distance/directionality aren't necessary for knowing some things (but bags, chocolate and popcorn as lsy raised at the root of this tree probably require notions of space). I can imagine a knowing agent whose senses are temperature and chemosensors, and whose action space is related to manipulating chemical reactions, perhaps. But I think action, causality and time are important for knowing almost anything related to agenthood, and these are structurally absent in ChatGPT UUIC. The RLHF loop used for Instruct/ChatGPT is a bandit setup. The "episodes" it's playing over are just single prompt-response opportunities. It is _not_ considering "If I say X, the human is likely to respond Y, an I can then say Z for a high reward". Though we interact with ChatGPT through a sequence of messages, it doesn't even know what it just said; my understanding is the system has to re-feed the preceding conversation as part of the prompt. In part, this is architecturally handy, in that every request can be answered by whichever instance the load-balancer picks. You're likely not talking to the same instance, so it's good that it doesn't have to reason about or model state.

But I actually think both of these are avenues towards agents which might actually have a kind of ToM. If you bundled the transformer model inside a kind of RNN, where it could preserve hidden state across the sequence of a conversation, and if you trained the RLHF on long conversations of the right sort, it would be pushed to develop some model of the person it's talking to, and the causes between its responses and the human responses. It still wouldn't know what a bag is, but it could better know what conversation is.

That presupposes that our existing tools for detecting the presence of ToM are 100% accurate. Might it be possible that they are imprecise and it’s only now that their critical flaws have been exposed?

And why do you think "feeling of a cat" cannot be encoded as a stream of tokens?

Fascinating, and seems like a plausible description of what's going on.

This feels related to the idea of the Chinese room. There I think the resolution is that the human following instructions does not understand Chinese but the room, the system of instructions + the human to follow them does. In a similar way obviously an individual neuron doesn't understand anything but brains do.

I guess it just feels like this general argument, that merely seeing things and making predictions that turn out to be right isn't enough to understand it will never go away. We could have a full fledged robot walking around having conversations and I could dispute its ability to really understand. It's just learned to imitate other humans I'd say. It doesn't really know anything, it's just following a statistical model to decide how to move an arm

I certainly don't see that as obvious, and I would guess that while you can learn _about_ their perceptual mode, you can't learn what it is like to perceive magnetic fields just through talking about it. I would consider the Mary's Room thought experiment, and the What Is It Like To Be a Bat paper from Nagel.

I think there's a relationship to the Chinese Room, but I want to be clear. In the original formulation, the person in the room follows a book of pre-provided instructions to produce a response. The LLM and person in the Thai text completion scenario must learn an equivalent set of instructions themselves, and for this I would claim that they are comparable to the human + book combination in the original Chinese Room. The person who learns to complete Thai text doesn't know what they're talking about, but they know more than the person following instructions in the Chinese Room. But clearly they still don't know what a Thai speaker knows.

> I guess it just feels like this general argument, that merely seeing things and making predictions that turn out to be right isn't enough to understand it will never go away. We could have a full fledged robot walking around having conversations and I could dispute its ability to really understand.

No, perhaps the end of my original statement didn't make this clear, but I think AI systems _can_ know things, and knowing is not a binary but part of a range. StabilityAI / DALL-e know quite a bit about the relationship between texts and images, and the structure within images -- but they _don't_ know about bodies, physical reality, etc etc. A system that has multiple modalities of perception, learns to physically navigate the world, interact with objects, make and execute plans by understanding the likely effects of actions, etc -- knows and understands a lot. I'm not arguing about a hard limitation of AI; I'm arguing about a limitation of the way our current AIs are built and trained.

In the Chinese room, the instructions you're given to manipulate symbols could be Turing-complete programs, and thus capable of processing arbitrary models of reality without you knowing about them. I have no problem accepting the "entire room" as a system understands Chinese.

In contrast, in GP's example, you're learning statistical patterns in Thai corpus. You'll end up building some mental models of your own just to simplify things[1], but I doubt they'll "carve reality at the joints" - you'll overfit the patterns that reflect regularities of Thai society living and going about its business. This may be enough to bluff your way through average conversation (much like ChatGPT does this successfully today), but you'll fail whenever the task requires you to use the kind of computational model your interlocutor uses.

Math and logic - the very tasks ChatGPT fails spectacularly at - are prime examples. Correctly understanding the language requires you to be able to interpret the text like "two plus two equals" as a specific instance of "<number> <binary-operator> <number>"[2], and then execute it using learned abstract rules. This kind of factoring is closer to what we mean by understanding: you don't rely on surface-level token patterns, but match against higher-level concepts and models - Turing-complete programs - and factor the tokens accordingly.

Then again, Chinese room relies on the Chinese-understanding program to be handed to you by some deity, while GP's example talks about building that program organically. The former is useful philosophically, the latter is something we can and do attempt in practice.

To complicate it further, I imagine the person in GP's example could learn the correct higher-level models given enough data, because at the center of it sits a modern, educated human being, capable of generating complex hypotheses[3]. Large Language Models, to my understanding, are not capable of it. They're not designed for it, and I'm not sure if we know a way to approach the problem correctly[4]. LLMs as a class may be Turing-complete, but any particular instance likely isn't.

In the end, it's all getting into fuzzy and uncertain territory for me, because we're hitting the "how the algorithm feels from inside" problem here[5] - the things I consider important to understanding may just be statistical artifacts. And long before LLMs became a thing, I realized that both my internal monologue and the way I talk (and how others seem to speak) is best described as a Markov chain producing strings of thoughts/words that are then quickly evaluated and either discarded or allowed to be grown further.

--

[0] - https://en.wikipedia.org/wiki/Chomsky_hierarchy

[1] - On that note, I have a somewhat strong intuitive belief that learning and compression are fundamentally the same thing.

[2] - I'm simplifying a bit for the sake of example, but then again, generalizing too much won't be helpful, because most people only have procedural understanding of few most common mathematical objects, such as real numbers and addition, instead of a more theoretical understanding of algebra.

[3] - And, of course, exploit the fact that human languages and human societies are very similar to each other.

[4] - Though taking a code-generating LLM and looping it on itself, in order to iteratively self-improve, sounds like a potential starting point. It's effectively genetic programming, but with a twist that your starting point is a large model that already embeds some implicit understanding of reality, by virtue of being trained on text produced by people.

[5] - https://www.lesswrong.com/posts/yA4gF5KrboK2m2Xu7/how-an-alg...

> you'll fail whenever the task requires you to use the kind of computational model your interlocutor uses.

I think it's important to distinguish between knowing the language and knowing anything about the stuff being discussed in the language. The top level comment all this is under mentioned knowing what a bag is or what popcorn is. These don't require computational complexity, but do require some other data than just text, and a model that can relate multiple kinds of input.

I think it's obviously no, because we don't have sensations of magnetic fields. It's the question of what it's like to be a bat raised by Thomas Nagel. The aliens can give us their words for conscious magnetic sensations which we can learn to use, but we won't experience them. We're basically p-zombies when it comes to non-human experiences.

> There I think the resolution is that the human following instructions does not understand Chinese but the room, the system of instructions + the human to follow them does. In a similar way obviously an individual neuron doesn't understand anything but brains do.

Searle's response to the systems objection is that we already know that brains understand Chinese. But we don't know this for the room. I would further say that brains alone don't understand anything, humans understand things as language users embedded in a social and physical world. One can invoke Wittgenstein and language games here.

A ion channel does not have even a tiny spec of conscience, no matter how you organize them, but our brain does indeed need those to be conscient (and incidentally it relies on a whole lot more "stupid" parts than that: try being conscient without oxygen, or glucose).

I would go as far as making conscience an emergent property of interaction with the environment: what does it mean to be conscious if nothing is there to confirm that you are indeed of a singular conscience? Is it possible to understand the concept of self if you have no concept of other beings?

> What this article is not showing (but either irresponsibly or naively suggests) is that the LLM knows what a bag is, what a person is, what popcorn and chocolate are, and can then put itself in the shoes of someone experiencing this situation, and finally communicate its own theory of what is going on in that person's mind. That is just not in evidence.

Knowing something about the patterns of word order in Thai is not the same as knowing about the world being discussed in Thai.

It also does not "know" that a pizza is an object in a world, because none of the words its working with are attached to any experience or concepts.

> is there anything that would stop LLMs from being able to do the same thing?

If you built an AI system which could hear/see/touch/move etc, and it learned language and vision and behaviors together, such that it knows that a ball is round, can be thrown or rolled, is often used at playtime, etc, then maybe it could understand rather than just produce language. I don't know that we would still call it an LLM, because it could likely do many other things too.

The point, for this thread, is not whether or not Socrates was correct.

Rather, it’s a warning that we must not confidently assume we are anything like a machine.

We may have souls, we may be eternal, there may be something utterly immaterial at the heart of us.

As we strive to understand the inner-workings of machines that appear, at times, to be human-like, we ought not succumb to the temptation to think of ourselves as machine-like merely in order to convince ourselves (incorrectly) that we understand what’s going on.

With that said, there is quite literally zero evidence for the existence of a soul, despite it being posited for thousands of years, and increasing evidence that consciousness is simply a product of a sufficiently connected system. I'll draw an analogy to temperature, which isn't "created", but is a simple consequence of two points in space having different energy levels. I'm sure there's a better analogy that could be made, but I think you get the idea.

More generally, what an LLM actually knows or understands is important if you're considering using one for anything other than generating first drafts which will be fact checked by humans.

Rather than your Thai text example, let's consider a friend of my sister H. H has been profoundly blind from birth. Not "legally blind" with the world a blur, her eyes actually don't work. Direct lived experience of a summer day is to her literally just feeling warmth on her face from the sun, her eyes can't see the visible light.

I've seen purple and H never will so it seems to me you're arguing I "know" what purple is and she doesn't, thus ChatGPT doesn't know what purple is either. But I don't think I agree, I think we're both just experiencing a tiny fraction of reality, and ChatGPT is experiencing an even narrower sliver than either of us and that it probably wouldn't do us any good to try to quantify it. If I "know what purple is" then so does H and perhaps ChatGPT or a successor model will too.

I do agree about grounding is needed. All our language is expressing or abstracting concepts related to how we perceive and interact with reality in continuous space and time. This perception and interaction is a huge correlating factor that our ML models don't have access to - and we're expecting them to somehow tease it out from a massive dump of weakly related snapshots of recycled high-level human artifacts, be they textual or visual. No surprise the models would rather latch onto any kind of statistical regularity in the data, and get stuck in a local minimum.

Now I don't believe solution is actual embodiment - that would be constraining the model too hard. But I do think the model needs to be exposed to the concepts of time and causality - which means it needs to be able to interact with the thing it's learning about, and feed the results back into itself, accumulating them over time.

It's an ironically apt analogy, because ChatGPT has the linguistic understanding of an entity that is deaf, dumb, blind, and has no working senses of any kind, and instead relies on a golem-like automated mass of statistics with some query processing.

We tend to project intelligence onto linguistic ability, because it's a useful default assumption in our world. (If you've ever tried speaking a foreign language while not being very good at it, you'll know how the opposite feels. Humans assume that not being able to use language is evidence of low intelligence.)

But it's a very subjective and flawed assessment. Embodied experience is far more necessary for sentience than we assume, and apparent linguistic performance is far less.

My sister isn't blind. H isn't my sister, she's a friend of my sister as I wrote.

Do you have concrete justification for your insistence that "embodied experience is far more necessary" ?

Much like when humans started experimenting with flight we tried to make flapping things like birds, but in the end it turns out spinning blades gives us capabilities above and beyond bodies that flap.

Back to the embodiment problem. For us as humans we have limits like only having one body. It has a great number of sensors but they are still very limited in relation to what reality has to offer, hence we extend our senses with technology. And with that there is no reason machine intelligence embodiment has to look anything like ours. Machine intelligence could have trillions of sensors spread across the planet as an example.

Note that this isn't just an exotic thought experiment. People like this already exist; the condition is known as "Wernicke's aphasia". People displaying this condition can speak normally. They can't understand things; they are missing a normal mental mapping from words to meanings.

Not really? They can speak in grammatically correct sentences, with connected speech, but what they say can be nonsense. I wouldn't call that normal. I think LLMs show that, solely with access to text, it's possible to produce a good enough model that what you produce is not only not nonsense, but so good that academic psychologists suggest it may have a theory of mind.

> However, often what they say doesn’t make a lot of sense or they pepper their sentences with non-existent or irrelevant words.

https://www.aphasia.org/aphasia-resources/wernickes-aphasia/

As long as our minds pops out appropriate thoughts for the given context we don’t even think about the magic machinery behind the scenes that did that.

When queried about our thinking we are mostly creating a plausible story, not actually examining our own thinking.

Also, blind people can talk sensibly about many visual phenomena, having learned about them through language

I think the new LLM are giving us all so many wow’s, because “understanding” is the only kind of compression that actually works at the scale of the training data

I.e. representations are being created that reflect the actual functional, as well as associative or correlative, relations between concepts.

But blind people can talk about color intelligently too, if not as completely as a sighted person. Despite not experiencing color qualia.

In other words, a LLM that is tied to a GAN that generates images, produces an system that can both describe to you what is a cat verbally and show you a picture of a cat. Does it, then, know what "a cat" is?

Edit: Furthermore, if you then tie this AI to a CV model with a camera which you can point at a cat and it will tell you that it is, indeed, a cat, and then it will also be able to produce a verbal description of a cat as well as show you an abstract picture of a cat or pick cats out of a random set of images, does this whole system know what "a cat" is?

If you, then, make a robot with a camera and hands, attach to the system a more complex CV model that can see in 3D, ask the LLM to produce you a set of code instructions that can be parametrized to produce a motion that would pet the cat, input those instructions into the robot to make it pet a specific cat that has the specific 3D point cloud (I guess that's currently difficult but solveable), and the system will then indeed pet the cat, would it then know what "a cat" is?..

The underlying LLM is still the same in all these scenarios. Where is the boundary?

In other words, the LLM wouldn't be the equivalent of the human brain. Instead, it would just be equivalent to that part of the human brain that processes language.

No, it's not the same LLM; you'd have to change the LLM in all of those cases. How does it receive input from the GAN? The typical LLM is constructed to literally receive a sequence of encoded tokens. There are vision transformers, and they do chunk images into tokens, and there are multimodal transformers, but none of these are fairly described as an LLM, and they're structurally different than something like ChatGPT. And after the structural changes, it would need to be trained on some new data that associates text sequences and image sequences, and after being optimized in that context you have a _different model_.

Does being able to identify images of cats mean the model knows what a cat is? No, and we could have said that a decade ago when deep learning for image classification was making its early first advances. Does being able to describe a cat from video mean you know what the cat is? Probably not, but maybe we're getting closer. Does knowing how to pet a cat mean you know what a cat is? Perhaps not if you need to be instructed to try to pet the cat.

But suppose 10 years from now, I have a domestic robot that has a vision system, and a motor control system, and an ability to plan actions and interact with a rich environment. I would say the following would be strong evidence of knowing what a cat is:

- it can not only identify or locate the cat, but can label parts of the cat, despite the cat having inconsistent shape. It can consistently pick up the cat in a way which is sensitive and considerate of the cat's anatomy (e.g. not by the head, by one paw, etc)

- it can entertain the cat, e.g. with a laser pointer, and can infer whether the cat is engaged, playful, stressed, angry etc

- it avoids placing fragile object near high edges, because it can anticipate that the cat is likely to knock them down, even if the cat is not currently near

- it can anticipate the cat's behavior and adjust plans around it; e.g. avoid vacuuming the sunny spot by the window in the afternoon when the cat is likely to be napping there

- it can anticipate the cat's reactions to stimuli, such as loud noises, a can of food opening, etc, and can incorporate these considerations into plans

Note, _none_ of the above have anything to do with language. If I add to the robot a bunch of NLP systems to hear and understand commands or describe its actions or perceptions, it may now know that a cat is called "cat", and how to talk about a cat, but these are distinct from knowing what a cat is.

Similarly,

- a human with some serious aphasia may be unable to describe the cat, but they can clearly still know what a cat is

- a dog can know what a cat is, in many important ways, despite having no language abilities