Charting the Path to AGI with Jon Krohn, Co-Founder & Data Science at Nebula

Today, we're talking to Jon Krohn, co-founder at Nebula and host of the SuperDataScience podcast, about the state of AGI and how he crafted a beer using AI. You're listening to Joel Beasley, Modern CTO.

I just listened to the episode where you guys took an AI, made beer, had the beer made and distributed, then went out and interviewed people on-site with how the beer tasted. Tell me about that project.

Well, Joel, thanks for having me on the show. That was a super fun project. I guess it's still ongoing. Hopefully, at some point, my Krohn and Borg beer will be available around the U.S. But yeah, as of right now, this AI-crafted beer can only be had at this SpeciesX Brewery in Columbus, Ohio, and it is fantastic.

I absolutely loved it. People that I invited, obviously, they loved it. People on camera loved it. That's a low bar. But on top of that, I organically heard people out on the street around the brewery talking about their favorite beer at the brewery being this Krohn and Borg beer.

And for our beer aficionados out there, it is supposed to kind of remind you of the Kronenbourg 1664 name, of that lager. But it's my last name, Krohn, and then it's supposed to be like the Borg from Star Trek. It's like the AI element. So Krohn and Borg beer.

I picked up on that pretty quick too. I was like, I think that's what he's doing. How did this come about? Were you just doing an episode and you guys said, "Hey, I wonder if we can get an AI to make us some beer?"

Yeah. So about a year ago, with the ChatGPT mania that was going on, I had one of my best friends on the podcast, which I've never done that kind of thing before. My guests are always luminaries in data science that I know professionally. I mean, they might be friends, but this was my first time just having a friend on the show because he was a heavy user of ChatGPT. And so he said, "Let me come on air, talk about how I'm using ChatGPT as a layman to help with marketing copy, to automate parts of my business, to deal with customer service requests."

And this guy, Zach, his business happens to be running a home brew supply store. So in this episode, we talk about beer a lot. We talk about how much I like beer. And shortly after that episode aired, one of the listeners to my podcast reached out to me and said, "I'm a brewer. I love brewing beer, obviously, and I love listening to your podcast. I'm interested in AI. I don't come from a formal AI background, but I have this project where I'd like to simulate crafting beer with AI." And that guy named Beau Warren, he now launched his SpeciesX Brewery in Columbus, Ohio that I already mentioned. And yeah, so I've been collaborating with him for the last year on this project. He's done all of the heavy lifting. The hardest part was data curation. But then for somebody who doesn't have a formal background in AI, he did an amazing job with the machine learning models as well. So I just provided a little bit of input, and yeah, the result was absolutely delicious.

So you've got the SuperDataScience podcast. You're north of 700 episodes. You're about where we're at, by the way. I don't know the official count. Maybe Josh can let us know. But we're around like 765. So we're like 760. Yeah. What's your main gig? You're an author, right?

I am also an author. So I wrote a book called Deep Learning Illustrated that came out a few years ago. I'm currently writing my second book, which is the Mathematical Foundations of Machine Learning. But my main job is I'm a co-founder and the chief data scientist at a company called Nebula. And so people watching the video version of this can see my Nebula hat.

And whatever we do, we're gonna mess with your green screen throughout the interview. It'll be a lot of fun.

Nice. Yeah. I wonder where I'll be. I'll be underwater or something. Yes. So Nebula is a machine learning company. We are automating white-collar processes, specifically in human resources to start. So we are using generative AI in particular to be able to automate and augment activities that people are doing to make them more enjoyable, more effective, and generate ultimately more commercial value for companies.

So as a kind of key example, if you've ever posted a job on Monster, LinkedIn, whatever, it's typical to get hundreds of applicants. And to go through each one of those applications and really read the resume and think, "Is this person a great fit for the role I'm hiring for?" Anecdotally, the experience that I've had of that is that when you start doing that, let's say you got 200 resumes to review for a particular job. The first five or so, I'm so methodical and careful, and I'm really looking for reasons like, "Hey, you know what? This person could actually be a good fit despite not exactly having the job title I was looking for, but they've got some really cool experience. It does seem relevant." And then after getting through those five, I'm like, "Oh, man. An hour has gone by. I've got 195 more of these. I've gotta pick this up." So then you start just kind of looking for reasons to—you're just starting to skim resumes and just looking for quick reasons to be like, "Nope. That person's not a fit. Person's not a fit. Person's not a fit."

Whichever of those two mindsets you're in, you're not making great decisions. So in the first mindset, you're going to be including lots more people than you should be for whatever your next stage, your shortlist process, when you're gonna have to review all those again. So, you know, you maybe get your 200 down to 20 in some kind of shortlisting process. So in the first mindset where you're really being careful, you're going to end up including people that you shouldn't have anyway. But then even more dangerous and more unfair to those 200 applicants is that most of the time when you're reviewing resumes, you're gonna be doing it way too quickly. You're gonna be throwing out resumes for no good reason, just because it's been too long since you've had a coffee or it's getting too late into the evening. And so what we've done at Nebula is we've built a series of AI models that handle different parts of the resume evaluation process, all the way from creating a job description with generative AI to taking a small amount of context that you might have around someone that you're looking for, like, "Hey, I'm looking for a data scientist in New York." And we expand that with generative AI into a richer query, like specific kinds of data scientist skills that you might be looking for. And we then use that generated description to search in seconds in our platform over 180 million U.S. professional profiles, and in seconds, rank everyone in the U.S. from top to bottom for the specific role that you're hiring for.

Nice. Yeah. And so, similarly, you could take if you have those 200 applicants that applied to a job posting that you'd put out, you could use our algorithms to sort those 200 resumes from top to bottom so that instead of going through them in a random order, you're starting with the people that are most likely to be the best fit. And, you know, so maybe it's only going to be in the first 20 or 30 that there are good people, and 170 of those you can just ignore.

So tell me a little bit about this Deep Learning Illustrated book that you wrote.

Sure. So Deep Learning Illustrated is an introduction for people to deep learning. It might be obvious. So deep neural networks is a synonym for deep learning. Deep learning is the technology behind 99% or more of the AI breakthroughs that we've had in the last decade.

So it's based on an idea called artificial neural networks that comes from the 1950s. And so this artificial neural network idea was trying to mimic the way that biological neural networks, your biological brain, works. And stretching all the way back 70 years into the 1950s, we already had a bit of an understanding how the brain works. Early computers were being programmed to mimic the way that the biological brain works. But it wasn't until the last decade when we had cheap enough compute, more than enough data being stored, that we've been able to leverage those concepts from 70 years ago to great effect across the broad range of applications that you see today.

So when your phone recognizes your voice, when your phone recognizes your face, when you see video generation tools like Sora recently being demoed, when you're using chat—

That screen's crazy.

Super crazy. Yeah. All of those AI advances are using deep learning models under the covers. And so my book, Deep Learning Illustrated, was an introduction to these deep learning approaches. And so the illustrated approach was this idea of trying to be highly visual, as opposed to too focused on the math. We do have some math where needed, but it's an introduction to this approach in a visual, largely intuitive way. There's 14 chapters in the book. The first four, if you don't have any programming experience, you can still pick it up and get an overview of the main application areas of AI. So you can understand machine vision. You can understand natural language processing. You can understand creative systems like text generation or video generation, image generation, as well as complex sequential decision making, where you have an agent that's actually making decisions in a video game or in a board game or in the real world. So across all of those four main application areas of AI, the first four chapters cover those without any code. And then in chapter five, we start getting into code examples in Python and showing you how you can use open source software libraries to build algorithms that do AI across all those application areas.

Did you see, you know, Tyler Perry? He put a hold on his several hundred million dollar investment into his movie studio when he saw Sora come out. Yeah. What's that gonna do? I mean, it's not a stretch of the imagination to think that movies will essentially be incredibly detailed prompts.

Yeah. It is absolutely possible. I had a guest on my podcast, SuperDataScience, I don't know, about six months ago, a guy named Ajay Jain. He runs a company called Genmo, G-E-N-M-O. And Genmo is in the business of taking text prompts and rendering them into video.

And so I think Genmo might stand for like generative motion, something like that.

I'd buy it. Yeah. That sounds right.

And I've kind of been guessing that on the spot, actually. He's never told me that, but it feels right.

We'll pretend. We'll just make it reality. Like, that's what it is.

Easy to remember now.

Yes.

And one of the really cool things that Genmo does as a stepping stone to that probably not too distant reality where you can be prompting whole scenes—they are allowing graphic artists to render 3D objects for films with a text prompt. So you can use their tools to generate an image, which probably lots of people have done with Midjourney or DALL-E 3. Their tooling also allows you to render videos, which now people have seen lots of amazing demos of with Sora. You can actually—you can go to the Genmo website and try, because Sora is not available yet, at least at the time of recording this episode.

Mm-hmm.

But you can go to the Genmo website and try out text to video through them. It isn't, at least again, at the time of recording, as amazing with Genmo as it is with Sora, but you can still get a sense of it. Another cool thing that they allow you to do is upload an image, and then they make it a video from that image. So you could upload an image of a bird, and then through their rendering, it will start to flap its wings and that kind of thing. But the point that I'm getting to is that in addition to those kinds of applications that your listeners might have already heard of, Genmo is also rendering 3D objects so that graphic designers for films or video games have these 3D assets that they can then put very easily into the sets, the digital sets that they're creating. Because just like I am today recording this podcast episode with a green screen behind me, a huge amount of films today are shot on green screens. And so, yeah, there's lots of ways that we are moving in the direction of having, you know, more and more components of films, of television, being able to be rendered easily with text-to-whatever prompting. Yeah. And in that episode, to say one last thing about it, Ajay, he gives specific timelines. I can't remember, but it was something he thought that inside two or three years, we could have whole feature-length films being created with text prompts alone.

I fully agree. I don't see why you wouldn't be able to break stuff down by scenes. I mean, anyone who's done video production where you get storyboards and see—I mean, the camera lens size is picked. Like, every scene is meticulously detailed out already in movie production. I got to talk with one of the co-founders of Runway. Have you had Anastasis on your show?

No. I would love to though.

Yeah. We'll make the introduction for sure. Super cool guy. But we were talking about, you know, the ethics side of things because people will misuse their generation engines and how they handle that and all of that. But it was a good episode, and they're really, really smart guys. Their cool example was The Late Show has this kind of cartoon, and it was several hours of editing work to make happen, like on a late-night show, a little segment they had, and they cut it down to like five minutes. So I imagine they're generating these cartoons now. You can train it on years of The Late Show cartoons and then give it prompts, and now they don't have to have animators sitting there for hours and hours and hours. They can get it like 80%, 90% done with a prompt.

Yeah. There was a paper that is in preprint right now at the time of recording this episode showing that creative workers' wages, rather—so kind of hourly workers working on creative things like those animations that you're describing—since the release of the original ChatGPT in fall 2022, there has been a decrease in wages for creative workers at the same time that, of course, we've seen wages largely exploding in an inflationary era. So that is one of the first noticeable in macroeconomic data impacts of AI on the value of human work.

Was that released by like a federal agency or just an individual?

It's university research.

Oh, cool.

And yeah, it was in a preprint archive. So in computer science and machine learning, there's an online journal called arXiv, A-R-X-I-V. And wherever preprint archive this one was in, it was an archive, which surprised me, but it was something—it was like social sciences something. It was like SSRN or something was the name of the journal. I can't remember off the top of my head.

Yeah. Absolutely. No. That is super interesting because this technology is gonna eat the world. Now AGI, is it here? Can we successfully say that artificial general intelligence is here? It's smarter than people that I know. So—

Definitively, artificial general intelligence is not here. So there's a really great paper from Google DeepMind on levels of artificial general intelligence. And I am quickly looking this up because I know I have it at my fingertips. If you check out my episode number 748, so you can go to superdatascience.com/748. In that episode, I review this levels of AGI paper from Google DeepMind.

(Jon Krohn at 00:17:34) And so what this paper does is it breaks down into different categories exactly what AGI would mean. So the idea here is for narrow applications—for example, like the way that GPT-4 or Claude 3 or Gemini Ultra from Google, the way that those algorithms can now compose messages, can compose text at the level of a human expert writer—that is not AGI. That is ANI. That is artificial narrow intelligence. That is human expert level or superhuman level on a narrow area, on a narrow piece of intelligence.

(Jon Krohn at 00:18:29) That same algorithm, it is not expert AGI because while it can generate text, it can't drive your car or pick stocks for you.

(Joel Beasley at 00:18:42) All right. Can I kind of push back and have some fun? Sure. First of all, I'd like to put a big asterisk on this is not my area of expertise.

(Joel Beasley at 00:18:49) My area of expertise was scaling business logic, like financial planning software and things like that. I just vacation here and have fun here. But as an industry and just kind of being on the outer fringes of it, like always being an engineer but not in that specialty, it seemed to me like there was a consensus that AGI was passing the Turing test. And then once we drove past that goalpost, right, then it's like, oh, we're going to get way more granular. Is that a fair assessment of it?

(Joel Beasley at 00:19:26) Because that's what it feels like happened.

(Jon Krohn at 00:19:29) There may have been people out there saying that passing the Turing test is AGI. I haven't myself. I don't know if I know anybody personally who has used that as a benchmark. There has not been a great definition of AGI historically, which is part of why this paper out of Google DeepMind is so important. So this one was published in November 2023.

(Jon Krohn at 00:19:57) The first author is Meredith Morris, so it's easy to find for your show notes. And what they've done in this paper is they have broken down AGI into different levels. And so I'm looking at the key table in this paper, table one right now. And so they define tools like GPT-4 as being emerging level one AGI, which is equal to or somewhat better than an unskilled human at a broad, broad range of tasks. The kind of thing that you're thinking of where there's an AGI system that is better than most people—they go and they define that. That would be something that's a level three AGI, so where it's at least the ninetieth percentile of skilled adults at a broad range of cognitive abilities, like learning new skills. So that would be when you're thinking about an AGI system that is something that is potentially dangerous or that is able to replace you in a lot of scenarios—it's hitting that expert AGI level, which we may be just a few years away from.

(Joel Beasley at 00:21:10) So that's just my lack of experience in the space. So my brain had just made those connections. My neural network, my custom organic neural network, made those connections between the Turing test and AGI. And so I like this more detailed resolution of here's the different levels, and I will go check that out because I'm personally kind of interested in it. But to one of the things you said, the artificial narrow intelligence—isn't that how humans work? Humans are narrowly intelligent in a specific area. Right?

(Jon Krohn at 00:21:49) Yes. So humans have to have a bit of both. Well, you don't have to have a bit of both. In order to—I mean, there's all kinds of edge cases. Right?

(Jon Krohn at 00:22:03) Like there could be people with particular disabilities where this kind of statement isn't true. But generally speaking, if you're going to be an adult in the world and you're going to be making your own economic decisions around buying an apartment, getting a driver's license, getting a job, holding that job—there's a very wide range of skills required in that. And typically what we see today is then people will often—many people will go to university or maybe a postgrad program—to specialize narrowly in some particular task. And so I think that's what you're referring to, is that a lot of people develop a narrow expertise.

(Jon Krohn at 00:22:53) Yeah?

(Joel Beasley at 00:22:54) Mm-hmm. But that's all that's needed to replace from a job perspective is a narrow expertise. Right?

(Jon Krohn at 00:23:02) From a task perspective. So if you think about, for example, there's a famous quote from Jeff Hinton, who is often called the godfather of deep learning. About a decade ago, Jeff Hinton said there's no point in training radiologists anymore because we already today—that was like 2012—we have machine vision systems that can exceed human capability of detecting tumors.

(Jon Krohn at 00:23:30) And while that now, twelve years later in 2024, is more sure than ever—we absolutely have tons of machine vision systems that are great at detecting tumors at or better than a radiologist could—it turns out that a radiologist's job, only about 10% of their time is actually spent reading radiological scans. They spend two-thirds of their time on the phone with insurance companies. They spend another 20% of their time managing a team. They spend another 10% of their time—I might be running out of percentages, but you're roughly getting the deal. They spend another 10% of their time talking to patients directly and consoling families. So while the radiologists spent twelve years or whatever after their undergrad specializing to become a radiologist, that narrow expertise was still only a small part of what they do just in their job, forget about in the rest of their life. And so this paper has some really great examples of how, in terms of narrow capabilities, there are things like being able to predict protein structure. So proteins in your body, which do almost all of the physical work of chemical activity in your body—proteins make up your eyeballs and your liver and skin and your brain cells.

(Jon Krohn at 00:25:10) And so there's lots of different ways that your genes can express proteins that do all different kinds of activities throughout your body. And these proteins, they're defined by a two-dimensional structure. So you can think of like just a single straight line of information. But in the real world, that single straight line of information called amino acids folds into a protein structure that's three-dimensional. And in some ways, it's actually four-dimensional because it can do work, but I won't get into the 4D. Let's just stick with a 3D shape. So this two-dimensional object—well, one-dimensional, because it's just a sequence. A sequence of amino acids, one dimension, turns into a three-dimensional object that can do work in your body. There's an algorithm from DeepMind called AlphaFold.

(Joel Beasley at 00:25:58) Mm-hmm.

(Jon Krohn at 00:25:58) Yeah, you know it. Which is an example of a narrow AI. And that is what they define in their levels of AGI paper at level five, superhuman narrow AI, because it outperforms 100% of humans at that task. And so there are increasingly examples for very narrow tasks of AI systems being able to do things that you can't even imagine as a human being able to do. And in the immediate term, the key for us as humans is to be able to leverage as many of those tools in our work or in our home life in order to have kind of superhuman abilities ourselves. And so I don't know. I've kind of diverged for a really long time here, and I'm conscious that I've been talking way too much, Joel.

(Jon Krohn at 00:27:01) But maybe this is kind of hitting on some of your question.

(Joel Beasley at 00:27:05) No. But that self-awareness is what I love after almost a thousand episodes, man. I was talking with somebody in the comments of something somewhere, and I'm curious what your thoughts are on this. They were describing to me—explaining to me—how machine learning essentially elicits a response. Like it's trying to figure out the next token or guess the next thing to say. Right? And that was their basis for how it's not human intelligence. It doesn't know what it's doing. It's just trying to figure out the next right move. My response to that is, is that not exactly what humans do?

(Joel Beasley at 00:27:40) I mean, that's what we do. We're observing a situation. We figure out what's the next thing to say. What's the next way to move this forward? And so is their description of the machine learning algorithm operating no different than me trying to describe the proteins and neurons firing in my brain and saying that's not intelligence. That's just these two neurons firing and these protein structures interacting.

(Jon Krohn at 00:28:05) Yeah. So just to really quick—I'm going to nitpick almost unnecessarily really quickly first—is that you described this as machine learning predicts the next token. And so it's specifically generative AI systems. Some generative AI systems are specialized, and all the most popular systems today are specialized in predicting that next word. Machine learning can mean something a lot broader, you know, like having a machine vision system that's recognizing whether a picture is a cat or a dog. It's fundamentally doing something very different from just predicting the next word. So I'm super nitpicking. I know that.

(Joel Beasley at 00:28:48) That's what we need. I'm learning. I'm learning. No, dude. Feel free. This is a conversation.

(Joel Beasley at 00:28:52) I love it.

(Jon Krohn at 00:28:53) So within this subpart of machine learning, generative AI, where we are predicting the next word, that is similar to the kind of intelligence that we use the most that you're describing there, Joel, which—there's a classic book on the way that we think called Thinking, Fast and Slow.

(Joel Beasley at 00:29:15) Oh, yeah.

(Jon Krohn at 00:29:16) By Daniel Kahneman. You know that?

(Joel Beasley at 00:29:16) Yeah. Yeah.

(Jon Krohn at 00:29:17) Great book and highly recommended to people to help you understand what's going on in your own head better based on decades of research that Daniel Kahneman did with Amos Tversky in particular, who was a huge collaborator for him for many decades. And Thinking, Fast and Slow, the book name describes these two main thinking systems that we have in our head. We have system one, which is thinking fast, which is what you're describing there, where you're kind of just—your brain is this kind of ongoing narrative. And it's just like guessing the next word internally or externally. You're just kind of next word prediction. And that's exactly what you're describing, Joel. And I totally agree with you on that. In addition to that, we have our slow system.

(Jon Krohn at 00:30:06) So when you sit down with a piece of paper and you work out your financial plan for, you know, that house that you want to buy, you know, mortgage rates, when you're working on a geometry problem—you're not using the thinking fast system. You're using system two, which is the thinking slow system. And generative AI systems cannot recapitulate that thinking slow. Well, we can, to some extent, prompt the algorithm to be able to go a bit in that direction. So you can do this when you're using ChatGPT. You can say, do it step by step. And then it's kind of going in the direction of thinking slowly.

(Jon Krohn at 00:30:55) It's allowing for some opportunity to have one pass of thinking and then going back. But that's the thing. It doesn't actually go back. The production live generative AI systems that you use today cannot go back and reflect and think about at this step, did I make the best decision and that kind of thing. But what you can do is you can check out some new AI systems that are coming out. So there's AlphaGeometry, which is actually publicly available, open source.

(Jon Krohn at 00:31:35) You can just Google AlphaGeometry. That is a paper out of Google DeepMind, where you can check out my episode number 756 of my podcast in which I describe this AlphaGeometry model. There's also perhaps even more excitingly this algorithm called Q* out of OpenAI. And that Q* algorithm is tied actually to rumors of why Sam Altman was booted from OpenAI because Q* is supposed to be so powerful a model that internal safety researchers were concerned. Anyway, we don't actually know whether that was involved in Sam Altman being set up, but it's part of the rumors.

(Jon Krohn at 00:32:20) Anyway, both of these algorithms, Q*, which you can hear about in my SuperDataScience episode number 740, or AlphaGeometry, which you can hear about in episode number 756—both of those algorithms are trying to leverage thinking that is more like the slow thinking that we do when we're working through math problems. So we're thinking about what mortgage to go with or planning our finances. So those kinds of AI systems are coming, and that will be a big step in the direction of kind of expert level AGI that we were talking about earlier on in the episode.

(Joel Beasley at 00:32:59) Help me understand this concept that they can't go back and reflect, because the other day I had three different spreadsheets that had guest information over the years, right, in production, and I wanted to combine them. And so I wrote out all the details of the project to post on Upwork, but instead, I just put it into GPT and uploaded the three sheets. And it took like two or three minutes, but it said, oh, I'm trying this. Hold on a second. That didn't work.

(Joel Beasley at 00:33:28) Now I'm trying this. Hold on a second. That worked. Oh, I learned this. Now I'm trying. And it did that probably four or five times, and then it ultimately spit out the perfect CSV that was everything I wanted, and it handled stuff. And then it felt like it was going back and thinking. Is that not what it was doing?

(Jon Krohn at 00:33:48) Yeah. So, yeah, again, that is an example of a step in that direction. So in that case, what happened there is you were using the GPT-4 built-in code interpreter that ChatGPT has, also known more recently—they call it advanced data analysis. But basically, it allows you to upload a file, and the code will be executed. Python code will be rendered and executed. And then you're right. There is a loop kind of built in there where if an error comes up or if an unsatisfactory answer seems to come about, the system does have a built-in loop to go back and check and see what's going on. And so while that system is still just generating either natural language or Python code word by word, it does have a built-in element of looking back over what it's done, particularly when errors come up. So you're absolutely right.

(Jon Krohn at 00:34:51) It is a step in that direction. These kinds of algorithms like AlphaGeometry and Q*, they had it built in more directly as opposed to just having text or code generate and then look back over that. It's a much more advanced process of, for example, with solving a math problem instead of going all the way through solving—so the equivalent of what you're describing there is to try to solve the math problem. The generative AI system of today that you can access would spit out a full answer.

(Jon Krohn at 00:35:30) And then after getting all the way to the answer, it would look back over from the beginning. With these systems, like AlphaGeometry and Q*, they are—every step of the way—they're reconsidering. So they try one tiny step towards solving the problem and then see how's that been going, and then take another tiny step and see how that's been going and try lots of different permutations. So it's very, very expensive, very computationally expensive for this kind of detailed reflection. But just in the same way, it is cognitively expensive for a human to be thinking carefully about everything that they're doing when they're solving a math problem or trying to decide on their mortgage.

(Joel Beasley at 00:36:14) And I want to be respectful of your time today. So we've got a couple minutes left to wrap up. What's the one—so our audience again is technology leaders or people that are interested in technology trying to improve in their career. They learn leadership skills. They learn scaling teams.

(Joel Beasley at 00:36:30) They learn all of this stuff. They always need to be aware of the latest technology and all that. Stay relevant, stay in the loop. Knowing that that's our audience, people are trying to grow and improve, what is the best advice that you have for them?

(Jon Krohn at 00:36:46) You've got to be trying these tools on a regular basis. This is such a fast-moving space that I highly recommend—you know, if they're listening to the Modern CTO podcast, podcast is probably a format that they like. It seems like your show is a great place to be keeping abreast, given the kinds of insightful questions that you had today, Joel. And from the episode topics that I've seen on your show in the past, it seems clear that this is one of the ways to be staying up to date on what's happening in this rapidly moving AI space. Another podcast that I can highly recommend is called Last Week in AI.

(Jon Krohn at 00:37:19) The hosts of that show are so funny. It's a weekly show. They do a great job of summarizing all of the week's research and not just research, actually—news announcements. So you can really stay up to date. Because it's a new show, you get a lot of superficial depth on the gamut of possible business-relevant AI topics.

(Joel Beasley at 00:37:49) Oh, nice.

(Jon Krohn at 00:37:49) Yeah, that's definitely a show that I recommend.

(Joel Beasley at 00:37:51) Have you been on that show?

(Jon Krohn at 00:37:53) I have. I've co-hosted that show a couple of times. Yeah, just as kind of a guest. They have two co-hosts, Andre and Jeremy, and sometimes when Andre is out, I'll sub in for him on the show. That's a really great show that I highly recommend.

(Jon Krohn at 00:38:11) And then, yeah, if you want—obviously, this is a shameless plug, but our Super Data Science podcast, we go in detail on specific topics. So we have two episodes a week, every Tuesday, every Friday. The Tuesday episodes are typically at least an hour long. We get experts from diverse backgrounds. So some of them are experts at implementing things commercially. Some of them are investors in AI. Many of them actually are people right at the forefront of machine learning research at top universities or top tech companies where they have great AI research labs. And so we go deep on specific topics in those Tuesday episodes with guests. And then on Fridays, those are typically shorter episodes where I'll pick often what is the single biggest news story recently or thing that I think people need to be aware of. And I will do a five- to fifteen-minute episode on that particular topic.

(Jon Krohn at 00:39:13) So I think Super Data Science is another way to stay up. But the point is, whatever poison you pick—and there's all kinds of newsletters out there, Last Week in AI has a newsletter as well if you prefer that over the podcast format—but the key thing is that you need to have some way of staying up to date. Don't drive yourself crazy. Don't try to have a subscription to every newsletter, and don't be obsessed with scrolling through social media.

(Jon Krohn at 00:39:35) Because if you do that, even though it feels like you're keeping up, what you're actually doing is just seeing a lot of tweets or a lot of LinkedIn posts about the same story of the week. Sora, for example. You're not getting a lot of depth. And so I'd say pick one or two resources that keep you updated on the news and stick with those so you can get a bit of depth on each of these key topics, and then experiment with things. So at the time of recording, Claude 3 just came out, or Gemini 1.5 Pro just came out. Typically, there's a free tier to be trying these things out, or it's like $20 for a month to try something out with no cancellation fees.

(Jon Krohn at 00:40:22) So just try these tools out. You've got to be trying these tools out. As a CTO, you've got to set time aside for that, because there are things—particularly as a CTO—things like GitHub Copilot. This is something that can be hugely transformative for the productivity of your organization, in-the-flow tools like that that are suggesting how your software developers, your data scientists can be improving their code or collaborating with each other on their code.

(Jon Krohn at 00:40:49) That kind of thing can be hugely beneficial to you and your organization. And so, yeah, you've got to experiment with these things yourself. So I guess step one, have high-quality resources for staying up to date—just a few of them. Step two, actually set aside time to try out some of these tools.

(Jon Krohn at 00:41:09) And then, I don't know, step three, you've got to try to find a way to implement some of the ones that come across as clear wins.

(Joel Beasley at 00:41:16) I say step three, we're going to tell people about Modern CTO podcast.

(Jon Krohn at 00:41:22) I think that's still step one for sure. Modern CTO podcast is part of step one. Definitely.

(Joel Beasley at 00:41:28) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you'd like to hear discussed on the podcast, either add me on LinkedIn or send me an email: [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.