[00:02] code is fantastic. Co-work is fantastic. But they are grains of silicon doing consciousness. They don't have an inner monologue. You take an LLM and train it monologue. You take an LLM and train it on pre 1916 or 1911 physics and see if
[00:17] relativity. If it does, then we have AGI. >> Just today, by the way, Daario allegedly said that you can't rule out that they're conscious. You can rule out their cost. I think I mean come on to
[00:30] get to what is called AGI. I think there are two things that need to happen. One are two things that need to happen. One is >> Great to be back. This is one of my favorite topics which is um how do LLM
[00:46] >> And I think that uh you in my opinion you've done kind of the best work on >> Thank you. For those that did not see the original um one, maybe it's probably worth doing just a quick background on kind of what led you to this point and
[01:01] that you've been doing. >> 5 years ago when GPD3 was first released, >> uh I got early access to it and I
[01:13] started playing with it and I was trying to solve a problem related to quering a cricket database. >> Yeah. And I got GPD3 to do in context learning, few short learning. And you know it was kind of
[01:26] the first at least to to me it was the first first at least to to me it was the first known uh implementation of rag retrieval augmented generation which I used to solve this problem of uh querying
[01:39] getting GPD3 to translate natural language into something that could be used to query a database that GPD3 had no idea about. I had no access to GPD3's internals, but I was still able to use it to solve that problem. So, it it it
[01:53] worked beautifully. Uh we we deployed uh this uh in production at ESPN in September 21. But >> Wow. Wow. You in 2021. >> No, no, no. In 2020.
[02:08] >> 2020. I got it working and by the time you talked to all the lawyers at ESPN and you know, productionize it, it took it took a while. But October 2020 we had it took a while. But October 2020 we had well I had this architecture working but
[02:22] after I got it to work I was amazed that it worked. I wanted to understand how it worked >> and I looked at papers and all the other sort of deep learning architecture papers and I
[02:36] >> Yeah. >> So then I started getting sort of deep into building a mathematical model. >> Yeah. And now you published a series of the one where you had kind of your matrix kind of abstraction. So maybe
[02:51] >> Yeah. >> work. first one which is >> you were trying to you were trying to describe you're trying to come up with a mathematical model of how LLM works.
[03:04] >> And you had which was very helpful to me which was um and at the time you were >> Yes. Yeah. >> And you came up with an abstraction for large matrix and you use that to describe. So maybe you can kind of walk
[03:18] >> Sure. >> Yeah. So so what you do is you you imagine this huge gigantic matrix where every row of the matrix corresponds to a prompt. And the way these LLMs work is given a
[03:33] And the way these LLMs work is given a prompt they construct a distribution of probabilities of the next token. Next token is next word. So every LLM has a variants have a vocabulary of about 50,000 tokens.
[03:47] >> So given a prompt, it'll come up with a distribution of what the next token sample from that distribution. >> Yeah. So that's the posterior right? That that's how LLM work. And so the idea of this matrix is matrix is for
[04:01] every possible combination of tokens which is a prompt, there's a row. >> And the columns are a distribution over the vocabulary. 50,000 possible tokens, it's a distribution over
[04:14] probability the probability that the next token should be this versus that. >> Uh so that that's sort of the idea and and when you start viewing it that way,
[04:28] and when you start viewing it that way, it makes things at least clearer to people like me who want to model it. uh what what's happening? So concretely let's say you have an example that uh let's say your prompt is just one
[04:43] uh let's say your prompt is just one word protein. Yeah. So if you look at the distribution of the next word next token after that uh most of the uh probabilities would be zero but you'd have non zero non-trivial probabilities
[04:57] have non zero non-trivial probabilities on let's say two words one is synthesis the other is shake >> right and now the LLM is going to sample synthesis sample uh this next token and man pick synthesis or shake
[05:12] man pick synthesis or shake >> or you as a human will give the prompt protein shake >> or protein synthesis. Now, depending on whether you pick synthesis or shake, the next that row looks very different,
[05:25] right? If you pick protein synthesis, the terms that would have a high probability would be all concerned with biology, right? But if you pick protein shake, it'll all be about gyms and exercise and all, you know, bodybuilding
[05:39] stuff. So, that synthesis or shake completely changes what comes next. >> Yeah. So this is an example of uh you can say bijian updating. You start with protein you have a prior that after protein this is going to
[05:54] happen. As soon as you get new evidence then the next term is synthesis or shake then the next term is synthesis or shake you completely update the distribution. So now you can imagine that the whole the the entirety of LLM is this giant
[06:10] matrix where you have every row protein protein shake protein synthesis the cat sat on the >> you know Humpty Dumpty blah blah blah >> you know Humpty Dumpty blah blah blah >> now given uh the vocabulary of uh these
[06:24] LLM let's say 50,000 and the context window so GPD for instance chat GPD the first version had a context window of 8,000 tokens. Yeah, if you look at all
[06:36] possible combinations of 8,000 tokens and 50,000 uh vocabulary, the number of in this matrix is more than the number of electrons across all galaxies. Right?
[06:50] >> So, so there's no way that these LLMs can represent it exactly now. Fortunately, this matrix is very sparse. Why? Because you know an arbitrary gibberish. We're not never going to use that in natural in real life.
[07:05] >> Yeah. >> Also, the columns are also mainly zero. >> Right. If you have protein, then you won't have lots of, you know, you won't have arbitrary numbers or arbitrary words after that. It's very sparse both
[07:19] words after that. It's very sparse both in rows and in columns. So I in kind of an abstract way what all these LLMs are doing is coming coming up with a compressed representation
[07:32] of this matrix and when you give a prompt they try to approximate what the true distribution should have been and try to generate it that that's what uh in my mind at least it boils up to >> just from my understanding. So if you
[07:48] >> just from my understanding. So if you have a row of uh protein >> Mhm. >> is protein shake a subset of protein or >> It's different. It's a continuation from >> I see.
[08:03] >> Right. No, but I'm just saying like the actual the actual posterior distribution >> You you can say it's a subset, right? Uh if you have protein then protein shake continuations from protein. So both synthesis and shake have non-zero
[08:18] synthesis and shake have non-zero probabilities. So you can yeah >> right >> you you know you use this approach to describe how in context learning works and so maybe first describe what in
[08:32] context learning is and then kind of the conclusion that you came from that. So conclusion that you came from that. So eight context learning is when you eight context learning is when you uh show the LLM something it has kind of
[08:45] never seen before. You give it a few examples of this is what it wants uh this is what you're trying to do. Then you give a new problem which is related to the examples that you have shown >> and the LLM learns in real time what
[08:58] it's supposed to do and solves that problem. And by the way, the first time I saw this, it absolutely blew my mind. And I actually I actually use your DSL >> when I was like first learning about it. So maybe like kind of like
[09:13] the DSL thing is just just crazy this works at all. >> It's absolutely, you know, mind-blowing that it works. And so going back to that cricket problem was you know in the mid '90s uh I was part of a group
[09:27] that had created this uh cricket portal called cricket info. >> Yeah. >> Uh cricket uh is a very start sport. You know you think baseball multiply by a thousand. at all kinds of stats and we
[09:41] had created this uh online searchable database called stats guru where you could search for anything any stat related to cricket and has been available since 2000 >> but because you can query for anything
[09:56] do you make something like that available to the general public well they're not going to write SQL queries >> the next best thing at that time was to create a web form unfortunately ally everything was crammed into that web
[10:11] form. So as a result you had like 20 drop downs, 15 checkboxes, 18 different text fields. It looked like a very complicated, daunting interface. So as a result, even though it could solve or it could answer any query, almost no one
[10:26] used it. A vanishingly small percentage of cricket fans use it because it it just looked intimidating. And then ESPN bought that site uh in 2007. I still know people who uh run the site and I always told them you know why
[10:41] don't you do something about stats guru and in January 2020 uh the editor-inchief of cricket info Sambbal he's he's a friend so he came to New York and we had gone out for rings and again I told him you know why don't you
[10:54] looks at me and says why don't you do something about stats guru he was joking but uh that idea kind of stayed with me and when GP3 was released I thought maybe I could use stats guru use GP3 to create a front end for stats
[11:09] guru. >> And so what I did was uh I designed a DSL a domain specific language which uh converted queries about cricket stats in >> No. >> And to be clear you created this it
[11:24] wasn't like part of like any training that was online that like could have >> Nothing GPD could have seen. I created it. I thought okay this makes sense. So I designed that DSL and then I did that few short learning things. So I would so
[11:38] I created about a database of about I would say 1500 natural language queries and the DSL corresponding to that query. So when a new query came in, somebody's
[11:50] asking a stats question in English. What I would do is I would go through the natural language queries, do a semantic search, pick the most closely matching
[12:03] >> Uh and then use that natural language query and its DSL and send that as a prefix. Now GPD3, if you recall, had a context window of only 2,000 tokens. >> Yeah. So you had to be very judicious about which examples that you picked.
[12:18] But you pick that and then you send the new query and GP3 would complete it in the DSL that I had designed which until milliseconds ago it had never seen. >> And I had no access to internals of GPD3. I had no access to the weights.
[12:32] >> But still it worked. So that that's how so >> so so it's not obvious to me given your matrix example of like a prompt and then a distribution how something like in context learning
[12:45] >> would work and so like I think your first paper >> right >> um and so maybe you could walk through >> um and so maybe you could walk through your understanding of how LLMs
[12:58] do in context learning. >> Yeah. So, so when you think about what >> Yeah. So, so when you think about what in context learning is is that as you see evidence. So, so you know in the first paper what I also did was I I
[13:11] >> Yeah. >> And I uh I depicted the next token probabilities >> mhm >> of the model as it was shown more and more examples. So the first time you
[13:26] show it this DSL the natural language and the DSL the probabilities of the DSL tokens were were extremely low because GP3 had never seen this thing. When it
[13:38] saw the cricket question in its mind it was trying to continue it with an English answer. So the probabilities that were high were all English words. >> Yeah. Once it saw my prompt where I had
[13:55] the question and the DSL, the next time I had the question in the next row, the probabilities of the DSL token started going up with every example, it went up and finally when I gave the new query, it
[14:09] was like it had almost 100% probability of getting the right token. >> Yeah. >> So this is an example of in real time the model was updating its posterior probability. It was upgrading its
[14:22] knowledge that okay I've seen evidence this is what I'm supposed to do. Now Beijian >> inference is. Beijian updating basically is you start with a prior when you see a new evidence you update your posterior.
[14:38] That's the mathematical division. But but in in English it's basically you see something you see new evidence you update your belief about what's >> Yeah. >> Right. So it was clear to me that LLMs
[14:51] are doing something which resembles Beijian updating. So in that first paper I had this matrix formulation and I showed that you know what it's doing. It looks like Beijian updating. >> Yeah.
[15:04] series of papers. >> That's right. So okay so I mean it it it seemed pretty conclusive to me at that time and then you went quiet for a while and then I still remember the WhatsApp text. You said Martin I know exactly how
[15:17] >> Yeah. Well >> and then and then and then listen you broke the internet. like you went super viral on Twitter like I mean people viral on Twitter like I mean people really noticed. Um uh and so I I want to
[15:29] get to that in just a second. But before that, um >> I remember when your first paper came out, people would be like, >> you know, these things are definitely not Beijian. Like, you know, you know,
[15:43] Beijian, but they're not. Like, why do you think that there was this reaction >> you know, there's something new, they're not Beijian? I mean, I felt like there's they're being characterized as >> Yeah. Yeah. I I think this whole world
[15:59] >> Yeah. Yeah. I I think this whole world of uh uh probability and machine Beijian and frequentists. >> Yes. >> And I don't want to get in the middle of that sort of political battle, but
[16:12] Beijian has become like almost like people had a reaction to that. It's it's >> I see. >> So, it's like the old Beijian frequentist type battle. Yeah. So, so people just had oh no you can say
[16:25] anything is Beijian right? So I said okay maybe they have a point maybe what we are saying is not really Beijian how do we prove that it's Beijian >> right >> so then first I have to thank you and
[16:38] >> so then first I have to thank you and and Harovitz for this uh >> you know when I when I when I said that I in my first paper I showed these probabilities >> uh it was because open AI had in its
[16:51] chat uh interface uh this option to displays those uh this option to displays those probabilities then they stopped what's happening for some reason they stopped openai
[17:06] I'm not going to get into >> the open and close but but they stopped so then we developed our own interface which could let you look not only at uh the probabilities but also the entropy of the next token was this on top of an
[17:19] >> yeah yeah so so you can load any sort of open source model but you know being an We didn't have access to compute. Thanks to to >> your generous uh uh donation, we got uh
[17:32] >> to run uh over what it's called token >> Is it still running? >> It's still running. It's still running and people come to it. Uh I use it in my classes uh to get students to do
[17:46] assignments. They write their own DSLs and you know they say that that it LLMs work. So I literally my understanding of LMS came from token pro just you know sit there and just look at the the distribution as as you filled
[17:59] enlightening. So for those of you that are listening um >> Token probe >> token probe.cs.colia.edu. >> Yeah check it out. It's actually very very useful way to I can actually see
[18:14] how the probability distribution gets updated as as you fill out a prompt. >> But then I cheated. Oh, >> I you know it was running >> but I also had access to the GPUs that
[18:26] >> Mhm. >> And then along with colleagues at >> And then along with colleagues at Colombia and one of them now is uh is at Colombia and one of them now is uh is at deep mind we started to sort of think
[18:39] about how do you really prove that it's Beijing >> Can you just explain it? Actually I I actually don't know the answer to this. >> Yeah. It seemed to me you proved it in the first paper like what was missing.
[18:53] It was empirical >> and you could see I see I see you could obvious to me that >> yeah it was even obvious to me but to convince uh you you could say you know people who dismiss oh anything can be
[19:08] >> we had to show it precisely >> Got it. Got it. So then we came up with this idea you know my colleagues at Namanagaral and Sedhad Dalal we the series of papers were were written with
[19:21] them. We came up with this idea of a Beijian wind tunnel. Okay so what's a wind tunnel? Well wind tunnel in the aerospace industry is where you test an aircraft in an isolated environment. you don't fly it and you test test it
[19:35] against all sorts of uh uh you know aerodynamic pressure then you see what what it'll withstand what kind of altitude pressure blah blah blah right you don't want to do it up in the air testing
[19:48] >> so we said okay why don't we create an environment where we take these architectures and we tested transformers mamba LSTMs MLPS all architectures we
[20:01] say why don't we create take a blank architecture. architecture. Give it a task where it's impossible for the architecture to memorize what the solution to that task should be.
[20:15] The space is combinatorily impossible for given the number of models. So it's difficult enough that they cannot memorize it >> but it's tractable enough that we know
[20:29] >> but it's tractable enough that we know precisely what the the Beijian posterior should be. You can calculate it analytically. So we gave these models a bunch of tasks where again we show that it's impossible
[20:42] to memorize. We trained these models and we found that the transformer got the we found that the transformer got the precise Beijian posterior down to 10 ^ precise Beijian posterior down to 10 ^ minus 3 bits accuracy. It was matching
[20:54] the distribution perfectly. So it is actually doing Beijian in the mathematical sense given a task >> where it has to update its belief. Uh Mamba also does it reasonably well. LSTMs can do one of the things. So the
[21:09] in the papers we have a taxonomy of Beijing task. Transformer does everything. Mamba does most of it. LSTMs do only partially and MLPs fail completely. >> So is this a reflection of the data that
[21:24] it's trained on or is it more a reflection of the mechanism? architecture. The data decides what tasks it learns. >> Right? So in the first paper we had these beijian wind tunnels and we show
[21:39] that you know it's doing the job where different tasks in the second paper we show why it does it. So we look at the transformers we look at the gradients and we show how the gradients actually shape this geometry
[21:54] which enables this basin updating to happen. Then in the third paper what we did we take we took these frontier production LLMs which have open weights so that we could look inside them and we did our
[22:09] testing and we saw that the geometries that we saw in the small models persisted in models which are you know hundreds of millions of parameters the same signature existed. The only thing is that uh because they are trained on
[22:25] all sorts of data, it's a little bit dirty or messy. >> But you can see the same structure. So the the whole idea behind the Beijian wind tunnel was unlike these production LLMs where you don't know what they have
[22:38] >> So you cannot mathematically compute the posterior. >> I mean it looks based in you know from >> From the first it looks Beijian, but you know. So the wind tunnel sort of solved
[22:50] that problem for us. We said okay let's start with a blank architecture. Give it a task where we know what the answer is. It cannot memorize it. Let's see what it >> so do you think this provides any sort of like indication of how humans think
[23:05] totally independent? >> No no it it does provide right. So you know human beings also uh update our beliefs as we see new uh update our beliefs as we see new evidence. Right. So we do in some sort
[23:19] of in some sense uh Beijian updating but we do something more than that I'll come we do something more than that I'll come to that but uh these transformers uh or even mamba do this beijian updating >> and uh but but but the difference with
[23:36] humans is you know we we'll update our posterior when we see some new evidence but the way our brains have evolved evolved over
[23:48] hundreds of millions of years is our optimization objective has been don't optimization objective has been don't die and reproduce. Right? That's been sort of the driving force and our brains have learned to adjust and so when we
[24:01] see some danger there's some something rustling in that bush. Don't go near. We know how to react to that danger. We know how to uh save ourselves. We internalize that
[24:16] learning and our brain cells or our synapses remain plastic throughout our synapses remain plastic throughout our lifetime. What happens with LLM is once the training is done those weights are frozen.
[24:29] when you're doing an inference for instance in context learning or anything during that conversation okay you're doing bijian inference but then you forget the next time a new conversation starts
[24:42] with zero context you don't retain any learning that happened in the previous cricket DSL that I was doing every invocation of it was fresh
[24:54] it did not remember the last time I sent a query what the DSL looked So that's one difference between uh how humans uh uh use sort of beijan
[25:08] updating which is we remain plastic all our lives our lives >> whereas uh LMS are frozen and there's another uh sort of difference which uh if you want me to
[25:21] >> tell me yeah yeah yeah >> so so the other difference is uh u well first you know our objective ive is don't die reproduce. LLM's objective is
[25:33] predict the next token as accurately as possible. Right? So all these uh scary stories that you you read about that oh the LLM tried to deceive and it tried to prevent itself from being shut down. That's not a function of the
[25:48] architecture. That's a function of the training data. >> It has been fed you know articles on Reddit or SMO or whatever. I mean, just Reddit or SMO or whatever. I mean, just today, by the way, Daario
[26:02] >> allegedly said that uh you can't rule out that they're conscious. >> You can rule out their I mean, come on. And I said, you know, Antropic makes great products. Cloud code is fantastic. Coco work is fantastic,
[26:17] but they are grains of silicon doing matrix multiplication. don't have an inner monologue. They don't uh they're not driven by the same objective function. Don't die, reproduce, right? They're driven by
[26:32] don't make a mistake on the next token. And that's driven entirely by the training data, right? You train the LLM with stories of ASMO or Reddit where you know to survive it's going to do this or that. It'll
[26:46] reproduce that. So it's it it's a reflection. It's not a mind. And and the results, just to say it for the 10th time, are perfectly vision. >> To the to the to the digit. >> To the digit. Yeah. I mean, I I trained
[27:01] it for 150,000 steps >> and uh the accuracy was 10 ^ minus 3 >> I could have trained it for you know this happened in half an hour >> on the infrastructure that you provided for token pro in the background. I could
[27:15] use those APUs to train. But uh so thank you again for that. But so no human >> Beijian, >> but we do something else. You know when
[27:27] >> Dodge it or >> do it? Yeah. >> Why will you dodge it? >> To avoid being hit. >> Avoid being hit. But your head is not doing a Beijian calculation of okay,
[27:41] this pen is coming. The probability that it hits me, it'll cause this much pain >> Correct. What you're essentially doing in your head is you're doing a simulation. >> You see the uh the the the pen coming
[27:55] and you know that it'll come and hit me. Your mind simulates and you dodge it. Right? So all of deep learning is uh doing all of deep learning is uh doing correlations.
[28:11] >> Yeah. Causal models are the ones that are able to do simulations and interventions. So you know Judea has this whole uh causal hierarchy >> where the first hierarchy and the first hierarchy is association which is you
[28:26] build these correlation models. Deep learning is beautiful. It it's extremely powerful. I mean you see every day all these models are like amazingly good. >> They do association. The second is intervention
[28:39] in the hierarchy. Yeah, >> deep learning models do not do that. Third is counterfactual. So both intervention and counterfactual simulation.
[28:52] You you build a model of causal model of what's happening and then you are able what's happening and then you are able to simulate. So our brains do that. The current architectures don't do that. Another example I think which will make
[29:04] Another example I think which will make it clear is uh the difference between entropy >> and kmogrove complexity. >> Sure. >> So if you look at the Shannon entropy of
[29:17] >> it's infinite. >> Sure. >> It's impossible to predict and learn what digit will come after. Yeah. So that's the definition of Shannon entropy and Shannon entropy sort of tries to
[29:31] build a correlation. It tries to learn the correlation. Deep learning does the Shannon entropy. >> Gulmagraph complexity on the other hand is the is the length of the shortest program.
[29:43] >> Which will reproduce uh the string that you that is under >> Yeah. >> Now the program to get the digits of pi are very small. >> Yeah. Thanks to Raman Jim and others you
[29:56] know there all sorts of really small program that can reproduce it exactly. So the colograph complexity of pi is very small. Shannon entropy is infinite. >> I think deep learning is still in the Shannon entropy world. It has not
[30:12] crossed over to the colog complexity and the causal world. >> Wow interesting. >> Right. So uh do you to what extent do you think this provides us research directions to kind of improve the state
[30:26] specific example you talked about >> human beings don't actually update you know the matrix they don't kind of update their weights >> but right now there's a lot of research on continual learning you know so
[30:42] >> does your work provide some guidance of how you might approach those problems and and in particular I've always had this question which is we use so much >> Yeah. >> To create these models like is it even
[30:55] the weights and actually have a meaningful impact you know with in in just need so much more data in order to do that. So can you start answering >> You you can start answering some of these questions and and one of the
[31:10] misconceptions that exists today is that scale will solve everything. Scale will different kind of architecture and this continual learning is a difficult problem. You have to balance the fact that you will learn something new
[31:24] against the risk of catastrophic forgetting. >> If you update the weights and you forget what what was important and what you have already learned then then you are you know you're not making progress.
[31:37] Then it'll just be some sort of random chaotic model. So to solve that problem is difficult. That's one aspect of it. So, so, so you know to get to what is called AGI, I think there are two things that need to happen. One is this
[31:52] plasticity which has to be implemented through >> Secondly, we have to move from correlation to causation. >> Yeah, >> that's uh uh I how much is this similar
[32:06] to what Yan Lun talks about with the >> so Yan Lun causality planning? action would >> it is it is related you know he he's the jp model right >> but it is related the the other thing is
[32:23] uh you know the first time I came on this podcast I I mentioned this test of >> I don't remember >> so I said you know uh uh you take an LLM >> so I said you know uh uh you take an LLM >> and train it on pre 1916 or 1911 physics
[32:39] >> and train it on pre 1916 or 1911 physics >> and see if it can come up with the >> Yeah, >> if it does then we have AGI. I mean it's a high bar but you know we should have high bars. It won't. And this is the
[32:53] same test that I think Demis uh mentioned at uh the India AI summit couple of weeks ago. It's created a lot of news. But why why is that and how is that related to this idea of Shannon versus Kro?
[33:08] versus Kro? So at the time of Einstein there were a lot of clues that Newtonian mechanics there was >> Yeah. >> Right. Uh people knew that Mercury's
[33:21] orbit didn't make sense. There was something off about it. Then there were these experiments done uh the Michaelelsson Mley experiments where Michaelelsson Mley experiments where they were trying to figure out uh
[33:34] they were trying to figure out uh uh this uh medium called uh the ether through which light travels. And they felt that if you know you And they felt that if you know you bounce light in different directions
[33:47] uh the speed might change and they they could detect a change in the speed of light. They tried several experiments. They had really precise instruments which could measure the speed and they found
[34:00] nothing. They found that that speed of light did not change at all. >> Yeah. >> Then gravitational lensing. So there were a lot of these signs that Newtonian mechanics
[34:15] >> Yeah. But until Einstein came up with a new representation of the space-time container, >> So if you had a model that just looked at correlations and so uh sees all of
[34:31] this, you know, all of these uh pieces of individual evidence and put together, it would not have come up with the beautiful equation
[34:43] that Einstein came up with. you know uh I'm forgetting exactly what it is g muv= 8 pi t muv some something like that >> where you know uh the the the equation of uh the rel the space-time continum that the tensor
[34:58] that the tensor >> so he came up with a new formulation >> so he kind of rejected the existing axioms he came up with a very short >> interesting >> the world
[35:12] >> one equation from that equation everything else follows >> right whether you're talking about gravitational waves or black holes or gravitational waves or black holes or mercury or how GPS works you know GPS
[35:24] the GPS that we use every day in our phones it uses the equation of relativity so do does this end up becoming like um you you you almost have to ignore the
[35:38] majority of previous data in order to do it which LLM can't because they're data. It's like you almost have like this kind of data gravity that's pulling you back. It's like it's like everybody said it's X.
[35:52] it's Y, but because everybody said it's X, like the LM will always say it's X. >> It'll always say it'll treat that Y as an anomaly. Actually this is actually a >> it's like >> I so now okay now I get your Shannon
[36:07] entropy versus like one of them is like >> the total amount of information there amount of information there which is what happens right now. what happens right now. >> Yeah. where you can actually describe
[36:22] another another motion. You can describe everything with a shorter description totally different motion, which would be like Yeah. You know, another way that I've always thought about these, I thought
[36:36] you articulated it well in the last time we talked about it, which is the universe is this very, very complex space and then, you know, somehow humans >> Mhm. >> That's less complex.
[36:50] >> And then that gets kind of written down and then the LLM. So that's kind of some still a very large space, but it's it's a bounded space. And the LM learn that a bounded space. And the LM learn that manifold and then they kind of use,
[37:05] and down that manifold, but they're kind of bound to that manifold. >> And then again, I don't want to put words in your mouth. And then, but like what they can't do is is generate a new manifold, right? Which requires
[37:17] works and then coming up with a new representation of the universe. >> Yeah. Exactly. >> Einstein had to create a new manifold. >> Yeah. If you just stuck with the old manifold of the Newtonian physics,
[37:30] but you could not come up with a manifold that explained them. So you need to come up with a new representation. So to me you know there are lots of definitions of AGI uh you know Turing test we have already passed
[37:43] that you know performing economically useful work every day you see you know >> Do we I don't know. No, I mean they are >> I mean I mean without human >> No no no. So that that's different but still you know it's like a car can run
[38:00] faster than humans, right? >> I mean that's a that's the that's a >> Yeah. So all these definitions do useful >> you know maybe you know in 6 months >> you know maybe you know in 6 months you'll have cloud or what a gemini do
[38:14] without intervention cing tasks which are well defined well scoped >> that's possible but to me AGI will happen when these two problems get solved >> elasticity continual learning properly
[38:28] and building a causal model from you know uh in a more data efficient manner Yeah, we we are hearing people now talking about you know seeing generality like Donald Kuth for example in the last few
[38:44] days right you know had this you know this you know aha moment apparently that think that that suggests that we're seeing generality or >> No no no so so that actually to me it validates what I've been
[38:59] talking about for a while now how >> so so if if you read what he did uh with the help of uh you know a colleague he got the LLMs to solve this particular
[39:11] got the LLMs to solve this particular problem of finding Hamiltonian cycles odd numbers we won't get into that and he got the LLMs to keep solving for one odd number after the other right >> what he also got to do is after it found
[39:26] a solution for a particular value of m he made the LLM update its memory with exactly what it learned in solving that problem. So the LLM's tried many
[39:38] >> You know, something worked, update the the memory. So that's kind of like >> Yeah. >> Right. It's learning what it has done as we went along. Again, it's it's a hacked version of it. You're not changing the
[39:53] the context. >> Right. >> Right. But you as you learned and even after that so this whole space of Hamiltonian cycles and the associated math is well represented in the
[40:09] trained on >> right you just had to find the right connection and LLMs I know compute you throw enough compute they will find the throw enough compute they will find the right connection so can was able to find
[40:24] the LLM's attempts And eventually it needed him to put together what he saw into a solution. It definitely helped him get
[40:37] to the solution but he had to create the new sort of manifold >> to come to the solution. The LLMs were after a while stuck right he you read what he has written. I mean it just hot up the press I think two days ago.
[40:52] >> Two days ago days ago but uh eventually he used the solutions and he came up >> Yeah. >> Right. So it's like you know it's like >> Right. So it's like you know it's like Einstein saw all these evidences
[41:06] then he thought what will explain he came up with a causal model. >> Yeah. So canut and his brain is sort of the >> that's in the chimograph is the human right and the llms are extremely
[41:20] efficient at doing the shannon part of it. It found all the solutions by trying you know various things and learning more and more wondering like do you think this again I'm going to ask the same question again
[41:32] which is do you think this provides some sort of insight on like the next problem >> like like is there a mechanism that will get the kagarov complexity get the kagarov complexity >> or not like is this it tells us which
[41:47] >> but clearly not how to do it like >> not how to do but even colograph complexity has largely remained a sort of a theoretical construct >> yeah for sure there's no algorithm There's no there haven't been practical
[42:01] >> the shortest program. >> We know it exists. You know, you can argue about it. It but so so that's where I think >> it's my bias. That's where our energy >> should be focused not larger models with
[42:16] >> Can you and can you can you tie the two things like how does that pair with doing simulation or is that simulation totally orthogonal? >> No, simulation is is related, right? So you think it like basically you do
[42:30] you think it like basically you do simulation and somehow that is a step towards doing the kagra complexity. >> It it's it's the simulator is the is the the perfect program. >> Oh I see.
[42:44] simulator that when I'm throwing the pen you know that it's coming at you right and you duck. So, so you're not computing the probabilities as it goes, very physical thing versus we were talking more conceptually.
[43:00] >> It's the same mechanism really. >> Right. >> I see. For most things, right? term. >> Yeah. you know
[43:15] add infinitum but here it it's making a difference in the way we view intelligence >> how how how has the last three papers been received >> no I don't know there well I mean I mean
[43:29] the archive >> versions will let me tell you it I mean >> um lot of great reception a lot of people read it I'm just wondering like >> I'm getting good feedback but I'm an outsider in this field right that's
[43:42] >> I'm a networking guy. Why is he writing about you know learning and machine learning and deep learning and basian so but but people who have actually taken the time to read those papers I'm getting really good feedback uh there
[43:56] was a recent paper by Google research >> which tried to teach uh LLM by some sort of RLF to do Beijian learning properly. I think people are coming around to the view that okay LLMs are doing Beijian
[44:12] learning. I know that some people also looked at the Beijian vent tunnel paper the archive version and they reproduced the experiments. they they did the trading and they saw yeah this is actually happening. So
[44:26] >> that's great. So what's next? uh what's next is uh you know these two uh what's next is uh you know these two parallel uh tracks I hope to make parallel uh tracks I hope to make progress there plasticity and causal
[44:41] mechanism >> and you've created a formal model how it works >> and so now you're actually interested in >> and do you think it's an entirely different architecture
[44:55] like part of the solution >> I think LLMs are definitely part of the >> But but there has to be something more >> and other. >> So you know I was not interested in sort of cataloging what all these
[45:07] >> LLMs can do. >> Was more interested in why are they and how are they doing it. >> I think now we have a good grip on >> the why and how. >> And the next step is to you know
[45:20] >> move them to the next level. We we now I I think we have a fairly good >> Yeah. Now how do you uh go to the next step? >> Is there an is there an equivalent kind of theoretical
[45:34] framework for causality that applies here like similar to like Beijian for inference? >> Well the Japal's whole causal hierarchy >> That's that's a very good one. You know the whole do calculus uh approach
[45:50] I think it's a good way to think about it. you know the the sort of association >> Yeah. >> It takes you from correlation to >> Yeah. >> In a mathematical way.
[46:05] appreciate you coming. This is awesome. So, we had you here for the first paper where you had the empirical results. >> Mhm. And then we had you back when you actually have like the formal proof and hopefully the next time you come back
[46:17] you will have a proposal for the mechanism that uh >> Hopefully. >> All right. Cool. Thank you for coming in. >> Thank you for having me.