907: Neuroscience, AI and the Limitations of LLMs, with Dr. Zohar Bronfman

26:49Jon Krohn:generative creativity was it yeah exactly so um combinatorial where it's uh combination of a couple of different i i'll give an example in a second i just forgot the term sorry combinatorial what was the other by bowden transformational transformational was the term You know, it's jargon. So I'll explain quickly combinatorial and transformational creativity, because I think the distinction is interesting and it indeed relates to what we just discussed in terms of specific versus general domain intelligence. Combinatorial creativity would be creating something new and unique, hence creativity, that is some kind of an assembly of existing things.

27:42Right? So you take a few interesting pieces across domains and you put it together. You create something new. And it's unique and it's creative and it checks the box of being creative. However, it's a relatively, quote unquote, simple form of being creative. LRMs can do it. They can take a couple of jokes that one stand-up comedian said, a couple of jokes that another one said, maybe something from a movie, and create a new joke that draws on those three. now obviously it's a continuum and it's not like a completely two separate categories because in some sense everything is being composed of other elements but transformational creativity again has this concept of meta approach to a problem where it's not like you just compose different aspects that exist.

28:40You take a completely new perspective to the problem or to the issue, and you get a completely new approach. And yes, the approach is inspired by something. And yes, it has elements of existing things for sure always, but it is very different. It's order of magnitudes different than how people thought about the specific problem previously. Obviously, You know, the great examples would be when Albert Einstein came up with relativity and said, no, we're looking at it completely differently. It's a matter of geometry rather than mechanistics. And, you know, that's an example of transformational creativity.

29:21And one of the arguments is that if you have domain-specific intelligence, it would be extremely hard for you to build transformational creativity and synthesize, basically, knowledge. This is something, this is like key, in my mind at least, for where we want to get one day with artificial general intelligence.

29:47Jon Krohn:I love that. That was such a fascinating section. When we think about generative AI capabilities in these terms, combinatorial versus transformational, probably most listeners use LLMs, have conversations with them on a regular basis. It's certainly amazing to see that combinatorial thing work where you ask for some science concept to be explained in the style of Snoop Dogg and it works unbelievably well. It does it at a level that would be superhuman for most humans to be able to reel off science in the style of whoever. But it's not transformational, as you described there. So we've talked about generative AI a fair bit here.

30:39But you actually, in a keynote two years ago, you declared that prediction is all you need.

30:49Jon Krohn:that generative AI shouldn't necessarily be prioritized by the AI community as much as it has been. So tell us about that perspective. Do you still feel that way two years later? I think to a large degree, yes. I think if you ask yourself, and again, I'm talking specifically about businesses, not about consumers and prosumers who obviously have a lot of value just even from the existing capabilities LLMs possess. I do think that specifically for businesses, if you're thinking about a business that is trying to, you know, transform how things operate within the business in a way that is going to be a real needle mover, then today the vectors of value that come out of LLMs are quite limited.

31:44I don't think LLMs changed entirely a trajectory of a business in a way that is even close to how machine learning and predictive capabilities did or are doing. If you think about the biggest companies, the most successful companies out there, from Google and Facebook to Amazon, Uber, Spotify, I think it's a fair argument to say it was machine learning and predictive modeling that got them to the place they are. But I don't see, so, you know, there's this joke that the ones that made the most money of the gold rush were the ones that were selling the, how do you say, the smaller - The picks.

32:36Jon Krohn:It's like picks and axes and maps. Exactly. The picks and the axes and all of that. Now, again, I'm not being cynical in any way, but today the big money still goes for companies that enable LLM, computationally, infrastructurally. And that's amazing. I'm all for it. We use LLMs throughout our stack in many different ways. But ultimately, the real business value of taking a certain process and optimizing it with your data still, I think, relies 99 % of cases on one or another form of making predictions or advanced analysis of the data. I don't think this will change. And therefore, I think we should for sure continue investing in large language models.

33:25We for sure need to find ways of tying them closer and closer to business value because we know ultimately this is where value lies. And I still think machine learning has maybe only 5 % of its potential uncovered in the sphere of companies.

33:45Jon Krohn:This episode of Super Data Science is brought to you by the Dell AI Factory with NVIDIA, delivering a comprehensive portfolio of AI technologies, validated and turnkey solutions with expert services to help you achieve AI outcomes faster. Extend your enterprise with AI and Gen AI at scale, powered by the broad Dell portfolio of AI infrastructure and services with NVIDIA industry-leading accelerated computing. It's a full stack that includes GPUs and networking, as well as NVIDIA AI Enterprise software, NVIDIA Inference Microservices, models, and agent blueprints. Visit www.dell.com slash superdatascience to learn more.

34:26Jon Krohn:That's dell.com slash superdatascience. data science. Yeah, there is, whether we are close to AGI or not, one thing that's for sure is that there's a lot of opportunity still in enterprises to be taking advantage of data that they have or that they can collect and automate things. And absolutely, a huge amount of that will happen in the predictive realm. There are particular places, and we're going to talk about this more later in the episode, particularly with respect to your startup, to PECAN. oh do you pronounce it pecan or pecan? It actually depends where you're from in the US or North America for that matter so usually in the east coast they say pecan I also say pecan for some reason but west coast and midwest would usually go with pecan or something like that So you're okay with either?

35:23I'm okay with either, I'm like a pluralist

35:26Jon Krohn:Yeah, I'm an East Coaster, born in Toronto and now 13 years in New York. And so, yeah, it's Pecan for sure. It sounds funny to me when I hear Pecan. But yeah, so we'll talk about how LLMs are useful in your platform. And a lot of that is around having an interface, a natural language interface with some underlying backend, some code being generated in the backend, some data being pulled automatically using natural language, that's really useful. And there's a huge amount of places in the enterprise where that kind of capability can still be added or, you know, having something convert from, you know, a tabular structure or, you know, some unstructured data, you know, converting that into some natural language that can be output to a customer support person or maybe to a user directly.

36:23Jon Krohn:there's lots of places where generative AI can be useful for an enterprise. But there's still, I think, as you say, there's vastly more opportunity for predictive. I agree with you 100%. So something interesting that blends together predictive machine learning and the neuroscience that we've been talking about already in this episode is we tend to So different ways of doing recordings from the human brain. For me, something that has stood out for me as one of the craziest things that I've learned in neuroscience are these experiments by a guy named Benjamin LeBay. L-I-B-E-T, yeah. And so you're familiar with these.

37:07Jon Krohn:And so I'll explain briefly my understanding of this, my memory of this to the audience. But basically, I think he was using EEG. So, you know, relatively simple approach where you have recording electrodes on someone's scalp, and it gets kind of this coarse reading, coarse spatially, because the skull kind of interferes with the electrical activity of the brain. So spatially, you get a bit of a coarse response. But in terms of time, you get very precise, you know, millisecond timescale precision on recording events of the brain. and something very interesting that Benjamin LeBay and countless others have shown is that you can have a neurological, the neural basis of some conscious idea that you have happens hundreds of milliseconds before you have the conscious thought.

38:01Jon Krohn:And this is a very disturbing thing to think about because you kind of, most of us go around through the day with this illusion that you have some kind of control over what thoughts come into your head or what action you take next. But in fact, what these experiments show is that you become aware of a decision after that decision has already been made subconsciously in your brain. And so, yeah, I don't know. There's plenty to dig into there, but maybe talk to us about this a bit more and then maybe tie it into your belief in AI's ability to anticipate user behavior? So I think we were talking about, you know, once you get exposed to something in the realm of neuroscience and AI, you lose sleep.

38:50This was probably the biggest sleep deprivation I had because it's mind-blowing, right? If we think about it, you know, deep, we might end up in a rabbit hole of, Am I just an agent carrying my neurons or something like that? Which I don't think is very easy to disprove, by the way, in all honesty. So it's mind-blowing. It's obviously a lot to digest. But yes, by the way, those experiments happened first sometimes during the 80s. Since then, it's been replicated and reproduced in different settings and in different environments, in different technologies, in different animals. so many times that I don't think it's any more even just an open question.

39:36It's a truism. It's given. Now, obviously, there's room for interpretation. But the fact that there are brain processes that are directly, causally related to decisions we make and that we don't have access, we don't have conscious access to those processes, I think is already completely agreed upon. Obviously, you can ask the questions of how elaborate these processes are, whether as something reaches consciousness, it can override or kind of change some of these or veto some of these processes and so on and so forth. But the fact that these happen is like hard fact. Now, it means that much of what we are doing, okay, as humans is predetermined by things that have nothing to do with our immediate desires.

40:24okay so you can put someone in fMRI like a functional MRI that basically shows the blood in your brain and you know which areas are active and you can tell 10 minutes or 15 minutes and they drive in a car simulator you can know 10 or 15 minutes in advance before they reach the junction whether they're gonna left turn left or right so we and and you know it it has many contributors to it. Maybe it's a question of your stronger side. Maybe it's something that happened in the morning. Maybe your neck was sore and it's harder for you to look to the right. By the way, it's my case at the moment. So there are many different ways that can contribute to these unconscious processes that end up affecting your decision or your action.

41:13But what it also means, and this is something, again that is quite known for many years, it means that as a consumer, your behavior is also affected by many things that you're not aware of. And it means that as a business that sells to consumers, you can probably know much in advance of your specific customers' behaviors before the event takes place. So you can predict the purchases that customer is going to make, the conversions or lack of those, lifetime value, best products, churn, and so on and so forth. And that ability to make those predictions based on their historical behavior is, like I said earlier, the biggest level we know in the industry for transforming businesses.

42:12So I'm basically saying if you collect data about your consumers as a business, there's a good chance you can start making predictions about their behavior in the future and you can optimize their experience. You can optimize your processes. You can basically just make the most out of those precious interactions the consumers have with your business. my you know my personal and this is what Piken is all about my personal mission I want to bring these capabilities to as many small and mid-sized businesses as possible because they also deserve quote-unquote that remarkable technology that basically you know tells you what people are going to do even before they know what they're going to do and that's why we've invested so much in connecting LLM's data and machine learning together in one nice package.

43:14Jon Krohn:Nice, that was so fascinating to hear. You definitely know a lot more about the kind of Benjamin LeBay stuff that I was talking about and subconscious neural correlates of decisions before we become consciously aware of those decisions. and it reminds me of how, you know, so that also is a reason that I ended up doing machine learning is because I actually got, I got a full scholarship to go to University College London to study the neural correlates of consciousness and I had a nervous breakdown writing that PhD proposal because when you, like you said, when you think deeply about these things, it is, it kind of becomes, and I just try to live the rest of my life like forgetting basically that I know this, but that it becomes very hard to see yourself as outside of a machine or having any real free will.

44:08I hear you. I think you went through a very similar thing. I just decided I'm going to jump to these waters nonetheless.

44:16Jon Krohn:Let's see what comes up on the other side. But these are remarkable disciplines, so interesting. And by the way, it's not an incident that so many of the AI researchers and leaders, and by the way, you mentioned earlier the forefathers of AI and so on and so forth, most of them are cognitive and neuroscientists. And it's quite remarkable and it's very interesting. It doesn't mean that you can't come from other disciplines, obviously computer science and other statistics and so on, but there's definitely something about brain dynamics and those functional organizations there that really contribute to our network thinking.

45:00Jon Krohn:For sure. And a large amount of neuroscience inspiration that goes into AI systems, for sure. Even if it ends up being a gross oversimplification, it can end up being powerful, scaled up. One last question for you before we get into some really PECAN-specific questions, one kind of big philosophical questions. You ended a blog post, which I will put in the show notes. You ended one blog post with a challenge, what question would a machine never ask? And I've been dying to find out what you mean by that and if you have examples of the kinds of questions that machines would never ask. Yeah, I mean, I think it relates mostly to the discussion we had earlier about value systems.

45:51so I think I'm expecting machines not to ask questions around why. Why questions? Not why mechanistically. I mean like write me a letter. Why are you trying to write to a CEO or to a board of directors? No, but like why? Why do you ask me to do that? Why should I do it? These types of why questions, they are the latest ones that children develop. but then when they do, I don't know if you have children, John, or not, but I can tell you for my children, they don't stop asking. And understanding why, not the mechanistic why, the real why, the value why. This is where when we assign them, the children, their understanding of the value system that surrounds us in society, in the family, as humans.

46:49and I don't think machines, for sure they're not asking it today and it's completely irrelevant for them, but I don't think they will ask themselves until we get very, very far in the AGI road.

47:02Jon Krohn:Yeah, this is something fascinating that fiction writers, movie creators have dug into for a long time. This threshold over which machines start to ask that question like your children, and I don't have children. just to, but I've seen, it's kind of this trope of like the kid asking why, why, and you know, kind of going down the rabbit hole further and further till you get to a point where you're like, I don't know. Exactly. Because so many times the answer you provide them is a mechanistic answer and they're not trying to get the mechanistic answer. It's not a question of what happens if this ball hits that one.

47:38It's the causal, it's the value question that lies in underneath. Why do we want to be kind? Why do we want to be supportive? Why should I love my relative? These are a core value. And then in many cases, what we provide are actually axioms, right? And they have to just accept those axioms or not accept them. And these are the types of conversations that machines are very far from being able to conduct.

48:07Jon Krohn:regular listeners will already be aware that i'm obsessed with anthropic's fable 5 model and it has taken over my working life i'm writing a technical book that includes latex files mathematical notation python code examples and fable 5 and clod code handles requests i make across whole chapters with accompanying jupiter notebooks end-to-end work that a few short months ago would have been dozens of separate requests with way more manual fiddling required with fable 5 it just works, essentially like magic, first time. Claude is the AI for problem solvers. It's the collaborator that understands your entire workflow and thinks with you, not for you.

48:44Jon Krohn:Whether you're debugging code at midnight, building a financial model, or strategizing your next business move, Claude extends your thinking to tackle the problems that matter. For problems worth solving, get started with Claude at claude.ai slash superdata. That's claude.ai slash superdata. And check out Claude Pro, which includes access to all of the features mentioned in today's episode. Claude.ai slash superdata. We eventually run into fences that we can't climb over and see what's on the other side. It doesn't take too many whys to get to a fence that is unclimbable with things like, why is there anything?

49:23Jon Krohn:Why is there matter? and how did it come to be that, you know, there's a few billion conscious monkeys talking to each other over podcasting platforms on a one rock that's floating through space. There you go. Here's another. Thank you. But I have a, just really quickly, there's a film I watched a couple of years ago, I guess when it came out, because it's only a film from 2021. It stars Colin Farrell and it's called After Yang. And it is an exceptionally fascinating. I absolutely love it. I think about it all the time. And it's a great, it's, it deals with this kind of idea of, of machines getting to a threshold where there's something, yeah, this kind of this depth of, of question or of, of kind of emotion of, of wondering why of trying to get down.

50:24Um,

50:25Jon Krohn:yeah, down, down, down deeper into things. And after Yang, I'll have that in the show notes. I really loved it. Um, anyway, let's get to PECAN. Um, so at PECAN, you've taken on the biggest challenge. Uh, yeah, you, you've said that the biggest challenge of making a model is not the training, but the data. And actually this is something that's come up in a number of recent episodes on the show, this kind of this idea of how critical and it's kind of obvious when you say it, like we spend so much time worried about model capabilities and performance on benchmarks, but in practice, none of that matters.

51:04It matters very little typically relative to the underlying

51:09Jon Krohn:data that you're training the model on or that you're trying to use at inference time in some consumer or enterprise AI use case. And so just as some other examples of episodes recently where we've talked about that, we've got episode 906 with Professor Jason Corso from the University of Michigan. We've got episode 901 with Lilith Batlia, who leads a section at some of the big conferences like NeurIPS on data-centric machine learning, she calls it, data-centric machine Learning Research, DMLR. And so this is, I think this is something that, you know, it's interesting because we haven't talked about that a lot.

51:49Jon Krohn:You know, we've done over 900 episodes of the show. And in most episodes, we are talking about model capabilities, but for some reason, recently in episodes on the show, so maybe that's some kind of indicator of where consciousness is more generally around this, this realization that data are so much, are often the key limiting factor in having an AI model today that works effectively in your organization. And so, and something, a reason why that should be obvious to a lot of our listeners is if you're a hands-on practitioner, the hard part usually in a data science and an AI project is getting the data into a place where it can be fed into a model, where it can do something useful.

52:31Jon Krohn:And the model itself is often, you know, you could be using open source model weights for a lot of it, maybe some fine tuning. And so, yeah, you have a quote here that if you analyze the work of a team of data scientists, 90 % of their time is going to be modeling the data as opposed to, you know, passing it through some existing model. So why is data modeling so challenging and how can we overcome the bottleneck it creates? Yeah, John, let's break it up to a couple of components. First of all, we are very much obsessed with performance and accuracy for several reasons. First of all, it's very easy to measure.

53:13It's a classic bias, right? It's very easy to say my model does this and that on this and that data set and reach this and that benchmark. I don't say it's not important. And obviously, especially for pure research purposes, you have to have benchmarks and you have to measure accuracy for sure. However, in the business context, the difference between a 91 % area under the curve or 91.5 % is usually, not always, but usually meaningless. It's absolutely meaningless. There's so many other considerations that are 95 times more important than the statistical accuracy of the final model. For example, what are you even trying to model?

53:57Are you trying to, for example, let me just give you a very simple example. Let's say you want to predict churn. Okay. Let's say that you predict churn a week in advance. So you're saying this customer is going to churn and it's seven days before they notify about canceling the subscription. In that case, you reach 95 % accuracy. But let's say that from a business perspective, those seven days leave you no time whatsoever to change their mind. And let's say that 14 days in advance, you only reach 70 % accuracy, but you can actually have time to give them a call, suggest a promotion, or send some nice customer experience gift or whatever.

54:45It's the actual business framing that is crucial rather than the accuracy of the model. In many cases, the accuracy would be just a vanity metrics. The other thing which relates to the data is also a hint into why the realm of generative AI and LLMs and all of that are still struggling when it comes to predictive modeling. The reason is every company, every company has its own data fingerprint. No two companies have the exact same data structure and data context and semantics and quality and so on and so forth.

55:39and every company has to go through a data transformation from how the data resides in the warehouse and the data stores and whatnot into the data that is actually being fed to the predictive model. That transformation, that one sentence, you just need to go through transformation, transform your data, make it ready for machine learning or for predictive modeling. That is by far the most challenging aspect because it unfolds the whole discipline of data science into it. How do you define the entity you're going to predict for? How do you define the label? How do you define the stride and the frequency of the data set?

56:29How do you consolidate different features and different attributes? How do you prevent leakage? How do you prevent drift? How do you make sure you don't have crucial anomalies? How do you make sure you have enough samples of positive and negatives and you don't have any skews that are going to completely mess up your model? How do you make sure it corresponds to the framework that you're going to use the model within? All of these questions make the work of the data scientist long and hard, as all data scientists will attest. And if you ask yourself, what is the real barrier for non-data scientists, for people who are just data savvy and builders, and they want to become data scientists in practice, it is that data transformation and structuring.

57:26It's not the modeling itself.

57:28Jon Krohn:Right, right, right. All right, so let's get into some mechanics why kind of questions, or how actually, I suppose, is more of what I'm getting into here. So you went through an impressive list there of the kinds of issues that data scientists run into when they are preparing data for a model or trying to get a model to work like the business would like it to in production. That was impressive in itself. but how does PECAN so things like you said, data leakage, concept drift, real time model validation how does PECAN how does PECAN AI

58:12Jon Krohn:mechanistically how does that work? Is somebody walk me through a user journey I go to your website, I download a tool and then how do I use it or maybe there's a couple different kinds of user stories you can go through to kind of practically give us a sense of how PKN is solving these kinds of problems for people. Yeah, so like I mentioned, our mission is to help everyone in small, mid, and large businesses to benefit from the remarkable capability of harnessing machine learning and building predictive models. What we've done over the course of the last seven years is develop very big and very deep and wide technology that automates all of those data structuring and transformations that are so crucial for building a real business valuable model.

59:05It includes a lot of LLMs, LLMs that go over your data, understand the semantic relationship, do a lot of the transformation and the engineering. It also includes data engineering, proprietary, actually patented data engineering processes we've developed. And we've actually built, think of it as if it's a vibe data science notebook. So it's kind of a co-pilot that has a conversational interface, and it walks you, the data-savvy yet not data scientist user, in the various, I should say, processes that are involved in building a predictive model. It understands the data. It understands the business goal.

1:00:01It helps you define your entities and your label. It creates queries. It shows it to you. It asks you whether it makes sense, whether the output makes sense. It creates metrics for you. And it basically just think of it as a little agent that works with you in all of those different steps. And it does it all super fast so that you end up with a model very quickly. Now you can evaluate that model. You can try and understand that model to a greater extent with our agent. And then this becomes a quick iterative process until you reach a model you feel comfortable with.

1:00:39Jon Krohn:This is really cool. And so the website is at pecan.ai. And actually, really quickly, what is behind that name? The reason we are calling ourselves Pecan is because we believe AI is a hard nut to crack. But when you do, it's very good for you, like Pecans. So we just want everyone to crack the nut with us and just get all the value they can from AI. I love that. I'm so glad I asked. I worried you were just going to say, well, it was a five-character domain,.AI domain we could get. No, we started before AI was so hyped, so we were lucky. Nice. It ties together a lot of things we've been describing in this episode already.

1:01:26Jon Krohn:It seems like it's tailored towards people who want to be building predictive models, where, as we've discussed earlier in the episode, that's where there's the most juice to squeeze in enterprises overall. But you figured out clever ways of including LLMs in a lot of places so that it allows people to, with even just natural language alone, be able to describe some problem that they're trying to solve and have the data be cleaned up so that it can go into a predictive model. and you can get that predictive model in production much faster and with much less effort. Exactly. You can think of it as your predictive agent, like an agent that really helps you become a predictive builder.

1:02:18Jon Krohn:Really cool. I like that a lot. And yeah, is there any kind of... So in terms of an organization, I guess any kind of organization could make use of this tool. It's enterprises, government organizations, I guess anybody who's trying to automate aspects of their business with AI, right? Yeah, I'd say if you could potentially benefit from machine learning, which is probably the case for most organizations out there, you could use PECAN and depending on the use case and the business process, see the benefit. Obviously, there are simpler use cases, there are more complex ones, but the platform is ML at large.

1:03:07And there are so many different ways you can do it.

1:03:09Jon Krohn:Really cool. And so something that is an important part of the platform to you, you talked about PECAN having kind of like a co-pilot and natural language interface. And this enables organizations to harness predictive analytics without reliance on data scientists. So you've said if only data scientists can use it, PECAN has failed. So tell us a bit more about this, especially our primary audience is hands-on data scientists. And so how do you see tools like yours evolving so that people like analysts, and I don't mean data analysts necessarily, I mean like financial analysts. Business analysts.

1:03:51Jon Krohn:Business analysts, marketers, operations teams, those are the kinds of people that you want to be using your tool, not just data scientists. And so, yeah, how does that work? How is that going? Are you getting the feedback that that's working? And how does this impact us as hands-on practitioners, as many of our listeners are? How are data scientists, data engineers, ML engineers, MLOps people, do you see them being less critical in the future with more tools like PECAN coming along? So I want to say, you know, right off the bat, to all of our data science listeners, guys, I love you. I love data scientists.

1:04:37I consider myself a data scientist. I have nothing against data scientists. However, we all know as data scientists that there are probably 100 times more use cases and potential for data science and machine learning out there. And so many organizations that just don't have data scientists and could benefit from data science. So what we are basically saying, yes, we obviously have a lot of customers who are data scientists and they love the platform and they use the platform. and it's a terrific match. And what we do for data scientists is usually helping them to accelerate because data scientists, and we all know it, have limited capacity.

1:05:20They can deal X amount of use cases a year. With our platform, they can just take a little bit more and accelerate some of the processes, prototype real quick. So there's a lot of benefits for data scientists for sure. However, when you think about the company's core mission, I sleep well at night whenever I see another customer who's all of the sudden running four different models in production and have no data scientist in their org chart because they're too small, because they didn't prioritize, because they couldn't hire, whatever might be the reason. So we serve data scientists, we're proud of our data scientist users, and we still also love getting other companies that don't have enough data science resources just across that threshold of predictive modeling.

1:06:15Jon Krohn:Nice. I like that. John, with regards to your question about what is the future work of data scientists and MLOps and ML engineers, I would say it's a bright future. I'm not from the camp that holds some kind of a catastrophic perception that we won't need data scientists in the future. I think it's very far before we won't need data scientists, even with a platform like PIKEN, because data science, and again, data scientists know that very well, has levels of complexity and has levels of nuances. The idea here is that you take the less complex use cases, the ones that are already very well defined and well understood, you automate those, and then you free up data scientists to deal with the more complex and nuanced things, right?

1:07:10It's like, think about lawyers or accountants. GPT can probably do 95 % of what a lawyer or counsel intern can do. But you'd still want to have lawyers because you want to deal with the very nuanced and complex and human-related aspects that LLMs can't really address. Same goes for data science, in my mind.

1:07:36Jon Krohn:They need more embodiment. and then they need to need robots in the courtroom before they can really replace lawyers. Nice. So one final question for you around PECAN is you've described its origin as a roller coaster. So I've got a why for you. So why PECAN? I understand there's something to do with missing a deadline in a data science competition that the company came about. Yes. So I would say, first of all, Noah, my co-founder and our CTO and myself, we started PECAN back to back from graduating our PhD studies. So first of all, we never worked in a company before that. I didn't even know what a payslip looks like.

1:08:27So it was a roller coaster even from day one when, you know, you were overwhelmed with everything. We raised$4 million seven years ago, and it was just for us like, oh, my God, what just happened? What are we doing now? So from that moment on, everything just became increasingly more insane. You know, it is a roller coaster. I'm experiencing almost at a daily level huge wins and success stories and obviously huge challenges. It's non-linear in the most non-linear way you can imagine. Specifically, the deadline was that we decided we want to test our system and we want to participate at the competition.

1:09:15And speaking of accuracy and benchmarks, we wanted to see what happens when we let PECAN solve some of the hardest data science problems. Specifically, it was around the customer next best offer use case without us intervening. So basically, here's the data. Do all of the structuring we talked about. Do all of the cleansing. Do all of the feature engineering. Build a model. Do the predictions. Send them to the competition. Let's see how well you are able to perform. And we got to the first place. And we were psyched. We were absolutely psyched. and then we realized a little bit later that we submitted after the deadline.

1:10:00So it was just an example of how things constantly change. Within the day, you get good and bad news every moment, basically. But I guess over time you get used to it and you just try to focus on the goal and constantly push forward.

1:10:21Jon Krohn:Well, congratulations on your success, Zohar, and to your continued success. May those daily successes marginally outnumber the daily challenges and you continue to see some great growth. Exactly. We just need them to marginally outnumber and be good. Exactly. So before I let you go, this is actually, I have a note that I'm supposed to warn people before we actually start recording that at the end of every episode, I ask for a book recommendation, but I forgot to give you that warning. So you're just getting, you're getting this cold. Do you happen to have a book recommendation for us? It doesn't need to be related to our field, though it can be?

1:10:59Wow. You know, first of all, I, so here's the story. I used to read every week. I had to finish at least one book. Otherwise I was upset with myself. Obviously it was during the philosophical and neuroscience studies and I have a ton of recommendations. So I will give you one mainstream recommendation, but then I will give you one fringe recommendation, which is very recent. So the best, number one book, my most favorite book on the planet is Samuel's Beckett trilogy.

1:11:39It's, I mean, you know, the way I define books, there are a couple, a handful of transformative books, books you read and then you become a different person. To me, that was a transformative book. I carry it in too many places and try to kind of revisit it. It's a very weird book, right? If that's my mainstream, what would be my fringe recommendation? But it's still a piece of art that I don't think any AI will ever be able to create. It's just amazing. Okay, so that's the mainstream recommendation. I will send you the name in English because I'm reading it in Hebrew at the moment but there's a book that I'm reading now I don't read a lot of books unfortunately now but now I'm reading one book and it's very fringe it's very freaky but it's extremely interesting and it's about aliens and it's about a person writing on his own encounters with aliens during the 90s and early 2000s.

1:12:51I have an alien thing. Not sure yet exactly what thing I have. But for those of us who question, you know, the day-to-day assumptions, I recommend reading that book. It's definitely eye-opening and it's definitely, it gives you an angle that, I mean, we don't come across every day. you just need to keep an open mind and don't be judgmental. I'm not saying it's true or false. I'm just saying it's a very interesting narrative and story.

1:13:22Jon Krohn:Yeah, there are interesting things going on there. So yeah, so you're going to have to send me, I guess, the name of this book for me to put in the show notes. Because it occurred to me, I was like, maybe you can just tell it to me in Hebrew and I can Google it, but I can't even do that. I can't Google things in Hebrew because I can't possibly, I don't have an internal LLM that can convert. I can show, I can put the, I have it on my desk now. So that's in Hebrew, that's the book. Yeah, see, I can't type that. You can't type that. I will send you all the information. And I mean, if someone is into aliens, it's an interesting story.

1:14:04Jon Krohn:Yeah, that's, I've been trying for the past year, and we'll see how it goes. I've been kind of, I've been developing a TV series idea related to AI. And so I've been talking to a lot of production companies and someone who's been involved with this over the whole year. She's an Emmy award winning documentarian and a project that she did relatively recently was a series called Encounters for Netflix. and yeah it's kind of it's you know there's I think there's four episodes and some of the first episode I thought was really interesting like they dig into kind of radar data that are quite anomalous around UFO sightings and some of the other episodes I didn't find had as much kind of substantial underlying hard data but yeah there's certainly when you watch something like that, or I guess read something like the book that you're reading, there's some, some unexplained things.

1:15:10And Hey, we are all here data scientists. We can assess that the likelihood of no aliens is probably as close to zero as the likelihood can be. Now, the only question is, you know, what to do with it.

1:15:25Jon Krohn:Yeah. Yeah. Yeah. Nice. All right. Thank you so much. Zohar. Before I let you go, do you have, uh, what are the best ways to be following you? I mean, this has been a fascinating episode. People want to get more of your thoughts. Yeah, where should they follow you on social media or PECAN on social media? I'd say probably my LinkedIn account. I'm usually active on LinkedIn more than any other social media. So everyone are more than welcome and sometimes share my thoughts and other interesting pieces. So by all means. Nice, fantastic. We'll have that in the show notes as well. Well, Zohar, this has been a fascinating conversation.

1:16:04Jon Krohn:I could go on for hours and hours, and maybe we will have that chance. Hopefully we can have another episode again in the not-too-distant future with you to check in on how PECAN is coming along and get more of your insights at the intersection of neuroscience and AI. It's been a fascinating episode. Thank you. Thank you so much, John. Happy to be here.

1:16:29Jon Krohn:wow wow wow thanks to dr zohar bronfman for such a mind-blowing episode in it he covered how llms lack the unified value systems and domain general learning capabilities needed for agi llms can't for example transfer learn from chess to business negotiation like humans can he talked about how bumblebees demonstrate unlimited associative learning by recognizing objects by sight that they've only touched before, illustrating that even simple brains can do cross-domain transfer that AI cannot. We talked about Benjamin Lebe's experiments demonstrating that our brains make decisions hundreds of milliseconds before we're consciously aware of them.

1:17:06Jon Krohn:With existential implications, as well as implications for businesses, they could theoretically predict customer behavior before customers know what they'll do themselves. And we talked about how while LLMs show impressive combinatorial creativity mixing existing ideas, they lack the transformational creativity needed for paradigm shifts like Einstein's relativity. And we talked about how PKN AI uses LLMs to democratize predictive analytics, automating the 90 % of data science work that involves data preparation rather than modeling. As always, you can get all the show notes, including the transcript for this episode,

1:17:48Jon Krohn:social media profiles, as well as my own at superdatascience.com slash 907. And yeah, thanks to everyone on the Super Data Science podcast team, our podcast manager, Sonja Breivich, media editor, Mario Pombo, our partnerships team, which is Nathan Daly and Natalie Jaisky, our researcher, Serge Massis, writer, Dr. Zahra Karche, and our founder, Kirill Arimenko. Thanks to all of them for producing another extraordinary episode for us today. for enabling that super team to create this free podcast for you. We're deeply grateful to our sponsors. You can support the show by checking out our sponsors links, which are in the show notes.

1:18:24Jon Krohn:And if you ever want to sponsor an episode yourself, you can get the details on how to do that by making your way to johnkrone.com slash podcast. Otherwise, support us by sharing the show with people that would like to have their minds blown by Zohar. Review the show on your favorite podcasting app or on YouTube. subscribe obviously but most importantly we just hope you'll 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 super data science podcast with you very soon