SDS 799: AGI Could Be Near: Dystopian and Utopian Implications, with Dr. Andrey Kurenkov

This is episode number 799 with Dr. Andrey Kurenkov, machine learning scientist at Astrocade. Today’s episode is brought to you by AWS Cloud Computing Services.

00:00:15

Welcome to the Super Data Science Podcast, the most listened to podcast in the data science industry. Each week we bring you inspiring people and ideas to help you build a successful career in data science. I’m your host, Jon Krohn. Thanks for joining me today. And now let’s make the complex simple.

00:00:45

Welcome back to the Super Data Science Podcast. It’s my great joy to be joined in this episode by Dr. Andrey Kurenkov. Andrey founded and co-hosts my favorite podcast, Last Week in AI, which is a weekly program that recaps all of the AI related news you need to know about. He’s also a ML scientist at Astrocade, an NVIDIA backed generative AI platform that converts your natural language prompt into a functional video game. He holds a PhD in computer science from Stanford with research that focused on robotics and reinforcement learning. Today’s episode should be of interest to just about anyone in it. Andrey details the genesis of the wide range of AI publications and podcasts he’s founded. 

00:01:26

What the future of text to video game generative AI could look like. Why AI as a product rarely works commercially, but what you could succeed at with AI instead. Why AI robotics is suddenly progressing so rapidly and in one of my favorite on-air conversations ever, how soon AGI could be realized and the potentially dystopian or utopian implications. All right. You ready for this outstanding episode? Let’s go.

00:01:58

Andrey, welcome to the Super Data Science Podcast. It’s awesome to have you here on the show. We’re in person. 

Andrey Kurenkov: 00:02:04

Filming in person from San Francisco. It’s a fun day to be here. Pride is going on. 

Jon Krohn: 00:02:10

It is. 

Andrey Kurenkov: 00:02:12

There you go. It’s nice to be on the show with you having been on our show a couple times. 

Jon Krohn: 00:02:17

Yeah. Last Week in AI, the podcast that you host, is my favorite podcast to listen to. Is invaluable for me to keep up to date on all of the latest news. Something that I think I’ve said the couple of times that I’ve been a co-host on your show is that prior to discovering Last Week in AI, I used to assume that it was impossible to keep up with all of the things going on every week, and so people would say to me, oh, you haven’t heard about QLoRA or whatever, some specific thing? You haven’t heard about RedPajama? And I’m like, you can’t possibly keep up to date on all of these things going on. But now since listening to Last Week in AI and I listen religiously, I never miss an episode, I pretty much always listen to the whole thing. It puts me in a position where all of these terms that come up, I’ve never been caught off guard. They’re coming on a year now, maybe 18 months, listening to it regularly. 

Andrey Kurenkov: 00:03:10

Very nice. Yeah. We have to go usually two hours just to cover a week’s worth of stories and we try to cover a variety like business research, things going on with generative AI a lot of the time, things you can use, all this stuff. We do our best to try to cover and make it so you can keep up. Probably we don’t quite succeed, but we do try. 

Jon Krohn: 00:03:38

Yeah. You have specific news sections. Do you know those offhand? Are you able to reel them off? 

Andrey Kurenkov: 00:03:44

Yeah, we have tools and applications first, which is things like news with ChatGPT or things that normal people can use to interact with [inaudible 00:03:56]. 

Jon Krohn: 00:03:56

A user interface. 

Andrey Kurenkov: 00:03:57

Then we have business advancements, which is more where we have funding news, news about OpenAI drama, which we have quite a bit of. Then we have open source and something else, I forget the second one where we cover new releases of models. Then research and advancements where we have research and advancements, policy and safety where we have laws and safety advancements, and finally synthetic media and art where we cover art duplications. Yeah. We try to be pretty furrow in the various sections. 

Jon Krohn: 00:04:36

Now, this is probably better suited to conversation off-air, but I’m going to hit you with it anyway on air.

Andrey Kurenkov: 00:04:40

All right. 

Jon Krohn: 00:04:40

Something that I’ve wondered about in the past is have you ever thought about having a headline news section at the top? What if the biggest news of the week is in synthetic media? You’re like, we should put that at the end of the podcast ’cause it’s for that category. But I suspect, and maybe you actually have the data, not everyone listens through the whole episode from beginning to end. You could have a headline news section at the beginning potentially. 

Andrey Kurenkov: 00:05:04

That is an interesting idea that somehow I haven’t thought of. We do that in the text newsletter where we have top news that covers four stories, three stories, and then we have sections. I suppose we don’t do that on a podcast because I assume that people have different interests. 

Jon Krohn: 00:05:25

Right. 

Andrey Kurenkov: 00:05:25

Someone is more interested in research, someone more in business.

Jon Krohn: 00:05:29

Right. 

Andrey Kurenkov: 00:05:29

We cover the main stories for each section, but perhaps that’s a good suggestion. See, you are a big fan. 

Jon Krohn: 00:05:37

You have them bookmarked. You have them timestamped with bookmarks. On, I assume most podcasting apps, I use Apple Podcasts, and with that it’s very easy for me. Every once in a while we’ll get deep into some story where I’m like, I already know this one, or sometimes it happens ’cause I’m catching up. I was recently catching up on an episode where you guys were talking about the Apple Worldwide Developer conference that was coming up soon. 

Andrey Kurenkov: 00:06:06

Yeah. 

Jon Krohn: 00:06:07

And I was listening to it after it had already happened, so you guys were having conjectures about what they might talk about now. I was like, well, it already happened now. 

Andrey Kurenkov: 00:06:14

Skip it. 

Jon Krohn: 00:06:17

When a section that comes up, I can skip it, go to the next section and move on. Yeah. That makes it easy. Just an idea. The way that we were originally introduced was your co-host on Last Week in AI for a couple of years now at least is Jeremie Harris, who has developed into a great friend of mine now. I think he’s one of the most brilliant, articulate, also funny. He’s very funny. 

Andrey Kurenkov: 00:06:46

He’s quite funny. Maybe more than me. 

Jon Krohn: 00:06:49

I don’t know. 

Andrey Kurenkov: 00:06:50

Yeah. 

Jon Krohn: 00:06:50

You’re both very funny. It’s great. I mean, you’re a great pair. You’re both extremely knowledgeable on air. I think it is clear that there’s some topics that one of you is more expert in others. Anytime hardware things come up, somehow Jeremie knows tons about hardware and chips. I guess maybe his physics background lends him to being really interested in that stuff. And then of course policy because his company, Gladstone AI, is a policy company. 

Andrey Kurenkov: 00:07:20

Mm-hmm. 

Jon Krohn: 00:07:20

When those topics come up, of course. 

Andrey Kurenkov: 00:07:23

Yeah, I’ll let him talk for five minutes typically on those sections. But yeah, I believe we met with Jeremie maybe when he co-hosted or we interviewed him. Originally we had co-host with Sharon, a friend from Stanford who also did her PhD when I did that. Then she got too busy with a startup as people do at Stanford, and we invited Jeremie and yeah, it’s been great. Jeremie is a fantastic co-host and he connected us, made it so you could be a co-host, which I think worked out really well. 

Jon Krohn: 00:07:59

Yeah. I can’t remember how I was originally introduced to him. Now Jeremie will have to remind us if he’s listening to this in the future, but somehow I got introduced to him and then he ended up being on the podcast at least twice now that I could find really rapidly when I looked it up in our index of historical podcasts. If people aren’t familiar with him and they haven’t listened to Last Week in AI and you want a Super Data Science episode to just get a taste of Jeremie, he was most recently on in episode number 668, which we got that episode out very shortly after GPT-4 came out last year, and we did a series of GPT-4 episodes. Episode 666 I introduced GPT-4, which I thought was perfect, and then we had an episode 667, which was on commercial opportunities that are arising out of it. And then 668, Jeremie came on to talk about the potential risks associated with GPT-4, which of course, yeah, that is one of his specialties. In terms of your specialties, I guess we’re going to end up talking about that a lot over this entire episode. Maybe we don’t need to dig into that right now. That is what we’re going to end up talking about this entire show. While you were at Stanford finishing up your PhD, which I guess you finished about a year ago now. 

Andrey Kurenkov: 00:09:15

That’s right. 

Jon Krohn: 00:09:17

That PhD was in computer science was the major, and we’re going to talk, we have a section coming up later that I have planned or researcher Serg Masis had planned on specific papers that you wrote then. Robotics focused, Reinforcement Learning focused. While you were doing that Stanford PhD, you created a number of publications and publishing entities. Tell us about things like we already just talked about Last Week in AI, which in addition, you even mentioned this. In addition to the podcast, there’s also the newsletter, and the newsletter predates the podcast, right? In the numbering, I think it’s weekly numbered, and the Last Week in AI newsletter is in the 200 somethings. 

Andrey Kurenkov: 00:10:00

Yeah. 

Jon Krohn: 00:10:01

It’s been around for four or so years, and the podcast is in the hundreds. It’s been around for over two years. 

Andrey Kurenkov: 00:10:10

We started at the very beginning of COVID, March of 2020.

Jon Krohn: 00:10:14

That’s only four years. 

Andrey Kurenkov: 00:10:15

Yeah. 

Jon Krohn: 00:10:16

Yeah, yeah, yeah. And the Last Week in AI podcast is intended, as you already alluded to, to a broader audience. It doesn’t assume Python coding knowledge. 

Andrey Kurenkov: 00:10:30

Right. 

Jon Krohn: 00:10:32

That’s a great episode. I know we have a lot of listeners on this show that come from any background. They may not be hands-on practitioners of data science. The Last Week in AI is a great show for that, but you also have the Gradient, which is older than Last Week in AI, right? And that’s also a publishing platform as well as a podcast. And that was intended more for hands-off practitioners, right? 

Andrey Kurenkov: 00:10:53

Yep. Yeah, it’s I guess in some ways a weird story. I started, even before Stanford, I had a personal blog and I did various writings. When I wanted to learn about deep learning, I wrote this 10,000 word blog post called Brief History of Neural Nets and Deep Learning. When I did get into Stanford, The Gradient came about in an interesting way. It was actually being done by a team of undergrads and it was done I think as a class project or something to try and democratize understanding of AI. I originally got involved as someone who wrote, I was invited to write a piece, ’cause as you say, it’s a magazine. And the team running it was more of a team of editors and publishers, but usually not the authors. And so I wrote a story pre the launch of The Gradient, and then it took a while for them to actually get launch to go public, and I had this article that I was waiting for them to release. 

00:12:05

And so after a while I just got involved. I was like, let me join the team so we can get this thing launched. I was among a pretty large team originally, 10 people who got it public. That’s one of them. And that launched I believe May of 2018. And then before that in 2017, this was a period in AI that is interesting. It wasn’t as big as it is now, but you did have increasing interest in awareness where OpenAI had been around and they had some splashy stories. We got AlphaGo, it was starting to be more mainstream. And at some point I remember I got annoyed with some things being a sensationalist. I think there was an example of Kissinger saying that AlphaGo is a sign of AGI coming or something, which from a technical perspective didn’t make a lot of sense. 

00:13:08

And so I wanted to start something where we could basically dehype AI that had this broader appeal where we explained to you this is the story, but maybe the mainstream media doesn’t quite get it from a technical perspective. We will write little articles. And so I started work on it in 2017 and somehow it just worked out that that got ready to launch with a few stories in April of 2018. We are one month apart. And then Last Week in AI was an offshoot of Skynet Today we had three categories of editorials or reviews and digest, where digest were these summaries of news. And so for quite a while I was involved in both. I was an editor that helped people from the broader AI community write articles for The Gradient. I was also part of a team that launched the Stanford AI Lab blog, I just remembered, in 2019 or something. I don’t know why I really liked doing this, but I guess I did. And so that was for blog posts that cover research papers, but mainly I was involved in The Gradient, which was slightly more technical and intended for conversations within the AI community. And then we had Skynet Today, which was explaining topics within AI and recent news stories. And yeah, that’s a background of those two. 

Jon Krohn: 00:14:46

Yeah. And so you just threw in Skynet Today there. Skynet Today, that’s the entity that publishes Last Week in AI? 

Andrey Kurenkov: 00:14:53

Yep. 

Jon Krohn: 00:14:54

Yeah. 

Andrey Kurenkov: 00:14:55

Yeah. Last Week in AI, I think originally we called it something else, but pretty soon after we published a website we started, so that’s why it’s in 275 or something. We launched the initial one in May 2018. Then for a while it was biweekly and it took us a while to figure out how to do it without it being a ton of work, but it’s been running for years, I guess, what, six years now? That’s crazy. But for a while we did everything. We published articles and we had the skynettoday.com as the main website with all of us. It turns out doing all this is a lot. At some point we wound down all the articles of day. We launched Last Week in AI as a separate entity, newsletter, and then podcast. And that’s currently… Skynet Today is you could say, not the main, there’s pretty much only Last Week in AI left of it. 

Jon Krohn: 00:16:05

Gotcha, gotcha, gotcha. Yeah. But it is interesting that still today when you go to the Last Week in AI newsletter page, which is a substack, anyone can sign up for that for free. There’s something like 30, 40,000 subscribers to the substack. And then so when you go to that website, it says brought to you by Skynet Today. And then same thing when you introduce the podcast every week, you say Last Week in AI brought you by Skynet Today. 

Andrey Kurenkov: 00:16:35

Yeah. I guess that’s the legal entity that we have since 2018 is Skynet Today. Maybe that’s why I do it or just out of habit. Although the website itself Skynet today mainly functions to lead directly to Last Week in AI these days.

Jon Krohn: 00:16:52

Are you stuck between optimizing latency and lowering your inference costs as you build your generative AI applications? Find out why more ML developers are moving toward AWS Tranium and Inferent to build and serve their large language models. You can save up to 50% on training costs with AWS Tranium chips and up to 40% on inference costs with AWS Inferentia chips. Tranium and Inferentia will help you achieve higher performance, lower costs, and be more sustainable. Check out the links in the show notes to learn more. All right. Now back to our show. 

00:17:27

And I guess that Skynet Today name is that in and of itself is a hat tip to the way that you are trying to dehype because you’re saying because Skynet is from Terminator two, I guess even Terminator one, I don’t know what happens in Terminator one. No one ever talks about Terminator one. Have I even seen Terminator one? Is Terminator two the first movie? Skynet is the system that causes World War everywhere, nuclear weapons being launched and machines do that. And so it’s the end of the world caused by Skynet. And so it’s a joke, Skynet Today, is that the world is ending, AI’s taking over today. 

Andrey Kurenkov: 00:18:09

Exactly. It was pretty much a joke and I was not too sure about the title ’cause it did seem like maybe people would mistake it for the opposite, but I wanted a title that connected to something like a newspaper and this format of something today is often associated with this journalism. Skynet Today just sounded fun. 

Jon Krohn: 00:18:38

Yeah. 

Andrey Kurenkov: 00:18:38

And so that was the initial thought I had that popped into my head in 2017. Went ahead with this idea and it just wound up sticking. 

Jon Krohn: 00:18:49

Mm-hmm, mm-hmm. And it’s doing incredibly. Last Week in AI has a lot of listenership. I know that because we went for a period of six months running Super Data Science podcast messages. Thank you for doing those sponsored messages on the show. I don’t know exactly what the numbers are, but something like you get 20,000 listeners an episode every week.

Andrey Kurenkov: 00:19:15

Yeah. About that. It varies. And it’s interesting looking at the history of it because for the first two and a half years, basically 2020 through the end of 2022, we were doing it and we were getting some traction, but it was one or 2,000 listeners and then ChatGPT happened and Sharon moved on, Jeremie came on. And so the start of 2023, without any advertising or attempts to grow, it just shot up because I guess now many more people are aware and need to keep up with AI. And it’s interesting because it’s always been a bit of a side project. It’s not a real official or corporate or media podcast, I guess you could say. It’s pretty much just me and Jeremie. I still edit every episode by myself. Yeah. It’s been super fun to do it and I guess it is very nice now to have a lot of listeners. 

Jon Krohn: 00:20:22

Yeah. You’re in the top percent or certainly you’re in the 98th percentile at least of podcasts globally, which is amazing. And tons of listeners, but I think part of the charm, and maybe it’s part of even why I love the show so much, it does have that raw feeling that this is two guys chatting about AI, not taking things too seriously. And I think that’s part of the charm of the show. It’s amazing because there’s shows out there with a quarter of the weekly audience that you have that they end up getting hoovered up by professional production companies. And so yeah, it’s great that you guys keep that grounded feel and it’s always a lot of laughs. I always enjoy listening so much as I’m learning. I’ve recently been developing a TV show idea and I recently heard this idea from a production company. They said, you’ve got to hide the vegetables, which is this idea that people may be getting a nutritious meal by listening to your show, but because it’s so humorous in the way you guys deliver it, it feels like you’re having junk food. The vegetables are hidden. 

Andrey Kurenkov: 00:21:32

Well, I guess for two hour long episodes, I guess you have to make it a little entertaining. 

Jon Krohn: 00:21:37

Yeah. 

Andrey Kurenkov: 00:21:40

‘Cause it’s a lot to take in. We cover usually 30 news stories per week, which is a lot, but we do have people who listen all the way through. And again, that’s quite gratifying. But yeah, that’s cool. And then since we have been doing this whole long conversation about it, it sounds promotional, this was not a paid thing, I guess it’s hopefully interesting to the listeners. 

Jon Krohn: 00:22:09

I mean, if people are listening to this data science podcast, I figure because these things like Last Week in AI, which is still running for people who are interested in, I mean not just still running, I mean Last Week in AI is thriving and growing right now. It’s a big focus of your life and Jeremie’s life is creating this show every week. And so that should appeal to people who listen to this show, who are interested in just an overview like I am of all the AI news every week. And then simultaneously The Gradient which you started and you hosted for a couple of years, I guess. You no longer host The Gradient, but The Gradient has had incredible guests. There’s quite a few guests that have been on The Gradient and I’m like, how do I get that person on Super Data Science? Incredible guests on The Gradient. And so people want that interview show and technical, that’s something else that you can be listening to today. Yes. It wasn’t intended as a, yeah, certainly no money has changed hands to make this episode. It’s not sponsored at all, but I assume that’s something of interest. 

Andrey Kurenkov: 00:23:06

Yeah. If I can do a quick humble brag, not so humble brag, brag I suppose. When I was doing interviews for Gradient, I guess the highlight was interviewing Yann LeCun, which I just messaged him on Twitter and it worked out somehow. But yeah, no, it’s a lot of fun. We’ve been doing it for I guess four and a half years and there’s a reason for that. It’s partially just been very educational for me and hopefully also for listeners. 

Jon Krohn: 00:23:37

Yeah, no doubt. All right. That’s enough for the promotion of the podcast of your shows, so let’s move on to what you’re doing now. You, in terms of your full-time job, which is you’re a machine learning scientist at Astrocade AI. Astrocade is an NVIDIA backed startup that leverages generative AI. I can’t believe startups are getting into that. NVIDIA backed startup that leverages gen AI to allow anyone to create games with natural language prompts regardless of technical skill. And I’ve got to say, prior to us having done any research on you and preparing for this episode, I imagined in my mind that what Astrocade was doing was assisting video game developers in creating assets for their video games. And so it was interesting for me to learn that what in fact Astrocade sets out to do is to allow somebody who has no technical skill whatsoever to, with a natural language prompt, create a full working video game, potentially not just for an individual, but for multiple players as well. Is that right?

Andrey Kurenkov: 00:24:51

That’s right, yeah. We are working on an app for your phone for just regular users and in particular the idea of kids and younger people to make little games and publish them. Really a platform for creation of games and sharing of games and socializing and making friends within games. That’s the goal. I’ve been there for over a year now, since I started, before finishing my PhD actually. And it does have a connection to Stanford. It started by a former lab mate, Amir Sadeghian, who was in my lab, graduated a couple of years before, and then worked in another startup as their lead of AI, where my advisor, Sylvia Subrese, was a major, I think, believe co-founder. And so in a way, it’s a little bit weird because my PhD was focused on robotics, not generative AI, certainly not games or anything like that. But because I’ve done all these side projects, which we’ve covered, I’ve written blog posts, I do podcasts, newsletters, I even posted YouTube videos, photography, a lot of stuff. The idea of enabling people to be creative with AI is something I really wanted at a startup.

Jon Krohn: 00:26:19

You also play video games, right? 

Andrey Kurenkov: 00:26:20

I do, yeah. Haven’t tried to develop them as much, but that’s partially because that requires coding and a lot of specialized skill. Yeah, near the end of my PhD, I wanted to join a startup rather than a big company to be able to have impact and really care about what I’m doing rather than just optimizing some tiny feature of some product. And the idea of having something that regular people use, not businesses, although that’s very, maybe ill-advised at [inaudible 00:26:55] startup because typically doing B2B, doing something for businesses is mostly what succeeds. Y Combinator, I believe, only wants things for businesses, not for consumers. But I like the mission, I like the people involved, so I joined and have been working there for a while. 

Jon Krohn: 00:27:17

Yeah, I think with B2B, you get big contract sizes. You’re not totally dependent on just network effects. You can potentially, in the beginning, just have a couple of big clients and that be enough to get you from a series A to a series B kind of thing. So yeah, you do have that B2B thing there, but B2C, if you can pull it off, it ends up being a really cool thing because it’s something that any of your friends or family, people that you’re lecturing, to your podcast listeners, they can all be signing up and checking out Astrocade.

Andrey Kurenkov: 00:27:48

That’s right.

Jon Krohn: 00:27:49

That’s something cool. It’s interesting. I also didn’t know until, even through our research, I wasn’t aware that it was a mobile app that you’re developing, so that’s interesting. 

Andrey Kurenkov: 00:27:58

Yeah, part of that is because we are still at the stage of not being super public. We are in, let’s say pre-alpha, where it turns out this takes a lot of work. It turns out, compared to a lot of startups where you train an AI model and then you just give someone a website to write a prompt and get the output, games are hard. Generative AI isn’t today sufficient to just output things. You need to really think about the user experience, the system that enables you to use AI to make a video game. And so, yeah, even though we’ve been working on it for a while, it’s still kind of a bit closed off, and we are hoping by the end of the year to have a wider launch, maybe alpha or beta, for more people to be aware of it. 

Jon Krohn: 00:28:50

Did you know that the number one thing hiring managers look at are the projects you’ve completed? That’s why building a strong portfolio in machine learning and AI is crucial to your success. At SuperDataScience, you’ll learn how to start your portfolio on platforms like Hugging Face and GitHub, filling it with diverse projects. In expert-led live labs, you’ll complete an exciting new project every week. Plus, through community-driven projects, you’ll tackle real world multi-week assignments while working in a team. Get hands-on experience with projects like retail demand forecasting, building an AI model from scratch, deploying your own LLM in the cloud, and many more. Start your 14-day free trial today and build your portfolio with www.superdatascience.com. 

00:29:26

Nice. Yeah, I don’t know if you’re able to give us a sense of how complex a game that you’re targeting being able to render, say in that alpha state or beta state because I imagine you’re kind of expecting over the years to have more and more complex games be possible, both because the technology, the underlying LLM capabilities improve, as well as your development of the product feedback from users. 

00:29:54

So for example, I recently, at the time of us recording, I had just this week recorded an episode. In fact, actually it should be the most recent episode by the time your episode is published, episode number 798, which should be the preceding episode before this one that we’re recording right now. It’s on Claude 3.5 Sonnet, and one of the cool things about that Claude 3.5 Sonnet release from Anthropic is that they coincided that with the release of a new user experience, Artifacts. 

00:30:27

And that is a big shift for me, allowing you to have, you have two panels side by side, where the left-hand panel is that now classic generative AI experience that we’ve had for two years through ChatGPT where you’re having the conversation on the left, back and forth dialogue with the machine. But instead of having the outputs like images, and Claude at this time doesn’t directly output images, but you can end up outputting images anyway by asking for it to generate SVG graphics or something because it can write the code to generate the SVG graphics. So you can get in this right-hand panel, the Artifacts panel, you can get images, you get your code outputs, you get rendered websites that work. 

00:31:13

And so, one of the things that I did in creating the preceding episode, episode number 798, is I recorded, sometimes my Five-Minute Friday episodes have a video component where I’m demoing something. And so, I was messing around myself with how this would work. I had an idea in my head of something that I thought might challenge it, which was to create a simple video game, and I said, “Create an interactive website of the shell game,” which is it let me know in the dialogue part on the left that that’s also sometimes called the ball and cup game. So this is I guess like a classic game. It probably stretches back a millennia of where you have three cups on a table surface, someone puts a ball under one of the three cups, and then you kind of shuffle the cups around on the surface, and you ask the person to guess which cup it’s under.

00:32:10

And it was able to render that game for me. It wasn’t the very first rendering. So in a few seconds, in five, 10 seconds max, it renders JavaScript code for both this interactive front end that works in that Artifacts panel on the right-hand side within the platform. So you don’t have to set up web hosting yourself or anything. You all of a sudden have this very simple cup game, shell game that you can play in the browser. And the very first time, the animations weren’t very compelling. So I said to it, can you make the animations look more like shuffling? And it did. All of a sudden, it looked like shuffling cups, and I could track. 

00:32:51

But then, the third time I said, well, I can’t see the ball going in. It didn’t render the ball. Can you do that? And it did. It wrote code to be able to render a ball going into one of the three cups, and then they shuffled around in a realistic looking way. And if you track it with your eyes, and you click on the right cup, it’ll say you got the right cup. And so, it is interactive video game experience, so that to me does feel very impressive. I was blown away to see those kinds of things happening. And so, I know that that is kind of possible, but I imagine that the kind of experience that you’re creating in Astrocade is more nuanced than the ball and cup game. 

Andrey Kurenkov: 00:33:42

Yeah. I don’t know how detailed I can get, but I can get you a sense, and yeah, it’s interesting you note that because I believe very early on, like a year ago now, either OpenAI demonstrated this. I believe maybe they did, this was one of the things they said is, “Here, take a look. We can develop Breakout or something with just some prompts.” And there are some companies working in that direction. 

Jon Krohn: 00:34:05

Oh, yeah. There was a demo of creating a spaceship moving around and shooting at asteroids or something like that, and it was interactive. It was exactly, I think it was with GPT-4 maybe, that was one of the demos was showing somebody programming an interactive video game that became more and more complex. The user is like, okay, render a spaceship that looks like this, and then have these controls, allow it to move around like this. So you can gradually describe iteratively a game where you’re shooting at asteroids, and you get points for destroying asteroids or whatever. 

Andrey Kurenkov: 00:34:40

That’s right, yeah. So the kind of north star, what we want to achieve at the end of, I don’t know how many years, but as AI gets better is pretty much that, text to game, where very little effort you can make anything. But turns out, that’s a little hard at present. You can make Snake. You can make Breakout. But when you get to this idea of let’s say even a Mario level or even beyond the Mario level, you want something kind of pretty flexible, and let’s say you want multiplayer in it, well, that requires a lot of infrastructure and code beyond what you can get with a single kind of output or even iterative output.

00:35:24

So we are aiming in terms of complexity in this kind of direction of having an avatar, being able to play with your friends, being able to play with everyone and either be competitive or cooperative in relatively small games where you do have your own character, so you can create an avatar with your face and play them. So it’s maybe not the intuitive approach you would immediately guess of just take an LLM and make it do all the work. There’s a lot more that is required at present today, but maybe that’s a good thing in the sense that we are doing a hard thing, I think, and hopefully we get there and people really like and have this experience of doing what you can’t do right now with just an LLM. 

Jon Krohn: 00:36:24

Very cool. Nice. So as a Machine Learning Scientist at Astrocade, what do you do there?

Andrey Kurenkov: 00:36:30

Yeah, I do various things. I guess I can say it’s, as you might expect, very different from what I did during my PhD. I did actual research was a focus. You wanted to publish a paper, that was the goal, to achieve some new knowledge, versus being in a startup, the goal is to make something people like. And it’s much more collaborative with bigger teams, and you play a role in that team. 

00:36:58

So what I do is dependent on the features we work on. Broadly speaking these days, it’s a lot of taking what a person asks for in the game creation process and then understanding what they want, processing this raw text input into things that we can delegate to different models to address. For instance, let’s say you want a pink dragon that shoots fire or that has a bunch of health. We then process that and create a game entity with the right parameters and satisfy as much as we can, what you asked for.

00:37:45

So that’s my role to a significant extent, although it really depends. I have done work also outside of machine learning on the game engine side where the game stuff, fixing bugs. I’ve done some work in the creative product aspects about actual engineering, just thinking through what we should be doing, which is a lot of what you do in a PhD. So that’s the broad answer. Machine learning scientist sounds like what I do all day is train models and read papers, but in practice it’s not quite like that. 

Jon Krohn: 00:38:23

Yeah, I think the reality in a startup is that that’s kind of the dream state. If you’re able to carve out some amount of time every week for that exploration, that’s great, but often there’s a long list of deliverables, especially as you’re trying to find your market and get to market. There’s a lot of pressure to make that happen in short order. And so, yeah, it ends up being commercially oriented. So I don’t know how much you can talk about this, but we’ll see as I ask the question, and anything that is proprietary, we’ll just move on, no problem at all. But do you guys end up leveraging third-party LLMs, or do you end up having to develop your own in-house in order to make this happen? Is that the kind of thing you could talk about? Maybe it is, maybe it isn’t. 

Andrey Kurenkov: 00:39:17

I suppose I can say we do leverage open-source LLMs, open-source and closed-source, and what I’ll say is part of why that is, is things move so fast. Every two months you get new open-source models, you got Llama 3, you got GPT-4.5, Sonnet 3.5. That for a company, for most companies, I think there’s very few companies training new models from scratch. That’s a sort of rat race, and we certainly need to fine tune, but we also do make use of APIs. 

Jon Krohn: 00:40:00

I guess that for me is a given, and it probably isn’t for all of our listeners. But when I say your own, I actually mean a fine-tuned open-source. It would be such a crazy large task for almost anyone. Given the broad functionality of these open-source LLMs and how you can quite inexpensively with low rank adaptation, parameter-efficient, fine-tuning approaches, be able to in some cases with just hundreds of input output examples, of labeled examples, being able to create a really high performing LLM for your particular use case. Yeah, it would be great. It seems like, and you guys talk about this on Last Week in AI all the time, but it seems like a lot of the companies that are creating the next open-source LLM, in a lot of cases it seems to make that headline, to be able to say on this particular benchmark, we’re SOTA for today. 

Andrey Kurenkov: 00:41:04

Exactly. Versus us, we have players. We want to have users. Right now, we are in closed testing, so we don’t have quite as much data, but the hope is once we do launch and get kind of a feedback loop, certainly we’ll do a lot more training than we’re doing right now. It’s kind of funny, we have this trademark, or I think we filed for a trademark, AGI, which stands for Artificial Gaming Intelligence. So clearly the goal is to have our own models that are very optimized, and I wouldn’t say we don’t have that currently, but we’ll have a lot more of that when we get more data and we’ll have many more users. 

Jon Krohn: 00:41:47

Nice. And this might be getting into another question that you may or may not be able to answer. I know I feel like I learned a lot from your last answer, even though most of it is something you can’t talk about. But when you’re trying to create an intuitive user experience with a game like this, how do you balance between simplicity and depth of creative control? That seems like something that would be very hard to get. 

00:42:13

So for example, probably a lot of our listeners have the experience of using a tool like Dall-E 3 in chatGPT where you ask for an image to be rendered. And what comes out, maybe there’s parts that you like, parts you don’t like. It is difficult to be able to, on your second prompt, get something that say looks a lot like the first image, retains the parts you like, while changing the parts you don’t like. Very often the experience is that you just get a completely different kind of image. 

00:42:48

And so, that kind of thing with a video game where there’s this interactivity that’s expected eventually with the outcome, it just seems like it’s even more complex. It seems like you’d kind of, in the way that OpenAI with that Dall-E 3 model is kind of saying, “For now, we’re not going to worry about it. We’re not going to try to have continuity across the outputs.” It seems like for your application, that’s essential. You can’t have after some iterations, you get to the third iteration, and everything that the person had asked for in the first iteration is kind of gone. 

Andrey Kurenkov: 00:43:26

Yeah, well, to your question, it’s hard. I think if you compare it to something like Midjourney or just general text-to-image, we have arguably a harder challenge because at least to those, you write what you want, and you get an image, and you can see immediately if that’s what you want. Versus a game entity, you can see what it is, but there’s stuff you maybe not be able to see, the logic behind it, the parameters behind it. It’s a more complex entity. Our goal is to really make it as simple as possible for people to create, even for people who are young to be able to do it, and perhaps even this to be their entry point, to understand how to use AI, how you do a prompt, how do you iterate with a prompt. So it’s been a path to try and understand that CX, and I don’t know if we’ve nailed it yet, but we’ve learned a lot about it, and we are making quite significant progress towards it. So certainly as we go broader, hopefully people will like it. 

Jon Krohn: 00:44:41

That’s exciting. Obviously you’ve tried using this engine yourself. And again, at the risk of asking question that you can’t go into too much detail on, have you had some holy crap experiences with using it yourself? 

Andrey Kurenkov: 00:44:55

A little bit. Certainly when we add some new features, just seeing the intelligence of it, how it can process a very broad request and then map it onto something we can do in our current game functionality so that essentially you can ask for anything. We can’t do everything, at least for now, but we can interpret it and do something that still hopefully satisfies your request in some sense. And this question reminds me of a recent discussion point. After Apple Intelligence came out, there was kind of a broad conversation by a few people, a few notable people that have observed that these days, what we are learning is AI as a product often doesn’t work.

00:45:50

So if you look at something like Rabbit r1, the Humane Pin, that was the idea to a large extent is you pay for AI. AI is what you get, and we’ve seen examples of those things not really working. And if you compare it to, let’s say Apple’s approach, where AI is a feature within a larger product or tool, and AI needs to be integrated in a smart way to really empower you, let you move faster, but it isn’t necessarily the only thing. You do need good UI and UX to empower people to use it in a way that they understand.

Jon Krohn: 00:46:35

Makes perfect sense. AI as a feature as opposed to AI as a product is the way to go. And I think, over the long term, that will become more and more true because the AI as a product companies will become increasingly, if they aren’t already, like a utility, like cloud compute. You can switch from AWS to GCP or Azure. There’s some headaches associated with making those changes, but you can do it. And in the same kind of way, if I… Right now, there’s production reasons why I’m calling the OpenAI API, but I actually in my personal life find that using Anthropic’s Claude for most use cases that I have far exceeds the performance, the capabilities relative to what I’m getting from the GPT experience over at OpenAI. And so, I am looking now to be switching even in production purposes from the OpenAI API over to Anthropic, and it’s this constant battle. And they’re also constantly battling on price. With the Claude 3.5 release, 3.5 Sonnet, that outperforms seemingly on most capabilities you would want at half the price and twice the speed relative to their Claude 3.0 model, Claude 3.0 Opus. 

00:48:09

And so, yeah. It’s like a utility. It becomes easy to switch. Lots of people who are providing AI as a feature, even huge players like Apple with their Apple Intelligence, they are going to be calling on in some instances with your permission, they’re going to go out to the OpenAI API. And so, I mean, that shows if a company like Apple, if it isn’t worthwhile for them to be developing frontier models in-house as one of the most valuable companies on the planet, $3 trillion in market capitalization, and they’re better off using a third-party service, even though that probably there’s people in there shaking their heads that the privacy implications of that, it goes to show you that trying to compete on that front is going to be very expensive and very hard. 

Andrey Kurenkov: 00:48:59

Yeah, a few companies can do AI as a product, you can say. I mean, OpenAI is an example of maybe the first one or one of the first ones that basically took a model and sold access to a model without a lot more. But we are seeing, even with OpenAI Anthropic, they are moving away from purely AI as a product. As you mentioned recently, Anthropic came up with Artifacts, which is a nice quality-of-life feature. We’ve seen a lot of features going to ChatGPT outside the API, things like memories that you can add it directly. So they are, I think, learning along the way the UI and UX to make it more useful for people while also being one of the relatively few companies that are trying to develop AI to sell, in the same way that there are a few cloud providers that still compute. 

Jon Krohn: 00:50:07

And I suspect also a lot of those features, just as you were saying it there, I suspect these companies, like having that history feature, having in the example of Google Gemini, which is another one of the few people operating at the frontier with developing frontier models and providing access to these LLMs as a service. One of the things that Gemini has done is they’ve created these Gems which are personas for different tasks. So I could have one, I don’t actually have this as one. It would be maybe a good idea for me to do this is to have a Gem for converting the transcript of a podcast episode into a good social media post. And then, I don’t have to remember or look back through my prompts to find a great prompt where I did that last week or last month, and I can just be like, “Okay, there’s my Gem. Pop it in.” 

00:51:02

And so, those kinds of features, like that Gem, like history in ChatGPT, not only do they mean that these providers are moving away from AI as a product towards AI as a feature within a broader product ecosystem, but simultaneously these are sticky features. Because once I have all that history with ChatGPT, and it’s kind of learned my preferences, I don’t want to have to start over with Claude and do that. 

Andrey Kurenkov: 00:51:28

Yeah, for sure. That’s one of the things I suppose that in 2023, even today to some extent, these three companies, GPT-4, they’re all kind of around the GPT-4 level with Claude and Gemini and OpenAI. They are very close in functionality. Some of them are a little bit better at something or something else, but as a user it’s pretty easy to jump between them. So that point about stickiness is interesting. And speaking of gems, it reminded me of something I consider interesting. So similarly, OpenAI has a GPT Store where you can have customized GPTs, and people forget, I think, that this kind of thing started with Google, with their LaMDA project. 

00:52:29

So they internally had this tool where you could take a chatbot that was an LLM optimized for chatting, and create variations of it. And then, what happened was this person, Blake Lemoine, I forget, claimed that it was sentient in a Washington Post interview, and there was a whole bunch of media and drama. And I think that’s part of why Google is seen by some people as being a late mover. It’s interesting to think what history might’ve been like if they didn’t have that kind of PR problem because that happened in May 2022, and then ChatGPT launched November 2022. And I bring this up only because I think this is an interesting bit of history of AI. 

Jon Krohn: 00:53:26

Yeah. It’s one of those, what they call sliding doors moments from, or closing doors moments. Was the film called Sliding Doors? I don’t know. They’re called sliding door moments where- 

Andrey Kurenkov: 00:53:37

One of those two. 

Jon Krohn: 00:53:38

Yeah, it’s the idea, there’s this whole movie in one way the movie plays out, somebody gets on a train, and they meet somebody. It ends up being their partner for the rest of their life, whatever. In another version, they don’t get on the train, the door is closed just before they get on, and they live this entire alternative. And the movie kind of plays out those two vastly different lifestyles that played out based on making it on the train just on time or not making it on the train just on time. And so the sliding door is closing and so yeah, it is interesting. That Blake Lemoine is an interesting character in one of these, it’s like, did Blake Lemoine have this religious experience and make a big fuss about sentience or not, and what impact would that have had on Google capitalizing on making public technology that it had? 

00:54:40

I think that it is interesting that we’re at this point where you and I both, I think see, while Gemini, Anthropic… Sorry, I may as well say the companies and the products, Google Gemini, Anthropic’s Claude, and OpenAI ChatGPT, their frontier model within each of those families is roughly the same for most use cases, and there might be some, like you said, where one is better for some particular application or another. But of those three, it seems like Google Gemini is the laggard. I have subscriptions to all three and I’m least likely to go to that one. There’s the fewest use cases. Real- time web search is one where I’m kind of more likely to use it there because it capitalizes on Google search. Bing search is not the same as the Google search, and so you do an OpenAI search with that, and I think Claude still at this time doesn’t do real-time searches. 

Andrey Kurenkov: 00:55:47

Yeah, and that’s an interesting point we just covered when we recorded yesterday the newest episode of Last Week in AI, Gemini, they are launching it as a feature across the product, so they launched it in Gmail. There’s going to be a new button in the top right, and I saw this yesterday using Gmail, that there’s a Gemini button and if you’re using Google Documents, Google Slides, that is also in there. And Microsoft has done this as well for Office Suite. We have integrated AI. So I think it’s true that Gemini, the chat bot is maybe not ideal, but for most people it might not matter exactly which one is the best, and Gemini is cheaper. At least you get a two-month free trial if you are a Gmail user, which I made use of. And really, I think there is an ecosystem thing of if you’re in the Google ecosystem and you use Gmail and you use Google Docs as I do, then you will use that chat bot just because there’s less friction if you’re in Microsoft world, if you’re in Apple world. 

Jon Krohn: 00:57:00

Totally. 

Andrey Kurenkov: 00:57:01

Yeah. 

Jon Krohn: 00:57:02

Yeah, and minimizing friction is the key because as you see from website visits to say ChatGPT around November 2022, and then again with the GPT4 release in March 2023, you had these huge spikes in interest in that particular interface that have since subsided. And so that kind of fits this overused Gartner hype cycle where you have the peak hype and then the trough of disillusionment that follows. In that trough of disillusionment, you’re probably less likely to be just going to one of these interfaces like ChatGPT to have a conversation for fun because you’ve worn that through. It’s not a fun thing anymore. I wrote in a Waymo for the first time in my life yesterday and that’s amazing and completely mind-blowing for a few minutes. 

Andrey Kurenkov: 00:57:57

I agree. I agree. That’s how I felt. 

Jon Krohn: 00:58:03

I know that you’re a regular Waymo user. I hear you talking about it. You’ve been talking about it on the Last Week in AI podcast for months. I’m sure it’s not something now when you get into a Waymo, you’re probably not phoning your grandmother on a video call and showing her what it’s like like I did yesterday. 

Andrey Kurenkov: 00:58:18

Yeah, exactly. And that’s what happens when… Waymo as an example is interesting because they just recently opened it up to everyone in San Francisco, so now everyone can try it and as you try things, I guess the thing, yeah, they become less magical and become normal. It’s kind of funny I think for a lot of people in AI, this idea of large language models has been around since 2020. We had research papers, we had the open AI playground in which you could already see to some extent the kinds of things you can use ChatGPT for.

00:58:59

And it wasn’t ChatGPT, but it was GPT-3 and then GPT-3.5 and there were a lot of papers where they used it for these kinds of things, and then ChatGPT hit and everyone else realized this thing and as someone who in a way knew already about large language models and the scaling laws and all these things, you didn’t realize or I didn’t realize at least how profound it will be in a couple years when everyone sees and uses it. And yeah, I still have my mind blown pretty regularly, most recently text to songs. 

Jon Krohn: 00:59:48

UDEO. 

Andrey Kurenkov: 00:59:49

Yeah, UDIO. 

Jon Krohn: 00:59:50

Lost myself in there for a while after you guys… because you guys started doing theme songs. Well, you had one episode that started with a theme song created by UDIO, and then I guess even that gimmick kind of wears off a little quickly even for you then and now you have an ending song, which is something I keep meaning to implement on the Super Data Science podcast because it’s really fun because you create that ending song based on what happened in the episode. 

Andrey Kurenkov: 01:00:14

Yeah, yeah, and it is a lot of fun. I think that’s an application of these things that makes sense, a fun little thing to include as part of a larger project or thing. Now that you mentioned it, I am starting to wonder if we should do little songs in section transitions. It is a lot of fun and those you can make a one-minute song and I have done it now a few times and I find it quite fun.

Jon Krohn: 01:00:51

Yeah, I think it’s the kind of thing though that over time, I think if you had a one-minute song transition between sections, I think that would probably get old for people after pretty quickly. 

Andrey Kurenkov: 01:01:00

Exactly, exactly. 

Jon Krohn: 01:01:01

If it’s like a five-second transition, then that probably makes sense. People will tolerate it, but I think that probably fits into this whole conversation of Waymo, of state-of-the-art text to text LLMs. This fits perfectly into this idea, which I guess I’m now going to get back to, which is that AI as a product rarely works, which is the quote that you had earlier, but AI as a feature does work well, and so as you’re saying, with something like now having Gemini in your Google office suite in Gmail, that makes a huge amount of sense, and so it doesn’t really matter if they’re the third-best frontier model because that third-best frontier model is still very, very good. 

01:01:46

It’s 99.5% of the way kind of on any given capability relative to Claude and to ChatGPT. And so if I can just have in the same way that in Gmail my next few words are anticipated and I can just press tab and have those autofill, I’m going to use Gemini if it’s that easy instead of copying something, pasting it into another tab, finding the right part in the output and pasting it back in. If all that can be handled for me seamlessly, I mean, that’s saving me quite a bit of cognitive load. 

Andrey Kurenkov: 01:02:24

Exactly, and I guess we’ve seen that over the past year is every single thing, every single tool has started integrating AI in one way or another. I think Adobe has been very proactive and impressive in doing that for Photoshop for instance, where they have more and more cool features. We’ve seen Notion do it. We’ve seen… I don’t even remember. Every thing you use will have AI in it. It seems like even some things that don’t need it, like it’s in Instagram now and it’s in Snapchat, these kinds of places.

01:03:12

But to me, I maybe just recently realized this, that this is as big a technological impact on the world. Certainly smartphones, you could argue the internet where everyone has smartphones now or most people. It changed the way you interact with technology and now it seems like everyone will use AI. It’ll become just part of how you do things, which is interesting to think about for me. We are living through history and we’ve had in the past couple of decades three technological revolutions of internet and smartphones and now AI and before that it was like computers and microcomputers in the ’80s and ’70s. So things are interesting. These are interesting times. 

Jon Krohn: 01:04:14

Interesting times for sure. I’m sure that’s part of what compels people like you and me to be creating shows that air once or twice a week with updates on this because it moves so rapidly. Actually, it’s interesting even in the four years that I’ve been hosting this podcast when I started in… My episode started airing in January 2021, and at that time, some weeks it didn’t seem like I had… At that time, Tuesday episodes were always with a guest and they were long and Friday episodes were always quote unquote Five-Minute Fridays. They’re pretty much always longer than five minutes, but the idea was to keep them really short. It was just me and my goal with those Five-Minute Fridays was to always have some important piece of news covered, some important paper. 

01:05:04

And in 2021, a lot of weeks I would end up talking about things like habit-tracking spreadsheets of things outside of AI that I found interesting and helpful and I thought listeners might enjoy these kind of productivity tips. I even started doing this series of habits. It was a numbered series that I can’t remember where I got to. I maybe got to number eight or number nine of these are I think the most important productivity habits that you can have, and it was intended to go on for dozens of episodes, but then I think ChatGPT came out, and then since then, it’s just every week I have an endless list effectively of extremely interesting, directly relevant AI stories that it never occurs to me anymore to be like, “We should do the habits episode this week.” 

Andrey Kurenkov: 01:05:57

Yeah, it’s interesting looking back, we’ve been doing the show for almost three and a half, almost four and a half years. We started March 2020 and I remember as a PhD student, as someone working on AI, it already felt like things are moving super fast at the time, at least within research, but now I kind of miss those days. It felt really rapid then, but now it’s moved out of academia and out of this niche circle where we are making a lot of fast advancements on beating benchmarks, on utilizing neural nets to do things you could not do before and now all those things are coming into the real world. Yeah, exactly. It was like it used to be that we didn’t have 30, 40 stories per week that are actually meaningful and interesting and now it is definitely the case. 

Jon Krohn: 01:07:09

Yeah, it’s wild. Something that I actually talk about in public talks that I give a fair bit is I open with how this is likely only going to get faster and faster. 

Andrey Kurenkov: 01:07:23

Yeah. I’ve reflected on this recently and I don’t know why it took me… I used to think that even going back two years, one year, my thinking was it’s hard to say whether we will have human level AI anytime soon. It seemed possible, but it didn’t seem definitely likely, which some people did already believe in, and I had particular technical reasons of scaling, of limits of length and memory, these different things. Now, a year later, a lot of those challenges we’ve made impressive progress on. I remember initially ChatGPT had a context limit of 4,000 tokens. Now you have context limits of 250,000, one million. 

Jon Krohn: 01:08:19

In Gemini, you have a million. 

Andrey Kurenkov: 01:08:20

Yeah, and- 

Jon Krohn: 01:08:22

That is also, that’s another great use case for people. Real-time search in Gemini or also if you’re uploading a huge media file, you can upload into the context window of Gemini today 10 or 11 hours of audio, an hour of video, the equivalent of 10 novels, and they’re testing a 10 million token context window version, which would 10X all of those numbers I just gave, so it would be 100 hours of audio, 10 hours of video or 100 novels worth of context. 

Andrey Kurenkov: 01:08:56

Yeah. It’s just I guess since ChatGPT, there was a lot of money going into AI to some extent already, but now it’s impressive to see what humanity is capable of if we take some of the brightest, most ambitious, most hardworking generally, let’s say, I don’t want to say the top or the most talented, but you have a lot of people pushing to progress to make progress and to solve problems, and we are doing that at a faster and faster pace. We got Claude 3.5, Gemini 1.5 Pro that are already better than the larger and more complex systems, and we’re still solving some challenges like making these passive models. We give it one input and you get one output. 

01:10:02

A lot of people working on the agent aspect of it where you give it one input and it can operate independently for some period of time and go off and do things, which I think would be required for human level intelligence to really… You can’t have it as a just passive model. You also need multimodality with vision and with movement ideally and audio, and that’s another thing that we’ve made a rapid progress of. What I want to say with all this is after just a year or two when I used to think, I don’t know exactly when AGI happens, it could be soon, it could be decades, now I feel like with all the money, the people and all the talent, I could easily see it happening in two or three years and that’s crazy, human level AI in just a couple of years, but that’s the world we live in. 

Jon Krohn: 01:11:01

I’m in exactly the same boat as you on exactly those timelines. The watershed moment for me was GPT-4 where yeah, GPT-3.5, the ChatGPT Experience, I was like, “This is very impressive. This is very impressive. No question.” I loved it. When GPT-4 came out, I was like I flipped from exactly what you just said, is something that is exactly the way I felt and I’ve said on air a number of times before on different shows including this one where I went from exactly as you said, thinking if it’s possible at all to have human level intelligence across all human tasks, it’ll probably take decades if it’s possible at all. I don’t know if I’ll live to see it. And now it is exactly the same. It could be a few years. 

Andrey Kurenkov: 01:11:51

Yeah. And I guess part of why that is, part of why you believe it is what is happening now is everyone is spending billions of dollars, right? OpenAI is spending billions of dollars on just training. Well, they’re not spending billions of dollars yet. They spent millions to train GPT-4 that we know, maybe 10 million, but Meta, it’s certainly in the order of tens of millions and it seems like people expect GPT-5 or something like that to take something like 100 million or 1 billion or whatever, but they’re doing it. They’re getting the money, they’re getting the capital. 

01:12:36

You have giant companies like Meta and Google that have the money anyway just throwing money and infrastructure and compute at it. And so that’s another aspect of it where if a big part of humanity comes together independently to make progress in a direction. It’s something that I’ve realized that there were these technical challenges, but so many people are working on different aspects of a problem and there’s so much capital and other things made to progress fast that this is what humanity does apparently. We just do amazing sort of miraculous things as we’ve done with computers and smartphones and everything. 

Jon Krohn: 01:13:29

And hopefully this will assist us in tackling the major challenges for society of our time. 

Andrey Kurenkov: 01:13:38

That’s the hope. We hope is it makes life better. And we have seen that with prior technological revolutions. We’ve seen that with computers, with the internet, with phones. There are some drawbacks of course, especially at the beginning. With industrial revolution, people lost jobs. It was kind of a tough phase, but in the long term, it improved things and we just got to hope that’s the case because it’s coming. 

Jon Krohn: 01:14:07

And that’s two wits. The next episode of the show, episode number 800, I interviewed my 94-year-old grandmother who was born in Ukraine and emigrated to Canada in the 1930s and had no electricity and I said, “Is there anything you miss about life back then?” She’s like, “No, this is way better.” And so it is interesting. Sometimes you have this feeling. There are things that happen, there are negative consequences, some that Yuval Noah Harari talks about repeatedly in his book, Homo Deus, is that the kinds of things that for all of the rest of human history or probably all of the rest of the history of living things on this planet, hunger, famine, war, that in recent decades particularly, things have flipped on their head where, for example, with hunger, for all of the history of living beings on our planet as far as we’re aware, just being able to eat and survive was probably the most fundamental issue, and today more humans on the planet die from diabetes, from overeating than from hunger.

01:15:38

And there’s all kinds of things associated with that, like what is the quality of the nutrition, what is the quality of the calorie, that kind of thing, but still is a general oversimplification, that fact that people are dying from overeating more than undereating. That’s a very recent thing in this world where in a completely new kind of paradigm and hopefully there’s kinds of reasons why you could be very then concerned because you think, “Okay, well maybe we’ve plucked the low hanging fruit.” We’ve been pumping hydrocarbons out of the earth, which has spurned a lot of this growth. We’re still today as much as probably most podcast listeners to this show, and I’m sure you and I hope that we’ll be able to phase out hydrocarbon soon, still today in terms of global energy mix and in almost all western countries, regions, we’re still super dependent, 90% or more of energy coming from hydrocarbons and things like even just extracting ore or other raw materials out of the earth.

01:16:47

We’ve plucked the low hanging fruit of the easily accessible things and now we’re like, “How do we do deep sea exploring for this ore?” And so there’s a case to be made. There’s a skeptical case of we’ve pulled out the low hanging fruit. Where we are today is the pinnacle of quality of life for humans, and from here on out it’s going to be a poop show. And that’s what that dystopian view is what pretty much all scripted dramas of the future play on that’s very rare. It’s fewer than 10% of TV shows or films about the future have a utopian thing. But it seems to me that’s where we’re going. And so it’s a weird thing. It seems like people in general are like, when we did New Year’s Eve this year, the overwhelming sentiment at the New Year’s Eve party I was at was like, “Wow, lucky we made it another year with all these things that are happening.” And it’s like, well, I don’t know. I think things are getting better all the time. I’m really optimistic. I think that AI can play a big role paired with us.

01:17:56

There’s lots of issues we need to navigate, but I think that we could be looking at a super abundant future, things like being able to crack fusion and energy, having a bunch of suns on the planet generating effectively limitless energy. I mean, we could have many, many orders of magnitude more energy. And then with that plus human level intelligence, that’s trivially easy to deploy wherever you want, problems like, okay, we’ve got too much carbon in the atmosphere, we have abundant energy and intelligence. Let’s just have the machines rig things together to take advantage of that fusion energy and pump it into the ground. 

Andrey Kurenkov: 01:18:27

Yeah. I think when I made this kind of thinking switch, and this was just a couple months ago, when you realize or believe that AGI human level intelligence is coming, it’s coming in a few years, it’s easy to just have… Well, of course you have a mind blown moment, but then it’s kind of weird where now you’re living with knowledge. You’re almost living with knowledge of a meteor is coming or you know that everything is going to change in a couple of years and the world is chaotic and overwhelming and there’s enough bad stuff going on as is, right?

01:19:10

But to your point, I think it’s important to at least remember that things will improve in some ways and things will get worse in some ways. And in general, things have improved in a lot of ways as technology has evolved, right? The last World War was a long time ago. The Cold War ended. My parents lived in the Soviet Union, and certainly even with all the flaws and badness and everything that we have in the world, most people don’t live in the Soviet Union anymore.

01:19:48

So anyway, I guess the point is human level AI is coming. It’s easy to think that we are heading towards a dystopia for sure, but just to keep your sanity, it’s important to think, well, hopefully it’s not a full on dystopia. It’s just some things are going to be bad and some things are going to be good, and that’s life. 

Jon Krohn: 01:20:13

Interestingly, your co-host on Last Week in AI, Jeremie Harris, whom we talked about a lot at the onset of this episode today, the Super Data Science episode today, he is really concerned about this AGI event. He thinks that there’s at least enough existential risk and exactly what existential means, you guys did a great episode. I’m going to link to it in the show notes for this episode. There’s a fantastic episode of your show where you two debate and you go into a more granular detail than I have heard anywhere else before or since. Yes, I’m sure there’s all kinds of resources that people could dig into more, but for me as a podcast listener, I thought that that was a great introduction to this stuff. 

01:21:02

Introduction to this stuff really in depth. And in that, yeah, because you guys have these different perspectives, it comes up on the air of your show periodically where you’re generally optimistic about a AGI and that unleashing what some people call Ray Kurzweil’s term, the singularity, a point that will change the world so much that we can’t predict what anything will be like beyond that. Despite all those unknowns, you are generally optimistic and I think Jeremie is generally optimistic as well. But simultaneously he believes that if there’s even a 10% chance or even if there’s a 1% chance of all of humanity being wiped out or a large portion of humanity being wiped out, we should be concerned, we should be putting some resources into it. And so, yeah, I- 

Andrey Kurenkov: 01:21:53

Certainly there are different perspectives and different things that people take away from the knowledge that this trajectory… if you assume that the exponential improvement trajectory is going to improve, which some people don’t. Some people say it’s a logistic function, it looks like the exponential will keep on going, but some things flatten out like airline travel as an example, some technologies, storage of compute, our ability to make energy. There are examples where for a while it’s exponential or Moore’s law, for instance. The exponential, but there are some fundamental physical laws or something that means that you flatten out. And it is true. I think it is important to observe that GP4 came out a year ago. We’ve had some improvements with G PT 4.5 and Claude and Gemini. But they’re all still at a GP4 kind of level and we don’t have GPT5 for quite a while. 

01:22:58

So it could be that you flatten now. We don’t know. We just don’t. My change was primarily driven by this idea that it doesn’t seem likely we flatten now and certainly there are people very intelligent, very reasonable people with an AI like Jeremie who believe things like in three years there’s a 20% chance of 50% of humanity dying because of AI, which again is mind blowing. It’s like if you live with this knowledge and this belief, how can you sleep at night? But you got to keep living and that’s why we have people working to at least avert that possible future, to make it so you can’t make nuclear weapons and various kind of misuses of advanced AI. And we do need those people, even if I don’t exactly agree with how concerned they are.

Jon Krohn: 01:23:58

Yeah. It’s a really interesting thing. And it’s interesting how if a lot of resources go into preventing say existential risk or really bad outcomes from AGI, and then when a GI arrives no really bad things happen, it’s like who was right? Because if it was because of the fuss made now that the funding went in and so disaster was averted, it’s like this cynical perspective is right either way 

Andrey Kurenkov: 01:24:38

We’ll see. We’ll see. There are examples of humanity getting it right with things like nuclear power where certainly there was a possibility of the end of the world. People were afraid of the Cold War and the possibility that we’ll have the Fallout series of games of we just nuke ourselves into dystopia. And we haven’t, right? So that is in a way kind of impressive. And hopefully it will not be quite so radical. But who knows? 

Jon Krohn: 01:25:15

Nice. Well, so we ended up talking about this for quite a while and this was not exactly on my plan, on my roadmap. But obviously what is a bigger issue to talk about? And so to some extent, and it probably made sense that we focused on this for a while, obviously something you and I both are really interested in. 

Andrey Kurenkov: 01:25:35

When you follow the news and you see how much is happening every week, I guess that’s what you start thinking about.

Jon Krohn: 01:25:42

Yeah. Yeah. Yeah. And it is interesting how as you start to really believe that this is coming, it does change… So for me, for example, it has caused a shift in the things that I appreciate in the sense that I have made over the last year and I’m continuing to make changes in my life that allow me to focus more on time with my family, with my loved ones, with close friends and things that are really important to me outside of work that I really enjoy like exercise. Prioritizing those things because it seems obvious to me that there’s going to be such huge paradigm shifts coming in the next few years that a lot of the things that we work on aren’t going to be that important or meaningful in a few years. 

Andrey Kurenkov: 01:26:42

Yeah. That’s I suppose a good reminder to just do the things that make life good, right? Go into nature, touch grass, do all those things that you should do anyway instead of scrolling, doom scrolling on the internet and don’t try to do things even as the world changes at a ridiculous pace. 

Jon Krohn: 01:27:07

The clock is ticking, folks. Touch that grass while you can. All right. So the next topic area that I plan for you, let’s just spend a short amount of time on it. I thought it might be a huge part of the episode, but I think it’s an interesting part of your education and background. So let’s just spend at least a little bit of time here. Before joining Astrocade where you’re now doing text to video game experiences, your research at the Stanford Vision Lab and the Socially Intelligent Machines lab during your PhD involved humanoid robots and deep learning. I am super into these embodied AI applications, embodied meaning where it isn’t just software, where there’s some embodiment of the artificial intelligence in the world. So this could be a robotic arm, it could be a humanoid robot. But in some way the AI system can interact with the world. 

01:27:59

And so last year you mentioned being excited by Google AI’s work with LLMs for robotics. Nvidia has been betting big recently on this with things like their Eureka announcement. For me personally, a huge thing that I’ve been so excited about is Covariant, Pieter Abbeel’s company Covariant with their robotic foundation model one, RFM one, which gives you a multimodal interface for robotic arms so that you can have text in or out, images in or out, video in or out. I did a whole episode on this recently if people want to hear more about that and why I think it’s so interesting. So that was episode number 774. 

01:28:46

But yeah. So unlike me, you have real world experience developing real world robotics applications stretching back to things like even during your undergrad at Georgia Tech, being involved in the solar racing car team, which in some ways is quite different. There isn’t probably, as far as I’m aware, an AI component to that. But nevertheless, you’ve been dealing with hardware, with real physical systems. I never really have. So I’m curious of digging into these things a bit with you. So yeah, so how do you envision LLMs and their injection of common sense of world models of interactive conversations with humans in a natural language way? What do you think are the big deals here that could be happening for us in the next little while?

Andrey Kurenkov: 01:29:41

Yeah. I think if you haven’t worked in robotics, it’s easy to underestimate how hard and how painful it is. 

Jon Krohn: 01:29:51

A phrase I got from your show that I now think about all the time and even say out loud all the time, I think Jeremie said it first, he said it’s something that they learn in Y Combinator, which he went through, which is hardware is hard. 

Andrey Kurenkov: 01:30:01

Hardware is hard. Yes, exactly. And part of why it’s hard within the context of AI is in a way you need in everything. You need computer vision. And a lot of what AI has been and what we’ve made progress in AI is you take problems and you isolate them and you figure out how to do them without a general purpose model. So things like classifying objects, things like translating language. For the longest time, basically until a few years ago, that’s what people worked on in AI, like narrow problems, seeing a model could do one thing and you could do one thing well.

01:30:45

And so for robotics where you need all these things to some extent, you need to be able to move, you need to be able to reason and plan, common sense, you need to be able to see things and understand where they are and you need to do that all while having a physical system with the standard issues of control of these very low level things, things that Boston Dynamics for instance works on of being able to just keep upright and compute all the things that you compute somehow effortlessly, it’s been hard.

01:31:19

And so as we’ve seen the rise of general purpose things, what people call either foundation models or frontier models that are multimodal, so they deal with images, they deal with texts, they deal with audio, it has seemed like that has a very direct impact on robotics because you have the common sense piece, you have a vision piece. What you don’t have yet is the emotion piece, the control piece. You can do the high level, but not the low level. But there’s many ways in which you can integrate known techniques and that’s what has been done for a while. There were examples of that coming out of Google, for instance, pretty early on where you took a high level reasoner and low level control via reinforcement learning and you were able to then do general purpose problems. And that’s a dream. Your dream is a general purpose robot that can do whatever. We haven’t been able to do that. But now as with artificial human level intelligence, it appears like we can do that way faster than most people expected. 

Jon Krohn: 01:32:36

Yeah. Exactly as you’re saying, part of what excites me so much about this space is how it seems like except for as you’re saying the control parts, the software only LLMs or multimodal foundation models… That is the thing actually for me, I’d be interested in your thoughts on this. I kind of use the terms interchangeably. To me, even a multimodal foundation or frontier model, I still often call that an LLM because it seems to me that the… I’m not aware really of big multimodal models that don’t have a language component. And so it still seems like at an encoding level, the language piece is still really important. And so I kind of always end up calling those LLMs anyway. Do you think that’s silly? 

Andrey Kurenkov: 01:33:18

I suppose it’s sort of saying that a tomato is a fruit or something. They do include a large language model as one of their capabilities, so you can say that. But yeah, it’s not quite precise enough. It does limit what it means. 

Jon Krohn: 01:33:33

Yeah, that’s good. I should be more careful about that. But yeah, so these multimodal foundation models that have language as one of their capabilities, I like that tomato is a fruit especially because I love catching people out on that. You can really blow people’s minds with how many things are fruits that they think are vegetables, like peppers, cucumbers, there’s so many things that they’re like… If you think about a Greek salad, a horiatiki salad, it’s just a fruit salad. Everything in there is fruit and cheese.

01:34:04

So yeah, so I go off piece a little bit here. But yeah, what’s really exciting about this is that these multimodal models, except for as you’re saying the control piece in terms of controlling the hardware, it does seem like we’ve cracked so many different [inaudible 00:34:22] things, world model things where you’re like, “Wow, it seems like there’s going to be so many embodied applications.” So that makes it really exciting. 

01:34:29

But simultaneously because hardware is hard, it seems to me like there are great opportunities for people out there who have the patience and the capital to be developing embodied AI systems because hardware is hard. If you can figure it out, you’re going to have a moat no matter what. Whereas in the software space alone, like we were talking about earlier, these companies, even OpenAI is spending a billion dollars let’s say on GPT 5 or maybe GPT 6, then someone else comes along and figures out how to do it at a 10th of the price and offer it at a 10th of the cost just six months later or 12 months later. It’s crazy how quickly your moat can dissolve at least on pure software LLM capabilities, whereas if you have some hardware piece, that’s kind of like where, again, tying back to the AI as a feature as opposed to being the product itself. 

Andrey Kurenkov: 01:35:20

Right. And that’s part of why you can be excited for robotics or expect robotics to not take quite as long as we thought. For the longest time, it was hard to get money for general purpose robotics. We had a couple players like Boston Dynamics doing R&D to do that for I would guess military applications and military funding. And you had some big automotive companies out of Japan doing some progress in human robotics. But you didn’t have many players and it was hard to come by the money. And in 2023 that changed. A lot of excitement happened around ChatGPT and that I think in part led to companies like Figure, like 1X, like… Tesla is a bit different. But there’s now several players that are spending a lot of money, a lot of talent to develop humanoid robots that you can use then to deploy and do whatever.

01:36:28

Certainly Elon Musk says that Tesla could be a trillion dollar company because of Tesla bot for instance. And who knows if we’ll get there. But Tesla bot as a roboticist when they unveiled it at Tesla AI Day Two, I think the reaction for me and many people was like, “Oh wow, this is actually pretty impressive.” And you have the same thing with the videos that are coming out from Figure and 1X. It’s pretty impressive how fast people are making progress. 

Jon Krohn: 01:36:59

Yeah, yeah. It’s very cool and I’m excited to see these things make it more into the real world. I think I heard you recently say on the Last Week in AI podcast that one of the things that prevents humanoid robots from being used in say residential use cases is that they’re still too strong, that there’s a chance of them really hurting someone or a pet or a child or- 

Andrey Kurenkov: 01:37:26

Right. They’re too strong and they have bugs. If you imagine someone has a bug, then it doesn’t necessarily matter how strong it is. It’s made of metal and it’s moving. So it’s sort of like a car situation where just by virtue of the mass of it and the hardness of it, it’s a little bit dangerous. And there are ways to work around that. You can have compliance, which Rodney Brooks’s company before they went under, Rethink Robotics had. So that’s an element of it. I think cost is another very real element of it. But certainly before we get them in our homes, they’ll be in factories and so on. 

Jon Krohn: 01:38:17

Yeah. So yeah, bugs make them risky in residential situations because they’re strong, heavy metal is the kind of summary point I’m taking away from there. All right, yeah. I had lots of other questions for you and maybe we’ll just have to save these for a future episode with you, but things like your AC teach paper, which I guess I’ll at least put a link to in the show notes here, but talking about [inaudible 01:38:40] actor critic method paper. 

Andrey Kurenkov: 01:38:43

Yeah. If I’m honest, I think talking about human level intelligence is probably a little bit more interesting than my research. I like my research, but it wasn’t necessarily groundbreaking. But yeah, it was really fun to just improvise and talk about whatever, which I guess is something I tend to do. 

Jon Krohn: 01:39:01

Yeah, for sure. I really enjoyed this conversation, Andrey. I hope our audience did as well. I have a strong feeling that they would have. As always, let us know. Feel free to be candid with us about how you think today’s episode went. But Andrey, for me personally at least, this has been a fantastic experience. It’s been amazing to meet someone that I listen to for a couple of hours every week for over a year now, to be able to meet with you in person and have a conversation with you, pick your brain, which we’ve done a bit. I’ve been on the Last Week in AI podcast now three times I think. I was thinking about this actually in the last time I was on. I misspoke because I misremembered. I’d said the last time I was on your show, you and I were co-hosting. And at the beginning of that I said that the preceding two times I’d been on with Jeremie, but in fact the first time it was the three of us.

01:39:57

So we’ve done me on Last Week in AI three different ways with all three of us, me with Jeremie, me with you. But that, because of your format where you’re getting through the 30 or so news stories, you don’t spend any huge amounts of time on the things that we’re just thinking about that we think are important. And so it’s been amazing for me to get this experience with you, this long interview, pick your brain about… Yeah. We seem to be in exact agreement on what the most interesting, exciting, but also potentially challenging items are, memes are, thoughts that you can have in the world today. 

Andrey Kurenkov: 01:40:39

I agree. It was a ton of fun. Good to met you in person for the first time today for the interview. I guess if I am to plug some things before we finish out- 

Jon Krohn: 01:40:50

Please do. Yeah. 

Andrey Kurenkov: 01:40:51

Last week in AI, that’s its style. You could probably Google it. You can go to lastweekin.ai for the Substack, the text newsletter. I’m not super active on any social thing, but I guess- 

Jon Krohn: 01:41:04

Yeah. Anytime I try to tag you in something on LinkedIn, it doesn’t even let me do that. 

Andrey Kurenkov: 01:41:07

Yeah. But I’m on LinkedIn if you want to follow me. I’m on X as Andrey Kurenkov, so feel free to look me up. 

Jon Krohn: 01:41:16

But you’ve recently started publishing YouTube videos of your Last Week in AI as well. 

Andrey Kurenkov: 01:41:20

That’s true. Last week in AI on YouTube is there if you care to take your podcast that way. But yeah, we would be happy to see some listeners. And it’s a lot to take in. I think it’s not for everyone, but hopefully some people from your audience will enjoy it as I am sure some of our audience enjoys your podcast after [inaudible 01:41:44]. 

Jon Krohn: 01:41:43

I hope so. Lots of crossover. I think when you do Google searches of us now, we’re pretty up there in terms of, “People also searched for.” Last Week in AI and SuperDataScience seemed to have a lot of overlap in what people are searching for, so that’s really cool. We now have kind of done the follow, which is my usual final question. The penultimate question that I usually have for our guests is now what you’re going to have as your final one, which is do you have a book recommendation for us, Andrey? 

Andrey Kurenkov: 01:42:09

Yes. I am going to go with a weird one. 

Jon Krohn: 01:42:15

Perfect. 

Andrey Kurenkov: 01:42:16

There’s some I love. But one book that is very interesting… And I mostly read fiction. I think a lot of people in the Bay Area read a lot of nonfiction because it’s productive and you learn things. I like art. And so I’m going to recommend House of Leaves I believe it’s called, which is a very unique book, in some ways of very challenging book, but I think a pretty profound and special experience. This something you can’t get twice. So if you like big heavy books- 

Jon Krohn: 01:42:51

You can’t get twice. 

Andrey Kurenkov: 01:42:52

Yeah. It’s a singular book I would say. You will not find something quite like it. 

Jon Krohn: 01:43:00

Wow. 

Andrey Kurenkov: 01:43:02

Yeah, it’s kind of a heavy one. It’s kind of a long one. But I love it. So that’s my recommendation. 

Jon Krohn: 01:43:08

Nice. House of Leaves. All right, thanks Andrey. We’ve kind of already done the whole like, “Oh, so great talking to you and great meeting you,” and all that stuff so we don’t need to kind go through all that rigmarole again. But it was great having you on the show. Thank you so much, Andrey. 

Andrey Kurenkov: 01:43:21

Thank you for having me. 

Jon Krohn: 01:43:22

Whoa, what a conversation. In today’s episode, Andrey filled us in on the genesis of Last Week in AI, The Gradient at Skynet Today, how Astrocade aims to launch an iOS app, perhaps an alpha as early as this year that will allow text to video game generative AI including cooperative or competitive multiplayer games with an avatar for each player. Cool. He also talked about how AI as a product rarely works, but AI as a feature can lead to terrific commercial success for many startups, that the release of GPT-4 last year flipped him from being an AGI skeptic to realizing a superhuman intelligence could emerge in perhaps as little as a few years. And we also talked about how multimodal foundation models imbue real world AI embodiments, including humanoid robots with common sense world models, making them vastly more capable today than just a few years ago. 

01:44:19

As always, you can get all the show notes including the transcript of this episode, the video recording, any materials mentioned on the show, the URLs for Andre’s social media profiles as well as my own at www.superdatascience.com/799. Thanks of course to everyone on the Super Data Science team our podcast manager Ivana Zibert, media editor Mario Pombo, operations manager Natalie Ziajski, researcher Serg Masis, writers Dr. Zara Karschay and Silvia Ogweng, and founder Kirill Eremenko for producing another outstanding episode for us today. For enabling that super team to create this free podcast for you, we are deeply grateful to our sponsors.

01:44:56 Oh yeah, you can support this show by checking out our sponsors links, which are in the show notes. And if you would like to sponsor an episode of this podcast, you can get the details on how by making your way to jonkrohn.com/podcast. Otherwise, please share this episode with people who might like to listen to it as well. Review this episode say on your favorite podcasting platform or on YouTube, subscribe if you somehow are not already a subscriber. And yeah, most importantly, I just hope you’ll keep on tuning in. I’m so grateful to have you listening and I hope I can continue to make episodes you love for years and years to come. Until next time, keep on rocking it out there and I’m looking forward to enjoying another round of the Super Data Science Podcast with you very soon.