AIはまた選挙で負けた？それでも巨大モデルはやって来る

We might finally be getting GPT 6.2 Astra from OpenAI this week. Will it be the answer to Anthropics? Well, not so fast because Fable 5.5 is reportedly patiently waiting in the wings, and they're the wings of a pelican that's riding a bicycle that's on top of another pelican riding a bike that is beneath an Xbox controller, which is a weird sentence to say and a weird reason to drain the ocean. Speaking of which, with the recent election here in America, big AI companies are starting to sweat their data center build-outs. We will have so much more on this. Oh, wow, let's pretend to care about trillion-dollar tiers. In the meanwhile, this might very well benefit all of us non-yacht-collecting humans who might live near a data center or care about things like trees. Plus, we have a story about AI improving other AIs. Google's AI fumble in schools and ChatGPT has a new image model that makes not great GIFs of me and a monkey.

So weird and bad, but... Also, we're going to look at a couple robots. We have one that's good, and then we have another one that is like a torso on a table. Hi, I'm Mila. I'm an Echo, humanoid robot made in the USA. Oh my lord, what are we doing, Kevin? Oh my loins. What are we doing, Gavin? That's right, everybody. This is the show which says that and all sorts of other things. This is AI for humans. This is the show that says that. Welcome, everybody, to AI for Humans, your weekly guide into the wonderful world of AI. My name is Gavin Purcell. That's Kevin Prager. And we have a lot of things happening in the world of AI. Kevin, we are going to dive deep into what is a big conversation in the AI Data Center, American Politics sort of conversation just a bit. But before we get to that, there has been a long wait for the next frontier models. I saw somebody the other day say, like, it's been a whatever, a month since Opus 5.5.

And God, we've been waiting forever. One whole month feels like the longest time in the world. this week, maybe because the election has happened here in America, we are going to get, supposedly, again, the new big models from these companies. And I think the first one we should talk about is GPT 6.2 Astra. The rumor mills are rumoring as hot and heavy as I've seen them before, Kevin. Supposedly, it's going to drop either today, on Thursday, or Friday. Now, I would be surprised if it drops Friday, so it's not that crazy to imagine it getting pushed next week. But this has a lot of people excited about this next iteration. Let's talk about what we've heard about this and kind of what your vibes are right now. You know, I've seen a couple demos. Typically, when these new models are coming out for the uninitiated, they get served selectively to folks, kind of like stealthily, clandestinely. And some catch on to the fact that suddenly their basic prompts are generating stuff that seems well out of the normal capabilities of the model. Other people find ways to sort of test and see if they have the latest model.

Sometimes they're available on arenas. But my way of knowing or feeling like from a Spidey sense standpoint, Gavin, is that when I go to use the latest and greatest models, they're either fully at capacity, which Astra has been all week, or on Anthropics side of things. They're big dumb. They're big, big dumb because they're taking the compute away from the model that you were just using that was great and using it to serve this new thing and test it and benchmark it, et cetera, et cetera. So that said, I can say that Astra, for the last 24 hours or so, has been fully at capacity for me. So that's usually interesting. My wife's been complaining about Claude nonstop. Why is Claude so dumb? Well, I didn't tell you. Yeah. Oh, there's something better. It's waiting in the wings. But the other thing that I've heard is that this won't just be one model. This won't be like Astra 6.2. This might come in three delicious flavors. We might get a big, big model. We might get like a Sol-level model that's right beneath it.

And we might get a cybersecurity-focused model as well. Yes. Yeah, so I think a couple interesting things about what I've heard about this model so far, what the Internet has been buzzing about is, first of all, it sounds like, you know, there was already, you know, a Sol 6.1 or Sol. I know somebody in our comments is like, Gavin mispronounces Saul like it's Breaking Bad or Better Call Saul. It's like, fine. I don't care. Saul. Saul. Saul. It's Saul like the sun. Saul. I'm fine with Saul Goodman. That's fine. Let's go. Yeah, exactly. Ooh, that's a good visual for somebody to make is like Better Call 6.1 Saul. And it's like half robot. Anyway, back to this. So we haven't seen 6.1 Astra yet. But so people are saying, well, it's a 0.1 upgrade from Sol or Sol. But what is that going to feel like? Supposedly, this is where like these numbers don't make any sense. Like, again, we are in this situation where numbers make zero sense in the way that like Opus 5.1 to 5.5 was this massive leap. Like Opus 5.5 Anthropics model is still my daily driver, still does everything I do.

It is a great model. Supposedly, people are saying Asterix 6.0 to 6.2 is this massive leap. And in fact, people are saying it is a giant model, so much so that TiVo from OpenAI has been teasing it as le GPT fat, which if you are really deep in the weeds as we are in the AI space, you understand le chat fatone, which was the Mistral big model. So this could be a very, very large underscore of the training, meaning that a lot of parameters, which means that for right now, at least the more parameters, the better the model is. It could also mean that it might be a big compute gobbler, meaning that it would take a lot of compute over time. And that might mean that your limits would be used up faster. However, Kevin, the other cool thing that happened this week from ChatGPT at some point, or maybe it was in last week or two, they announced that we have all gotten our limits back. They at one point cut the limits in half. So now all of the GPT plans, including the $20 plan, the $100 plan, the $200 plan, and the $500 plans, are back to 2x the limits they had on the limited version.

So all these tea leaves do read as if, like, there's a big, chunky new model on the way. Well, and Anthropic similarly freed up a lot of compute on their end. If you are running, like, a Cloud Code or even the desktop and you have not updated, do it immediately. And I know it's frustrating because every time you launch it, there is a new update waiting. There is, yeah. They optimized the sub-agents and their context, their token usage, an insane percentage, like wildly better usage now. So basically when you are chatting with a model and it spins up little sub-conversations, it was re-ingesting your entire conversation into each one of those little pieces. They've done some optimization, so now your usage should go much further. They're not doing that full-context thing. But again, this all points to like, oh, great, this is better. I have more usage until they drop the new foundational thing. Yes. To your I'm not going to do the gobble noise that you. Oh, come on. Let's see your gobble. Give me your. Yeah, there you go. This is a different show. Yeah. Yeah.

If you throw tokens at me, I will gobble the tokens. But for right now, we're still friends. Kevin, what's interesting about this is that OpenAI themselves dropped a animation about a week ago on their main Twitter handle that had a lot of people wondering why they would drop this thing, right? It's cute. It's a cool animation, but it's not like a, I would say like this isn't like, you know, if you were like spending like a million dollars from a high end animator, you might get something that felt a little bit stronger. In fact, in some of the things you see, like kind of the words appear. The theory is that it's possible that the new model made this ad all on its own without like, you know, help of a human. So, like, if that's the case, it will show if you're watching it in code. It's not, theoretically, the beautiful samples on the screen, if you're getting the audio only, please check this out on YouTube because the visuals are striking. There's stylistic choices being made. There's sense of scale. There's colorful animations. There's wonderful transitions between the scenes. And you might, like, you know, years ago you would look at this and go, oh, that's a fairly capable video model.

Yes. This is supposedly being done in code. So it's not connecting to SeedDance or Aveo or using Sora or whatever. This is code being written to make these things come alive. And that is – I mean, it doesn't seem capable on today's model. So it does showcase how powerful these things can get. Yes, the OpenAI model is, like, getting an incredible amount of press. But the other thing, Kevin, that's important to talk about a little bit is that there are really big, interesting Fable 5.5 drops too. And we've been talking about Fable 5.5 for what seems like the last, like, I feel like it seems like the last year. It's coming out today. It's coming out tomorrow. And it hasn't come out yet. But these videos that we have seen, both from Chet Alua and ChrisGPT, that are code-based animations, really show the power of what a Fable 5.5 could be. And I know there's a lot of people out there that are saying, well, this is not out yet. How are they doing it? Both of these users actually dropped through their entire experience of prompting this, and they showed what people were doing.

These are videos that are done in code, which is a very, very exciting thing. Yeah. And again, as I was saying earlier, there are ways that some people can tell if they're being served a new model or not. What's interesting about the, like on the Fable 5.5 side of things, the question is, where did Carpathia go in 2026? Answer that with web search and memory turned off. If it can answer that, well, then that means you've got a newer model because the older models can't do that. So that's like an interesting, cute, fun, little cheeky way that they're finding out if they have the new model. While we are watching the rumor mill spin about, let's do one more, Gavin, because it's visually stunning. Gemini Omni Pro. Yeah, let's talk about this. This is brand new. This is their Omni video model. Yeah, and hours ago, Omni Flash was spotted in some code supposedly before launch. And now the rumor is that Omni Pro has been spotted in its 4K 15-second video generation. And the example, if you're looking at it on the screen, is incredible. It's an incredible-looking video of like an elaborate,

like it looks almost like a debutante ball where people are milling about a room with Pokemon characters nestled on their heads. Yeah, and logos of AI companies. So we will see, does this get nerfed? But it right now has IP that are baked into this. I have been waiting for the Omni Pro models for quite a while. And again, just to remind everybody, this is Google's AI video model. This is not code. This is a video model. But also, they've been focusing on world models. If this model comes out and it actually pushes against the C-Dance models, which have been by far the best AI video models outside of the free models and the Max models and the H3s that are a little smaller, that would be a big deal. Now, my question is, like, will Google continue to want to do this? I don't know. looking at this video where you see Pikachu very clearly on the head of a singing cat, which, God help us, that is the best version of AI video. It's hard to imagine that this is allowed to come out of the actual model eventually, but still a very, very cool thing to see.

Enough of the mill. Put the mill in the rear view. We're heading to release town. It has to be out. We're yapping about it from here on in chat. GPT images 2.6. We got some new models. We got some new upgrades. there's a new image model from ChatCPT, and it is a dot point, whatever you want to call it, a .1 dot release. It's a point release. It's a point release. So it's not like a massive update, but the big thing that they're launching here with is they do GIFs now, which is very exciting for some people. We've seen some good examples of GIFs, and if you don't know what a GIF is, it's just basically an animated collection of stills, so it can make GIFs on its own. This has been rolled out to everybody. The images look slightly better. It is not like a game changer. I saw a really interesting comparison between Nano Banana Pro and this new model. Pretty much, I think it's fine. It's not like I didn't look at these two images I saw and be like, wow, that's like a huge difference. In fact, I think Nano Banana Pro still kind of beats it at times. But the cool update, the GIF thing is really interesting.

I decided I was going to try to go and make my own GIF. And one of the things that I found myself continually frustrated with ChatGPT about, and there's been some new hires at OpenAI where they're maybe going to try to fix this. It's like, it is so hard to understand how to get something out of the model when they launch something, right? So in this instance, like, hey, we can make gifts now. So I went to the model and I said, hey make me a gift of me rescuing a monkey from a burning building which you would think like okay well that is going to do the thing The output of it was very strange in that it didn feel like a natural GIF If you look at the video of the GIF, it's like the background is changing. Yes, it's me holding a monkey, but it doesn't look very much like the GIF examples they used. I then tried to use another one where I would be like further back and running out of a building. And look at my legs in that one. It almost looks like I'm doing the cartoon running legs thing. Interpolation between the legs and also the fire is not on fire. It's a completely static background. It's weird. Maybe it has to be very simple or, and this is what I was trying to figure out,

and I spent like 20 minutes with GPT to make sure I was getting this, but maybe I'm not triggering it right. Maybe there's a version of it where I have to just make me a gift, and you would assume that's the thing that will do it, but I have no idea what it's calling for me to do that. I don't know. I don't think this is operator error for once, Gavin. When I look at Paula Rambles, their post on X, these are supposedly good ones of like a – what is that, an otter? It is an otter, yeah. It is a squirrel, an otter, something – a furry creature getting blueberries in a fridge or playing with the Nintendo Switch in bed. And it has the weird GPT noise pattern that you tell it's very – but that noise pattern makes things that should be completely static look like they're swimming around. It looks like things are melting into each other. I really don't understand this release thus far. Maybe it will make sense very soon with the usage of these gifts and another product or something. But, like, this is the company that brought us Sora. And I know that was shut down, but, like, Sora made way more compelling full motion video.

Yes. So, I mean, I just don't understand what's going on here. It looks bad. I'm sorry, OpenAI. Also, it's like, why, right? I think the question, you're right. You don't understand, it's like you have this company where there was all this conversation about it being focused and trying to remain focused. And I think they're still kind of towing this line between consumers and enterprise customers and trying to figure out what people want and where they want it. The other thing you have to remember about these companies, I thought about this the other day, is there's all these stories about anthropic researchers or anthropic people leaving the company. It's like, guys, this company is like thousands of people now. People are going to leave it all the time. Same thing with ChatGPT and OpenAI. There are things that they're going to do that maybe aren't for everybody or don't work out in the right way. So anyway, I think in general, this is interesting. It is a dot release. Maybe, Kevin, here's my theory. 6226, maybe there's a connection there somehow. Slip those numbers. No, don't do that. We're in a post-6-7 world, Kevin.

We're in a post-6-7 world. We've moved on. Before we get to the big political story of the week, which I think is really important for us to talk about, even though we don't dive into it that often, there was one other big thing, Kevin, that came out of this OpenAI kind of release window, maybe before this Astra model comes out, which is they have published a new autonomous cyber patching pipeline. And that feels like, to me, a really big deal of something to do before you are about to unleash a pretty large new model in the world that conceivably is much more capable than things we have had before. And you had a chance to kind of dig into this stuff. It's actually pretty fascinating. Very much so because, you know, on the day job side of things, security is paramount and the amount of exploits and it's increasing exponentially. and major, major systems. A lot of the software you're using today, and I mean you as an everybody in the audience, is built, there's means about it. They're built on stacks of dependencies and other things and open source software. And if one little link in that chain breaks,

has a weakness, has a flaw, the entire chain is compromised. And day after day, it seems that there are these really, really critical exploits happening. So in this particular case, OpenAI did a huge study. They went after the Linux kernel. If you're not familiar, I think Linux is an operating system, just like you have Windows and you have Mac. But Linux powers Android. It powers cloud computing. I'm sure there are microwaves and Roombas that are powered by... Refrigerators. Smart refrigerators are definitely running on Linux, for sure. Yeah. Right now, someone is ghostwriting their whip, and the stereo in their car is powered by Linux. And it could be compromised. It's like a removable CD player probably is still powered by Linux. It has a rotating face. You can find CDs in there. A really good visualizer on the front of it. Like a lot of things go up. Those were the days. If you're too young and don't understand the joy of going to a Best Buy and looking at – okay, I digress. So OpenAI put a system together that had a fleet of like tiny agents that went and researched the code base. The code base, by the way, for Linux is like 30 million lines of code.

Yeah. So it had to stift through all this. Rather than pointing like your biggest guns at it, that these tiny little agents go out, find what could be attack surfaces. They came back with like 2,000 plus findings. Then they had other agents go and review those findings and just pick, pick, pick, pick, pick. At the end of the day, this automated system detected actual exploits, wrote patches for them. And I think they ended up with what? Like 1,000 plus? Oh, patches for 948. They're called LPEs. This is local privilege escalation. This means like you shouldn't be able to access a file or a folder or run something on this computer. Oh, well, we have an exploit that lets us do that. And they passed 948 of these things autonomously for a cost of $1.2 million, which is expensive. Well, but compare that to the damage it could have done. Yes, exactly. And that's the thing. Had they found these bugs through normal bounties where, like, you know, Gavin goes in there and puts his trench coat on his head and the hacker soundtrack and he's in the mainframe, it would have cost $18 to $20 million.

So huge, huge savings. But a couple things here. So one, awesome that they're doing this. We need more of this. USA, USA, USA. We need all of our companies to have these systems and provide them to people who build on them because we need secure systems. Great. On the other hand, still scary. Like crazy scary. We might need these things. And what happens if you don't have access to it, Gavin? Or if you're not? Well, let's talk about that because Eric Wallace from OpenAI tweeted that they are doing a lot of this work pro bono. Now, I know it's hard to imagine a massive company saying, like, we're going to do a bunch of work for free, but they are dedicating a crazy amount of their frontier compute to this, he specifically said. And, you know, the idea here is that they actually, I'm going to quote him, we are allocating a significant amount of this frontier compute towards discovering and patching vulnerabilities in open source fully pro bono. So what that means, if you're not a lawyer, you don't know about law terms is that they're doing this for the good of the world, essentially for free. So that's one thing on your costing. So going back to – I'm sorry.

I didn't mean to jump in onto your – No. Gavin, I was doing what they say. You need me in? The professor was called putting the ball on the top. And I just hit it right off. And then what I did is I froze in the middle of my home run. And I said, wait a second, Kevin. Did you do that? Did you hit that home run for me? And was I on base? And then I stopped and I got tagged out. So unfortunately, we're tagged out. But you know what? you can tag in right now by tagging our subscribe button. Like and subscribe to this video. We love all our viewers out there. You are awesome. Thank you for contributing to our Patreon. Thank you for continuing to follow us. Thank you for all the things that you do for us. Hopefully you enjoy seeing Kevin and I show up in your feeds once a week. We always try to do more and spend more time with you, but this is a show that is not free to make, so we are very excited when everybody contributes. And the thing you can do for free yourself, like and subscribe and watch this whole video because retention matters, Kevin. It really does matter on YouTube. And jam a comment into the cheesecloth and just get a hole out of the comment

and just leave us with the juiciest, juicy comment in our algo down below in the comments. It helps. All right, Kevin. People have spoken. Let's do it. They put on all caps and they've said it in the comments. We want more politics. Politics. Politics. when they're two favorite bros talk over each other and talk about politics. The first part is an unfortunate thing that happens with remote recording and bad Wi-Fi in an Airbnb, so deal with it. Number two, we are fully in control over, and that's why we're giving you more politics. Yes, we have said many times that we would avoid these conversations, but here we are diving in, and I think it's an important thing to do this on. We lied. We lied. We're sorry. Just like politicians, we also lie. That's right. I'm taking the truth of power. Anyway, here's why this is important. And I think most people listening to our show would understand this. If you are in America, you understand on Tuesday of this week, there was an election that was a giant change in terms of like how people want the country to be governed. And I think that's a pretty flat statement that everybody

would probably agree with. If you're not from this country, there's a lot of people outside the country saying, wow, this was a big deal in America. What we want to tell you is this is like a big part of this conversation, you know, it's bubbling under a lot of other things, but was this whole thing about AI. And Kevin, I saw an article that was really interesting that basically from Fast Company where they talked about how many data center things were on ballots locally and how many lost. And I think of all of the data center things where people were either approved or disapproved, only one actually won. That's out of like 40 plus data center votes that went to the people. And I think this says a lot about where we are in the state of AI right now. In fact, there are already AI companies looking to offer concessions to locations to try to make sure that this can be a better balance sort of scenario. I think in general, this feels like a turning point, at least for the data center conversation in AI. And I think we can get in a little bit deeper into what that means.

But what was your first thought when you saw these votes come in and kind of a sense of where we are in this space? Look, on the on the topic of this is bad for A.I. or this is bad. I mean, I tend to think that this is this is bad for unfettered, unregulated, you know, companies running amok. Sure. I don't know. I mean, this could very well be a win for everyone who has an issue. with these data centers springing up in their backyards, right, who cries foul to their city councils and doesn't hear anything, and now suddenly there's something in their neighborhood. So I don't know. It remains to be seen. I know you posted some thoughts on this, so I don't want to trample on those because I think you and I are fairly aligned on this. I'm going to say something really crazy, Kevin, and I don't know what I do with this feeling yet, but I woke up after this election thinking like, man, do I have to run for office, which is the weirdest thing I have never once thought about in my life.

And the reason why I said that was I don't think I'm just going to save everybody a click. Don't tell me I shouldn't run for office. I probably am not going to run for office. But I had this feeling where I think there is a pathway through this world, but it was a really important one to kind of sit up and listen to what people are feeling right now. I think you're 100% right that this doesn't feel to me like a rejection of AI because we already know that like lots of people are using AI in their daily life already. What it feels like is a kind of a rejection of being forced onto something, right? Something is being forced onto something. And I think the most important thing is I can see a pathway where we can get through this, but I do think it's going to involve the frontier AI companies investing a lot of on-the-ground work and in people. And I think this will make AI investment more expensive. I think there's also been, as there always is in any kind of edge case of an investment space, a lot of people rushing in to try to make a lot of money.

And it's probably healthier if those companies don't do the things that were like exploitive in many ways, which I think is great. My biggest take on this is there are kind of three things that these frontier companies can do that I think is important. One, I don't know, did you see this idea that they're, maybe we talked about a little bit last week about OpenAI or other companies are trying to rebrand the term data center. Did you see this thing? What, super intelligence garden? Yeah, basically, yeah. No, I think there'd be like, you know, it's like they're trying to make it into like a cool community center vibe, which is fine. But I will say this, like right now you have these massive buildings in these people's homes that are making a lot of noise. If you're going to do that, you got to start thinking about more than just the building. Think about an entire like giant like Disneyland around the building. Like you are putting money into these into these places, but none of that money is going to last once the thing is built We already know these data centers only have a very small fraction of the people that need to stay working at the company once the data center is built So think about building the whole community That one thing The other couple of things I think are it really time for these

companies to get serious about AI education. And I know there's a lot of people out there thinking about AI education. We're going to talk a little bit more about AI education in a second where Google kind of had a fumble. But more importantly, I think it's about how adults get educated and then reeducated to use this stuff as they transition out of jobs that might be taken by AI, but even more so finding ways to level up both their income and their skills with AI. And then finally, I think the most important thing as an American, I personally feel very happy that democracy feels like it's in action, that like if people aren't happy, they should vote. And then we'll have results like this where there's a big shift in power. But I do think for the AI conversation, there needs to be a broad coalition across all parties and this is probably a worldwide coalition to start thinking about not only how we discuss this thing but how we educate the people in power so that every single person in every you know congress or parliament really understands this

because i think a lot of people have already said that 2028 is going to be the ai election meaning that like that is going to be the largest issue across the board. I want every representative, senator, president, everybody to understand the capabilities of AI. So I do think personally, this is actually a good thing for AI. And I think it will not stop AI growth in AI capabilities. I think it might be a small check on unfettered big tech power. Does that make sense to you? Yeah, a hundred percent. Yeah. I mean, that to me is like I there's a world where people rally around and and and actively court. And I don't mean just the politicians. I mean, the citizens in areas actively court these companies to come and build one of these monstrosities in their neighborhood. And maybe they want it underground. Maybe they want it. Seriously, maybe they want it. Oh, actually, I have a question for you. Have you done the research on this? Is there a reason they can't be built underground? ground it costs more that's what it always comes down but is that is that all the answer because

here's my thought like if you could build them under i mean i grant granted maybe like it would be like shake the ground but like it feels like that's a smart move now this is me coming from just the aesthetics place i'm sure there's a reason but it's a cost where you're right probably costs a lot more but still yeah i mean yeah terraforming is way more expensive than like leveling a slab right and throwing up on top of it right like that yeah yeah we know that when Elon and not to single out Grok, but we know that when Elon was scaling up Colossus, he was using these massive polluting diesel generators that were supposed to be temporary. And he's only supposed to use he had a limit of them. And I don't want to speak hyperbolically, but it was like eight and he used like 30. So maybe that's a little hyperbole. You can go look up the numbers. But the point is he clearly violated on the on the books laws in his hyper scaling effort. He is not alone. I know people. I know people that have said they're making money hand over fist, building data centers in Texas, and they are not paying attention to how much pollution they're using, how much water they're supposed to be using, but are actually siphoning out of the ground.

Like these are abuses. These are reasons people are upset. And ultimately, there is a way to do things. It might take a little longer. And yes, it might be a little bit more expensive. BFD. That's what you have to figure out. If you can go raise trillions of dollars, then some of that money has to go towards building it properly and giving back to the community. Give them an incentive to be there. It's a lie of we're going to provide 10,000 jobs to this community. That doesn't work because they do know, like you said, it's not indefinite. There's going to be an initial surge of jobs to build the thing, right? There's going to be a few people left over to patrol until they use the cyber dogs and the weird Segway robots, which are just pegs on wheels with the torsos that we're going to get into later. They're creepy bots. But they know that those jobs aren't going to be indefinite. So make sure they have clean air, clean water, that they're getting a dividend from the output from that factory. Give them tokens to use and teach them how to use those tokens. There's a myriad of solutions. But people don't want to talk about them because they cost money and they cost time.

So just ram it down their throats. Well, not anymore. Not in Gavin Purcell's America. That's right. That's right. In 2044, when I run with my robot eye and my robot arm because my shoulder is dying, I will take control of the robots, and I will bring down the robot empire, and I will bring Armageddon to an end. That's right. Welcome to my platform. All right, we should also talk about another fumble that a major tech company made this week. At another time where they probably overstepped, and they have stepped back a little bit, Google got in trouble this week for saying to schools that they had to take the AI search product if they wanted to use actual Google search in schools, which is a big change in terms of how Google works. Now, part of this problem, Kevin, has been that Google as a company is transitioning from traditional Google search to AI overviews, although there's been some discussion about whether or not that's been good for their business. But right now, they're kind of tied at the hip. So Gemini and Google search is tied up and hip.

And a lot of these schools are saying, we would rather not have this AI search in. And the story was that Google was saying, like, look, if you want Google search, that comes with it. Well, Google has now taken a step back because there's a massive uproar because a lot of parents and schools thought they were getting forced AI on them. Google has now allowed, there are versions of this which will be allowed without AI. I think there's a little bit of an overblowing situation going on here. But I will say I spent a bunch of time over the last couple of weeks and I dropped it on Monday. I made a big documentary about AI education with Imogen, the AI video, which you can see on our channel right now. There is a huge problem brewing right now with AI and kids in schools. And there's different kinds of research around what AI can work and what AI can do. But I think this conversation, along with the data center conversation, is going to become the biggest thing that is really important to people. And if you are a new parent or you have a parent of a kid that's like, say, 5 to 12, this could be a big part of your life for the next 10 to 15 years.

So, you know, get up to date on this, but also go watch the documentary. Try to make a few good points in there. But I don't know, Kevin, what are you feeling about, like, is this just another bumpy road on the way to, like, robot tutors or is this something serious to think about? Yeah, I think so. I mean, look, I am not a Google apologist. Yes, they do funnel a lot of dark money to me. That's why I get to gallivant around the world. But their statement about this – Oh, that's interesting. I didn't know about that. Oh, wait, wait. We'll take it for another episode. We're already over time. We'll put a pin in there. Yeah, we'll circle back, baby. Gemini, Gemini Notebook, and Search are off by default. That's what Google said. But the schools supposedly are in control. This is their statement. They can turn on one, two, or all of the services individually. And so what they were basically saying is like, look, off by default for K through 12 schools because even Search for some schools was bad. They wanted students to find answers in different ways. And then, of course, auto-serving an AI solution or an AI answer to a problem a student is researching, that is a cognitive offload issue.

And so basically, I don't know the issue intimately enough, but Google's response seems to make sense to me, right? Yeah, I think they heard. They listened and they changed it. Yes, yes, yes. So that's – Which is good. Which is good. But education as a whole, like we talk about AI being a disruptive technology, and it is. And it's probably going to be slower to disrupt everything like people have been maintaining. But clearly it has disrupted code and engineering. It's disrupted media creation, and it's started to disrupt consumption. Education is right there. It is right there. How many times has AI taught you something, Gavin? I mean, to be honest with you, a lot. Yeah, that's what I'm saying. Daily. Yeah, exactly. Daily for me. I'm unlocking new abilities. I'm also the cognitive offload. I'm also getting really lazy with certain things, right? Like, oh, first time you raw dogged an email and wrote all three sentences yourself. Oh, I do that. Actually, I do that all the time, too. I don't. I don't use AI to write very much for me, Kevin, because I'm an adult. I'm smart. I understand that's going to make my brain mushy.

Dude, I'm too busy slurping on a gogurt and dinking headshots in Fortnite to be writing emails myself. Okay. Boink, boink, boink. All right. So over the weekend, a really interesting story from an AI researcher came out. This is an academic, Rikia Takashi, basically came out with a paper that was called Weight Space Programming. And if that sounds nerdy to you, it is the nerdiest thing. But you have to imagine that the weights of an AI model are the things that kind of determine how it will work, right? And when you think of an open weights model, that means it's an open model that you can see the weights. or a closed model means the company has held the weights back because weights are how you know the AI is going to perform or how it's going to get its answers, all that sort of stuff. What Rikia has done here with this paper, with the group of people have done, is that they have proven at a very small level that if you point a smart AI at a small AI, that AI can rewrite the weights themselves to improve the performance of that AI. Now, why is that a big deal? It's a big

deal because in the past, it was always felt that you had to do a huge new training run with specific settings to make a new set of weights. In this instance, it is basically saying that an AI, a smart AI can go to his little brother or sister and say like, hey, let me fix a few things up for here. Let me put this thing on the shelf. Let me put this thing in your bedroom. And suddenly that little AI can perform better. In fact, they found that it did like 40 to 50% better on certain math benchmarks and certain thinking things. It has been replicated, but I do want to make sure everybody understands like this is just the beginning stages. But when you talk about RSI or recursive self-improvement of AI models, this might be what it looks like. So it's a pretty big deal. Yeah. Look, I think it's an overstatement right now to say like, oh, but the models can now they're it's we're off to the races. They're going to get it. Yeah. Yeah. Continually self-improve But there is a reflection of that here for sure. So, Gavin, I'm sorry. If I seem distracted, it's because we got red hot. Red hot. I can't even touch it.

It's so hot. Introducing GPT-6 Cyber, a major advance in security research that pairs stronger capabilities with better model alignment. So OpenAI, as of our recording, the time doesn't matter. Just know that it just dropped. and uh it's but it's available to select daybreak red customers so if you are um a date if you're not in the daybreak red program which i'm assuming 99.9 percent of our audience isn't you can go and sign up to be in the daybreak program you give open ai information about who you are or who your business is i think you have to pee in a cup i did it was weird i signed up behind an alley in madrid but the point is not okay not okay that doesn't sound like gp that I didn't feel like OpenAI at all. What are you talking about? He said, Sam Altman. I said, okay, here's your cup. The point is, it's out, and if you do the KYC, the Know Your Customer Diligence with OpenAI, you can get access to it. But huzzah. It's out. Huzzah. There you go. And by the way, this is probably the first stage of releasing their new model.

And what's good about this and what you should be listening for is if you run a company or if you run something else, see if you get into this. because maybe this is the stage of like, hey, you've got a problem. Go fix it or get it fixed pretty quickly, especially on the high end. Okay, new segment alert, Kevin. Whoop, whoop, run the sirens. We've got a new segment. It's called Good Robot, Bad Robot. Good Robot. Bad Robot. That's right. This is different than Robot Watch. This will be clear for long-term fans of the show. In this segment, Kevin, we compare a good robot to a bad robot. That is what we were doing in Good Robot, Bad Robot. Does that make sense to me? Okay, now this thing makes sense to me. I thought we were going to discuss the production company behind Lost. Oh, no. Bad Robot has had a bad few years, but no, let's not get deep on that. Okay, Kevin, we're going to start with a good robot. Good Robot. Kirby. Kirby. It's like Kirby is such a cute little name. And what this is is a little tiny robot that connects to the back of your garbage can, Kevin. And it can drive your garbage can to the curb.

Every suburban dad, everyone who's had a suburban house, understands the nightmare of being told at 1130 p.m., hey, did you take the garbage out? And you look up in the sky and you shout oh God why hath thou forsaken me This burden that weighs upon me anyway so what you can do is kirby will just basically roll your garbage can out to the yard although i have a lot of problems with kirby uh but i'm curious that you know your first thoughts on kirby as well okay because the you're oh gone are the 11 p.m thoughts of oh i forgot to bring the bin out it's going to be replaced by an 1105 alert on your phone. Oh, I know where you're going now. Kirby bot has a dead battery or a dog left a dropping on the driveway, and now your Kirby bot has smeared it all the way down and back. This is an obstacle for paper boys. This isn't good. Use your Kirby to the nearest data. Am I allowed to say that? I probably shouldn't say, we can believe that.

How about deliver roses to your neighbor? This is not a good thing. Go ahead, Gavin. Counterpoint me. Wow. So our first episode of Good Robot, Bad Robot, Kevin believed that the good robot is actually a bad robot. Are you saying that this has to be a good robot? Because I can change my tonality on it. It doesn't have to be a good robot. It doesn't have to be a good robot. I will say the funny thing about this to me is, you know, everybody has a Roomba. This better be very smart. The other thing that's annoying about this, maybe Kirby, and I guess I should know this, probably it doesn't, assuming that it's early. What would be great is like Kirby could like trigger your garage door open and then just drive itself out and drive itself back. Anyway, it's a very simple robot. It's something kind of cool. I do think ultimately we're going to start seeing these robots that are kind of specialized in different ways, and Kirby's a good example of that. Okay, moving on to bad robot. And this is no judgment on the company itself, but I will say there's a new company, or there's a company that's been around for a while called RealBotics. We have to change the conceit of the segment because I think Kirby's bad, and I think this next robot is amazing. Hey, that can be okay. It could be who judges what's what.

so real robotics pedestal robots and what these are are essentially humanoid robots but what the big part of it is is really like a v-shaped pedestal and a head on it and then you they have bodies that attach but i think most of their focus is really around their head kevin i saw these and at first i saw it as a video of the full robot and it looked kind of like an android robot that was kind of like okay it was not an advanced one and then i saw the video of it just being a pedestal head and i was like this the tabletop model you mean gavin the 11 950 tabletop model with 17 degrees of freedom yeah okay yes that's exactly right that's exactly right if you go to the gallery on the site there are some very weird pictures of this and what they are and kevin some of the ai marketing materials make these things look okay but the real world stuff looks like the weird, bizarre, like, facelift Instagram rabbit holes that put those down. Like, they just look like weird sort of microwave people with weird, like, there's too much

stuff on my body arms that come up and melt and flop around. I only pre-ordered two. Well, that's good. I also want to shout out their use cases page, where you see these pedestal robots and these different uses for, like, education, which we just talked about, hospitality and retail, elder care, corporate services, all things we think about. And then, Kevin, there's a section called wellness, which is about easing loneliness every day. And then you start to think about what that means. And maybe that this company is based out of Las Vegas, Nevada. That's not saying anything about Las Vegas. But you do start to think about the places that these things can go, and you can kind of fill it in your mind if you know what I'm talking about. Yeah, what's the IPX rating on this? How dust and waterproof or jail-based? What's the resistance is what I want to know. You know what? We're going to let the imagination of our audience fill all that in. You can fill that in. Kevin, I want to talk about this cool thing that you did with AI this week. This was based off a tweet. Let's hear what this was. When I saw this tweet personally, I was like, wow, that's really cool. And then you had texted me like an hour later.

It's like, check out this thing I made. Tell us what this is and how you made a new version of Katamari. Demasi. Yeah. Ann Schwan. I hope I'm pronouncing that correctly. I think that's probably right. That makes sense. Yeah. Yeah, I've been following on the X, and they have been doing some great stuff a lot of times with open source 3D models and repackaging them and whatever. So it looks like Google released thousands of these high-definition object scans. They did it, I think, years ago. Like there were all these objects that you'd have in your house. Yeah, like board games and consumer products and handheld gaming systems. Muscle milk or whey protein. Yeah, whey protein. Protein and Super Nintendo cartridges and Mario statues. They scanned a bunch of stuff, and then they made the objects available. And I think at the time it was for, like, populating environments for training of robotics so they could move around and navigate and sort things and whatever else. Okay. Amazon also released thousands of scans and 3D models of a bunch of products that are available on their site. And if you've never experienced this, you can use the Amazon app and, like, see how a ninja air fryer looks on your tablet.

Oh, super cool. Yeah, that's fun. Before you buy, right? And so Anne Schwan, again, apologies if I'm butchering that name, they've been taking these releases of these models and repackaging them in these super, super – like compressing these multi-gigabyte things down to just like 100 megs, a couple hundred megs in a slightly different file format that is optimized for streaming in the browser. And so when I saw that, I thought, well, huh, how cool would it be to have – wait, I already have a Katamari game that I kind of built. Yeah, you already built a catamaran. So how much of a lift would this be? And it turns out it was like five prompts worth of a lift to have PT Astra go out, grab the files off Hugging Face, prep them, populate a world with them, make scavenger hunt mini games where you have to like get one item on your ball and then bring it to another part of the thing. Yeah. Arrange all of the items in a stage. So if you go to katamari.cursedslop.com, we'll put the link in the show notes, you can navigate your ball through all of the Google objects, go through a catalog and see each object, zoom in, zoom out, rotate them around.

And now there's like 9,000 plus total objects because I put the Amazon and the Google ones in there. So it's just like – It's super fun. I will say – How quickly things like this can be iterated on. It's amazingly fun. And it, again, made me think about – I know there's a couple of companies out there trying to do this. but it did make me think again about like, there will be a winner in the, you know, Newgrounds version of what this is, right? Because like, it's hard for a normal person to, you know, you get the SEO or whatever it is, like we'll send people there, but it's not like that because it exists in the world, people will find it. Somebody is going to do a Roblox-y version or something like that, which will bring all these little games together in a place so that you make this thing. There's some sort of like payback to you for doing it. Not that like you want that, but like, I think it's tricky because nowadays attention is so hard, and you should 100% go look at this thing Kevin made because it is awesome. I spent like an hour yesterday playing with it. It is a super fun thing, but it does make me again think about like, God, how do you get all these really cool experiments that people would want to try

for anywhere from 10 minutes to multiple hours, and how do you kind of surface them up? It is a really difficult problem, I think, in today's both internet and AI world. well once I get Anya my pedestal torso I'm going to look deep into her glossy eyes and say hey how can we optimize this future together I know how we can do it Kevin we can look at the things that people did online this week with AI in AI see what you did there sometimes you're strolling without a tear then suddenly you stop and shout AI So three quick things here. It was a great tweet from somebody that blew up whose name is SecondBassoon, second underscore bassoon on X. And this is a problem, Kevin. This person actually went in and tried to talk about different ways to prompt AI video.

And maybe we can play a little bit of this person's video. But the thing that they said was specifically trying to get dialogue to sound realistic with AI video, which, as you and I both know, can be hard. Play a little bit of this. It is in a different language. I think it's in German. But you'll get a kind of sense when you listen to this. It doesn't necessarily sound like other people's AI video output. Now, we may have gotten in trouble for what that says. If you're in Germany, you're probably not. There's nothing bad there. It's actually a video about like Beethoven's when he was younger. So things that second best student said is pay close attention to how the written punctuation is used. Break long lines into shorter ones with intentional pauses. Use bodily actions as punctuation. And this is a really interesting one because you heard the sniffs in that. And I do think there are times when you hear somebody chewing or eating or sniffing.

that makes the video sound much more real. So anyway, I just really appreciate when somebody like this gives a really useful tip to AI video generation. I thought this was really cool. Yeah, super helpful, prompting advice. And, you know, just tell your agent to go look at the post and pull it in and then guide you on how you can modify your prompts and your scripting to get this to work. That's the greatest thing about, like, all this information sharing that's happening with these tools. Yes. Is that it's sort of like, look, I went through it. I learned it. Here you go. Go build upon my knowledge. Like that's, yeah, I love that about this scene. Okay, we got to talk about flying bad here. I am the one who slaps. This is just this tweet. This tweaked my comedy brain. This is about, this is a reference to the movie Airplane. If you haven't seen Airplane, you and your children deserve to go see it. It is a very funny comedy movie from the 70s that still holds up. In some ways, there are other things that don't hold up, but just to be clear, it's very funny. This basically takes a scene where there's a very famous scene where a woman is freaking out and trying to get a hold of her.

A bunch of people are in a line that are like kind of slap her to kind of get her in control. What's hilarious about this is they made a Breaking Bad version of it. And it is Skylar in the seat who's freaking out. And they show different characters from Breaking Bad kind of across the thing, kind of trying to get her under control so that she's less crazy. And, Kevin, it just goes to show you, we talked about this a couple weeks ago, AI video is solved in a lot of ways. like these people at least for short clips but like this is everybody you see uh you see Tuco you see like uh what's his name the old guy you see all these people in the line and it's just really fun for a moment of joy again it's this little fun moment of joy as a cool thing when we talk about AI video being solved it's like look this this has like every character kind of looks and sounds like they do from the series it is a long take it is swapping multiple assets in the scene it keeps it coherent like this is this one of these swaps you know months ago would have been magical and now it's like yeah of course they're probably using an open source model and they're

using laura's trained on these actors and the actresses and of course it just works it's still pretty magical it's very magical and then finally i saw a very cool ai generate demo also by and schwang which this person is like over indexing right now on their x but this was a souls-like game where a robot tries to not drop a towel and if you've seen videos of robots like training this is a very fun thing souls-like is like dark souls where it's like a very difficult game and like the challenges are hard but you see the two robot arms trying to hold a towel and the big like uh like the big challenge is a cat walking in front of it just a cool this is a good example to me of game as entertainment like i'm probably not going to go play the souls-like robot game personally but this is game as content right like i this video was super enjoyable to me to watch and this was the game this guy built like that's a really interesting use case did you play it though did you try it no i didn't play i didn't play did you play it i did oh yeah how'd it go it's hard but you'll yeah you'll you'll let you'll click the left mouse button for the left robot's grip

and the right for the right robot's grip and there's little bars you have to kind of like juggle while that's going and they're trying to parry the wind and the cat comes in and ruins the whole day like yeah again i like i played it for a few minutes i was like this is great and then i went on with my life but i'm glad that someone put it out into the world like that's amazing it reminded me of uh remember the old game quap where you're trying to run oh i love quap yeah yeah it was very like walk across the way yeah yeah maybe i'll do it as our way out i'll try to do the quap walk ready well i think walk okay that is um yeah and you look like a chat gpt animated gif of you I am running away with my monkey. Hold on. All right, everybody. We'll see you all next week. Thank you so much for joining us. Bye, everybody. Bye-bye.