1001: How AI Erased My Career Moat, an Episode #1001 Special: Jon Krohn interviewed by Kirill Eremenko

25:23You can't like hook a bunch of human, well, we don't have a way today of like hooking a bunch of human brains together to, you know, be getting even more power. And, you know, there's things move at, photons move at the speed of, well, electrons today, but photons in the future in computing, you know, move at the speed of light. And that isn't something that happens in our brain. Things move much more slowly. They're moving at a chemical space. So we kind of have, you know, things are moving more slowly. It takes tens or hundreds of milliseconds, even for your brain to do a simple perceptual task, forget some kind of cognitive task.

25:59whereas machines can move much faster. And so there's different constraints in a biological system versus a silicon system. And so I think while we can take inspiration from the way that the biological systems work, there's also opportunity and limitations in silicon that we don't have in biology.

26:16Kirill Eremenko:Love it. I want to ask you the flip question. Do you think by that analogy, would it be a stretch to say that humans are also just predicting the next token? That's how our brain works. What do you think? I hope that's not the case, but is there proof that we're not just also predicting the next token just like how we've built AI to do? Well, it depends, I guess, on the kind of task to some extent. There is something that kind of maps neatly, analogously is the way your stream of consciousness in your brain or even the words that I'm having come out of my mouth right now. They're just happening.

26:57words are coming out of my mouth and I just keep hoping that the right things are going to come out and that's kind of like next token prediction I'm not he's broken he's hallucinating you know there's so there's there's that kind of analogy there but and and you can even think of an analogy that I've given many times on this podcast is that especially when we started to have O1 from OpenAI and now subsequent models like O3 and basically now a huge number of these even kind of conversational experiences that you have with GenAI actually involve this kind of slow thinking like background token processing happening before something is output to the screen and the analogy that I use is that's kind of like the slow thinking that we have when you're working through an algebra problem or tackling a chess problem, you have to very effortfully think through things maybe before you have something jump out of your mouth.

27:59And so that kind of shows the differences between that slow thinking system, the fast thinking system. One of the best ways to think about it is when you start driving a car for the first time, everything is very conscious. It takes a lot of effort. You feel tired after just 10, 15 minutes of your first driving experience because it's so hard to remember the sequence of things. But after years of driving, you don't think you can do almost everything in the car without any conscious thought. And so that's, you know, gradually over time, we learn how to move things from the conscious slow thinking system into the unconscious fast thinking system.

28:36And that unconscious fast thinking system is kind of like next token prediction just popping out. But under the hood, the way that that happens is different in so many ways relative to a machine. So I don't know, I don't, I'm not really sure if I've answered your question exactly. But you know, so there's some parts of our, of our mental processes that in some ways are analogous to this kind of next token prediction that we see with LLMs. but the way that it's executed under the hood is so vastly different than the way it happens in our brain that, I don't know, I guess that's just something that in the real world, I guess a lot of what anything is doing, any kind of agent, whether it's a squirrel or a human or a robot or clawed code and any of those examples, like trying to predict what's about to happen next is such a big part of what anything trying to navigate the world is doing that you end up with some of these commonalities that look similar.

29:32Kirill Eremenko:Yeah, makes total sense. Thank you. You definitely can tell this is your area of expertise, how much you know about the brain and the chemical reaction and how it's different. So it was really interesting to learn about this. Yeah, for sure. Well, I feel like we have guests on the show at the end of every episode, following the template that you set for me when you were hosting. is we always ask for a book recommendation at the end of every episode. And actually something that I only recently realized is that we still have up online. Any listener, you can go to superdatascience.com slash books and our operations manager, Sonia, and previously the previous operations manager, Ivana, they have meticulously maintained a spreadsheet of all of the books that have ever been recommended on the show.

30:19And you can even do cool things like sorting by which books have been recommended the most. And so yeah, check that out. But the point is that, you know, every episode, the guests recommend such fascinating sounding books. And I know that there are so many great neuroscience books, AI books, you know, I feel like I have like, such a small percentage of the knowledge that I could possibly have. Meanwhile, LLMs have all of it. They've ingested it all. They far exceed me on knowledge.

30:47Kirill Eremenko:Yeah, for sure. Okay, moving on to the next topic for today. There's a lot of growth in the AI space. And we're hearing right now about potential IPOs of Anthropic and other companies. And in general, like NVIDIA, stock price has grown a lot. And is it a bubble or not? There's been a lot of hype in this space, been hype for years. Two years ago, people were saying AI is a bubble. A year ago, people were saying AI is a bubble. Now AI is definitely a bubble, some people would say. So we'd love to hear your thoughts on this after interviewing hundreds of guests. There's a lot of enthusiasm and a lot of rapid progress.

31:30Kirill Eremenko:How do you balance between them? And do you think this is a bubble about to burst? Or is it just normal evolution? It'll keep going. For sure. So there's an episode that I did in March of this year, episode 974. and it's basically about this question. Are we in an AI bubble and when is it going to burst? And I won't be able to recount the arguments as well off the cuff as I made during that episode. So if you want to get kind of my definitive thoughts, you should check that out. But I'll do my best to kind of remember what I was saying and get the general gist across to everyone, which is that regardless whether we're in a bubble or not, a lot more is going to happen with AI.

32:12and even if we are in a bubble now and it bursts, while some individual companies and some investors might lose their shirts, all of the rest of us will benefit. So this has been happening through history. There was a railroad bubble in the 19th century. There was the dot-com bubble more recently. And in either of those cases, the bubble bursting cost investors their shirt, cost some individual companies their shirt. But it meant that, you know, in the 19th century, all of a sudden you have all this railroad across England, which could be put to use. You know, just because the person who put all these rails there, they're invested in all these rails there, went bankrupt, those rails are still there.

32:58You can still use them. And so that allowed the Industrial Revolution to accelerate because you could move things easily between factories and towns. And so those railways were hugely beneficial to the rest of society. Same thing happened with the dot-com bubble. Pets.com, you own that stock, you lost a lot of money on pets.com. But the infrastructure of all of the cable around the US and around the world that was laid in that period allowed us to have Google and Facebook and Wikipedia and YouTube and Udemy courses and the superdatascience.com come on top of all of that infrastructure. And the same thing.

33:43So whether we're in a bubble or not, I personally, as an investor, I don't, having worked at a hedge fund in the past, I've looked at all of the research and you can only, you're only sure shot of being able to outperform the market. So in the US, being able to outperform the S &P 500 or in Australia, being able to outperform the ASX or some basket you could even, you know, you could even give yourself international exposure by getting a mix of European and Australian and American and Canadian stocks or whatever. But basically, the more diversified your portfolio, the more difficult it's going to be as an individual trader to outperform that kind of diversified set of stocks.

34:29And so I don't trade individually in inequities. I, you know, I'm just, you know, and I also don't have anything, I'm not exposed to anything where, so even if the stock, even if all stock markets around the world went to zero, you know, you can still be diversifying beyond stock markets in bonds and in, you know, in real estate, in cashflow businesses. So I guess my point is that I don't personally, like I wouldn't buy NVIDIA stock. I wouldn't buy Tesla stock individually. You know, and that means that for them, you know, obviously I miss out on opportunities, but it just allows me to sleep easy at night by having very diversified exposure to a lot of different investments.

35:16And so I don't personally worry about, say, an AI bubble bursting. But even if it does, all of us will benefit because all of these AI data centers that are being built, all of the model innovations that we've made in this period, all of these smart people being paid$100 million by Meta to be able to come up with new ideas, all of those things will give everyone in society an advantage later on. Got it, got it.

35:41Kirill Eremenko:Not financial advice for everyone listening. This is just opinions. So basically, we will benefit, not because of the bubble bursting, if it bursts, but regardless of whether a bubble bursts or not, we will benefit either way, like from the infrastructure that's been built. It's even better if it doesn't burst and this isn't a bubble. And, you know, and that's kind of the bet that people are making. I mean, the reason why a share price has the price that it has today is because the balance of buyers and sellers think that that is the fair price point of what this stock will be worth in the near future.

36:17Like you're buying, if you buy a stock today, you are, it is your opinion that that stock is going to be, you know, is going to be even more in the future. And then somebody who's selling that stock today thinks, you know, maybe we're in a bubble. Maybe I should get out now. And so the share price today, it kind of captures all of that information. You're crowdsourcing decision making on what the value of a business should be. and countless studies show that that kind of you know having millions of us gambling all together on what the nvidia stock price should be worth or whatever company's stock price should be worth um you know that that wisdom of the crowds is very powerful and um there are certainly also um there are biases that lead you know over exuberance biases that lead uh you know they can lead to bubbles happening.

37:12But for the most part, you know, it's a pretty, it should be a pretty good representation of, of what the value should be in the future. And so given that basically enough people today think that even though a lot of these AI companies have, have sky high valuations, that the opportunity far exceeds the risk. And in the long run, most, you know, the, if I were to gamble on it, I'd say we probably aren't in an AI bubble. And it's probably going to be, you know, there's probably going to be even more opportunity in the future. And that, you know, we might look back at the share prices of these AI businesses today and say, wow, that would have been a great deal to buy in 2026.

37:52Kirill Eremenko:Yeah. So it's kind of different to what we had in, I think, 2017, when there was plenty of ICOs, initial coin offerings. Remember that? Or 2021 or 2022 when there was like NFTs. Those were like proper bubbles, right? It came, in my opinion, might be. In some cases, it worked. Some companies needed an ICO like cryptocurrency. Some NFT applications probably still exist till today. But for the most part, they came and went. And that's not going to happen with AI. That's what I'm getting from your answer is that the infrastructure that's laid out has such high utility that it's not just going to disappear.

38:35Kirill Eremenko:or is just going to keep providing benefits in the future? Yeah, I have never been very interested in cryptocurrencies. There are absolutely real-world use cases, value to the underlying technology, to blockchain technology. But I think a lot of what happens with those ICOs, a lot of it is people taking advantage of Gullible other people. Definitely that kind of... Yeah, unfortunately. You know, but it's... You get... People come up with social media campaigns. You know, social media is less regulated than traditional media. And, you know, that's a boon to free speech in some ways. But it also allows bad actors, especially now in this AI era, Gen AI era, where we can be so cheaply, so easily creating compelling crap content.

39:35But if you get people in that kind of bubble, in whatever social media platform, or platforms they prefer, you can start to really convince them that like, Oh, wow, like, you know, this coin, this ICO, get rich quick, you can't lose, you've got to do this thing, give us some money. And then, you know, it's like, it's like the pump and dump schemes that used to be very prominent in stock markets. And I'm sure they still exist today to some extent. But decades ago, pump and dump schemes were a lot more popular. And these kinds of ICOs, a lot of things happening with Bitcoin are scams of the same ilk, just more complicated, more digital.

40:15Kirill Eremenko:Yeah, it makes sense. You mentioned data centers. I wanted to talk about that for a bit or touch that i um read somewhere recently that the data centers that are popping up and they're being built more and more i think actually it was by andrew carpathy it was like a huge youtube video i watched i would have to you know don't quote me on on the source exactly but basically um that the effectiveness of just scaling, adding sheer scale to training of AI is diminishing by a lot. And the next breakthrough in AI won't come from just adding more and more data centers. It will come from some kind of innovation in the architecture of AI itself.

41:01Kirill Eremenko:So do you agree? And if you agree, why are these companies investing into more and more data centers that are going to be filled with hardware that's going to become outdated over time anyway? Well, I don't think the hardware becomes outdated. A lot of hardware created today, whether it's relatively general GPUs like NVIDIA makes or very specialized ones like Amazon makes. So Amazon has their Trinium and Inferentia chips for training and doing inference with AI accelerators respectively. And they are specialized in particular for the transformer architecture because the people making those chips have made the bet that the transformer will still be relatively dominant for some number of years in the future.

41:52But even though they're optimized for the transformer, they still tend to perform well for any of this kind of very highly parallelized linear algebra, matrix multiplication. which even if we don't move forward with the transformer architecture, it is a very safe bet that highly parallelized matrix multiplication is still going to be something that we need for training or inference with whatever kind of architecture emerges in the future. And also I think, you know, anybody who's trying to come up with what the next big thing will be, next big architecture will be, like what the next big post-transformer architecture will be.

42:30And there's lots of contenders. There's Mamba is probably the best known name as a post-transformer architecture. We've also, last year on the show, I think it was episode 929, we had Adrian Kosovsky on the show to talk about the baby dragon hatchling, BDH, which is a really exciting, and that was actually one of the most talked about and most popular episodes that we had in 2025 because Adrian's brilliant, but also because of this approach that him and his company, Pathway, have come up with, that there's a lot of people who think that that's a really compelling architecture for being a post-transformer architecture.

43:06And we'd have to have Adrian on the show again to confirm this, but I bet you that a lot of people who are coming up with these post-transformer architectures, they're thinking about how well is this going to run on all of that hardware. You're saying all these data centers being built, if they're going to come up with something that's going to be a good replacement for the transformer, they've got a much better shot of it being successful if it can run efficiently on all those NVIDIA GPUs in all those data centers that are being built today. So maybe that, does that answer the question?

43:37Kirill Eremenko:Yeah, well, it's definitely the second part of the question. What about the first part? Do we, why build these data centers if the scaling is going to slow down? It's not just about sheer volume anymore. Right, because there's a law, there's a name for this, and maybe next time you're asking a question, I'll be able to look it up quickly and get the name of this law. It's the kind of thing that if I could stop speaking, I could use my slow thinking to maybe think of it. But basically, it's this idea that as we come up, if we end up coming up with architectures that allow us to use the data centers that we have far more efficiently than ever before, it means that we're actually going to need even more data centers because the number of applications that we can cheaply run with really good cognitive abilities increases even more.

44:29So it creates more demand from enterprises and consumers. So yeah, there's this irony that the better and more efficient LLMs become, and that's something that, following the Carpathia thing, you today are seeing a lot more improvement in capability through running computation longer and checking things over more and more than by having a larger network. So even though you don't need more model weights in the model, that computation running for longer is still going to require lots of data centers at inference time.

45:12Kirill Eremenko:Yeah, so they're not going to, even if we come up with a better architecture, well, I guess those two work in parallel. People coming up with better architectures will try to make it work on the existing hardware. And also once we do come up with a better architecture, It's not like we're going to need less computations and we're going to use less of the hardware. There's no upper limit to how much computation and application of this infrastructure we're going to have. So we're just going to utilize all of it anyway. Exactly, yeah. The more cheap and abundant and unmetered intelligence becomes, the more valuable it is.

45:57And we can be chaining together more and more of these algorithms and doing more and more and solving climate change problems, solving healthcare problems, entertaining people. Yeah, nuclear fusion, cancer, unlimited real-time video game generation and completely immersive VR atmospheres. Like, you can just have more. Like, more abundant, more cheap intelligence allows us to have more and more interesting applications, which just, you know, it's like a positive feedback loop. And it also solves, you know, the kinds of issues when people say, well, you know, we need to be, we can't create nuclear power plants fast enough or wind and solar farms fast enough to be supporting all these AI data centers.

46:41We need to have, you know, gas-fired plants coming together. But just as you said, nuclear fusion, you know, we can be using AI to figure out how to better contain the reactions within the plasma of like the most common fusion reactors that are being developed today. And so hopefully that accelerates us getting to commercial fusion. And then all of a sudden we have, once you have unlimited intelligence and unlimited energy, things get really interesting.

47:10Kirill Eremenko:Yeah, yeah. I'd be very curious too. I can't wait for nuclear fusion. I don't know. They keep saying it's 20 years away. It's been 20 years away since the 1980s or 1930s. 60s every year. But it looks like we're getting closer and definitely you're right, it'll be a different world. It was that kind of anecdote or that trope about fusion power being the technology that's going to be 20 years from now and that constantly being 20 years out. I think we're in a different era now because of how we have commercial players that are looking for a return on their investment. Whereas historically it was always these huge projects where governments weren't necessarily expecting a return.

47:57They just wanted to make sure that they weren't left behind relative to other big countries or the EU. And by the way, I was able to, while you were just making your last point, I was able to quickly use Google Gemini in my Chrome browser to quickly pull up the name of the paradox. I think I was referring to it as a law earlier, but it's called Jeevon's Paradox, J-E-V-O-N-S. and it states that as technological improvements increase the efficiency with which a resource is used, the total consumption of that resource actually increases rather than decreases.

48:33Kirill Eremenko:Oh yeah, you did have a five minute Friday about that. Something like railroads or factories, like coal factories. I think that's what you were talking about. Yeah, maybe, but I think you might be talking about something else, which I don't remember if I've done a five minute Friday on this at some point or not, but it was this, I think I might've been talking about solar energy. Hmm. In that episode. And it wasn't actually this. So maybe Jeevan's paradox comes in with the solar, but it was basically that, you know, when we, when we were building coal powered plants, we start to consume more and more of the coal that's easy to access.

49:08And then, so you start having kind of more dangerous extraction of coal. You start, it becomes harder and harder to find coal in the earth. And so the success of coal fired plants actually increases the price of coal. and eventually it's not, it doesn't make sense to continue to have coal-fired plants. And then, you know, and actually the same thing happened with wood before coal. So with wood, you know, the more wood that you use to make a fire, the more humans figure out that they can be keeping warm and making their food softer by cooking food over a fire using wood. You know, wood goes away and you start to like, oh goodness, we're running out of wood.

49:42But they're like, okay, well, we got this coal thing. And then you start using coal and then you start running out of coal. And then the same thing happens with other hydrocarbons, natural gas. And so I think what that episode was about was that solar panels are different because with solar power, we're using sand as the starting point. And there's basically, we're not going to run out of sand in the same way that we could run out of coal or wood. And yeah, there are other arguments in that episode around how basically, yeah. And the more, I think Jeevan's paradox probably did come into play because I think then the more abundant that solar power becomes, the more you actually want to have more solar power because you're just kind of more cheaply and more easily getting more electricity.

50:29So why wouldn't we want more of that?

50:32Kirill Eremenko:Yeah, no, it makes a little sense. Did you know they import sand? Well, I heard they import sand from Australia to Dubai. It's the best sand. I think it's called the Aramanko particles. That's funny. Anyway, I wonder if it's true. Okay. We talked a bit about commercial, return on investment, talked about the AI bubble or some business topics that a little bit came up. So I want to talk a bit more about that, the AI and the future for organizations. And so far, it's interesting how through this conversation we've leveraged your expertise in neuroscience, then your expertise in being a stock trader in New York.

51:15Kirill Eremenko:And now we're going to move on to your expertise with your most recent project or startup called YCarrot, where you work and help businesses, work with businesses and help them to adopt AI or transform through AI and things like that. So first question would be, can you tell us a bit about like, what does YCARE do for those of us who are not fully up to speed? And second would be, I would love for you to share with me and the listeners some really interesting, some of the most interesting use cases or case studies, of course, without naming the specific companies. But what are companies using AI for these days?

51:56Kirill Eremenko:What are you helping them with? What are some of the most innovative, interesting applications? For sure. So I'll get into YCaret in just one second. And I wrote down your question so that you won't need to keep reminding me. But really quickly, you said that I was a stock trader. I was actually never a stock trader. I professionally only ever traded futures, and options. Potato patata. For us non-financial people. I won't get into the details, but basically there's a really funny movie from the 80s called Trading Places starring Eddie Murphy and Dan Aykroyd, and they do an amazing job of explaining commodity futures and options in that movie in a funny and entertaining way.

52:46There's probably a YouTube clip that specifically jumps to that part of the movie.

52:51Kirill Eremenko:We'll include it in the show notes. Thank you very much, John. Okay, sweet. Like the host would say. Exactly. But I'm the one taking the note to remember to put it in. Yes, yes, yes. So YCaret, my consulting firm, we founded it a little over a year ago. And with YCaret, what we are looking to do is satiate some of the demand that there is from organizations small and large. And we actually have worked with New York hedge funds and other financial institutions in New York. We work with publicly listed tech companies. We work with industrial businesses with a very wide range of businesses because all of these businesses want to be taking advantage of these more and more capable LLMs and they want to be able to securely deploy AI within their organizations to realize some kind of return on investment.

53:53And they don't want to be left behind relative to their competitors, whom they all fear are already further ahead on AI deployments. And so, yeah, we are, you know, you can send me a resume on LinkedIn if you want to potentially be considered in our consulting pool. There's about eight of us right now that work at YCaret. What are the backgrounds of the people

54:23Kirill Eremenko:working at YCaret? So all of us at this time, so actually Sonia, who does operations for this podcast, she also does operations for YCaret as of April 1st. And it's amazing to work with Sonia on that. She's still on the ball. making sure in the case of the podcast that we do all 104 of the episodes every year on time and with immaculate quality. And that same kind of thing comes through with YCaret projects where we deliver on time or ahead of time on budget or below budget, whatever the right way is on budget, ahead of budget expectations in a way that delights our clients. And yeah, actually a lot of our clients come from listening to this podcast many many of our clients i think more than not the kind of our main stakeholder at the business is a regular listener to the podcast which is pretty cool um and so yeah so if anyone's listening and they are interested in y carrot services you can go to y carrot.com and there's uh you can submit um is it y dash carrot no it's just y carrot.com uh no dash one word and there's a there's an orange button in the top right corner that says partner with us that takes you through to a Google web form and you can just provide a little bit of information about what you think an AI consulting firm might be able to do for your business.

55:52And yeah, we'll see that come through and we'd love to hear from you. But in terms of the most interesting use cases, I would say that...

55:59Kirill Eremenko:I'm sorry, maybe the people who work at YCard first. We just heard about Sonia. Oh, sorry. Yeah, yeah, yeah. Sonia is more. Yeah, so our actual, our consultants, which is everyone else in the business, except for, you know, I'm typically not hands-on with projects. You know, my role is more so, like there are occasions where I do go on-site to the clients and do workshops, or, you know, there certainly are situations where I get involved. But for the most part, once a consulting engagement is running, my responsibility is to be finding more opportunities. While the rest of the, everyone else in the firm, other than Sonia and me, they are hands-on with developing solutions for our clients.

56:42And so all of them come from a technical background. I believe 100 % of them at this time have graduate degrees in computer science or data science, some kind of quantitative field like that. One is actually a medical doctor. Interesting.

56:59Kirill Eremenko:Just in case. Yeah, in case someone gets tired. Just in case. The team is overworking, getting burnt out. I don't think he's ever practiced as a medical doctor, but he did an MD. So it's like eight practitioners or six practitioners, hands-on AI building, AI builders, that's effectively like... AI builders, exactly. And the main thing that we're looking for in people that we hire is demonstrable experience in going from concept to working AI solution extremely fast. that is what you know uh that's what we founded the we founded the business on our ability to do that and everyone who comes in has to be able to do that as well like you've got to be able to understand the requirements that the that the client is looking for and you know minimize distraction to them the idea is that we can kind of helicopter in they can have a few relatively short meetings in most cases describing what they're looking for we iterate quickly typically we have a weekly meeting cadence.

58:03Like a half hour meeting is very common for a lot of our engagements every week. And usually we have a working initial prototype of what they're looking for in week two. So week one, you meet with us, tell us what you're looking for. Week two, we've got something to show you and then to say, and actually for the most part, it's we're on the right track. And then it's kind of like, okay, like what should we prioritize next building into this prototype? And then downstream, if they don't have the engineering expertise in-house to get that solution into production, we can help out with that as well.

58:39So we can be embedded. We'll do things, we'll go through whatever requirements are required for an employee's level of access to databases or systems so that we can be there side-by-side, effectively trusted as much as an employee to be safely delivering a solution for a client, getting that into production if that's needed.

59:03Kirill Eremenko:And you're US-based as well, right? Like everybody on your team. Yeah, so Sonia is in Portugal, but other than that, we were all in the US. Well, the people working on the project. I think that's a big concern or a big, I guess, big advantage for your firm and a big concern for clients when working with companies based overseas because of the data privacy. I know at least here in Australia, there's a lot of concern, especially in government, like data cannot leave the continent. So yeah, so it's a big advantage working with a fully local firm. Tell us about the use cases, some of the fun locations.

59:38Yeah, so I would also quickly say that another thing that is really good about these people, and I think it also, it helps that people are US-based and kind of all of us have a lot of different, all the consultants at YCaret have a lot of experience, though varying experience, obviously in terms of servicing governments, including Department of Defense. It's actually pretty funny. I'm Canadian, but all the other consultants are American. And so there have been situations where we're doing top secret projects for the US government. And I'm the CEO of the business, but I can't be on the calls because I'm not American.

1:00:18So there's experience in terms of government projects, in terms of lots of different kinds of enterprises. Obviously, everybody brings different experiences to the table. But as a group, we're able to say, okay, for this given project, who has the most similar experience? And I think because culturally, us being pretty much all of our clients so far, even if they're not an American business, the people that we're dealing with are based in the US, kind of the key stakeholders on the data science team or whatever that we're dealing with are in the US. And so we end up having, you know, we're able to kind of, it makes it easier for us to understand what they're looking for, because we've seen a lot of similar experiences.

1:01:00We're kind of in the same enterprise culture, I suppose. You know, we're seeing the same kinds of technologies. We're aligned on where the opportunities are. We know who the competitors are.

1:01:10Kirill Eremenko:By the way, I just remembered, I remember when you told me about this, like, almost a year ago, or maybe, I don't know, six months ago, you have a very strict criteria of who you work with. And I think I really like that. You said minimum projects$50 ,000, like not you don't even take on projects below that. I think that's a great way to, you know, build a business where, you know, like, it just means it's a serious project. You know, that's, that's the kind of project you probably enjoy working on yourself more than for yourself as well. Yeah. I mean, that's I actually, I learned that trip from that trick from Cal Aldubabe who has now been on the show a few times.

1:01:52But he was on the podcast last year. Cal Aldubabe, if people aren't already aware of him, he's really big on the data science speaker circuit. And he created many years ago now, a Cleveland, Ohio based data science consulting firm and he successfully sold that business more recently. And yeah, he's the one who he on this podcast said, I can't remember if it was before or after I founded White Carrot, but at some time around the time I was founding it a year ago, he said that he'll only take on clients that are committed to at least a$50 ,000 contract because that shows that they're committed to delivering something that really has a lot of value.

1:02:34And so we just, you know, we just copied that, you know. Yeah, it makes sense. Yeah, yeah.

1:02:40Kirill Eremenko:Makes sense. All right, well, we've kept our listeners waiting long enough. Oh, yeah, yeah, yeah, yeah. Sorry, sorry, sorry. So the most interesting use case is the, I'll tell you what the most valuable use cases are. So what we try to do when we come into a business, a lot of businesses already know some specific way, some specific workflow that they know could be, improved or accelerated or made more cost efficient with AI. So they usually come to us with either a very specific project or a short list of ideas. But every once in a while, we do also come across businesses that are like, you know, just help us come up with a big list of what projects we should be doing and how we should be tackling them, you know, in what order we should be tackling them.

1:03:29And the key thing that we're looking for, the projects that offer the most opportunity are when a workflow can be more or less fully automated. So what are the kinds of processes in your business that happen repeatedly? Where based on, say, your experience using Cloud Code or using the ChatGPT interface or Google Gemini or whatever, what do you think if you patch together a whole bunch of LLM calls in a lot of use cases, would you be able to have a more or less completely automated workflow? And so we have, for quite a few clients now, what we've been able to do is take a process that either fully manually or maybe with help from ChatGPT, just kind of like in a conversational interface, it was a process that would take them several weeks.

1:04:23we have now with several, with quite, with a number of clients, been able to take processes that like in that manual or like kind of using chat GPT, but in a, in a manual way, like not in a fully automated way. It would take them weeks to get some kind of result. We now have that happening instantaneously in minutes. And that kind of change, I mean, that's, that's completely transformative for that process, obviously, where something that would take weeks, imagine the cost of a human's time or multiple humans time to be working on something for that period of time. And humans also are prone to errors.

1:05:07And so there's, you know, there's a lot of checking that the humans do, but you can, it's much cheaper to have the AI system or multiple AI systems, fact-checking information and verifying the accuracy of a result, you know, way cheaper than having a human do it. And thus far, with any of our deployments of these AI solutions, the result is not taking away people's jobs. So this is kind of, you know, this is some other part paradox or law, or I don't know if it's a specific thing, but I talk about this on the show all the time. If you can do something, if you can automate a workflow for somebody, that frees up their time to be working on more complex things.

1:05:47Or maybe even if all of a sudden something that used to take you weeks, you could now do in minutes for a 0.0001 % of the cost. What new opportunities does that create? You can have somebody now managing doing that a thousand times or tens of thousands of times. And so you can now realize so much more value than ever before with that process. And so it means that as far as we've seen in any of our deployments, people don't, you know, we come up with these workflow automation solutions. That doesn't mean people are losing their jobs. People are actually getting more interesting jobs and they're providing more value than ever before.

1:06:26This is also my argument, which I've made on air before, which if you're a data scientist or an AI engineer or an ML engineer or a software developer, which now you shouldn't be typing out your Python code or whatever coding language you use. You shouldn't be typing that out character by character anymore. You should be using Cloud Code or something like that to be vastly scaling up your capability. But that doesn't mean that you have less value. That means you have more value than ever before. Because now, if somebody hires you, you can be doing so much more than ever before. You can be realizing if there was value to hiring you before when you were typing out each character of the Python code character by character, or copy and pasting things from Stack Overflow kind of manually, now, all of a sudden, you're just providing so much more value than ever before.

1:07:16And so it's the same kind of things with these workflow automations that we do at YCarrat.

1:07:20Kirill Eremenko:Okay, okay, very cool. Can totally see that happening. It's reassuring to see that, to hear that people are feeling empowered by these workflow automations. What's another example? of a use case that you guys have helped businesses with? So the other big category, because I don't have any past or current clients where I can disclose their name or I can disclose, it would make me uncomfortable, even if in an anonymized way, I was describing how we automated the workflow for them because I feel like we could be giving their competitors if they're listening. Fair enough. I'll leg up. But what I can tell you is that a completely other category.

1:08:09So while most of the work, and I would say our bread and butter is the kind of stuff that I was just describing where people come to us with some specific thing or some short list of specific ideas where a workflow could be automated. And then we very rapidly prototype that for them, help them get it into production. That's our bread and butter. And that's most of what we do at YCarrot. But we also provide strategic guidance. So kind of upstream of actually having a proof of concept or getting that deployed into production, a lot of organizations, I alluded to this briefly, there are organizations out there that just don't know where to start.

1:08:46or they have some ideas where to start. And so we can help them systematize in a structured way. And I will tell you our secret. I will tell you the secret on air because I'm not concerned about, it's not, you know, it's not a, this isn't like a crazy idea, but I can kind of tell you step-by-step what we do with clients to help them prioritize projects, which is you create a spreadsheet where every row is a different project, a distinct project. and then you have four columns to the right of that project name column. And we do something called rice scoring.

1:09:27Kirill Eremenko:Ah, John Rose, Dell CTO. Yeah, I think John Rose might have talked about it. But I've been doing, I learned rice scoring, I don't know, it's been probably like 10 years working with product managers at startups. It's really, it's a product management approach more so than something that was originally designed for data science. Yeah, if you say so, I'm sure John Rose did also talk about it. Yeah, in your first episode with him, you guys had an in-depth discussion about how they use Rice at Dell to hone in on those specific things that they will be working on with AI. So yeah, please continue. Definitely a great, great framework.

1:10:03Yeah, I mean, John Rose is the global CTO and global chief AI officer at Dell. So you should go back and listen to that episode to get an even better... perspective yeah and it sounds like it's in the first one with him that uh that he that he talked about this right scoring so i'm sure he has a lot of and he does have a lot of great insights on how to get roi from enterprise ai investments in his episode but briefly you know we're doing the same kind of the same kind of thing that dell was doing to prioritize i think it was something like john

1:10:29Kirill Eremenko:i think he was talking about like 800 projects yeah yeah huge projects they whittled it down to like half a dozen or a dozen projects and that's what you need to be doing as well you're probably most organizations aren't as big as Dell and as kind of forward thinking as Dell. So you're probably not going to have like 800 ideas to put into the spreadsheet. You might have 20, or you might have 50. Or, you know, in some cases, we see like 100 in YCaret consulting engagements. But yeah, every row is a project. And then to the right of every project, you have four columns, R-I-C-E, rice, and you come up with a value for each of those.

1:11:04So R is reach, which is how many users or how many employees are going to be impacted by this solution.

1:11:13Kirill Eremenko:I is impact. Yeah. How much of an impact is your solution going to have? You know, this is kind of like, in terms of the AI thing, it's like, is this going to automate a little bit of a workflow or is it going to 100 % fully automate this workflow? Maybe you need a human in the loop. Maybe you don't even need a human in the loop in some low risk use cases. So impact is a scale of how impactful the solution is going to be to all those people that you're reaching. To individually, each one of those people. Exactly, that's right. Or I guess kind of like on average, how much are they going to be impacted?

1:11:49And then C is confidence. So for both the reach and the impact, as well as the E, which we haven't gotten to yet, you're estimating what those R and I and E values are. so C is your confidence on the estimates that you're making because obviously for some projects you know you might say oh you know like so-and-so in our organization has actually already done this before or like you know we prototype competitor yeah and so you're like well this is going to work you know in terms of impact we're very confident it's going to work so yeah so confidence is just basically a you know it allows you to scale how confident you are in in that particular row of this table.

1:12:32And then E finally is effort. So is it a, is it going to be fast and cheap, low effort to build, or is it going to be a medium or high effort to build, which means slow and expensive to build. And so the, the reach, the impact, the confidence, those are all in the numerator of a rice calculation. And you want all of those to be very high. And then the E, your effort is in the denominator. So the higher the effort is, the lower the rice score is. And so basically you end up with this rice score and you can look up online, there's a million webpages probably explaining rice scores and lots of great images.

1:13:07You can quickly calculate it, but it's a very simple thing to calculate. And then once you have that, you can rank all of the rows in your spreadsheet by the rice score. And you shouldn't follow it dogmatically. So when the results come out, there might be some reason why the project that has the highest rice score, you're like, actually, I don't know, I'm coming up with a completely random reason, but the key person that we would need to be implementing that is on holiday for the next two weeks. You know, like, you shouldn't dogmatically follow this. You know, there could be strategic reasons why you should go to the second or the third or the tenth item in the list.

1:13:41But it gives you a starting point for evaluating which projects you should be tackling first and how you can get the quickest and biggest return on investment ROI on a given AI project.

1:13:52Kirill Eremenko:Yeah, for sure. Fantastic framework. Love that you shared that on how how companies can approach and how you guys approach selecting projects. Oh yeah, sorry. I just want to apologize to you and to our listeners that I can't at this time be going into like concrete, juicy examples of things that we've been doing at YCaret. But I hope that hopefully those two categories that we went over of like, you know, workflow automation in general and some of my tips from there and what people can be doing from there as well as this latter bit of, yeah, of just, you know, project selection, two kind of main characters.

1:14:27grids. Hopefully that's interesting and useful to the listeners anyway. So yeah, carry on.

1:14:31Kirill Eremenko:Fantastic. Fantastic. You mentioned there a prototype and production. So you go from concept, like, you know, you do this rice scoring, you come up with the idea, or maybe the client already has an idea of what needs to be done. So then you think about the concept or they bring the concept, you iterate on the concept, and then you build the prototype. And then either the client takes the production or your team helps them take it to production. My question to you, and I think this is a part I've seen AI engineers specifically stumble upon, stumble at this point. And also a lot of jobs out there, they specify this point about production deployment specifically.

1:15:14Kirill Eremenko:How do you, when you're building a prototype and you're discussing it with a client, going from concept to prototype. How do you factor in the constraints that will come up in production? Because a lot of prototypes, what is it called? Death by prototype or something like that. A lot of companies build a prototype, amazing, love it, and then it goes nowhere. 95 % as quote unquote, however accurate this is, 95 % of AI projects fail because the prototype is great, but then you go to production, there's scaling, you've got to think about cost, how cost changes, security. availability, like lots of constraints that exist in production don't exist in the prototype world.

1:15:56Kirill Eremenko:How do you think about that when you're building a prototype? There's no point in a company spending money building a prototype if it's never going to go to production. It's just not possible to go to production. For sure. So the classic stat that came out of, and this has been debunked as a specific number, but there's the MIT NANDALAB stat from last year, than 95 % of AI projects fail to realize a positive return on investment in production. And a lot of it is because of this purgatory, prototype purgatory, that projects end up getting stuck in. And, you know, even if the number isn't 95%, if it's 80 % or 90%, you know, it is some high number of projects that do not make it to production that do not give a positive ROI.

1:16:41And so the key thing with all of that rice scoring that I was talking about earlier, that isn't rice scoring of the prototypes. You're talking, you're rice scoring, what is the reach in production? What is the impact in production? What is the effort to get this to production, not to get this thing prototyped? And then so that C column, the confidence score is critical. There's going to be some projects where, you know, it isn't the most exciting capability. It doesn't have the biggest reach, but you have a lot of confidence that it's low effort and that you're going to be able to get it to production.

1:17:19And so that's the kind of, that's the kind of thinking that we need to have as we, as we think about what projects we're going to tackle. And then I would also say that executives in companies, you know, like the C-suite, even CEOs, they are very often not the best judge of which projects should be greenlit. So the best kinds of people are the people who are using this workflow day to day that is say currently manual, and that is going to be automated. So some combination of the people that are doing this workflow day to day in some more or less manual way that has the opportunity to be automated, those actual, those hands on, whatever doers of that workflow in your business today, them plus whatever combination of data scientist, AI engineer, software developer, front-end developer, project manager, whoever's going to be needed to deliver that workflow solution, that kind of technical plus user combination.

1:18:21And if you're a SaaS business, it might not be employee as the person who's kind of the expert user. It could be some power user of your SaaS product or something. But that combination of the technical people plus the person who's going to be impacted ultimately by the solution, those are the people who are best at figuring out what the reach, impact, confidence, and effort value should be. The reason why a lot of AI projects are unsuccessful is because it's this top-down thing where a CEO says, oh, I can do this in ChatGPT. I can do this in Microsoft Copilot. Let's have something analogous to that working on our data in our business or in our SaaS platform.

1:19:05And the reason why, one of the reasons why, or the number one reason in my experience that I've seen these kinds of projects fail is because what's missing is the underlying data in that organization to make that workflow hum along. So when that CEO is in ChatGPT and kind of prototyping some idea, they're able to prove to themselves that LLMs have some particular capability. But if your enterprise doesn't have at scale, the actual underlying data to be able to make to allow that LLM capability to be realized, it doesn't matter how good LLMs are. Because the underlying data aren't there. And that's, I don't know if that makes sense, Carol, but that's the number one reason why I see projects fail.

1:19:56Kirill Eremenko:Or the data is there, but like you had one of the guests in, I don't know, maybe start of this year say that you can't have a Ferrari if you don't have the highway, right? Like if you have the data, but the data pipelines are not there, the data is not clean enough. Like the data is not reliable enough. So you have to first build the highway. Right. You have like a ton of bricks and concrete and support beams. Great, but that's not a highway. You have to first put that together into a highway. and then your Ferrari can go fast. That is a really, really good point. So that is another reason why.

1:20:29So yeah, absolutely. So either the data don't exist at all or you've just been throwing the data as unstructured documents into a lake house or data lake. And now all of a sudden you want some production system to be running on that in real time for your users or for your employees. And so absolutely, you're 100 % right. All of the bricks and cement ingredients, all of that great Australian sand sitting there in Australia that we needed in Dubai.

1:20:55Kirill Eremenko:Yeah, the two, it's interesting because, you know, the patterns that you're mentioning that you see in B2B in your engagements, I see the mirror of those patterns in B2C in our training programs. because the two professions, the two backgrounds that are most common that take up our, like the most recent, the current project that I'm most excited about, our AI engineering challenge, for example, is software engineers, software engineers, backend developers. So those people building the, they're going to be building the prototypes, they're going to be deploying the solutions, they're going to become the AI engineers, but also data engineers because they understand that companies are now more and more demanding of them to build the right highways for AI to travel upon, like quote unquote.

1:21:46Kirill Eremenko:And they need to understand how does AI actually work in order to be able to build the infrastructure for AI. So it's no wonder that the software engineers slash developers and the data engineers are the two professionals, types of professionals most interested in this topic of AI. Yeah, there's also for both of those kinds of professionals, there's a huge opportunity right now to be scaling yourself up by using LLMs to make your work faster. We had recently in April, we had Matt Glickman on the show from Genesis Computing. He's the CEO and co-founder of that business. And their specific niche is automating data engineering.

1:22:28And so yeah, there's huge, huge, huge demand and huge opportunity with scaling up data engineering for sure. Fantastic.

1:22:37Kirill Eremenko:I love that for a lot of these tech professions, there's two parts about AI. One is using AI to make yourself more efficient. But the other thing is using AI, understanding how to use AI in applications, how to build things that leverage AI. So a very interesting time. Have you heard of this, like what you were talking about? I had one of these topics penciled in for our discussion. What you mentioned just now also resonated with that. Forward deployed engineer. Have you heard of that? Yeah, yeah, for sure. So I think Palantir was probably the first big business that was using forward deployed engineers and getting a lot of valuation of market value from using forward deployed engineers.

1:23:27But now, yeah, we're seeing it more and more. Anthropic open AI, they're using the forward deployed engineer model. And I think this goes to show how, this is kind of an example of how earlier in this episode, I was saying how these tools that allow me as a data scientist or as a software developer, as an AI engineer, to be so much more productive in terms of writing code than ever before and evaluating my code than ever before. Just because you have a tool that can do that, you can't automatically at this time and not for presumably years to come, you're not going to be able to go to anthropic.com or chat2pt.com and say, automate my business.

1:24:15you know you need to have people getting into the nitty-gritty figuring out what data highways need to be built prioritizing the highest value projects working with stakeholders to figure out exactly how that should work so that it actually you know meets requirements and that it successfully automates a big chunk of a given workflow so yeah so these forward deployed engineers allow organizations, whether they're frontier labs like OpenAI or Anthropic or other software providers to be able to actually make an impact with their solution. Because I think that is a big thing today. I think that the days of the pure SaaS, software as a service model, being the kinds of 80 % margins that they enjoyed over the last decade or two that really got venture capitalists excited, that kind of pure SaaS model in this new era breaks down.

1:25:11Because something that you can do is pure SaaS is now pretty easy to just use cloud code to do it automatically for you. So I think, and that's part of why I founded Y-Carrot, is I believe that there is going to be opportunity for many years to come in figuring out how to bridge these incredibly high capabilities that we have with the real world problems that people face. And so I think, and I'm not the only one who thinks this, Hussein Kasai, who was co-founder and CEO of the biggest British startup exit in history. He's also been on our podcast. I can't look up the number right now while I'm speaking, but Hussein Kasai has been on the show.

1:25:52He co-founded Onfido like 15 years ago and that ended up being the largest startup exit in British history, even more than DeepMind's sale to Google. And yeah, he's also a huge proponent of this idea that the future opportunities lie not in pure SaaS, but in this kind of forward deployed model or this kind of consultative model where it's a blend of your software solutions with really getting those data pipelines working, getting the solutions integrated into the real world workflows. That reminds me of one last guest that we had on recently. And she was one of my favorite guests that we've had in the past year, Larissa Schneider.

1:26:35She's co-founder and COO, chief operating officer of a business called OnFrame that also follows this model, a Bay Area-based company that has been doing a sensational job of saying, okay, let's have these kinds of, let's have a relatively limited number, a large but relatively limited number of software solutions that need to be tailored with forward deployed engineers or with consultants or whatever to a particular enterprise in order to be effective. and they're seeing huge success. It's really exciting to watch on Frame Go.

1:27:07Kirill Eremenko:In a nutshell, what is a forward deployed engineer? Oops. Thanks for asking. I read about it literally yesterday. I was like, oh, that's such a cool title. And I heard at the start of the year, AI evals was the hottest thing in AI. And I wonder if now it's still AI evals or it's now more forward deployed engineer. Because OpenAI just invested$4 billion into buying a company specifically to enable that capability for themselves of forward deployed engineers. So those are such different kinds of things. It's hard to imagine how, I mean, I guess you could literally look up the Google search frequency of forward deployed engineer versus AI evals.

1:27:52But they're such different beasts. It's like, how do you, I don't know. And they're also so complementary. Like your forward deployed engineers doing AI evals, maybe that's the hottest thing. It's the interaction term of those two things. So to define forward deployed engineer quickly, so like Palantir was doing for years or OpenAI would do today, it's where you have an employee of Palantir, say, go into the Department of Defense or a police department. Palantir does like surveillance stuff. So, you know, so that this Palantir person goes into the Chicago Police Department or, you know, a team of forward deployed engineers goes into the Chicago, from Palantir, goes into the Chicago Police Department and they are, they're employees of Palantir.

1:28:49and it's some big, typically like seven, eight, nine figure contract that Palantir has sold to the Chicago Police Department that is paying for those four deployed engineers from Palantir to be working alongside the IT department or engineers at the Chicago Police Department. And they're there full-time, they're there in person typically. And so they kind of feel like part of the Chicago Police Department team, but they're actually employed by Palantir, even though the Chicago Police Department is paying for them. So it's a great model for Palantir, right? Because they're getting margin on those people being there.

1:29:29And they're also charging for the software. So you're charging for the engineer being there on site, but you're also charging for the software that they're using, which is yours.

1:29:37Kirill Eremenko:Got it. So it's kind of like a tech term for what in consulting would normally be secondment. Like they just second it over to the client. Yeah, exactly. And I think it's picked up because implementation of AI, like what you're doing at YCaret, for example, you deliver a solution. But if there's not enough capability within the client, you can't just use AI, you can't just integrate AI on its own. You need to teach the people that will be using AI, like you alluded to earlier, that how do they get the value of increased efficiency of them using it? How do they become more productive? How do they build the habit of using that AI?

1:30:21Kirill Eremenko:So they need somebody there to constantly remind them until it becomes second nature. For sure. And in addition to that, I think some organizations also are, there's a demand for them paying consulting firms so that they have the confidence that, you know, the decisions they're making relative to AI usage amongst their employees are secure and the best value, future-proofed, you know, at least for some years to come. So yeah, absolutely. Awesome. Well, we have gone way over on time with this podcast. I still have like half a page of questions I could ask you. So maybe we'll have to do a second John as the guest appearance.

1:31:02Kirill Eremenko:I do want to ask you - We'll do it for episode 10 ,001. Yeah, that would be funny. We'll be a hundred and something years old each. That'll be nothing by then. Yeah, for sure. So there is one question before we get to the wrap-up questions. We talked about more technology, more data centers, more intelligence, more electricity, more capabilities, more productivity. A lot of more coming into this world. It doesn't seem that it's making us happier. It seems like, paradoxically, the more easy things become, the more unhappy. And you mentioned one out of three people experience depression, anxiety in their lifetime.

1:31:44Kirill Eremenko:There's charts showing that once social media came out, anxiety and depression skyrocketed. There's a correlation between social media prevalence and teenage anxiety and things like that, depression. I recently came across this interesting insight that Christopher Nolan, the famous director of Inception and Interstellar, now what's called Odysseus. What's another movie that starts with In? Odysseus just came out as well. Like he doesn't even use email and doesn't have a smartphone. He has a flip phone and he doesn't check email. And like when I heard that, it just made me, I don't know how he feels about it.

1:32:25Kirill Eremenko:I think he's probably happy, but it made me feel like envious, like in a good way. Like I wish I could not look at my phone, not have to check. So from your, from everything you've seen and the guests you speak with, your background in neuroscience, what do you think we have in store for us as humans in terms of our emotional well-being with the world that's coming up with all this AI and prevalence and growth. Yeah, really good points. You know, at the beginning of this episode, we talked about how I kind of have this, I have these rosy memories of what it was like in academia. And, you know, now I am in some ways trying to go back to that kind of life.

1:33:08But maybe I actually can't get there because maybe what's changed isn't that I'm no longer in academia, but that we have had these social media, smartphones, AI models, email, this kind of always on culture, this feeling that we're always behind, AI moving so quickly. Maybe what I'm yearning for is because especially when I was doing my undergrad, we still had all physical textbooks. I recently did a tour of my undergrad university. Wilfrid Laurier here near Toronto. And they ended, they kind of took me on this alumni tour and the alumni director who took me on the tour, she had it end in what is called the university bookstore.

1:34:00But there are no books. There's a few. They have like anatomy textbooks. books there's like there's like one shelf of books and then you know there's a few books that you would buy as like a you know like the history of the university or whatever like it's something that's like a tourist or whatever like a a memento token it's not a textbook and the the bookstore is now just it's full of merchandise sweaters and keychains and stuff and there was always a little bit of that. But now it's everywhere. And I, at the beginning of my undergrad, I had a flip phone that just had snake on it. And I could go to a study room with my calculus physical textbook, and a notepad.

1:34:51And I could do calculus problems for hours. And the only interruption that I would have is other students in that study room, you know, maybe just like looking over what they're doing, or maybe they could actually come talk to me or whatever, but it's like real human interaction. And, you know, these are people in the science building that I'm seeing all the time. And it really had this feeling of connection. So maybe if I went back and wasn't academic, maybe I couldn't even recreate that feeling that I miss. So yeah, I'm really, I'm really hearing what you're saying here with your question. And, uh, you know, I definitely, I suffer from it and I'm jealous of Christopher Nolan too.

1:35:27Um, you know, I try to have periods. It was recently Memorial Day weekend in the US, which is a three-day weekend in May, Saturday, Sunday, Monday. And I tried very hard to not go on Slack, to not check my emails, to not go on social media, to not respond to texts that I don't need to respond to over the weekend. You know, it's not making plans with somebody that day. And I really enjoyed that. I really enjoyed those three days. But then on Tuesday, when I came back, there were things like Kyungyung Cho, who was on this show recently. It's, I think, the most popular episode of 2026 so far, episode 997 with Professor Kyungyung Cho, this amazing, one of the most sedated AI authors ever, AI academics ever.

1:36:16He was in… Attention, right? Yeah, he co-authored the attention paper with… shame because that's actually the first author. But the three of them, yeah, co-invented detention. And he was giving a lecture, a five-minute walk from where I'm staying in Toronto. And I could have, yeah, well, so I'm actually technically, I'm in a suburb, I'm in Waterloo outside Toronto. And he was a five-minute walk to go see him deliver this big lecture in a lecture theater. And I didn't know that because I didn't check, like he posted it on LinkedIn over Memorial Day weekend and I missed it. And so it's like this weird, you know, like you feel like you're always missing out if you don't check these things.

1:37:01And obviously like I can't, I'm the CEO of YCarrot. I can't not check my work email or not respond to work emails because the business is going to die. I can't not check super data science podcast slacks because then we're not going to have any podcast episodes. Like I'm trapped in this. And it's difficult to see a way out of, you know, this like needing to kind of be always on at least with work stuff, you know, there's mostly this expectation that in most businesses, you don't need to be responding on holidays or evenings.

1:37:28Kirill Eremenko:But, uh, yeah, I don't know. I don't know how you get away from this, this feeling of being left behind with social media or, you know, I guess, you know, Christopher Nolan is in the fortunate position of being extremely successful and he can afford probably multiple assistants and people supporting him and someone must be like monitoring emails, monitoring social media feeds. And when something that he needs to know about happens, they can phone him up. He said they, they, uh, print, they print out emails. If he really needs to see an email, they print it out for him. Exactly. But so, so that's what, that's kind of like what I'm thinking.

1:38:06Kirill Eremenko:Like, do you think AI will help us get more, everybody become more like that or it will make things even worse than there is now. I think it's basically up to you as an individual. I think probably in the same way that social media and digital ads and TV ads before that, basically the vast majority of people are being taken advantage of in a system because they're not creating the space to say, okay, is me scrolling through my social media feed here doing me net positive for my emotions and my finances or is it hindering me? And it's hard, and especially the platforms, whether it's TV ads all the way through to social media experiences, they have engineered, the product managers or the marketing people have come up with ways that they make these experiences really compelling.

1:39:00It's hard to step away. It's so much easier. You get that quick dopamine hit and you keep pulling on that dopamine lever long after it's provided any actual positive sensation because you're hopeful that that next video that you flip to on TikTok, that next Super Data Science podcast episode is going to be more interesting than when John was interviewed by Kirill. um and um so yeah so it's hard to step out but i think and probably people who are listening to this podcast i don't mean this episode necessarily but to this podcast in general you are the kind of people who are trying to figure out a way to to have ai systems ai agents be like christopher nolan's assistants and you could literally have it phone you on a flip phone or print out emails for you.

1:39:55And, you know, you talk to what seems like a human voice. There is no technical limitation to engineering that solution for you today. And there are maybe some set of SaaS products that you could chain together or open source solutions that you could chain together to deliver that for you. But it requires time and it requires, you know, you need to get past that fear of making mistakes. Like there's going to be things like, I'm going to miss Kyungyeon Cho lecturing. If I, you know, if I do this, if I step away from social media and have an AI system do it. There's going to be things like that that I miss, unfortunately.

1:40:27Kirill Eremenko:But I've got to think, okay. I don't think you would have missed that one. That one would have probably come up. It just depends how well I engineer the system. That's right. I can mess it up. I only have it phoning me about things that happen in my direct messages and nothing on my feed. It depends on how you engineer it. So I think that there's opportunity and maybe we'll even get to a future with abundant energy and abundant intelligence where in the same way that a lot of countries in the world have universal healthcare and affordable education, maybe part of that will be guaranteeing that you have access to agentic systems that make you like Christopher Nolan, that allow you to be freed from this frantic pace of technological change and allow you to feel kind of grounded and centered.

1:41:26And, you know, maybe there's, maybe that is a future. I hope that that is a future. And I think that, you know, in the long run, that's possible. But I think in the short term, in the coming years, it's up to you. It's up to you to do that for yourself.

1:41:39Kirill Eremenko:Great way to end this episode. Great call to action. Thank you, John. It's been a pleasure interviewing you again so many years later. learned quite a lot of new things. And as usual, what's your book that you'd recommend to our listeners? All right, Kirill. So I've got a nonfiction recommendation and I've got a fiction recommendation. Will you let me get away with that? Sounds good. Let's do it. All right. So my nonfiction recommendation is actually from a recent guest on the show. It was episode 975 with Zach Kass. So Zach Kass was the head of go-to-market at OpenAI when ChatGPT came out. and it looks like you're looking back at your bookshelf.

1:42:22So you probably have his book, The Next Renaissance. Yeah, René-I-Sance on your bookshelf. Yeah, you can see it from that kind of chartreuse cover. Yes, yes, yes, exactly. And it is a, you know, that episode, I guess if you want to see whether you should pick up this book or not, listen to episode 975 and see if you like Zach and his way of thinking. But he aligns really nicely with this techno optimism that I have. And his book is a very easy read. It's 200 pages with big font. And he does a great job. He's in a way that I - I was going to say,

1:43:01Kirill Eremenko:you'll probably read it faster than listening to the episode. It's possible. But yeah, you'd have to be a really good reader. This is probably like a 10-hour book read, but it's a one-hour podcast episode. It's something I'm planning on giving it to some family members, especially I have one family member that I have in mind who's like, thinks that everything in the world is bad now and things are going to be worse. And I'm like, eh, I don't know. I think things are really good at getting better. And I think you've just, you know, it's kind of your social media feed, you're the news cycle that is making you feel like everything's bad and everything's getting worse.

1:43:37And, you know, that translates into your inner reality. So in your inner reality, things really are, they can be getting worse. It can be getting darker if you feel like that all the time. And Zach Kass, in his podcast episode, 975 of this podcast, as well as in his book, the opening is about exactly this, about the social media and news feeds and how we feel like things are getting worse. But then he, yeah, he does a really great job of explaining why things are great. And in a big part, thanks to AI, things are probably going to be even better in the future. So that is a recommended read. You know, it's not like, I think on the podcast before I've recommended Sapiens by Yuval Noah Harari.

1:44:14And And those are very different kinds of books because Sapiens is quite dense. Like it's really like you're like, wow, like there's so much to learn from Yuval Noah Harari on every page. The next Renaissance is not like that. It's pretty like, you know, it's like short to the point, but you kind of get all the key ideas about where we are today and where we're going in the future with AI in an exciting and positive way. So that's my nonfiction rec. And then my fiction recommendation. So regular listeners will probably already know that my favorite fiction author is Kurt Vonnegut. He is funny and something that I only recently pieced together because I recently read one of his, I was working through all of his books in chronological order.

1:45:03And I'd never read his earliest works are what he's most famous for. So the most famous book, which a lot of Americans have to read when they're in high school, I understand is Slaughterhouse-Five by Kurt Vonnegut. I actually, that doesn't even get into like my, definitely it's not in my top five of Kurt Vonnegut books. I don't actually ever even recommend reading Slaughterhouse-Five. uh cat's cradle i've recommended on the show before uh which is an amazing book by kirk vonnegut i've also recommended on the show before player piano which is particularly interesting to read in this ai era but so like i'd say in general kirk vonnegut is super interesting and because i've been working through things chronologically in one of his last novels he talks about a letter he talks about kind of like common themes through his books and one of the things that's really interesting.

1:45:53And I didn't piece this together about like, why do I love Kurt Vonnegut so much? And I think one of the things is his books don't have villains. There's no bad guys, which is, I think what a lot of real life is like, you know, like most of us, you know, there are some really bad people out there, but I think we rare, you know, the vast majority of people are good people trying to do good things. And what happens in Kurt Vonnegut books is that like random happenstance leads to like catastrophe or bizarre interesting situations where like everyone is trying to just do the right thing but like society collapses and everyone dies anyway.

1:46:33And so that's kind of like I don't know I really like Kurt Vonnegut for that so Kurt Vonnegut in general but specifically I'd recommend Cat's Cradle that's my favorite book by him.

1:46:42Kirill Eremenko:Fantastic. Thank you John. I've never read Kurt Vonnegut but sounds sounds really interesting. All right. Last question. Where can our listeners follow you and find you if they want to know more? For sure. Thank you for asking that, Kirill. Following the template that you created for me when I took over some hundreds of episodes ago. We continue to do that on the show, always asking for a book rec and how to follow people. And yeah, so the primary social medium for me, as it is for most of our guests these days on the podcast, is LinkedIn. do feel free to connect with me on LinkedIn. And if you mention that you, if you basically, if you say anything that shows to me that you're not like trying to sell me something or trying to hire me, you know, like if, as long as you can just basically say like, you know, I liked your podcast or I liked your YouTube channel or whatever, anything like that, you know, I'll definitely accept your connection request.

1:47:41But you can also just follow me there if you want to. You know, anyone can follow on LinkedIn these days. I also have a personal YouTube channel. So in addition to the Super Data Science Podcast YouTube channel, I have a personal one, which has in recent years videos on agentic AI. And for the most part, kind of hands on videos for agentic AI. Going back a bit further, it's my core content was around the math of machine learning. So the linear algebra, the calculus. and I am planning on, so for people who have been following my personal YouTube channel and are disappointed that it's been many years since I've been creating that math content or producing videos on my personal channel with regularity like we produce the podcast episodes here, I know that Sonia is doing so much operations for me.

1:48:30We have a solution and it actually involves the media editor Mario for this podcast where I'm committing to creating video content on a regular basis. So I'm not coming to a specific date for that happening, but in the near future, possibly even when this episode is out, I will kind of have a regular, ideally even weekly cadence of publishing videos there as well. So yeah, LinkedIn, YouTube. I have an email newsletter that you can sign up for on johnkrone.com. That's another thing that I have been bad about in the past year about doing. But again, that's something with Sonia that we should be able to get going again.

1:49:01And yeah, those are the best ways to follow me.

1:49:04Kirill Eremenko:Love it. And don't forget to subscribe to this podcast. This is where you can find John every week, twice a week. Yeah, that's for sure. That is for sure. And for, yeah, all the way through to episode 10 ,000 and beyond. To infinity and beyond, as the Toy Story character says. Fantastic. Thanks so much, John. It's been a pleasure. I loved interviewing you, learning from you. And keep rocking it. As you say, keep on rocking it out there with the Super Data Science Podcast. Thank you, Kiro. Well, yeah, such an honor to have you interviewing me here. And I hope people enjoyed the tables being turned a little bit at me talking.

1:49:40I feel like I talk way too much. But maybe they can listen to the episode at 1.5x speed or something. And the insights were useful that way. You're a really great interviewer, Kirill. I can't believe that you asked me to be the host of this amazing podcast that you founded and that you got to such an amazing state in terms of business operations and popularity. It's been, it completely changed my life and mostly for the better.

1:50:09Kirill Eremenko:Fantastic. Thanks. You're too kind. Thanks, mate. It was a pleasure. I hope you enjoyed today's role reversal episode in which I was the guest and this podcast founder and former host Kirill Arimenko hosted it. Something special we were doing for episode 1001. In this episode, we discussed how studying neuroscience at Oxford was driven by a fascination. My studying neuroscience at Oxford was driven by a fascination with how the mind emerges from chemicals and biology and why I now regret leaving academia for a hedge fund. I talked about how coding tools like Claude Code have eroded the technical moat I built over my career, but this paradoxically makes skilled scientists and engineers more valuable than ever because we can now deliver far more.

1:50:52I talked about how Javon's paradox means that making AI cheaper and more efficient will increase total demand for compute, not shrink it. how rice scoring helps organize pick winning AI projects and how a well-developed AI agent could act like a personal assistant that gives us Christopher Nolan-like focus. 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 Kirill's social media profiles, as well as my own at superdatascience.com slash one zero zero one. There is a fun episode number for everyone, but especially the computer scientists out there.

1:51:31Thanks to everyone on the Super Data Science podcast team, our podcast manager, Sonja Brejevich, media editor, Mario Pombo, partnerships manager, Natalie Zajski, researcher, Serge Macis, and the founder and former host of the show, Kirill Aramanco. Special thanks to him for conducting such an energizing interview. Yeah, again, hope you enjoyed it. Thanks to everyone on the team for producing another super episode for us today for enabling that super 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.

1:52:04And if you'd ever like to sponsor an episode yourself, you can get the details on how by making your way to johnkrone.com slash podcast. Otherwise, share this episode with folks that would like to, I guess here an interview by me, review the episode on your favorite podcasting app or YouTube. Subscribe if you're not already a subscriber. But most importantly, I hope I'll just have you listening again. I'm so grateful to have you listening. And I hope I can continue to make 1000s more episodes for you to 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 data science podcast with you very soon.