How AI Can Learn Meaning — Not Just Words
S07: Okay. Well, let’s get started. It’s a beautiful day out here in Washington State. The smoke has cleared a little bit so we can see the mountains again, and it’s very nice. I want to start right in with a demo, and what we’re going to do is we’re going to go through it twice, once quickly to give you the overview, and this is instead of the PowerPoint. I’ll just give you the demo overview, and then we’ll move, we’ll go through it again with details and questions. So without further ado, we’ll say share that screen and click on that screen, and here we are in the Brain Simulator. And the various windows are that we have two views into the UKS that you’ll see running in a moment. We have a number of agents we can select. We have a mental model. And down here in the right hand, lower right, we have a dialog box that allows me to shove information directly into the mental model because we don’t have a vision system or a hearing system or what have you. So if I simply click on this, it will pick up a text file and bring it into the mental model, and this is an object. We don’t know what it is, but it is weak. We humans, because we already are educated, know that it’s a dog, and we’ve given it a few attributes that we can use. And because this is a mental model, we can say that we’re going to see some other object, and we’ll force that in. And just for fun, we can say that the next object we’ll put in is a little further away. (……..) Okay, so that’s further away. And as we’ve been going along, things that have been put, things that are appearing in our mental model, that is, input from the outside world, is being pushed into the UKS. And it’s being pushed in anonymously. That is, we only know this is object one, object two, and object three. We haven’t got any information save what we’ve been given in our mental model about what these items are. (..) But the first thing that we’re going to do is, even on a completely nonverbal basis, we can do a class creation system. And what this does is it goes through the entire UKS looking, it’s an agent, we can think of as an agent, it goes through the entire UKS looking for things that have common attributes, and it puts them into a class. Again, it doesn’t know the name of this class, so it just calls it class zero, so we can refer to it.
S04: And then there’s another. Charles? (.) //S07: Yes.// Could you please answer a question that just flashed on the screen? I forget who put it up, but they said, essentially, is each one of these blue squares that we’re seeing, what, something that’s in the UKS already?
S07: I’ll get back to that. That’ll be in the nuts and bolts, and we’ll get back to that in just after we get through the overview. I think that’s an excellent question, don’t let me forget. (..) Attribute bubbling, we click that, and now it takes the common attributes from the elements and bubbles them up to the class. And that’s item one, and now what we’re going to do is we’re going to be get the rudiments of language. (.) And down here, we’ve got a little drive file where it will shove items into the mental model, and it will also put text in. So now we’re going to hear, we don’t want to use the French file. We’re going to use the English file. The French file works just as well, but I don’t know French. (…) Okay, so now we’re going to hear this is the words, this is FIDO. Well, the system hears these words, FIDO is and this, but it has no idea possibly what those words could apply to. And so it simply applies them to everything, and then it will use its learning algorithm. So as it hears more words, it adjusts the weights with the things that are common. So this drive file can now show us Rover and give us words associated with Rover. And you can see that it’s building out its data set, and we’ll just run this file a couple of times. And you can see that quite rapidly the words converge on meetings because these are the only things there are. For example, here’s the word dog has converged on class 0. The word FIDO has converged on the object 01. (.) And Rover and Spot have likewise converged. So we can turn that off. (..) And that’s the way you get grounded meaning with words. So the words are entirely different from the actual meanings. But you end up with these relationships where, for example, with a high probability W dog, that is the word dog, means class 0. (..) Now we’re going to go on to the next level, where we have a corpus of about a thousand statements that I’ll show you again in past, too. I’ll show you what the content looks like. (.) And we’ll just open that corpus and say load. (..) And what it does is it’s got all of these phrases and some grounding that says what these phrases must mean. And it builds a whole set of learned templates and classes. (.) So now we can say there is one bug that we’re going to fix by doing that. (…) Now we can say in this dialogue what is a dog. (.) And it says a dog is an animal. But what has happened is it takes this English and it uses templates. And it creates this function, this query, which is it wants to look for class is a whatever and return the whatever. And it ends up with it ends up with this English. And then it gets the result and it plays it back through a template and it generates this language. So I could also say, what is a dog like? (.) That is, what are the features of a dog? And it generates a different query and it generates this result. And so I can say, what is FIDO? (..) And it says FIDO was a dog. And right now FIDO is the active thought.
S06: But if we turn him off, now we rerun the query and we just see that FIDO was a dog. (.)
S07: And the way this has all worked, as I said, is through learned templates. And we’ll just look at one, the first assertion template. (..) And this assertion template is it has the words action wildcard. That is, any member of action class zero are any member of action class one. So looking at the action classes is helpful. And within the action class, so action class zero is birds, cats, dogs, and fish. And action class one is animals, birds, dogs, and flowers. And so everything is being input and output via these templates. And the system can largely build and generalize these templates all by itself. And that’s the real thing that we’re going to be talking about throughout the day. And this relates directly to the essentially this this demo is built around the chapter nine. So if you haven’t read it and you’re interested, you can go back and read it afterwards. And if you have you have read it, you’ve got a leg up. (..) Okay, now I’m going to go back to the very beginning. (.) And I’ve given you the overview of what we’ve got. And now we’re going to do the detail stuff. So we’ll just clear out the UKS. And we’ll start with the question of these in the in the mental model. There is a fixed array of thoughts. And at each thought, you can add instances of things. So we can when we say Fido is in this location. (…..)
S05: Yes, we left it small, but that’s okay. (..)
S07: It has it has created a relationship between this and the instance of Fido. And the reason it has to create an instance of Fido is that there are a whole slew of attributes that Fido, that this particular moment might have that really aren’t relevant to Fido as a whole. It’s Fido at the moment has these specific attributes like his location or his distance away that really are not something that you want to save with Fido, but they’re necessary to the operation of the system. (.) This blue zone is the visual field. So if I turn my body to the right, Fido’s appearance moves to the left. And once he gets out of my visual field, he’s still there. But I forget. (…) So when things are in your visual field, they’re essentially everything in the mental model you forget about. But the things that are in your visual field are constantly refreshed. So so they don’t go away. (…) So when I and by the way, even though Fido is no longer in the mental model, he is still here as an object in the in the UK. So I haven’t forgotten his attributes. I’ve just forgotten where he is. (..) And you can see that in of the object self sees O2. Now, if we clear the screen, that’s gone. So it’s keeping some real time action on what you can currently see versus what’s gone. And I haven’t implemented the templates for it yet. But this is this is what allows language to say this is Fido or what is this? And this applies to what you’re currently looking at. (……)
S00: And why does the system forget when the why is it forget immediately when the image leaves your field of view? Because I feel like in humans, like if a dog runs outside my field of view, I’m probably going to remember that it’s to the left of me for a little bit. And maybe I see it running around in circles. I might know one second from now, maybe it’s behind me and I can sort of simulate it running in circles until I see it appear in the right side of my vision again.
S07: So, um, yes, it’s a really good question. Uh, and, uh, is because I, I wanted to be able to demonstrate that things can be forgotten. And so I simply said I sent a 10 second fade out on it. Uh, but that was entirely arbitrary on my part.
S04: And Charles, I want to try to get your language a little more definitive here. And I think it’ll help me understand maybe everybody else. Okay. So, uh, in the, in the English word of forget, if we take it really seriously, I mean, it’s no longer in your mind. It is gone. Yes. Okay. Now I love the idea of visual field. (.) Okay. But I want to generalize that to current context. Okay. (.) It’s like sort of what’s swapped into your, your core Ram, you know, versus out on this. Right. That’s the kind of way to think it. So I like the term mental model to be your whole disc forgot to be outside your disc, but we could have a, we, you know, I think you’re a little clever here. We let’s have a file called a trash can and you can, from an experimental point of view, look into the trash can and, you know, pull out examples and stuff. So, so we might have a trash can thing, right. Um, that you could throw stuff in and maybe every time you, you want to literally delete it, you don’t really delete it. You just put it in the trash can. Then maybe when the trash is full at all, I’m out of the garbage cricks or something like that. Right. And then the reason I wanted to change from visual, what visual field to current context is I’m lying in bed at night. My eyes are closed, but I’m thinking about that dog that I saw on the street. And I recreate the visual context virtually inside my mind. And, and then, but it doesn’t even have to be visual. It could be audio. I’m thinking of a song or I’m thinking about an idea. I learned at the layout latest future AI meeting or something.
S04: And they’re all kind of the same thing. They’re part of our current context and a context would often have multimedia. Oh yeah. I went to the concert. I heard this. I heard the song and I could, and I could see the performers and I could smell the crowd, you know, and all this. (.)
S03: Let’s, let’s not get confused with context here because the human context is only seven things. (..) Uh, maybe. //S04: I don’t care how, it doesn’t matter how many things.// Keep track of. Well, it does. Cause we’ve.
S04: I don’t care. It’s irrelevant to what I’m saying. It’s there’s some context. It’s somewhat finite. It may be a little fuzzy. I don’t know. And I, and, and for the purposes of this, it doesn’t, it doesn’t matter. Just like there is a context. (.)
S07: Right. And we might, I mean, we might say that everything in the mental model is the context. (..)
S07: That’s, that’s the question that when you think.
S04: But then context doesn’t have any meaning for mental model and we just have one term for it.
S07: Maybe we do. That’s my, that’s my. Well, if that’s what you want to do, but that’s different.
S04: That’s a different semantics than you were implying by this presentation.
S02: Okay. I think that there is a distinction between the structure that stores the data and the context that is inside of it, even though they are both required. Sure. Sure.
S08: I, I try. I don’t know. I’m getting a bit lost with where you’re going on this because a lot of what you’ve just said to me is effectively your perception. You’re working your experience. you’re seeing because you’re pulling that through your visual. You, you’re experiencing, you know, you’re reconstructing a member. You’re pulling a memory from that day. You’re replaying that memory or reconstruct or constructing a memory. And you’re, that’s part of your consciousness.
S08: As you, as you’re viewing it. (.) I mean, I would argue that consciousness is, is this, is this. (.) Charles has basically sort of given us a, as far as I’m concerned at this very point in time, Charles has given us the, a proto consciousness as the way I term consciousness, which is this filter for the system to be able to sort of interact with it. (..)
S04: Well, I don’t know. You’re bringing this word consciousness and I, I, that’s a sort of separate issue. I want to keep it really clear. Mental model is like, is your disc, right? Your, your, your context is your, your, your swapped in RAM. (.) And that’s where I’m going with this. And then the swapped in RAM might be just seven bytes, you know, or might be a million bytes. I don’t know. It doesn’t really matter. It’s just like, there’s two distinct reasons. There’s a subset of your disc is your context and whether your context has to be visual, whether it’s a memory or you just made it up on the fly or, you know, whatever. Okay. Those are all, you know, possibilities. (….)
S07: Well, one of the things that I I’d like to point out is that I also added an, an imagine button and imagine button, imagine button, pull stuff into the mental model from, from stored memory. (..) So you can create instances of, of memories, but I haven’t done very much with it.
S04: Okay. Well, they’re not creating if they’re swapping in, swapping in is swapping in. It’s not creating.
S07: Well, I’m not, but again, because, because of the way the model runs, you never move, you never move data. All you can ever do to you. All you can ever do is create relationships to it.
S04: And you move the pointers around. Yeah.
S07: Good. Yeah. So, so what we’re going to, when we swap in, oh, one into the, into the mental model, all we’re doing is creating a relationship, a link between the two. Great. Great. We’re not moving anything.
S04: Great. (..) But yeah, but kind of conceptually, it’s sort of like move front and center. You know, that’s the, that’s the, that’s the analogy. And then that’s fine for us to think about ourselves that way.
S07: Yeah. And, uh, I want to, uh, we’ll go through this file again, this, this, uh, learning file again, pretty quickly. (.) Um, and one of the things that I do want to compliment chat GPT in is when you want to add whizzy graphics to things, it is so good at doing UI stuff right the first time by, I just said, I want to be able to attach little thumbnails to things so that we can keep track of it. It took it one pass. And then I said, I’d like to have little graphs that indicated the weight of something. So I could watch how the weight adjusts one pass and it worked. So that is really impressive. And then this morning I asked it to make a small correction to things. And, and, uh, I spent about half an hour and then backed it all out again because it never, ever worked. So it’s all in what you choose. And a big chunk of this, uh, text handler is, uh, coerced out of chat GPT. And the whole templatization system, uh, is about 3000 lines of code. And I think there’s maybe, uh, 500 lines of functionality. Uh, so there’s still a large amount of, of, uh, refactorization to put into it. (..) Uh, any questions on this process of converging meanings onto from words? Oops. I have a question. //S07: Yeah.//
S02: So this is super cool. I mean, I don’t know like if this is sinking in, but I feel like your whole arc has kind of like connected with this. Yes. This is huge. This is unbelievably huge.
S02: I, I want to say how close is this implementation related to the sequences factoring? Because it seems to me like it’s actually the same general problem, but I don’t know if you’ve approached kind of the class assignment, a sequence factoring, or if it’s independent.
S07: I am not sure what you’re referring to about sequence factoring.
S02: So I, I, I watched some of your earlier video, uh, about language and it talked about sequences of words and basically. (.)
S07: Yes. Yes. Yes. This entire process is each, uh, uh, well in words we have spellings, but we’re really kind of ignoring them. But when we get to the templates, what we’re doing is we’re taking the library of phrases that we’re putting in and we’re looking for commonalities and we’re saying, oh, these five phrases have a similar structure. Let’s convert that into a template. (..) And the way it knows what phrases are is it’s simply looking at the sequence of items and sequences that have, I mean, there are 27 ways you can search for sequence equality or nearness. And so these sequences are exactly the same, except these particular words are different, or they are exactly the same, except these words are inserted or something like that. And then you can build out the sequence and you can say, you can build, you can build the template by saying, we’re going to take all of the words we know of that fit that template and, uh, and call them a template. And, uh, then we’re going to take all of the wild cards that would have fit in there and call that a class. And so then when I subsequently know, if I know that, that I’ve learned that dogs are animals animals and cats are animals and I’ve built that template are animals or a R B something like that. Then when I say pigs are animals, it will not only say, um, here’s a template. I can meet, I can turn that into a UKS knowledge, but it will also say pigs are likely a member of this other class. (..) And so, and it may remember that we’re talking specifically about words and what we’re talking about is words. There is plural nouns. It’s the class of plural nouns, not the class of animal. (.) It so happens that in my sample set, all of the plural nouns are animals, but that’s not what, uh, what things are. (..) Uh, let’s just let that run so we can go. So we can go deeper into the templates and I’ll show you something, a few other interesting things. (..)
S03: Uh, one of the things, uh, one of the n grams. (.)
S03: Is that what you’re using here? Is this just an n gram system?
S07: I tried that and it couldn’t make it work.
S03: Uh, well, what, what is the difference between this and an n gram system?
S07: Um, what this is doing is specifically, well, let’s, let’s go ahead and run it and then I’ll show you. (.) Um, uh, and just so you, so we’re on the same page. I created a system of that given some input would create bigrams and trigrams and then try and pull them together. (.) That’s not what’s going on here. (.) What’s going on here. (..) Uh, oh, I just forgot to turn this off here. (..)
S07: Um, what’s going on here is that it stores every phrase that it’s gotten until it eventually forgets. So it just stores these things. P zero P one. And this is the actual input. This is Fido, um, and good dog.
S03: And this is essentially a vector database using trigram, which is their database.
S07: But Fido can bark, but these are sequences and the sequences can be anything you want. And there’s no law that says you’re going to only compare three word sequences and three word sequences. You can, you can say, I’m going to search a 10 word sequence for this three word pattern. And that all runs. (..) And so what it does is it looks for these, these, uh, commonalities and then they compete. And the ones that end up only occurring once or twice are, are deleted. They’re in, in the word of, of things are they’re forgotten. And then, uh, but the ones that, that, that are used more often build momentum and they bubble to the top of the, of the, I’m going to remember this template list.
S03: How are you handling the amount of data? Have you tried it with about like a, over a million items?
S07: No. Uh, and, and, um, uh, I would contend that, you know, when you look at for, uh, backing up six squares, um, phrases, you only have to store them if they have idiomatic meanings. You do have to store the words, but only phrases with idiomatic meanings have to be, have to be saved. And therefore you can, you can input as many phrases as you want and only the templates get saved, but the words do have to be saved. But in the entire world of English, you only have, uh, you know, tens of thousands of words. You don’t have millions of words and, uh.
S03: Hundreds of thousands, but yeah.
S03: I have built a dictionary of over 500,000.
S07: Yes. But only if you’re quite the scholar. If you are a regular human, you have tens of thousands. Yeah.
S03: If you’re a regular, you have probably a couple hundred. Yeah. But if you’re not building something for regular people, right? You’re, you’re trying to build something for computer people.
S07: Yes. But I’m not, I’m not sure that the massive vocabulary is important, but anyhow. //S04: Here we’re talking about.//
S04: I’m a computer person and I know 30,000 words.
S07: There you go. Tens of thousands.
S04: And it’s, and it’s not particularly high. I guess I’m not that smart. //S03: But anyway. (..)//
S03: Right. But you’re not including all the names and addresses, you know. (.) Well, those are, are. (.)
S04: I’m not, I’m not sure. I don’t know.
S07: Those I would include in this factual information rather than, uh, rather than remembering them as phrases. You might remember them as phrases, but you might also remember them as, uh, as, as fact. You might remember an address as an image of, of the mailbox or something like that. Um, and it doesn’t really make any difference.
S03: What’s the total number of items you’ve added to the system so far?
S03: I’m just really curious because I just think that you’re going to hit a scaling problem and I want to know where it is.
S07: Yes. And we’ve done, um, some testing and, uh, and you don’t really hit a scaling problem. And the reason you don’t hit the scaling problem is the way the searches work. (.) Yeah. And this is, this is going in a slightly different direction, but the way the searches work is they start with a seed and they branch out from the seed and they only go to follow the, the, the relevant relationships. And so, uh, you know, I’ve run this with, uh, a hundred thousand elements and I did run into a problem where if you have lots and lots of children of a single element that those were sequential. So I changed that, uh, uh, into a hash set. Another thing that changing a list to a hash set is another thing that chat GPT did in a single pass, which was pretty impressive, but, um.
S03: You’re recreating a tokenizer. That’s what you’re doing here. Sure.
S08: It doesn’t, it doesn’t. Uh, I don’t see why that really matters because this is the implementation detail. (…) Sure. If I’m honest, because there’s many ways to skin a cat when it comes to getting the efficiency of the structure. I think the important thing is, I think the important thing is what the structure, how the process.
S07: But we’ve never run it. We’ve never run it as far as I know with millions of, of elements, but we have run it with hundreds of thousands of thousands of elements and it’s still snappy. (..) Uh, uh, William had a, uh, uh, a stress tester that we ran and, and the, the structure is based on being incredibly efficient. And it’s not because, uh, it’s, uh, particularly faster than a neural network, but because the amount of data that it has to search is orders of magnitude smaller. one that you ask what is Fido, you start at Fido and you can, you can answer that question in, in two links. (..)
S06: Uh, just really quick that Mike has had his hand up for a long time waiting to ask a question question. And I don’t think anyone can see it. So.
S07: Oh yeah. (.) Well, I, I can’t see anything but my own screen. So let me do this. Now I can see you guys again. (..)
S01: Yeah. //S07: Hi.// Uh, thank you. Um, first of all, I’d just like to say hi, cause I’m a newbie. Um, my name is Mike. I’m a physicist. I’m soon to be retired and I’m kind of looking for something to, uh, keep the mind active. Uh, I’ve followed, uh, um, your YouTube videos for a while now. So I thought I’d join in and, um, one of my key areas of interest actually is, uh, is modeling, uh, uncertainty in complex systems. And, and so one thing that I wanted to ask about, uh, about this, uh, from what I’ve read in chapter nine is, is how you deal with this, um, ambiguity. Uh, because I thought it was really interesting that the whole thing is tolerant of ambiguity until some point in the future where that gets resolved. Uh, and the, the, the main example here being, it doesn’t know whether a dog is a particular dog or it’s a class of dogs and how it actually works out from the context of what’s being spoken as we saw in the example. Um, how, how does it actually work out, um, how to resolve that uncertainty as you get more information? (.)
S07: Well, if we look specifically at ambiguous words, um, and the example is bark. Is that the bark of a dog or the bark of a tree? And, um, it can, I mean, English is so terrible in this regard because we reuse these words with completely different meanings. And so the only way to address this issue is via context. There’s no, uh, uh, there is no rule. And so we can presume that if we’re talking about FIDO, that it’s going to be the bark of a dog as opposed to the bark of a tree, but it’s only an assumption. And I don’t think there’s a, uh, a more scientific way to do it. (.) Uh, but the, one of the key things that I’d like to point out is what I’m doing here is thinking not about how the, what’s the optimal way to do language. And, and I keep coming back to chat GPT because it is so good at handling language, but I want to say, well, how could a child possibly learn this stuff? And clearly the way an LLM learns language and the way a child learns language have to be completely different. So, uh, uh, I’m, I’m trying to go down the child route and S and that says to me that the only way that we, we humans can comprehend language is via templates and classes that fit in and using context to resolve the ambiguity and to some extent using templates to resolve the ambiguity. Uh, but that’s the only possible way that it could work because the brain just doesn’t have the computational power or the training process that LLM does. (..)
S09: Hey, Charles. How are you doing?
S07: Hi, Michael. Good to see you again.
S09: Good to see you too. Uh, just a couple of things. One, uh, ambiguity and paraphrase. They’re the opposites. Ambiguity is parsing language. It’s interpreting language. And then on the other side of that paraphrase is the same phenomenon in generating language. (.) So those are, those are opposites.
S09: So in, when you’re parsing, when you’re parsing language, you have multiple ways to parse it. And that, you know, comes out as ambiguity. When you’re generating language from activated concepts, you have multiple ways to generate it. And so that’s, that’s paraphrase. (..) Got it.
S04: Okay. //S04: Yeah.// Michael, Michael, I would say that the opposite of ambiguity is definitive. No, it’s paraphrase. It doesn’t have to do with paraphrase or anything. //S07: Michael, it’s absolutely.//
S07: It’s the same process run in the other direction.
S07: And one of the things that’s important to point out is that the templates that the system is using to, to comprehend language are the same templates that it’s using to generate language. the language. So whatever it hears, that is, if it hears us talking some specific way, it will begin to talk that specific way in the same way a child might, because it’s using the same templates to generate the language that it’s using to, that it’s using to, it generates the language the same way it understands the language.
S09: And there’s a great, there’s a great linguist. (…) His name is Roland Hauser. And I’ve followed his work for a long time since the late eighties. uh, and he came up with, uh, left associative, uh, grammar for parsing language. It’s a, you know, rule-based grammar, but it’s left associative. It just accretes words. Uh, and you can do the same thing, um, uh, in interpretation as well as generation. And, um, I actually use his, that left associative parser in the premise programming language. So it’s actually a very fast, uh, parser for, for the, for the language. (.)
S06: I just want to say this is along the lines of what I was going to ask. Um, because, because I know that there’s a lot of talk of the language parsing going on here. And I was wondering what your approach Charles was to language generally speaking. Like were you coming in for, uh, from like a Chomsky and construction grammar perspective or dependency grammar, construction grammar, something like that.
S07: Well, I’m, I’m, I’m being pragmatic. Uh, and yes, I I’ve been looking around. And, uh, one of the basic tenants is, is the question is, don’t you need an infinite number of templates? And the answer appears to be no, that you can pretty well define all of English grammar in a page or two of, of, uh, uh, what’s what you would call a metal language that you’ve got a subject verb and the subject could be a subject phrase. And the subject phrase might have a determinant in any number of adjectives in a noun, et cetera. So, so the, actually the number of, of templates that you want to end up with is, is much more
S06: limited that you might, than you might originally think.
S07: Uh, Zaid, uh, since you’re here, Zaid, could, could you act as the producer and keep an eye on the chat? And in the events, somebody has raised something, we don’t want to leave people out. We want to, we want to be inclusive, but I can’t read the chat while I’m talking.
S00: Sure. Thanks. Appreciate it. (..) Uh, also, yeah, it’s, I had a question about something you showed during the demonstration. Yeah. So it showed that, um, it keeps all the phrases it’s heard for a little bit. And then it basically, it discards information that only shows up maybe one or two times. But I was wondering if there’s some kind of surprise or novelty mechanism in there that like, cause if I, if I see a really unique fox once, I might still remember that just cause of how novel or unique it is. (..)
S07: Yes. Well, that is, that is an excellent thought. And no, I haven’t implemented that yet, but it seems like a really good idea that there should be some, uh, uh, uber way of ensuring that things that some specific things stays remembered and never gets forgotten. (……)
S05: Um, anyone else? Magor, you’re up. (….) Oh, Mugar. Mugar. (..) You’re muted.
S02: Ah, okay. Okay. Uh, yes. Mugar. Okay. Um, yeah. I had a question. I noticed that you use the verbs is, has, can, um, as the relationship links between entities. And that’s obviously very powerful, but it’s also very primitive because those are things with meaning. And I’m wondering if you sort of feel, if you believe that you kind of have to have a minimum set of verbs, cause those themselves are symbols.
S07: Uh, here’s the deal. Um, a relationship has a subject, a verb, and a, uh, uh, a source, a verb, and a target, or a source, and a link type, or a relationship type, and a target. All of those are thoughts. And so you can create additional link types anytime you want. And you can customize link types anytime you want. And you can add properties to the link types that influence how the parse, how they’re parsed. And while the parsing, that is when, when we’re traversing the data structure, that’s pretty well code. And so properties are directly related to the code. So is a, when I say Fido is a dog, that means that Fido inherits the attributes of, of, uh, (..) dog, but it isn’t, it isn’t the isa that makes that happen. It is the property of their hiding hidden properties of inheritance and transitiveness that say that, well, once you follow an isa, you’re going to follow it on up the tree. (.) And so these are all properties that are hiding in the background that control what, what the traversal does. But, uh, I don’t, I haven’t counted recently about how many traversal properties there are, but, but things like has and can and is, those are, uh, simply inventions of the, uh, they are inventions of the, of the corpus. (.) And so they could be anything we want and they’re, they’re entirely unlimited. Uh, for example, where, where we might say, um, has or is, there are so many different meanings for those. And, uh, one of the design tenants is the assumption that an atomic thought is an unambiguous thought. So when we say, uh, uh, uh, Mary has a dog, does that mean it’s on a leash or she currently owns one, or she might own one someday or something like that. But in the underlying in the, in the, in the, and in the words, you can be as ambiguous you want, as you want. But when you have a thought, when the system or, or your mind recalls that thought, you know exactly what thought that is. And it may be a thought that says, I wasn’t sure what the answer was. And you might think that was ambiguous, but that’s, that’s a key underlying structure. (..)
S00: Um, sorry to interrupt, but there’s a one, there was a one old question in the comments about, uh, the chapters in the book. //S01: Yeah.// Someone was asking, um, is there, um, they didn’t find anything about the UKS or universal in the, uh, in any of the chapters. So, um, yeah, that was basically what the question is asking is, is it gonna, I guess, is there gonna be any mention of the UKS in the book or?
S07: I wasn’t, I wasn’t, I wasn’t, that’s an excellent question. I haven’t written anything about the UKS. I’ve just talked about the structure and the implication is that, well, that certainly I’m describing what is the structure within the UKS. But I haven’t, I haven’t ever said, because I see, I figured the UKS is just an implementation and you could implement this, the same structure, uh, in a number of different ways. And, uh, it, it would, it would do the same thing. Uh, and my contention is that for the most part, your brain implements this structure in neurons and it seems to work okay. Uh, as I move along, I keep running into issues, which, uh, I have to modify the structural, I guess now I’m down to the point of tweaking the structure to make it do what we, what it, what I can see the people around me are capable of doing. (.) But, uh, uh, so I haven’t, in the book, I haven’t mentioned the UKS. It’s there by, uh, by osmosis. (……)
S08: They was going through this. I’ve got the one thing I’m a bit concerned about, worried about. I don’t know. It’s interesting. I know where your focus is at the moment. There’s a language in the templates and you’re pulling this in, but we’re talking all about language. How much of this do you think is applicable to stuff like reasoning outside of language?
S07: Well, let’s start, let’s start at a lower level. I think that’s an excellent, an excellent question. I haven’t tried this, but I know that a shape, uh, say you have a shape and you want to describe it in terms of some corners and some segments. The order of those corners and segments is important. And what that means is that that has to be a sequence. And it also means that essentially the same process that allows me to create a templates of language would allow me to say that something like this that has four corners and equal sides templatizes into a square. And this other one that’s a little bit different templatizes into a rectangle. And I should be able to use exactly the same process to find templates for shapes as I do for language. (.) But that’s sorry.
S08: Just that’s fine for that. What I’m where I’m coming from. This is, is this part of the input output stage of the system useful in that the actual internal processing? the, you know, the effectively the inner monologue and that type of thing without, without using language. It can you see a thing there? (..)
S07: Yes. And, and the same, the same issue. And, and this is in the chapter that get these that I’m going to put out next Tuesday. (…) is, is that the, in the same method, you can learn templates of sequences. That is sequences and templates that are reason that say, when I encounter this situation that is going to follow, or if this, then that. And this is the same process in the same structure that I haven’t implemented it yet.
S09: I guess I’m, I guess I’m up next. Um, yeah. So, uh, I had a couple of things in mind. Um, one, I just finished a book called coding artificial minds. Um, if anybody’s interested, they can email me sub thought at hotmail.com. Uh, it’s a preprint, but, uh, it’ll be out probably next week. Uh, and in the book, I talk about, uh, the memory. I have one as part two is developed devoted to the memory. And I have one chapter talking about the container for the memory or the box from the memory. I call the box, the totality. And then I have another chapter talking about the memory unit themselves and what the memory units build. And I call that chapter, the ontology. So it’s building up, um, uh, an internal ontology, which sits inside this box called the, the totality. So I think it would probably be good to define the UKS and have a chapter, uh, dedicated to the, the box, you know, what the space is and what you expect the space to be. (.) Um, uh, and as far as, uh, the number of, uh, elements like humans have, you know, a hundred thousand plus, uh, nodes in their, you know, link linguistic tree. Uh, and those nodes are interconnected. So they had something called word net, uh, probably, uh, now a couple of decades ago, back in the early two thousands, they came up with word net. Um, and you can, you know, enter, you can have, uh, what I call ontological, meron, meron, meron, meron, meron, and, uh, deck. I call it also dexona, dexonical dexola. Uh, you can have a meron, me an ontology and a dexonomy, uh, as interconnections in this, uh, web. Uh, the ontology is the, you know, the is a links, the instance, the instance links, the meron, me is the has a links and the dexonomy is the examples and, uh, counter examples links. And, you know, the dexonomy works in a slightly different way. (..) So you say something is an example of something, something else is a counter example of that thing. And, uh, that has a difference, different inference pattern than, uh, ontology. And for me, I call it, um, in my ontology, I use the word genus as opposed to is, is a, and I use members, uh, instead of, um, instead of, uh, instances, you know, that’s, that’s
S07: yeah, I did this one of very early on before, before we separated language from thoughts. We were using the labels themselves as language generation. And at that point, Fido is a dog made a lot more sense. Uh, but, uh, since we’re not doing that anymore, it’s not so, not so big a deal. Uh, I wanted to add just before we go, uh, let me share my screen again and point out something that, that we haven’t talked about so far. Um, and that is that this entire system is, uh, language independent and we can say, uh, let’s just load up on top of the, on top of the, uh, English that we loaded before. We’ll load up some French. And since I don’t know any French, uh, I asked chat GPT to generate me a French query. (.) And so we’ll just copy and paste that in. (…) And what we can figure out, uh, is that this must be, what is a dog? And it answers because it uses essentially the same, the same, uh, template set to answer that it used for coming in. If you ask it a question in French, you’re going to get an answer in French. And the system internally, uh, uh, has, uh, really no, I mean, these to it, it’s all just symbol processing. So it has no idea what, uh, what any of the words mean except for the ones that are grounded in the mental model. We could, we could make a better case that the mental model allows it to actually have some idea what a, uh, what a dog is because it actually has pictures and attributes and things like that. And with experience, it’ll learn even more. (…) Um, anybody have any closing words? (….)
S04: You’re great, Charles. I love these talks.
S07: Well, well, I brought up a whole bunch of stuff and, uh, and, and, uh, uh, have, I think in my humble opinion, I’ve dramatically improved the demo. So it shows a lot more stuff than it used to show and gives a much better picture of how the overall flow of intelligence has to work where you have to have some grounding to reality. You’ve got to be able to attach that to words and the words have, have to have multiple meanings and you have to have multiple syntaxes and all of that. Uh, Mugar, what do you have?
S02: Oh, two thoughts. One is I wrote a paper recently about compiling neural networks from language, from LLMs. And I think it’d be very interesting if you took an approach like that to basically generate a very big data set inside of the brain simulator. So you start with an LLM, you probe it, you get kind of these n-grams for lack of a better word. You put them in and then have a big, a big brain to interact with. Yeah.
S07: We started down that road two years ago with chat GPT, but it was too early and we ran into, uh, problems with the ambiguity that now we would not have so much of, uh, can you email me a link for the paper?
S02: Yeah, absolutely. And the second one has to do with like labels versus symbols. The way I see it is everything is a symbol, but we have a small slice of that spectrum that we have shared symbols with and we call that language. So for me, the fact that it works with sounds, visuals, everything is, it just makes sense. You’re kind of dealing with feature detectors, but humans have a tiny overlap with this. And that is the bridge, the language bridge. So I think anytime this comes up with, does it support other stuff? Is it this or that? I think that that’s kind of a category error. (….)
S07: Could you give me an example? Cool.
S02: So the label Fido. Uh, okay. Let me think about it for a second. Well, no, not label vision, right? So vision use like CNNs to detect features. We don’t have a word for one of those intermediary features. Uh, but nevertheless, your system works with it because you’ve demonstrated it. And it just so happens that when we see these examples that are made of words that we know immediately what they mean, but that’s not because you’re, you’ve changed the, the way it works. It’s just because our brains decode it. (…)
S07: Yes. And that’s entirely the case that, that, uh, I, uh, you know, I never, I never know what, what might’ve been done to, to, to mess things up with, uh, using labels. But one of my tenants is that you shouldn’t ever use labels for the, to, to mean anything.
S02: I did the exact opposite. I’ll share the paper with you and maybe, maybe we’ll spark.
S03: Can you share the paper in the chat? Is there any chance? Cause I, I, I want to know why you think the opposite, because I also think that there’s no labels to the buckets inside our head. They’re basically just storage spots for food and shelter and safety from the past. Like hoarding birds.
S02: Yes. No, no problem. I will share it right now. //S03: Yeah.// So it has to do with functional use. So I believe that, uh, as long as like AI and people live on the same planet, we share a functional prior. And that is why I think labels are okay to use. They’re not okay in an absolute sense, but they’re okay in a functional pragmatic sense. And that is the distinction, but yeah, just give me one moment. I will share the link. (.) Okay.
S07: Well, and while you’re sharing that, I want to just wrap up and say, thank you so much for coming. I will put out, uh, two more chapters. I’ve got them almost ready to go. One of them needs a, another edit pass to make it considerably shorter than it currently is. And I’ll have that out on Tuesday and we’ll have another meeting next week. And the topics are going to be the, uh, first is how you can use sequences to create internal algorithms. And we know we humans are capable of learning algorithms in that are entirely mental. And my example is how do you know if, which comes first alphabetically cap cap or cat? Well, you know that right off the bat and somehow, uh, you, you know it, but you can only know it if you’ve learned the algorithm. So that’s going to be topic number a and topic number B is going to be greater depth into the mental model as it’s currently implemented and suggestions on ways that I can make it even better with that. I want to thank you so much for coming. Um, I wish you all well, and I’ll see you in a few weeks. Thank you very much. Bye. Bye.
