929: Dragon Hatchling: The Missing Link Between Transformers and the Brain, with Adrian Kosowski

29:56Nothing much has changed. If you look at all the family of LAMA models, LAMA 3, LAMA 4, all of them, this is like GPT2 almost unchanged. Yeah, just scaled up. Just scaled up. So here, the scaling of the transformer is actually a bit of an interesting story because there's no one single recipe for scaling the transformer. It ends up being done differently. You have a different number of attention heads. You have a different number of layers, et cetera, et cetera, et cetera. So the transformer scales in different ways and it's a little bit of know-how how to scale it. Or actually it's from the point of view of a theoretician or somebody who would want to apply something like computational complexity and ask what's the limit of a transformer.

30:38It's actually rather hard to decide how we should scale it up. But there's one dimension which appears to be fixed, to have converged, and that's the size of the attention head vector dimension in a transformer. So this does not scale even as the models get larger. It has stopped scaling. And here the intuition is that basically all concepts that the transformer works with have to be mapped into a vector space of about 1 ,000 dimensions. I see.

31:20Jon Krohn:I see. So we have this constraint that's appeared in the transformer that will limit its ability to have nuanced reasoning because it doesn't seem like we can scale the dimensionality of the vector space of the tension heads beyond a thousand. So this is actually a point that we address a bit more mathematically in our paper because we take the opportunity to analyze both, you know, baby dag and hatchling and also kind of take the reverse approach to history and look at transformers and approximation of baby dag and hatchling because it's actually easier to take this direction to say that we're coming up with a simpler and somewhat cleaner mathematically architecture cleaner in terms of the number of moving blocks.

32:06So here the advantage is we can kind of look back at what the Transformers are doing from this perspective and somehow see the way it organizes concepts and maps them into a vector space. The kind of question which comes up, and this is, I think, for a large part of the audience who are familiar with vectors, with vector spaces, with notions even quite far from the transformers, such as, let's say, preference vectors, you can manipulate them as if it was a linear space. So you can add vectors together, you can subtract vectors, you can have the opposite elements, you can have a negative vector.

32:52And this is the essence of a vector space. By contrast, if you go into the sparse positive spaces, the way you actually compose concepts is somewhat different. You don't work so much with linear combinations of vectors. You work more with bags of concepts, so bags of words, combinations. It works a bit like a tag cloud. It looks a bit like an association set, so a number of elements put together which form a whole. It's a bit how you form sentences by putting together words to get the meaning correctly or in some languages which have a bit more of a tendency to play, especially Germanic languages, German in particular, where you compose multiple words to create a new noun, for example.

33:53This is a place where you have this kind of compositional effect. And there are a number of differences. One difference to this type of representation, apart from the points that you discussed, John, that efficiency and so on, is actually that you don't look so much towards negatives or opposites. Starting with a simple example, if you show somebody a piece of work that's been done badly, and you tell them, look, this is what you're supposed to do, but do the opposite to get a good effect, there's no such thing as take the opposite and find you know the the opposite to it likewise we don't like in reasoning patterns there's no symmetry between being attracted towards a certain reasoning pattern and being repelled from it so if i tell you now don't don't think about the color blue don't think about the color blue you'll be consciously thinking about the color blue just to just to try to compensate for it and avoid it.

34:59But it's not the mechanism of switching off, of damping down. So we are in a different vector space, in a different space.

35:09Jon Krohn:Right. And so what you're saying there is that, so transformer architecture, it doesn't behave in the same way as a brain with these kinds of negative activations, but your BDH does, your new architecture. is that so there's a there's a term that you mentioned earlier uh and we kind of glossed over it i wouldn't mind trying to dig into this a bit more you described the bdh architecture as positive sparse and so now and now we've been talking about negatives can you explain a bit more about this positivity thing the question is actually a profound one um because the uh the correspondence between the two worlds, the world of dense vector spaces versus the world of sparse positivity.

35:54It's like two worlds, which are sometimes complementary. Sometimes one is seen as a view of the other, because it's possible to go mathematically between the two worlds. For the audience who is familiar with, again, with like hands-on machine learning, the vector world is the world where the L2 norm rules. That's the kind of king or queen of norms, the L2. If you go into the sparse positive world, you suddenly go into the world of probabilities. The world of probabilities, the world of chance, because the concepts that you're working with start to have interpretations of likelihoods, or at least some value between 0 and 1, which reflects how much you are drawn to a given concept.

36:50So in L2, you don't have it. In L1, in the probabilities, you have like an L1 norm interpretation.

36:57Jon Krohn:Okay, and then just to quickly say this, so in either of these cases, whether we're talking about L2 norm or L1 norm, these are ideas that have been around for decades in machine learning. And in either case, they're an additional factor that we add into our models that allows our models to generalize better to data that they haven't seen before. This is the idea that you introduce a concept that hasn't been seen before. And the question is, where do you place it compared to other concepts? So we can walk through this. I will be using as props a number of pastries. The first two... Apologies to our audio-only listeners.

37:39Jon Krohn:This segment, Adrian brought out a set of delicious treats, which smell fantastic, and it's very hard not to eat his props. And we have three different plates or serving dishes of desserts or of food. And they're critical to this explanation, I suppose. Well, they are at least they set the scene. And I will explain as best I can to our listeners. So the debate, what's better, like L1 or L2 from the point of view of working with basically any kind of dynamical system representation. This is a debate which is prevalent in many fields. And so what would better look like? Like if one is better than the other, what is the outcome?

38:36So, there are differences, and I'll explain the differences. It's hard to say which one is better, and that's why I'm bringing in cakes, because everybody has their own preference for cakes. And here, it's kind of, you know, there will be applications, there will be situations in which every cake is suitable. The first two cakes in the selection are actually not due to me. They're due to one of the more renowned figures in quantum information, quantum computing, Scott Aronson, who had this idea that the L2 norm world compares to a round potato cake like Latke. And the L1 norm compares to this triangular cake, the Ham and Daschen.

39:26and the explanation a little bit is to understand you know how you can of course it's it's it's trivial taking a high dimensional space and reducing it into into what we can see what we can feel but you can try to do it so in a in a vector space you'll have you'll have this kind of round thing it's like like a ball in which you have vectors and you you move around inside this kind of circle, you have positive, you have negative. In fact, if you move to the L1 spaces, then at this point you have corners. You start to have sharp corners. I see. And these are the kind of concepts that you are connecting.

40:07So the kind of well-defined entities are in the corners of a triangle, for example, if you take the lowest example possible, the smallest example possible. and the combinations of these are in the center. So you have this effect of combining corners to mix them in. I see.

40:30Jon Krohn:So the potato latkes are representative of the L2 norm. Yes. And so it's not a coincidence then that potato latke is relatively homogeneous, that it's all kind of the same kind of substance throughout. whereas with this other kind of treat which I must say is new to me latke is very familiar with I think they're delicious these are ham and tushen? yes nice and then that's interesting so the tushen part at least that's pocket in German yeah so I think both of them obviously come from the diaspora cuisine but you can get both in central Europe and then of course in New York and New York obviously I mean it's like we can get everything in New York but anyway I'm bringing up these two because as long as you stay in discussion between these two, it's a bit of a debate, as some of you may have seen, like debates between the two kinds of cakes, academic debates at most.

41:26Academic debates. Academic debates. So you have to reduce your problem to the debate between the two. But the kind of thinking here is that, you know, if you are in this L1 norm world. A commentation world with the pointy corners. Yeah, that's kind of like, okay, it's fine, but it only starts to get interesting when you go to higher dimension because sparsity needs higher dimension because you want to be choosing a few concepts from a very, very large group of concepts. So this is, and at this point, we would need like high dimensional pastry to be able to. to have to send this.

42:11Jon Krohn:They need a baker that can bake in more than three dimensions. That's it. That's it. So technically, yeah, technically here we're talking about three concepts, right? And we have a triangle. The three corners on our triangle. The best I could do was go one dimension higher and here we are escaping Central European cuisine to the best of French patisserie. Still available in New York. Of course. Of course. So here we have a structure which is a triangle but one dimension higher at least okay so yeah so in this case here for people who aren't watching the video version we have a stack it's kind of like you know people order those seafood platters with like lots of layers of seafood where the the bottom layer is the biggest and maybe there's like some lobster or some crab on there and then you have a medium tier that's a little bit smaller and you got the shrimps on there and then a top tier with scallops that's the smallest or something.

43:08It is a kind of representation of a pyramid. At least that was the objective to have like a pyramid. And the interesting thing is that when you look at this kind of structure, first of all, the first thing you notice is that you don't eat the center. In the center, you just have a plate. It's the things that are kind of outside on the walls that are interesting. So you have combinations of smaller numbers of corners, which give you the interesting elements. In fact, although we went from what is essentially a structure in two dimensions to a structure in three dimensions, we are not combining four concepts, but we're still combining usually three concepts to get the desired outcomes.

43:54Also, one thing about the specific type of construction...

44:00Jon Krohn:I don't think we even mentioned here is that this is... So Adrian didn't bring a seafood tower into the studio, which might not have been the most considerate thing to do. That would have been pretty interesting. But it's a tower of macarons. It is a tower of macarons. And these macarons are beautiful round objects. So you can think of each macaroon as having a certain radius, because indeed it does have a certain radius. And you can have concepts which somehow have a certain radius of its action. Right. Again, for those of you who are more into into machine learning, this has a vibe of a nearest neighbor's kind of detection field.

44:43Right.

44:44Jon Krohn:And that sounds similar to the idea earlier of neighboring brain cells being the thing that are of most interest to a given local brain cell. Yeah, so it is the kind of representation that we would be looking at. So again, this is kind of a little bit trivialized, but the other thing that we are trying to represent is how you can have a combination of different concepts and how it gives rise to kind of a new concept and how you can represent it in this past representation. And perhaps the most visual way we can think of it is through colors, because it's not like these concepts are positive or negative.

45:23I would want to leave the vector space world in which we talk about location and length and so on, but more about mixing colors. If you mix yellow with red with blue, you'll probably get something very brownish. But in some cases, if you mix too many colors, it's a bad thing. And if you mix just the right colors and just the right amounts, you get exactly what you want to get. So the whole kind of effect, the visual impact comes from the fact that you are just not mixing the whole palette together, but you are actually picking a small number of colors to mix in a palette, to place in one place in the painting.

46:01And that way, these sparse mixtures of colors allow you to represent the concepts that you want to represent. And only then do you kind of place them in the space.

46:15Jon Krohn:And so is that why there's some specific color ordering to your tower of Macaron, to your four-tier? Is that providing us with a fourth dimension of information? So I would say that the two, ideally, in an ideal world with, let's say, perfectly arranged towers, there would be a certain duality between the color and the location, meaning every color would have its location on the pyramid. So you could use one information or the other. but somehow the thing I want to hint at is that the location is less important than the actual, you know, the colour that this carries. I expect you could make the same analogy with taste and mixing taste.

47:02Again, if you're a good chef, you don't want to mix all the different possible tastes but you either want to have a sparse combination of tastes and you won't have a macaron with everything in it from chocolate to blueberries but you will be trying to isolate individual one or two tastes and put them together. Nice.

47:18Jon Krohn:Would have messed with your tower if I took a bite of a macaron. I could take one from the back so it doesn't change anything for the audience. Nice. That's one of the things, and this is, you know, how we go towards sparsity. You don't have to represent all of them, but I can still tell you the one that you've taken should have had that color. So that's... Those are very good macaron. Thank you, Adrian, for bringing those into the studio. So I'll dig into the latkes and the hamantushin later. We'll be demonstrating sparsity. Nice. Fantastic. Okay, so I think we've hit on a lot of the key ideas around your new architecture, the BDH.

48:00Jon Krohn:And so hopefully we'd have a lot of theory now under our belts. So then maybe some of the questions I have next can maybe, we'll see if they can be a bit more rapid fire. um in the biological brain there's been an idea there was an idea that was in vogue a few decades ago it's fallen out of vogue a little bit now but it was the idea that you would have like a grandmother brain cell that would fire when you saw your particular grandmother or heard her voice that this you know particular neuron would fire and in subsequent years it's become clear that But for something that complex, you don't just have one brain cell amongst the 90 billion in your brain that fires, but you have a set of brain cells.

48:46Jon Krohn:And those set of brain cells firing together gives you this representation of your grandmother. And so this seems similar, this kind of having a grandmother cell seems similar to something that you talk a lot about in your BDH paper, which is your discovery that there are specific neurons that fire for the idea of a currency or for the idea of a country. And the implication there is that it might be much easier with BDH to interpret what the neural network is processing than with a dense activation like a transformer-based architecture. Definitely. And definitely this is two. The way to look at it, first of all, again comes back to positive activations.

49:37One of the things to note about positive activations is that combinations are easy to express. You don't have to work so much with doing things like projections or finding negative and positive coefficients to say that I want this part of my network to activate in a positive way and that one in a negative way and balance it all together. You have full interpretability because you just say I want this set, I want this combination to fire and then this represents my grandmother. Interestingly enough, if you dive into optimization of this type of architectures and find the natural patterns that they optimize for, there's a pattern that the more important concepts are represented by generally smaller sets.

50:34So for those in the audience who are familiar with concepts of network science, of systems, this general search for power laws in systems, power law distributions in which the more important concepts are somehow more compactly represented. And you can actually, for the more important concepts, even in a relatively small network, even at like 100 million scale or even below that, find in our network an individual synapse which is sufficient evidence for a concept being mentioned. So this touches on a notion known as monosemanticity or being responsible for one concept and figuring on one concept.

51:18And this we see. Perhaps one more thing, John, to say is when you mentioned the notion of grandmother cells and how they appear, they appear spontaneously. And in architecture, it's not architected. There's no genetic code that says, hey, I want to have cells responsible for different operations. They evolve, they emerge in the course of training. We have no control over where they will be, but we find that they emerge and that you have this very clear location of signals passing through the artificial brain as a function of what we are talking about. Right.

52:02Jon Krohn:So kind of like your multicolored tower of Macaron, you don't have control over where kind of the orange and yellow Macaron end up in the tower. But you can bet that they will kind of end up aggregating together somewhere in the structure. You just don't know where. Yeah, it's pretty fascinating. And as we continue with this tower and it becomes sparser, you also end up with patterns in which the most important kind of locations are filled. So those more important concepts are filled and the less important ones are not that much filled. Maybe one thing to say for the audience who is kind of interested in technical details, but one fascinating technical detail here is that in our model, which we've been able to find and would actually shed some light on how these things work, is the concept of a grandmother synapse rather than the grandmother neuron, which means that if you think of how the state of the system works, the state of the system, the context that we are listening to, is represented by synapse activation potentiation.

53:12And those synapses that are responsible or that are related to specific contexts activate in those settings. So we have specific synapses which react to specific notions.

53:26Jon Krohn:Yeah, and just for our listeners who don't have a neuroscience background, a synapse is where two neurons meet and chemicals pass between them. And that is where learning must happen. Chemicals cross over this tiny little gap, the synapse, between two brain cells. And it's kind of analogous to the idea of the parameter in an artificial neural network model. Yeah, this is how we see it. And like being able to align the two is actually. Nice. All right. So next question for you related to this, something that I found fascinating about your paper, about your BDH paper, is you were able to concatenate, literally just like a concatenate operation.

54:16Jon Krohn:You could have one neural network trained on one language, let's say English, and you could have another language trained on, let's say French in honor of the macarons here. and with your architecture, and this seems like a rare thing to be able to do with an architecture that could be the building block of a large language model. You can just concatenate those English and French language models together. And because of the sparse activation, it just works and it's a multilingual model. That's the spirit. And I think this touches on so many different aspects, which I think are good to highlight because it's something new.

54:57It's new in many senses. As I mentioned before, the transformer, while obviously being an amazing breakthrough in the focus of machine learning and AI in general, does have its limitations in the way we understand its scaling. So if you have like two transformers and you put them side by side, there's no really clear way how to connect them. In BDH, this is much easier in the sense that the model scales in one dimension. We call it the number of neurons, N, and it's like the size of a bane. And then if you want to put two such banes together, you can do it. Depending on what you do, it will be a little bit like a mix of the skills that you had.

55:43or you can also do some post-training for the combined Bain and make sure it coordinates properly. But definitely, if you just put the Bains side by side, you have a model which out of the box has understanding for the different languages or is able to map them into concepts in English, for example, and to work with them.

56:02Jon Krohn:That is very cool. All right, so with all of these incredible novel capabilities of BDH relative to Transformers, So the positive sparse activation that we've talked about, this ability to concatenate that comes out of that, the energy efficiency that comes out of it and compute efficiency that comes out of it. Where are you today? It kind of sounds like you've, you know, with this paper, with BDH, with the Baby Dragon Hatchling paper, we're talking about a billion parameter model, which is about the size of GPT-2 from OpenAI, which is now some years old. and it performs comparably to GPT-2 despite requiring far less compute.

56:46Just to reassure the listeners to this point, we are looking at models which at a given scale are on par with models of a given scale. So really it's given all the progress that has happened in the state of the art, we use that progress obviously. So the 1B models that we produce are comparable or outperform the 1B models out there. The reason why we focus on this 1B scale for demonstrations is that this is a scale at which we are able to achieve instruction following and to start testing other capabilities of the model which is able to actually follow instructions and to have the basic capabilities that we would expect of a language model.

57:43And this is really for the ease and speed of experimentation. There's nothing particularly stopping us from releasing a super large model like in the 70-80 billion scale larger. The kind of question which is super pertinent is why do it? because if you're in the world of language models, just language models, there's a certain market, which we could call a bit of a commodity market for the kind of chatbot-like applications, discussions, and so on.

58:18Jon Krohn:Right, so your Claude, your Gemini, your ChatGPT, they're all kind of competing in the same space. I think the switch that most of us are kind of most aware of is if you are working with a reasoning model or not. Usually you are kind of explicitly aware of the switch, especially with models like GPT versus 0103 with Claude, etc. You have this option to go into reasoning mode. And this is the place where we don't want to just yet launch a non-reasoning model, which is super large, because there's actually not our objective here. What we are doing is we are entering reasoning models. We are entering it from the moderate scale, obviously.

59:09But this is a scale where we can display the advantage of this architecture. I see. I see. Notably, yeah.

59:17Jon Krohn:So yeah, so the most promising avenue for you for moving forward with this baby dragon family is into reasoning models. So models where you don't just have tokens output being spit out to your screen immediately, but there's multiple phases of reasoning happening in the background, refining your answer, ensuring accuracy. Yeah, that's where you see the most potential. That's it. Lots of consideration, lots of introspection, and also something that we see as extremely pertinent is the ability of reasoning models to work with contextualized inputs and to process them. so if you think of breaking the barriers the limits of like 1 million token context but you have a reasoning model which goes through billions of tokens of context here you're in a space in which you can for example ingest a contextualized data set private to enterprise like a documentation of an entire technology which is like 1 million pages of paper, 1 million sheets of paper that's 1 billion tokens you ingest it in a matter of minutes given enough hardware in this architecture.

1:00:31And with that in hand, you can start actually making sense of large data sets in the way you would expect of reasoning models. Again, maybe for the developer audience out there, I'm sure you're familiar with the use case of AI-assisted coding in general. and this is perhaps for currently the frontier use case. We are looking at the next generation of use cases like this, but to focus on this use case for a moment, the complexity of having an AI code assistant increases with the amount of pre-existing code with the size of the code base. And usually it's much easier to have a model which contributes a piece of new code to just invent things without actually having internalized everything that was created before its action.

1:01:30So it's basically doing a project on the side of its own, then to have basically a model which is able to control and contextually operate in an environment which requires understanding of a large code base. And again, code bases are perhaps the frontier example, but they're still the easiest kind of example that we're looking towards.

1:01:57Jon Krohn:Exciting. So for people who want to get their hands on this architecture or on models based on the BDH architecture today, can they do that? Are you providing these publicly? We are providing the architecture publicly, a simplified version of the architecture, which nonetheless performs reasonably well, comparable to the transformer as provided in our paper. We have our internal enhancements, of course, and especially those that allow this architecture to work faster, the transformer, especially in the inference generation and so on. These we keep internal, so you can lay your hands on the architecture.

1:02:33You can play with it. I believe it's an excellent playground for anybody interested in understanding models, so all topics of introspection, of observability, of getting a feel of how the model works. and what we will be releasing in the near future. So kind of follow on announcements, just hinting at them, will be related to inference efficiency and reasoning on especially enterprise datasets. Excellent.

1:03:08Jon Krohn:And so fantastic for you to provide this for our audience so that they can implement the baby dragon hatchling themselves. so there's there's been hubbub before about a kind of architecture that could potentially replace the transformer as the go-to building block within a large language model so for example mamba was something that was pretty big a couple years ago and you know there was hype around you know this being superior to the transformer i can't even now remember why do why was the mamba considered superior? Was it compute efficiency? It was compute efficiency in the case of long context and bypassing limitations on long context.

1:03:50Jon Krohn:Right. But yet today, I haven't heard somebody talk about Mamba in over a year. As far as I know, there's not a single frontier model that uses Mamba instead of a transformer. So what makes BDH different in your view? How could BDH actually be a transformer killer? So there are two aspects at least which I'd like to say in which we are more fundamentally different than the transformer. The first is something that I hinted at before, that in order to get a non-transformer architecture to work, you have to make a number of changes to make it work. You can't just say, I'm changing one thing. So the change that's most often talked about is taking attention, as we discussed attention before.

1:04:38Attention is an operation which viewed in vector spaces allows finding approximate closest neighbors due to the due to the soft max operation used in the transformer at least as it is implemented in the transformer and one attempt was basically to take the soft max out of it that's that's an operation which is called linear attention and linear attention brings us into the world of state space models so this is like one change but there are in fact five changes that you have to put together to make it work not one but five so being in sparse positive large dimension and all this has to be put together to make these models work and to find ourselves out of the ridge like in crossing the valley so to speak between two to from one place of local optimum which is the transformer to a place further out and there things start to happen so this is kind of one thing a second point is also that you can be completely radical with use of architecture there's a lot of talk of you know of supplementing the transformer with one or two layers which are different which which which help a little bit so like the hybrid transformer and mamba for example a hybrid transformer and something here there's no need for this you can call all baby go all the way baby dragon you can have like this entire architecture all consistent and what this means is that you don't complicate your picture compared to the transformer, you simplify it.

1:06:10When you simplify it, you simplify the way it works with hardware. Whether it is GPU, whether it is a GPU alternative, so one of the AI accelerators out there that do tensor-based processing, it's DPU or Lithuanian or others, the layout of memory is radically simplified with BDH. And you find that some of the bottlenecks in memory transfers of the transformer disappear because of this architecture so in some sense even if um if you take an improvement of the transformer which doesn't have a bottleneck but you still keep stick to the transformer you will have the transformer based bottlenecks with bdh you can start refresh and start sharding start arranging things differently for more optimal performance

1:07:00Jon Krohn:Well, really exciting times, Adrian. You and the Pathway team must be delighted to have made this discovery, this invention. It's a first. We're super excited about it. More news coming up soon. Definitely, definitely it's a good place to be. Congratulations. It's a big deal. It's incredible for me to think about all of the disparate knowledge that you and your team at Pathway concentrated together to be able to come up with this breakthrough. It's inspiring. And yeah, I look forward to seeing BDH in Frontier LLMs all over the world a few years from now. Exciting times indeed. Before I let you go, you'll surely remember this because you've been on the show before.

1:07:46Jon Krohn:I do always ask my guests for a book recommendation. I have no choice this time, John. I mean, I could think of different books, but since we're talking about Dagens, I have to tell you the backstory, like why dragons? I see. So these dragons are due to Terry Pratchett. Terry Pratchett, gotcha. Yeah, so it's the Discworld series. And here it'd be kind of hard best to pick one specific book. I think you can start from the beginning. So it's Color of Magic Light Fantastic. The Color of Magic Light Fantastic. The first two books, The Color of Magic and the second part. Very nice. All right, thank you for that recommendation.

1:08:25Jon Krohn:I'm sure we have some Terry Pratchett lovers out there listening already. And yeah, and then finally, where should people be following you or Pathway to get the latest on your breakthroughs or Pathway's breakthroughs, your latest thinking after today's episode? We'll be sharing on our website a series of updates. Please don't hesitate to subscribe to that. That's pathway.com. And of course, follow us on social media. Fantastic. Yeah, we'll have the links in the show notes for you. Yeah, what an exciting time. Crazy how quickly these innovations come, and it must be a real trip to be experiencing those innovations come up within your consciousness, within your biological neural network, and your attention mechanisms.

1:09:14It is exhilarating, and the interesting part is that you actually get to understand yourself better as you work on this type of topics. So that's kind of at a meta level, it's kind of all in-folding as well.

1:09:31Jon Krohn:Love it. Well, I hope to get the opportunity to check in with you and Pathway again on the podcast in the near future and see how these brilliant innovations are coming along. Thanks for joining me here in New York today, Adrian. Pleasure anytime, John. Thank you. Wow, what a mind-expanding episode with Adrian Kosofsky on the brain, on machine learning, and how Pathway's exciting new BDH architecture could provide the missing link between the two fields. As always, you can get all the show notes, including the transcript for this episode, the video recording, any materials mentioned on the show, the URLs for Adrian's social media profiles, as well as my own at superdatascience.com slash 929.

1:10:11Jon Krohn:Thanks, of course, to everyone on the Super Data Science Podcast team, podcast manager Sonja Breivich, media editor Mario Pombo, partnerships manager Natalie Zyski, researcher Serge Massis, writer Dr. Zara Karchet, and our founder Kirill Aramanko. Thanks to all of them for producing another outstanding episode for us today for enabling that outstanding team to create this free podcast for you. We are deeply grateful to our sponsors. You can support the show by checking out our sponsors' links, which are in the show notes. And if you yourself are interested in sponsoring an episode, you can get the details on how to do that by making your way to johncrone.com slash podcast otherwise help us out by sharing this episode with folks that would love to hear about it review it on your favorite podcasting app or on youtube subscribe if you're not a subscriber but most importantly just keep on tuning in i'm so grateful to have you listening and hope i can continue to make episodes you love for years and years to come until next time keep on rocking it out there and i'm looking forward to enjoying another round of the super data science podcast with you very soon Thank you.