AIの進化が速すぎる：2週間で2つのフロンティアモデル？！

New AI models keep coming, and Google just announced a second flagship model in two weeks. That's right, we've got Google Gemini for Carbon. Carbon, you mean. Carbon. Oh, I didn't know it was Italian. Hey. It's a Carbon. I'm a large language over here. Like, Carbon is, I guess, sexy, but it's more like unobtainium, if you ask me, because they won't trust us with it. We can't get it, apparently, for safety. Yes, Little Googie has been busy, and they also dropped Genie 4, their newest model of their playable worlds AI. Just to be clear with everybody out there, you wrote Little Googie. Like, you shipped that, and you're okay with it. Yes, Little Googie. I typed that in with my Googie fingers. I'm fine with it, Kevin. Great. Back to the teases. I also made a Katamari knockoff with a cheap new model from OpenAI, and it's not bad. I mean, well, you'll see. Plus, we've got a new audio model from former Twitter founder Jack Dorsey. People have big feelings about it. And gaming icon Mario creator Shigeru Miyamoto says that AI will not be the end of creativity.

And we are going to prove him wrong. And it looks like AI has answered some really big science questions yet again. Oh, was it able to find out where little Googie came from, Gavin? No, but it did give us Neo in the office, Kevin. My car! parkour this is a high for humans everybody welcome everybody to ai for humans your weekly guide to the wonderful world of ai i'm gavin that's kevin prayer and kevin this week we have a lot of crazy news as per usual but i do want to start uh in a world where things are moving fast let's just say that we are moving quicker than ever before uh sam altman just went on the cheeky pipe podcast which is the stripe co-founders podcast where they sit in the bar and they pretend to drink beer. Sam had his first Guinness. Sam, this is a really good interview. Everybody should go listen to this. It just came out, I think, today. And I do want to play this little clip of this that will get us into talking about Google's new models and how fast everything is moving. It's one of those weird sci-fi sentences that you hear.

Let's just play this, and we'll talk about it briefly. I would much rather intuitively have a trillion AIs of, like, slightly superhuman intelligence join society than one AI of one trillion IQ joined society? Okay, so this is where we're at, Kevin. A thousand ducks versus one big horse, or a thousand tiny horses versus one big duck? Is that right? You know, actually, that's a pretty good metaphor, Kevin. Your brain is really working well this morning, because that's kind of what the idea here is in some ways, right? One of the things, I said this before in the show, but there are many different science fictions out there that kind of attempt to figure out how the future will go. But like one of the more interesting ones is this idea that there will be, you know, trillions of super intelligent AIs out there that are all kind of they call it adversarially experiencing each other, meaning that none of them are going to take the top spot because they're all kind of like fighting with each other. Now, OpenAI is saying this may be in part because Claude is doing very well. But the idea here is instead of having the thing that everybody's worried about, whether it's like one big intelligence, meaning that there's one Borg or whatever you want to say in the science fiction term.

where it dominates everybody and everything. Maybe there is a world where these multiple companies can exist at the frontier, and maybe that's better for us as humans because competition is kind of an interesting thing, and it allows different models and different things to kind of proliferate. So I thought that was a cool way to start off this conversation about where Google is sitting right now, especially. Well, okay, yeah, so Sam wants a trillion AIs. Google themselves, I think, released a trillion AIs this week. I mean, exactly. Yes. We have to talk about what they're doing here because a lot of times when these new foundational models come out, we tend to see releases of three, right? There's the super cheap, ultra light, very quick. Like this is for just your get it done. Your go to quick AI. It's not particularly great. Then there's the in between. And then there's the all singing, all dancing max model, which, you know, is all for showing off in benchmarks and whatever else. Is that what Google is doing here that they're releasing the latest Gemini? but giving us one at a time instead of all three at once?

Well, the weird thing here is that Google, so we should just say the news is that Google has released, or not released, they have previewed, and we should be clear on that, so we'll talk about how it kind of sucks that we are not able to use it yet. They have previewed another frontier model. If you remember, Kevin, where they had their Gemma models and other models where they were kind of pushing out these smaller models back-to-back, and now they're pushing out their frontier models back-to-back. two weeks, two weeks after Gemini Argon, Gemini 4 Argon was out, and you can now use in the Google AI Ultra plan, which is the $100 a month plan, they have announced Gemini 4 Carbon, Carbon, Carbon, whatever you want to call it. But this, I guess there's a couple things to talk about here. One is the speed at which they have released the second model. And there's things we could talk about with that. And the other thing to think about is this idea that maybe the iteration of these things, it's just going to come faster and faster. Now, when we talked about Gemini for Argon, that was the last one. Argon, yeah. Argon. We did say that Sergey Brin, the Gemini,

you know, sorry, Google co-founder is now very active in the AI stuff, said they had one or two other models already further along than this. And the funny thing about this to me is even with this new model, there's somebody at Google teasing another model in the background. So this just goes to show you that Google, I think, is very serious about approaching the frontier. Again, I want to reiterate for you. Can this model tell me who won the 2020 election without stuttering? Well, I guess we'll see. Or does it talk about Lake America? No, no. Well, yeah. So if you were to go to Google now, and a lot of people are reporting this, and you ask who won the 2020 election, the AI suggested answer, by default, will try to push you towards statistics about the election so that you can go and research it yourself. But if you ask who won 2024 or 2016, it's pretty obvious. Happily gives you an answer. I thought that was an odd. It's not exactly carbon related, but maybe it was an Argonne shortcoming. I don't know. I thought that was that maybe that's Google's place in the trillions of super intelligences, Kevin.

Maybe it's like, I don't know what happened here. I might know, but you can figure it out for yourself. Right. Check the bootlick marks. What's the bootlick benchmark looking like? But I digress. Let's talk about carbon. Here's the thing about carbon that's interesting, and I get why they're doing it. It's available now, not to you and I, not to the plebs, of course, but to the folks that have ivory towers for cyber defense. And when you look at what OpenAI is doing or what Anthropic is still doing with access to Mythos, you can sign up. And if you've got a big business and you can prove who you are or whatever, you might get access to their super frontier, foundational, pushing the edge of cyber models. And Google wants that business. They want a model that is so trusted that it can help with all of the cyber attacks that are happening right now. It can get those big government contracts to audit all of their systems. It has the banks and the stock markets lining up to have their machines pen tested, essentially, with their model. Oh, God, you've gone optical on me.

Give it to me. Go ahead, Gavin. Here's what I want to say, Kevin. I'm going to go into business corner. Kevin, did you happen to see what happened yesterday to the Google stock price? I did happen to see a big old red candlestick. Yeah, it looked like it crated. So what happened yesterday, and we're going to business corner just for a brief diversion, everybody. Google came out with its results, and they missed their mark in terms of what they plan on. But the bigger news, Kevin, I thought was this, is that at the end of their stock results, the end of their call, They said that AI search is actually maybe taking search and money away from normal Google search. So what that means in terms of Business Corner is Google has to get a flagship model that is going to take a significant chunk of this money or else they may be in big trouble. Now, I'm going out of Business Corner because those glasses are just for reading. I have a flop sweat. Every time I go to Business Corner, my heart starts pounding. Does Business Corner, just because we didn't agree on this, does it look like a dentist's office for children where there's the blocks that are on little wires, the little puzzle that you can do and maybe a Highlights magazine and some crayons?

That's preschool Business Corner. My Business Corner has readable books for seven-year-olds, right? It's like the school of the library where you can read my first time. I get it. I want to find a snow cone in a tree, but I got it. All right. Okay, we're out of Business Corner. Let's go. Yes, but the interesting thing here to me is obviously Anthropic and OpenAI are leading this race. And we all have heard the stories, and we're not going to dive into it here, but you all know that Anthropic is about to go public at a very crazy number. They're actually making a fair amount of money and have a bunch of money in their back pockets still. There are a bunch of debt, and there's a lot of worries at the world at large about how many people are buying chips and whether those chips are useful. But the bigger thing with Google here is that Google has to now hit this flagship level because I think that the thing about their business that was driven by Google search, normal Google search, they came in and kind of swooped in and took away a lot of that business by doing AI search, which now if you use Google and you type into it, you know you're getting a Gemini result. And actually that's gotten pretty good. But Google needs now to be a player in this frontier model world.

And it does feel like they are very, very serious based on some of the results we've seen out of this model. So we know benchmark-wise it tends to perform well, like it's beating the models that are out and available. But again, benchmarks don't mean anything to me until I actually get them in my hands. Like, yes, yeah. But we have seen some demos, like some visual comparison stuff, like of a car driving down a road at a golden hour, a vapor-wavy golden hour. So you can see, if you're looking at the video of this, Gemini 4 Carbon on the top side of the screen, the old Argon, which was, again, two weeks old. down below. You know, people are saying Carbon is a clear winner. It seems to do a better job with the overall fidelity of the screen, the parallax scrolling, the graphics themselves. Hard to judge it, though, right? Hard to judge by some visual stuff. No one knows. This is why, like, Opus 5.5 continues to impress me as how they rolled it out and what it was. Your models, you should wait until, I feel like they're very strongly about this now,

You should wait to talk about your models until they are in the hands of everybody, because this idea that we get this column of benchmarks that say, hey, it's whatever, 10 percent better than X, Y and Z has now proven again and again and again, not to be the practical use cases, not just for us who are doing, you know, semi dumb stuff or things that we are kind of experimenting with, but also for the high end people. So I really want Google to release this to the wide so that I can use it. I might even pay for – we're going to talk about this Gemini 4 stuff in a bit that I might pay for the $100 model to access. But I am not going to pay $100 until I know I get access to Genie 4 and this new model. Obviously, maybe they've rushed out this new AI because they were worried about the stock price thing. They were worried about having to do results. But for right now, Google, show it to me first, and then I'll believe you, if that makes sense. Well, the genie is out of the bottle now. The genie is out of the bottle. From Google DeepMind. Do I write headlines for The Verge or what?

No, you really got it. They're claiming it's their most realistic and interactive world model yet. And we've covered every iteration of genie. This one looks particularly fascinating. It's a world model. It simulates our world. You can prompt yourself or use an image and drop yourself into an interactive version of that. It's basically like live streaming video where you can use your keyboard, your WASD, your W-A-S-N-D, your arrow keys to sort of move around or provide a prompt that changes the scene that you're looking at. So you could be riding a motorcycle on the highway and then say, I want a big Katamari to come rolling down the road and picking up all the cars on the side of it. It's going to try to make that happen. Yeah, there's a lot of really cool examples out there. There's a couple of people that were in the kind of trusted users version of this. And again, it is not rolled out wide yet. But Ethan Malik, who has a great new book out, you should go get Ethan Malik, was able to like get a mouse serving tea at a table and had a porcupine walk by. There a really cool example of like you can step into a photo And like one of the people I think at Google actually showed a photo of his mom who had passed away and it was kind of walking along the beach with her I think in a really nice setup

The one difference here, Kevin, from previous Genies I think is really interesting is that if you remember Genie 3, all of this was kind of about playable worlds and Demis Asabas tweeted about this and the idea was always like, hey, it's like a video game that gets generated in real time. This time, it's an interesting combination to me of those kind of like worlds, the kind of like real-time video models that we've been talking about. Like you see things like Fowl H3 Max, which isn't real-time, but it's fast. But like, remember, I think it was CREA or one of the other models had this thing where you could type in things in real-time and watch the AI image change in the background. Yes. I feel like what this is is like a video version of that, and it's really interesting in that way to think, you've said this before, right, Like prompt a thing happening in real time in an interactive experience would be really a fantastic thing if it worked. And suddenly that thing, because like the porcupine thing is really interesting. We're like, oh, there's a porcupine. What if I go to interact with that? Can I jump on it and go ride on it? Like, is that like suddenly become a mount that I can use in this game?

Then it starts to, you really start to change the idea of like how you think about what AI video can do and interactive AI experiences as well, I think. Yes. Can this AI video system be an interactive game? Can it be a live streaming event that multiple people are involved with? Yeah. I'm seeing the same thing as interacting with it. Can it be the metaverse? Can it be the Grand Theft Auto 40? Seriously, where I'm driving the car and then I say to my buddy, hey, but let's go watch a movie about the heist that we're about to do. And you go into a Cineplex and now it's just streaming the video. Like, I don't know. Every pixel is going to be streamed and probably generated by a model like this in the future. There's no reason this can't be your desktop, and then it turns into a game world or something else. There's no reason it can't visualize all of the things in your life. Yeah, you know, that makes me think a lot about how one of the things people don't realize sometimes is that if you've played games, video games for a long time, you know that graphics cards are really why, in a lot of ways, modern AI is happening, right?

Because people created these cards that were specifically for graphics processing. That was NVIDIA and all these companies, and now NVIDIA is the most valuable company in the world because of AI. It's another interesting angle of how video games might change the way that we use AI and interact with the world. Like Demis Hissabas is a very famous video game designer and now is the whatever chairman of Google DeepMind, but it was a big part of this process. If world models become the next iteration of this, Kevin, our video game training is going to be so great. even if we're in our 70s, you and I will just be imagining the most incredible worlds. So valuable. So I'm so glad I grew up a hashtag gamer. So glad I got bullied and jammed into lockers for so long. By the way, not just because these things might power the future video games, but they're actively, like all of the major companies, are using video games to train their world models right now. In fact, they're offering money to indie developers who can develop their own engines and develop their own games and simulations because they're going to pipe that. back into the machine. It's a weird kind of snake eating its tail moment of the games powering the video models,

which could power the models in the future. Yeah, yeah. But I'm here for, I do think, I think the future is probably a hybrid where there are some core game engine mechanics for inventory management or coherency or multiplayer networking or whatever matched with the video model that's just going to stream the whole world. So I'm not going to be the one who builds it though. I'm not going to be. How come? Why not? I'm not big smart. I don't know if you know this about me. I'm not big smart. That's what the trillions of super intelligence are for, Kevin. I just want to say one last thing about Genie before we go. In the blog post, which you should read, it's very long, a lot of really good examples from Google and Genie 4, they show video of a robot training in this model. And this is what it's really, really being created for. We can talk about riding porcupines all day, and that's fun, and the idea that we would want to do stuff like that. But really what it's being created for is to simulate the real world. and in this video they show like a robot like putting pipettes into the like different you know um i know a lot about pipettes because of you can see everything now i went deep on uh elizabeth

holmes's story so i know a lot about pipettes but like if you can set up a system where i keep thinking there's going to be a job where like the job is like is like imagination engineer and the idea basically is human brain only it's kind of like a version of prompt engineer but it's a better version for most normal people where your job is to come up with weird scenarios that the ai wouldn't think of on its own so like in that video where like the pipette's happening like what if an earthquake happened well then you have to create this simulation where you know there maybe there's like a bio maybe there's a bio weapon that they're trying to test and like you can't have the bio weapon fall anywhere so like i think that's a really interesting aspect of this genie 4 stuff that like it's that entry level point somebody said like genie 4 feels like the GPT 3.5 version of this interactive world model and it kind of does. Yeah, it's a smart, I think that's kind of a smart analogy. I like that. I think it is a watershed moment for these models and we have now reached a watershed moment not only for this particular show Gavin, but I think our podcast at large because

I just found out don't yell at me for being late to the party, okay? Just be glad that I showed up. I found out that people can like and subscribe. Leave a comment. What? They can back us? We have a thing called a patron? You can get in there? You can get in there. You can throw a couple dollars in a tip jar. You can help us out. And I didn't even know. And shame on me. I mean, really. But thank you to all of you who have taken the time to do that because it's the only way our podcast grows. Literally. Well, that and Gavin making AI documentaries that people care about. Outside of that, there is zero growth potential here. I have a new one coming on Monday that I think is interesting. It's about AI and education. and children's education, but I do appreciate every single person who has given us a dollar on our Patreon and all that stuff, so please, thank you so much. Also, we have a Discord where we show up from time to time, and if you have any questions for us or you want to know anything, that's a good place to go to. We will put all the links in our show notes. The moment I'm done with getting those milk and cigarettes, I'm going to be in that Discord.

I'm going to be there. Fair enough. All caps, and so excited, but thank you. Please, juice the algo. Leave a little algo juice in the comments below. All right, let's get to GPT 6.1. Tara. Yeah, tell me. So this is, the very basics of this is the GPT-6 models are now getting upgraded to 6.1. No Astra yet. We know that that was a problem. But Tara is one step down from Sol. So if you remember, the GPT models right now look like Astra top, Sol second. So Sol kind of makes it like opus level. Tara is kind of their sonnet level. And then there's Luna, which is very much like the haiku lowest level. So, Kevin, you actually spent some time with this. This model is, people are saying it's actually pretty good for computer use, and it's very good at running your computer fast to do a lot of things like Blender or graphics and things like that. Tell us your experience with it. I mean, that's precisely what I built this Katamari demo for, was to show off how good it was with computer use, specifically with making models in Blender,

because that was the big thing that they showed off with this. I mean, their launch video, OpenAI's launch video has the Terra model going in, using a 3D modeling piece of software to make like a little bust. Makes like a little thing. It's like, oh, it sees it and it builds it. Okay. So I went with it. I had it running as like a goal on a loop for almost two days, maybe three, inching on three days. Wow. Oh, that's how long it worked on this. It worked and worked and worked. And so two things, one, like I was so annoyed and frustrated when I would give it a creative direction that would work with a soul level or an astral level model, or even a sonnet, uh, if you're going for anthropic, like it would, I know usually how to whisper to these models to get something back from it. Right. And it was just not understanding the mission at all. So at one point I gave up and I was just like, listen, go ask Astra how to communicate, like what, what I'm asking you to do, just go ask it how to fix it. And Astra came back and was like, oh, you should build it in this way. But Terra spent some time doing the ripping and the Terra to make all of the 3D models and all of the world that are in this Katamari demo, which you can go play right now.

I'm playing it right now. Katamari.perswap.com. I like how the guy is like kind of waddling. That's a different look. Yeah. The Katamari Prince is a little waddly. At one point, I did have to yell at Terra and say, like, listen, all of your this is supposed to be a showcase for you, which is what I reminded it. All of its models were coming back like the knockoff pinatas that you have, like a Spongeb or Sanic, the Hanghog or whatever. Like it just looked kind of a slightly microwaved meltish version of whatever. But when I really yelled at it to like try hard, you'll see that the concept is you're rolling your Katamari ball through the different eras of video games. That's a hard idea. Pong. Thank you. Thank you. Yeah. And then you'll go through like SimCity, whatever else. You'll see like Chun-Li and Ryu in there from Street Fighter. and they're actually modeled pretty well. You'll see some really terrible Goombas from Super Mario. What I really wanted to do, Gavin, was Katamari MCP. And the idea was that you could connect your agent to the Katamari world and say, let's add this into this area.

Or let's make this. And it would use your tokens to contribute to a huge Katamari world, which everybody can roll around. And there's no way Terra's going to build that. But if you want to see what Terra was able to make, and here's what I will say. So it was quick with computer use. It is very fast for computer use. So kudos to that. Despite me running it for almost two days on a loop in the background, my $200 open AI plan, like, I think went down about 10% of usage. Wow. No kidding. That's amazing. That was good to see. Yeah. So it is an efficient model. It is not the gee whiz, let's make the best looking pelican on a bicycle, like a crazy huge kind of thing. But for what the model is supposed to be, I could see why people would run it and, again, use it specifically for computer use tasks. So, by the way, if you want to make Katamari MCP, please, Internet, do it. Do it. Yeah, we should tell people you can go to katamari.kurslop.com. We will put a link in the show notes and make sure people can go play. It's very cool.

You know, that makes me think, Kevin, of one thing. So I've been working on this AI doc again, and one of the things that is really weird is trying to get the system to create graphics on it so I want to make the graphics look interesting. One of the things I thought about is you went deep on this one specific project and let it go work in the background for a while. The other thing you can think about is breadth, right? Like if you said, like, hey, make like 50 different versions of these graphics. I want to just be able to see a bunch of them, and I can give you feedback on them. If intelligence gets that cheap or it can just continually work on stuff, that then opens the door to thinking differently about how you might iterate on ideas, right? So if suddenly it's like, okay, well, here are 50 fonts and 50 different, like, interactive ways that these graphics could play out. I pick these 10 as my favorite and I'll merge those into one. Like, that's another interesting way to think about really cheap intelligence and using it in different ways. I'm still fully agree, and I'm still baffled by the speed with which we are using this intelligence. Like we are still very much in dial-up phase, right? Yes. It is not – when this stuff becomes instant and on demand, and I'm talking about like the models now that we marvel at being truly instant, thousands of lines of code or thousands of polygons into a 3D model or the instant music composition to be able to in real time iterate on your idea.

And like you said, go give me 100 versions of this graphic. Okay, version 73 is the one we're going to focus on. Make it more this. Make it more that. Add a little more carbon. whatever it's going to be. I get very, very tickled for that. I don't know that it's what's trading the ocean for, but I get very tickled. Oh, well, you know, that's a whole different thing, Kevin. We'll talk about that on the Draining of the Ocean podcast, which is coming next week from the AI for Humans channel. Yeah, our Steve Staros side channel. We're going to fill the oceans back with our tears. We know AI hate works well on YouTube, so we're going to flip our personality and just do that there for fun. One other quick thing from OpenAI, They have now turned dots into a messageable thing in iMessage, and you can now text it. This, Kevin, this might get me to actually start using dots more. I just saw the news come out. Duet is hopefully coming to me soon. I just got a text message from a text chain that we're on where somebody else got into Claude Duet, which I hope I'm getting into as well. Duet has been one of those things that I have been waiting for because, like, I will say

in terms of personal agents Instinct was so cool in the beginning I was like this is great I can do all this stuff It saved me a bunch of money on these recurring payments And now I in this place where like ah I not using it that often It not being proactive enough I feel like if one of these agents does lend itself into my regular AI work and my regular stuff, that would be a way to be more proactive about it. I'm very curious to see what dots will feel like and what Duet will feel like as part of my work life rather than part of my personal life right now. So I think that's an interesting thing too did you use dots yet or have you i used it in the beginning when we first came out i used it a little bit and i was pretty disappointed what it was not since no have you been using it at all no no i mean listen i use gpt voice mode i use their live model all the time all yeah i just don't see the differentiation with yeah regular dots versus using that mode i think that's the big thing putting it in messages like i think that is a much more normal place for people to interact with this sort of thing. So, like, putting it in messages does change slightly.

Okay, Kevin, let us change over to a couple quick news stories. First, there is a new AI audio model from Jack Dorsey's company, Block, and it got a lot of people excited about it. Do you want to play, let's just play one of these examples to kind of give us a sense of what this is. How you doing, Renaldo? I'm good. How you doing, Renaldo? I'm good. I'm good. I'm not Renaldo. you're mixing me up with another guy with that no i'm i'm to be here man i'm not ronaldo no you're not ronaldo no i am not how you doing man i feel like i'm listening to an old jerky boys right now but i'm gonna just jump in this is the guy who's saying like oh you're not ronaldo that is dessa that is jack dorsey's new ai model that's the ai play a little bit of this that's the ai yes Nah, man, when I get to work, bro, I have to get in there and pick produce and stuff, man. It'd be tiresome. Work is tough.

Yeah, work is tough. Then I got to set up. I got to put all that stuff away by myself with no help. I feel bad for this day. Okay, so let's talk to this day. My initial reaction to this, too, was funny because I tweeted this out. I was like, I felt all of these, you should go listen to all these Dessa interactions. So the one thing that people are talking about is very specifically, these feel like real human interactions, right? These feel like there are interactions with like real people. It feels much more realistic. There's some people who are saying like, it is weird that these have different accents and different things. And that might be a problem depending on who, you know, who's making them and what their vibes are in terms of like, is it too far in one direction for a specific race? Is there a feeling that there is a level of people saying, hey, that's not okay to do this? But they're clearly achieving some sort of level of humanness here that is new. The thing I thought, Kevin, when I listened to these is like, okay, but like part of these demos are all like, hey, this is, it's pretending to be a person, right?

Like in the one you just heard, like it's like, like. He has a whole life, a whole job, a career. He's got a wife and kids that he talks about in the demo. I know. It made me think, it's like, well, is this just a hallucination of this kind of experience? I guess the question becomes is like use case for all these different things, right? Like I, you know, there is a level of what it feels like to talk to a virtual being, an AI, whereas now you're getting this kind of very like kind of edges sawed off, kind of very vanilla, almost like kind of corporate voice from the opening eyes and clouds on purpose, right? There is this whole world where like for AI companion apps or maybe for some sort of gaming system or things like that where you want a realistic interaction. Like I was thinking about the GTA 6 stuff where they talked about, like, there's just crazy stats coming out about GTA 6. There was an Edge preview that came out, and one of them was just some insane amount of NPC interactions. Like you could imagine a game where, like, a character like this could be in a video game and actually could be finally that thing where AI makes video game characters much more entertaining in some form.

Yeah, and instead of 50,000 lines of pre-written dialogue, it's just the actual personality, the prompt for that character. Anyway, they're not out yet. They're not out yet, but you can – hopefully we'll be able to try these soon. Again, it's like there's a lot of people pushing forward on all different aspects, and voice is one of them. So one of the things that they mentioned was they managed to train these models, like, on a very slim system. It did not require all of the compute that Colossus has. It was like one NVIDIA small server, basically. And I keep coming back to, like, look, they announced new MacBooks the week of this recording, right? There's a new M6 they're up to now. What is it? M16? What are they at now? I don't know if I'm touching my screen. You can touch the screen and it'll waggle. Look at this. Oh, God. Oh, no. Jeez. Wow. That's our first real camera blooper for a while. Well, I got a pillow behind it for acoustics. So, yeah, they announced the new MacBooks. You can touch the screen. But part of the big selling point, as much as they want your fingerprints on the screen, was AI. It was AI.

You can run AI workloads. And I'm like, that is really cool, except while they're over there in this ecosystem, which is getting better, I'm a part of it. NVIDIA just keeps releasing cool thing after cool thing for free that runs on their hardware. And so two things I wanted to shout out. Did you see the video for Nemotron 3.5 Transcribe? Very cool. The Transcribe is very – I saw this. It's very cool. Yes. We'll put it on the screen because the audio is a bunch of different people talking at once. but you're watching a model that can run basically on consumer hardware, identify up to like 10 different speakers in one scene and instantly transcribe what they're saying and put it in a color on the screen. Like that is just crazy interesting. And I can think of some very fun use cases, but you got to have NVIDIA if you want to play. And then they released OpenH3 Regenerate 2K. I hate these names, but basically it is an upscaler that you can run locally. And if you're, I mean, I know you're doing a ton with AI video, It gets expensive to take low-quality, quicker-draft video and make it something that is enjoyable to watch, even on a small screen.

Well, this is a free upscaler that runs really well, but you've got to have NVIDIA to do it. And so I'm like, that is an interesting play for them. I mean, NVIDIA is in a fight a little bit because there's all these – another big company, I can't remember what it was, but there was a big TPU announcement, which is Google's chips, that another company is going to spend a lot of money on TPUs. It might even open AI, but one of these companies is spending a lot of money on their – We should speak about, speaking of interesting things happening with AI, one of our heroes, I think my hero, I do want to speak for you, but, like, one of the most creative people I think that's been born in my lifetime, Shigeru Miyamoto, the guy who is behind Mario and Link and a lot of, like, amazing stuff from Nintendo. And I got to meet him a couple times when I worked at G4, really, like, one of the best, like, one-on-one experiences of my life. Good guy and a nice guy. You, sorry, coming out of that, like, he basically was asked about AI. and I think that this is one of the best answers I have heard about it. Like, it's a lot of fear, but, like, I think one of the things he says is, he says, as I mentioned earlier, there's a long interview,

so you should go look and read it, but this is an era where people who are not game designers can create games. If we broaden the definition of game, I think the era has already arrived where everyone can make games. I had always hoped for a world where those boundaries disappear, so I think this has become a very good era. And I think that is, like, this human-first idea. about how AI enables people to be creative. Now, I think there's things in here where he's not saying, like, it's the greatest thing ever for games, but, like, the idea that many, many more people have the chance to now make things that are interesting, first and foremost to them, and feel that kind of creative urge and stuff coming out of them, that's just a really positive thing for humans on planet Earth. And I'm very excited that this guy, who I admired for a long time, has come out and said what I assume would be his thoughts anyway. You know what I mean? Yeah, and I didn't see the collective internet calling for his resignation. No, it's because of Miyamoto. They're not going to do that with Miyamoto.

Right. That's what I'm saying. I'm glad he's saying his truth, but it's interesting that people are, oh, there's some nuance in there. He even discussed how at one point you were called a designer if you knew how to place text on a screen because that was a big deal. could you put font onto paper or then eventually get into a computer and have the layout to go. And you were a designer. Now you might not call somebody designer simply for honing that particular craft. They might need to have more tools in their tool belt. So he doesn't seem concerned. I mean, he doesn't specifically mention, like, the generative aspects of it, but he is taking the step back and saying that we're in a better world when anybody can create, which I disagree with. And I think, on the other hand, Kevin, we're in a much worse world when we have a leader of the free world asking everybody to call something by what it actually isn't. I just wanted to rant about this briefly, and Kevin has something you'd like to talk about in this world, too. And, again, we're not – don't go political, but in this case, it is a very political thing that's important.

The President of the United States right now is convincing everybody that we must call artificial intelligence superintelligence. And the reason I want to talk about in the show is that in my comments for both these AI docs and for this show, there are things in those docs where I have said, you know, super intelligence. And I want everybody out there to know super intelligence is actually a real term than artificial intelligence in the definitions. If you go to Wikipedia, look it up. Super intelligence, as we heard from Sam Altman at the top of this, is a intelligence that is vastly smarter than human intelligence. And when the President of the United States makes a decision randomly to say, like, everybody has to call something something, it is one of the just most frustrating things in the world to have, like, somebody dictate their idea about what something should be called, and then everybody else has to follow along. So, Kevin, this is a very dumb thing that is happening right now, and I think we just need to kind of, like, push back on it in a big way. Yeah, I mean, I'm a patriot, so I'm not really going to get into that.

I mean, you've got your view, and that's fine. There's a beautiful thing about this melting pot we call country, Gavin. You're right. We're all coming from different ideas, aren't we? I joked that you and I were going to change the name of the show last week, and people in the comments, I think, were big mad. Like, don't you dare? Or, you know, it is odd to say we're officially banning something. So stupid. Because, by the way, some of this was about, like, well, no one wants artificial intelligence because you don't like the talking point was artificial sweetener. No one likes anything that's artificial, right? And so I guess this is going to make people feel better about data centers in their backyard when they don't want them. That was sort of the thing. But it's for super intelligence. Oh, okay. Super, I get it now. Like the Nintendo. We don't get political, but this is kind of political. Something that I made and I'm just going to release into the world and see what happens. Yeah, this is actually very, I think it's good. I think this is a good thing. Tell people about what this is and kind of what your idea behind it was.

Yeah, so on the backdrop of I think a lot of people expressing frustrations with, let's say, data centers going in their backyards or their employers surveilling every correspondence or keystroke or flock cameras, let's just say, going up everywhere, even though communities are banding together saying we don't want them here. It seems like people are getting very understandably fed up with having artificial intelligence kind of jammed into them like they're making foie gras. And they don't seem to have that power back. And that's where I started to flip it. I was like, listen, if this tool set, if these technologies are good enough to surveil you and run analysis on you and sweep up all of your data and give big government a clear picture of what you're doing or what you're thinking or what you're experiencing, Well, why don't we use it against them? Because they technically work for us. And I think we forget that all the time, right? There's this weird balance of power. There's some magical things. So I created a concept called Seveil right now, and I'm giving it out to the world.

And I hope someone who actually has time and talent will do something with it. I would love for that. But the idea is let's use AI to capture the communications, the business dealings, the receipts and write-offs, the budgeting and the bids that companies make for everything targeting small communities for now. There's no reason. The technology is very cheap. It's securable. It's easy. It's an easy system to build. I've built systems like this for companies as a consultant. It easy to spin something up where you can go in and check the work of your local elected officials and see how this project doing What were the bids like Why did we go with this What has stopped What happening What were the Slack communications or the email chains that led to a decision being made I want the notes from your meetings. I want to see the receipts and the write-offs. Like, why wouldn't we have access to all that if our tax dollars are going for you to do your job? And because it's a government, that stuff is available in the public domain. A lot of that stuff is public domain. And that's where you can use AI to actually access this stuff through Kevin

School, but through other stuff too, right? So it is an empowerment tool. And I think that a lot of people are feeling disempowered, right? This idea that billionaires are deciding our lives or different people are doing stuff. We have to use AI to empower ourselves as other people might be using it to disempower your community or other things. You can also empower yourself. So this is, I think a really cool thing at large for people to go explore and also just maybe get fired up about the idea that you have control over aspects of your life that you may not have before because AI tools are available to you. Yeah. So surveil.curseslop.com. We'll put the link in the show notes. There's also a Reddit that I made where you can go and have a discussion about it because there should be a place where people can value their communications. And for those thinking, why not right off the bat, there is a page on there called, could this work, which goes over the top five pieces of pushback for no one's going to agree to this, or communication has to be private. This is more than me just saying, hey, GBT, make a wage against the machine for me.

GBT, make a bomb track so I can fight back against government. There's actual ideas in there and systems in place and thought with how we can keep things private, but also save taxpayers money with this. So whatever, I'll get off my soapbox now. I hope people check it out. I hope it inspires anybody to, again, think about flipping the power dynamic using these tools because there's no reason that we can't, especially if you're from a small community that's pissed off about data centers. Rally some folks together and use these tools against the people that put it in your backyard. And, again, it's not like we're saying like this is – it's not like, you know, one side or the other is bad. It's just information that you need to empower yourself with, right? Like use this time right now to use these things that are out there doing stuff to not only make your life better but also inform yourself better. And, Kevin, that transitions us pretty nicely into Science Corner. I just got to be clear, it is a different corner from Business Corner. It is a different corner than Business Corner. And we have a different background briefly. Now we're back to where we were. So this is now no longer Business Corner. We're in Science Corner. Kevin, I will say, optimistically, there was a really cool post.

I know some people are feeling weird about it. But in general, OpenAI dropped a post last week where there are three brand new discoveries that are very complicated, very interesting things. But in the past, we talked about how it had done all these advances in math. Well, this week, they had an advance in math. They had an advance in physics, which is the first time we're seeing some of that. And then they had a giant advancement that I know got a lot of, like, academic nerds feeling in their feels about computer science. And the one about computer science, I think, is probably the one that's worth talking about. I think I understand a little bit about what's going on. So in computer science, there is a thing called the exponential time hypothesis. That's ETH. And in the past, there was this idea that basically all computers in the world had a limit at the size of the sort of problems they could solve. Right. And there was this upper limit. And what opening eye has basically said now, this is like the very entry stage to this is that upper limit may not exist. Now, there's a very simplified way of explaining this. But I know tons and tons of computer scientists and academics were kind of like, you know, running around saying this is a crazy thing.

It's going to open up the doors to a lot of things. So I do just want to make sure everybody out there knows that, like, we talk a lot about, like, the crazy data center stuff, or we talk about, like, hey, you can change your work. But it's also really going to fundamentally change what's possible in our world at large right now, which is a pretty cool thing. This ETH thing, I went deep on it because I could not understand it. Again, hamster on wheel running very slow. Couldn't understand it. And I think I grasped it enough. Basically, like if you imagine a row of switches, there's three of them, right? I say one needs to be up and two need to be down or whatever. You could go through the permutation. OK, switch one is up and switch two and three are down. OK, switch one is down and three is down, but two is up. And you could figure out the possibilities of that, right? OK, I can figure it out and I understand how long it might take a computer to figure that out. But when you start adding like up to eight switches or 16 switches, suddenly the permutations. Complications. Yeah, the amount of complications. It's more atoms than we have in the universe. And the thought was for years since the 70s was that that's a fixed number.

Every time you add complication, it's going to take any algorithm more time, predictably more time to get there. And so there were so many math problems, so many physics problems, so many everything that got lumped into this ETH thing, which is just like, ugh, there's no way to elegantly solve this. Yeah, you just can't solve it without a ton more compute and a ton more time. And what OpenAI, their hypothesis is, is that, yes, you can actually. There are ways that you can manipulate the algorithm so that as the complexity grows, the amount of time to solve it doesn't. I still barely understand that. Me too. Me too. That was pretty good, though. That was pretty good. I feel like you did it spiritually, emotionally. But what it could mean practically is that all of these problems that the science community has just sort of swept under a rug or put in their science corner, now they might need to revisit all of them. And that has huge implications for people much smarter than you or I. A company called Physical Superintelligence has dropped five new physics possible solutions in quantum information. And then Anthropic has talked.

It did a long post about drug discovery, Kevin. And this is a really personal thing that I thought was interesting. They talk about how allergies, drugs that are allergens might be able to be discovered quicker based on some research they draw. In my family, I think I've said this in the podcast before, except for me. my wife and my two daughters are celiacs, right? And for celiacs, you are allergic to the gluten molecule. And I know some people think of gluten allergies still, which is crazy to me after experiencing it that like, it's like, oh, it's a choice you make, you don't eat bread. My family can't eat bread because it actually affects their immune systems. And one of my daughters, especially if she has even a little bit of bread or a little bit of gluten gets very sick very quickly. This could literally change our lives in the future, right? So this is a good example of a very personal thing where if they can figure out allergies, but you think about people who have peanut allergies or shellfish allergies where like it is deadly, that if somebody can start that, then we're starting to get that feeling that, oh, AI isn't just about these big boxes in my backyard that are loud and are making noises, or it's not just about the AI

slop videos that I see, which we're going to tell you, we're going to get to it. But it starts to change the world and then you're starting to feel like oh i get it my daughter now can eat whatever she wants you can have a chocolate croissant again because there's a way that they figured out uh to control the celiac molecules in or the root molecules that is a huge huge thing now we're at the baby steps of this but like even you know rune the one of our favorite uh users of of twitter who works has worked at open ai for a while said something really interesting which he He said, there will be, within the next few years, stunning advances in biology. There will be whole new ways to think about what biological study even means. I would only be mildly surprised if there were near-term solutions to problems like cancer or neurodegenerative disease, et cetera. That is pretty awesome if that comes to fruition. That's pretty awesome. That's what we wanted. That was the utopia from all of the BS that we're now wading through with myriad problems from all of these systems and politics and business.

That was what we all wanted. And I'm trying to just put aside the double standard because, you know, my 20s and most of my 30s were all about drug discovery as well. And nobody cared. Well, now you've got the remnants of that still there, right? So you understand what it's like to meet those little people that live in the metallic world, right? Isn't that what those people – what are those little green goblins that decide to ascend to the next level? Yeah, Gavin, I think I know what they're called, okay? Hey, let's talk about what you did with AI this week. It's time for AI. See what you did there. Sometimes you're strolling without a tear. Then suddenly you stop and shout. AI, be what you can't do. AI, be what you can't do. Okay, Kevin, speaking of the opposite, we are talking about high art today in the AI slop world. And I first want to discuss one of our greatest AI creators that lives out there.

We've been talking about him almost from day one of this show is Demon Flying Fox. And I don't know, Kevin, have you been following the female Draco Malfoy story? Have you been following this at all? No. So there's a whole world right now, mostly in the fan fiction world. There's a whole world in the fan fiction world. They are shipping a female Draco Malfoy. and I think her name is Draconia or something like that. Yeah, yeah, yeah. Yeah, yeah. Okay. Okay, so Demon Flying Fox, who made the original Harry Potter Valencia video, has now released a third video in which the female Draco Malfoy takes prominence and a main role, and the whole video is very awesome and good, but you just get to see, like, Demon Flying Fox use today's modern tools, and this video is so fun. I really encourage everybody to go to YouTube and watch it. It's just, like, six minutes of great, great video. I'm laughing so hard because like your passion for memes sometimes is so real. And like your explainer on exponential time hypothesis, like, look, I love you, but it was fine.

But you going deep on demon flying fox and female Draco Malfoy Balenciaga mashup is you found your home, buddy. I mean, this is the beginning. I want your tent topic. It's fantastically done. So go watch that. But there's another great thing I found on X where a guy took all of the power of Opus 5.5 and put it. Oh, this one's cool. Instead of, like, developing a world, developing a game, he just made it the best possible simulation of a single eye he could. And, Kevin, there's a little bit of a creepiness when you look at this eye coming to look at you. But what's cool about this is you. With the Pringle wedge of skin around it. Yes. That rotates about. Yeah, it's a little unsettling. But it was just a cool way of, like, some random curious guy was like, hey, I'm going to make, instead of going broad, and big and make a whole person or a world. I'm just going to make the best version of an eye I can. It's just very interesting. It's worth it. It feels like a Peter Molyneux game that was announced in 2004 that never came out. God's eye? What shoots out of the eye? Oh, it's a laser now.

Oh, it's a cloud of happy feelings. Do you have your people down below? Oh, that reminds me, by the way, I want to make this at some point. And I've been thinking about, for a long time I thought about what would be a cool thing is like, you know, GTA, obviously GTA 6 is coming soon. GTA is so big and broad. I still want to make a game where it's set in one block or one building. But instead of going, preach. You know what I mean? I think you and I talked about this. I think we did a long time ago. Which was a long time ago. We were like, just simulate one building. Just one block. And give everybody, because like the Shenmue from Dreamcast was like, oh, everybody's got their own schedule and their own business. It's like they kind of do because they show up on a street at the same time. No, but you go deep on every single person on that block. And I want to be able to go into every room and open up every drawer and have every little trinket mean something instead of just being a polygon to pop. Yes, all day. Yes, let's do that. Yeah, let's just think about that. Maybe it starts with the eye.

Maybe it's just the eye at the beginning of that. Maybe we have to spend three weeks on every body part. That's a problem. Anyway, last thing. We've teased this in the beginning. There's a great clip going around. And, again, we talked about a couple weeks ago about how Higgs field has populated the world. This is just using the office parkour clip but putting Matrix, Keanu, putting Neo and Agent Smith into the parkour clip. The Oracle's in there. Like, just so good. Yeah, Morpheus dancing around. They got, like, handheld camcorders. I love it. Again, it's the remix of the world that's happening in real time. We saw a touch of it when Sora 2 came out, but now we are here. just a very fun moment of chuckle for me when I saw this come across my timeline. And maybe there's some things that's real value in that, right? Like it's not going to change the math of the universe, but like I got a smile and that feels like something. Maybe that's what you do when you watch us. So thank you for watching us. We appreciate it. And we will see you all next week. We appreciate it, everybody. Hope you had a fun moment of chuckle. It was very calm and very quiet. Okay, bye, everybody.