AIForHumansOpenAIGPT5SamAltman

Kevin Pereira: [00:00:00] GPT five is officially here and across the board. It is the best model you can use today. We're talking agentic coating research capabilities if you need help. But Kevin,

Gavin Purcell: Kevin, look at the colors chat. GPT can now talk to me in whatever color I want. Chartres baby. This is a GI.

Kevin Pereira: Sure. But we're gonna show you why.

This is probably going to be your new go-to ai. Significantly better writing,

Gavin Purcell: way less hallucinations, and maybe more importantly, new colors for my little chatty. Please don't call it your little chatty. You're right, Kevin. Those colors are perfect for certain things like misleading graphics. There was a crime that came out of the charts on this one.

There were a couple unforced errors. Sure. But there were also a lot of benchmarks. Yeah. And the benchmarks are pretty good.

Kevin Pereira: I don't know if they're

Gavin Purcell: 500.

Kevin Pereira: Billion dollars. Good. But they're good. Okay. What about the vibes though, Gavin? How are the vibe markets reacting?

Gavin Purcell: The vibes are off the charts. We're rounding up all the early impressions so that you don't have to,

Kevin Pereira: plus Google just teased Genie three.

It's an AI powered interactive [00:01:00] world engine, and honestly, it is one of the most impressive things we have seen in this space, probably ever.

Gavin Purcell: That and 11 Labs just released a new artist friendly AI music model plus. Some not So family friendly, uh, image and video models from Uncle Elon.

Kevin Pereira: Gavin, you promised no yfu content at the top of the show. I lied. This is AI for Liars.

Gavin Purcell: All right, everybody. It is the biggest week, and I don't know how long actually it is here, Kevin. I look back on our old shows and we talked about where is GPT five Over a year ago, one of our YouTube videos was titled, where is GPT five? Today, OpenAI announced and we'll be rolling out to some places.

Kevin has access it within Cursor. I do not have access yet in my GPT plus account, GPT five. Kevin, let's talk about this. [00:02:00] What, what is this? Where are we going with it? How does the world change? PT

Kevin Pereira: five Bummer is here. Gavin, wait, what did you call it? DPT five. Hummer five bu.

Gavin Purcell: What's Hummer

Kevin Pereira: mean?

Gavin Purcell: The five Hummer.

What is that? If

Kevin Pereira: you put the, the number five and then Hummer, it looks like summer baby. It's a hot boy. Five Hummer. Wherever your brain went is not appropriate.

Gavin Purcell: Before we jump right in, let's just see a little bit from Sam at the top 'cause he had some thoughts right away.

Sam Altman: Today. Finally, we're launching GPT five.

GPT five is a major upgrade over g PT four and a significant step along our path to a GI.

Gavin Purcell: Alright Kev, so today, G PT five, three models. Um, we're gonna get into talk about the benchmarks. We're gonna get to talk about a bunch of other stuff, but the vibes

Kevin Pereira: as we did it kind of feel like for a multi-billion dollar company, Gavin, the unveiling of what should be the next gen foundation or something that's earth shaking.

It kind of had the same energy as like, Hey, my, my son just learned a new dance from TikTok. Everybody gathered the kitchen. Let's watch 'em do the do the

Gavin Purcell: thing with the floss kid. I personally appreciate [00:03:00] this as somebody who has given and done media training that they're willing to just kind of, these are people that work hard and they're out there doing this stuff, but Kevin, it's very important to talk about what was talked about.

So first and foremost, GPT five. Brand new model. It is replacing all of the other models. All the former, uh, AI models from OpenAI are gonna be depreciated. So goodbye, GPT-4 Oh, goodbye, uh, GPT-4 0.5. Everybody gets this, but Kevin, I think the biggest thing that kind of came out pretty early is this is going to be free for everybody and that.

Is a transformative moment. Uh, a lot of people have talked about the idea that when deep seek happened, that was a big deal because people had never seen a reasoning model before, and they're like, why is this so much better than GPT-4? Oh. We are talking the fact that students are gonna get it. In fact, I think that's why they dropped it.

Now they're dropping it in August 7th, which is when most schools go back. So they know that they're gonna get students. But more importantly, worldwide, this is a really big deal that suddenly everybody's gonna have a state-of-the-art, a reasoning model in their pockets. To do

Kevin Pereira: anything. [00:04:00] They bought chat.com for a reason.

Gavin, I think if you talk to most people, it's like, well, have you used an AI and something like, oh, I, I chat. I use chat, or I've chat gtd. Well, this becoming the default, like you said, for everyone else, this being, uh, non-pay gated, this opens the doors for everybody to experience it. It is best in class.

Right now, largely across the board, and we'll get into benchmarks versus the, the sincere sort of vibe check of it all, which is becoming more real as these benchmark benchmarks for sures become more irrelevant. Oddly enough. Yeah, yeah. Um, but th this is, this is the go-to model. And in the past, you know, people come to us all the time to say, what model should I use for this or that or the other.

I think now. By and large, maybe outside of video generation or music generation, it's, well, you're just gonna go to GPT five. You're gonna go for now?

Gavin Purcell: For now. For now. We know we reserve the right to be

Kevin Pereira: wrong by

Gavin Purcell: this time Next Thursday. Yes, because Gemini may be sitting on Gemini three and Logan Kilpatrick has kind of teased that idea.

And if you remember, Kevin, I mentioned this, uh, um, on the AI for [00:05:00] Humans, uh, Twitter channel. Uh, what was it two years ago when the four oh, um, advanced voice thing came out? They scheduled that right the day before Google's io. I bet Google wants to do a little flip flop. Anyways, let's talk more about this model.

Okay. I do wanna show a tweet from Ethan Molik that I think encapsulates this. Ethan had early access to it. Open ai. If you're out there, we always would love to get early access as well. We can't promise we will do the great things that Ethanol does, but Ethan said this and I just wanna quote it because I think this is kind of a general sense of it.

Um, he said I had access to GT five and I think it's a very big deal at his very smart and just does stuff for you. And he's showing. A procedural brutalist building creator where he drags and edits building something he did. And he is a professor, not a programmer, but this is like an a version of what I saw.

Lots of people talking about this idea that you do stuff and in the past it might have not done things for you, but it just does stuff. And Kevin, maybe that translates us to, like, talking a little bit about these benchmarks because I think yeah, if you were to look at the numbers, they might seem disappointing to [00:06:00] some people.

Kevin Pereira: Uh, well, it depends on what chart you're looking at Gavin. If you're looking at a chart where the, uh, the Y scale is just sort of. A vague suggestion, you might see massive leaps. Yes, yes. In some cases you might see a number in the fifties dwarfing a number in the seventies because benchmarks are made up.

But bad charts aside the benchmarks largely most people aren't gonna care about. They're a weird like, that's exactly right. Measuring contest.

Gavin Purcell: You know, who cares what? Who already cares? Who cares. Elon, he already tweeted out, oh, oh, we beat this on that. But anyway, that I, I think in general, most people could give a crap about benchmarks at this

Kevin Pereira: point.

Well listen, benchmarks for a while, were a way of, of, of really, yeah, showing the differences between the models. Now again, across the board, the benchmarks are fairly good here. What are the benchmarks actually testing? How well can it accomplish things like using tools? Um, to accomplish tasks, reasoning through, uh, uh, critical math problems or even, you know, language, uh, problem.

Mm-hmm. Problems. Um, the quality of the code that it's spitting out. They have all these different benchmarks, which are just a [00:07:00] series of tests that all these models go through. Some models do well because they're trained to do well on the benchmarks. Others are poisoned, meaning, you know, their data sets have some of the test problems in there.

So. There are issues with benchmarks, but by and large they give you a general understanding of how these things work. We are now at a point in time where, say for a few benchmarks, yes, the models across the board perform well at them. So when you see something go from 72 to 79. Well, yeah, that, that number's higher.

That's good. Right? Would you rather have $72 or $79? Probably 79. But at this point, like not me. I,

Gavin Purcell: I'm, I'm more for less money. That's my goal.

Kevin Pereira: Well, you're in a, a rarefied tax bracket. Down, down. I really am. That's 70. I need, I need to bring my, of your own, I need to bring

Gavin Purcell: my income down.

Kevin Pereira: But no, so the benchmarks did go up, um, as expected.

Yes. I, I also expect that next week there might be another model that SLI is slightly better with. Benchmarks. There was one benchmark that went

Gavin Purcell: down, and I think this is a really important one to talk about. [00:08:00] Oh, yes, yes, yes. Absolutely. So, so in my mind, one of the biggest benchmarks they showed was the reliability and accuracy benchmark.

And, and this is testing hallucinations. And if you look at this, this chart shows you basically how much it's making up crap. You know, everybody out there in the world. When they talk about ai, they're like, it's not trustworthy. You can't see it. It makes up things all the time. This shows that it's still gonna probably do some hallucinating, but these are massive reductions.

And they did say in the livestream that they have specifically been attempting to lower hallucinations. And I think this is where another thing kind of like it being free, the world at large may not be ready for an intelligence that actually works and doesn't make stuff up. Now again, this is not perfect.

You can see some of these numbers. It's still hallucinating some things, but. Right, especially on the health bench, a hard hallucinations number you are seeing oh three was hallucinating at a 12.9 percentage rate and GPT five with thinking is at 1.6. So when you think about use cases in health and healthcare, which they did talk [00:09:00] about on the um, livestream, they had a woman who came out who had cancer, who had used chat GPT to help her understand it better and then played with the new models.

That is a really big deal, I think.

Kevin Pereira: We have specifically been saying for the last, almost year and a half now, I think, uh, maybe a little bit more, that health specifically is a sector that these models really excel in. These were the older models that did halluc. Yeah. Yeah. That did get things wrong, and they were already, in many cases, outperforming medical professionals with the domain expertise and whatever they to be a professional on.

So you're right. I'm, I'm glad you you put extra light on this. Particular benchmark because this is, this is a incredibly important one. And when you talk about all of the sectors that will rely on this intelligence for hardcore math Yes. Or physics or medical diagnosis. Yes. This is something that does have to come down and it's there.

And also like, again, somewhat benchmarked. Is the speed. Yes. E even many are remarking that even if the intelligence and the capability of GPT five hadn't gone up, which it [00:10:00] has across the board, the speed versus something like an O three is so much faster now that it is making them already five to 10 X more productive because they can just keep trying over and over again with things until they get the result they need.

Gavin Purcell: What do you think is the max human X productive like. When you start getting to like a 50 x more productive, do your fingers burn off like when you're typing or does your voice go hoarse, like Yeah, I kind of have, we gotta have our own benchmark here. Like what's max productive, like if we can make somebody in our audience maybe make a character or also will, could attempt to make a character of.

Max productive. So he's like a superhero with like giant forearms. We need our own benchmark, Kevin, it's called? Yeah. He should look like

Kevin Pereira: a, a monster Energy can just vaped Guy Fieri and he's, he's Max performance, he also sounds like a pill You get over the counter at a gas station. Ooh, max. Productive. But to the rhino.

To the rhino blood, it's right. He rides two rhinos at the same time. He's Max. No, he doesn't.

Gavin Purcell: No, he doesn't. That's not a family friendly. The other thing I wanna talk about Kevin, quickly is the vibes. In this [00:11:00] instance, when you hear people talk about the vibes of a language model, it really means like how it's performing for actual people, right?

Indifference of the benchmarks where you have these kind of like questions that are set up. Vibes are a reason to like really understand what's going on. And I do wanna shout out, um, there's a lot of different things we've talked about here, coding and all sorts of other stuff. But there's a great, um, YouTuber named, uh, Theo, who has a pretty big YouTube channel, but he says, I don't care about intelligence benchmarks.

Now I'm post eval, which means he's posts that, uh, you know, he says GPT five does what it tells you to do. No other model behaves that well. Uh, this while. Trust me, don't judge it until you try it in editor, which Kevin did. Yeah. Give it tools. Give it instructions, watch it cook. So the, this is very much a vibes comment, but

Kevin Pereira: Kevin, maybe tell us a little bit about your experience.

I will and, and, and Matt Schumer. Similarly, Gavin, he's had access since uh, July 21st. So shout out to Matt. Whoa, that's a good, he got

Gavin Purcell: deep early access then he's

Kevin Pereira: had it for a minute and he said like basically at first he was underwhelmed and then he threw a problem at it that him and his [00:12:00] engineering team were struggling with and thought it might take weeks to engineer and it.

Essentially one shot at it. Wow. It came up with a, an elegant solution to a very complex engineering task and sort of did it without any fuss. And you know what we're hearing again, it's better than oh three intelligence. It is super fast, long context handling, which is a thing which, you know, yes, if you are.

Giving it a novel and saying, help me write the next chapter. Or you're giving it a, you know, a 10,000 line code base and saying, here's a problem, or all of your medical history. Ooh, all your financial documents. Whatever those things are. This thing doesn't lose context. You have, what was the name of the guy we

Gavin Purcell: just made?

Who was the, uh, max, what was his name?

Kevin Pereira: Uh, well, max. Max productive.

Gavin Purcell: Max Productive. What if he's the star of a, our own novel? If somebody wants to take on that. Max productive as a novel in some form. We write a novel with this in some Gavin, I've

Kevin Pereira: been handing out homework to the audience left and right. We need a million dollar donation and someone needs to write Max productive.

Yes, please. Yes, please. As long as he is writing three rhinos on the cover. Yeah. Um, there are, uh, several modes to this, right? [00:13:00] There's the auto mode, there's a thinking mode, and then the pro mode, which Matt didn't even have access to then. Yeah. Oh really? And promo will basically spin up a bunch of these things instead of just running one instance, but, oh wow.

So, yeah. So here are the vibes. Here are the vibes. I have been struggling with a little secret project that you and I are working on. I have been struggling with a particular task for the last three and a half, four days. Mm-hmm. Using, uh, Google's Gemini Pro and using Claude Code Max, uh, the, their Opus model, going back and forth trying to get this problem solved.

Varying degrees of success. And then something else breaks. I just, before we did this podcast, without even giving full context, said, Hey, GPT five with Max within, within Cursor, which is a a a A coding app. And I said, here's the kind of issue that I'm trying to solve. See what you can see. And it very quickly did a deep dive through the entire code base, identified potential issues, suggested several solves.

Took it upon itself to make one of those [00:14:00] solves and said, go ahead and try it now. Let me know what you think. And it worked. Wow. This is, this is pure anecdotal sauce, so it up, but like it. It's there. Like this, this felt transformative to me. And so then I was like, okay, thanks. Can you, um, also check out this other little thing?

And it sort of did it. And then I, there was an issue. I copied and pasted the logs. I just fed it, the logs. And I was like, oh, I see why this is happening. Here's a fix here. It, it's just, it's just doing it. It's not asking that's questions. It's just solving it. And so I think like maybe for the ultimate test, Gavin.

You and I right now could try to one shot your benchmark. Well,

Gavin Purcell: since we are talking about Max productive as a character, Kevin, should we try to do a, a mini three j, like a three js kind of game, sort of prompt around max productive and maybe max productive is in a bodega. And his job is to swat away rhinos as they try to, uh, knock stuff off the shelf behind him.

Kevin Pereira: [00:15:00] You asked. I, I thought we were doing animal brackets, but that's okay. That's all right. And there's no, there's no bad ideas as we know this. No bad ideas. No bad ideas. But I feel like in max productive lore. Yeah. He is a friend of the rhinos. He rides the rhinos. Okay. Right. So now you have, oh, then let's make, oh, there's Rhino idea.

Gavin Purcell: I was thinking there's like a series of rhinos writing. It's almost like Frogger. And he has to jump from rhino to rhino to get from one side of the screen to the other. So there's rhinos that are going left and right. And his job is to jump and there's one bad rhino in each mix that turns him into, uh, less max productive.

And he only, he can only hit productivity. He can only hit three less max productive rhinos and or else he goes into a stupor and plays video games all night. Okay. At what

Kevin Pereira: point does his, Hey, that's not gonna be in the show. You better stop saying that. So Gavin benchmarks are, uh, I guess, a little unhappy and saying benchmarks.

Our artificial general intelligence or super intelligence has been pushed back to 2075. Yeah, sorry. We're gonna get robots in the home before we get super [00:16:00] intelligence. Others that are surfing the vibe wave seem to be much happier. Yes. About these models and again, anecdotal, I'm impressed thus far, but I think we need to rip around.

A nice juicy let's it one shot or prompt for Max. Productive,

Gavin Purcell: hopefully

Kevin Pereira: Will made

Gavin Purcell: something that we can use on, put on t-shirts 'cause that we, that we're gonna have to do it. So Will,

Kevin Pereira: I'm really trusting

Gavin Purcell: you now.

Kevin Pereira: I, I, yeah. I, I, I think he's a cross between, uh, a, a slim gym, you know, like the, you know, the slim gym mascot.

Yep. Sure. Yeah. Well, macho man. No, there was a, there was a, they had, they personified the slim gym not too long ago, like a, like giant stick with arms, like in the nineties. Yeah. And he had like cool big, tall hair. Am making that up. Tone. That guy, like the Italian bra rock

Gavin Purcell: guy, Aming that. Hold on. Oh dude. A hundred percent type slim gym mascot.

Hey, there is, all right, let me Google it. Lemme type it in. Oh well, yeah, he is like a

Kevin Pereira: human with giant hair of slim s. Yeah. Okay, so well listen, I have so, so the folks at home know we are going to try to vibe code a Game one shot. We'll see if it runs. Okay. [00:17:00] I told this new GPT five in max mode in cursor to make a single page.

HTML five game, like a 3D game for max productive. I gave it a, a brief description of Max. I said it should be a frogger like world where you leap from rhino to rhino. I even misspelled the second rhino. Okay. I said, you're trying to collect giant pills, which enhance productivity. The more pills you get, the faster everything goes.

Ooh, I like it. And you collect five pills. A rhino explodes randomly. Okay. At at 20 pills you explode and win the game. Oh. Oh, wow. Okay. We'll see what it does with that. I said nothing else about camera controls or Okay. Uh, keyboard or commands or power ups. I'm gonna let it do its thing. We'll see if it works.

Yeah.

Gavin Purcell: Let's rip through a couple other things that were announced here real fast. Um, I do wanna talk briefly about their, uh, new personas. There are going to be four new personas in chat PT, which is kind of a big deal. The reason this is a big deal, Kevin, because now you can determine, do you want a.

Cheerful and adaptive persona. Do you want a cynic, [00:18:00] critical and sarcastic? Do you want a robot efficient and blunt, A listener, thoughtful and supportive, or, here you go. Nerd, exploratory and enthusiastic. So those are the four new personas, meaning that you can tell Chad, GPT how you want it to respond to, which actually is kind of cool because there might be certain things where like, I want, uh, I wanna a role play chat.

GPT as Kevin, or I'm sorry, not Kevin. Uh uh. Uh, Kayvon. Kayvon is who I sometimes role play with Chachi. See, I wanna make sure that they're listener, they're the listener thing. I definitely don't want the cynic because I get enough of that in my real life. So I would use the listener one. Wow.

Kevin Pereira: Wow. Shots fired.

Yeah. I, I think that's good. The also like tiny little things for developers, like, uh, veracity as, yeah. The something you can dial in if you've ever worked with these things at length, you know, sometimes they get, uh, unnecessarily yappy, right? Yes. And again, as someone who gets enough of that in their personal life.

Sometimes it's nice to just turn the dial down. Mm-hmm. Isn't this therapeutic for everybody?

Gavin Purcell: Isn't this great? Maybe didn't talk about the new voice updates right [00:19:00] now as well, because maybe I need to talk to somebody that isn't somebody else. And, uh, there are new voice updates. Uh, it's much more steerable.

You can get it to respond to you in single words, which is pretty cool. So if you really want voice to just say yes or no, or specific things, it won't go on forever, which sometimes is annoying with voice mode. And also also. Be better speed and control. Right? There was a really interesting demo where they asked it to speed up how it was speaking in Korean.

Maybe we could show a little bit of

Kevin Pereira: that here and slow it down, which is very helpful when learning any new language or diving into a new topic.

OpenAI: Could you speak that part really slowly? Because I'm a beginner, I want to hear every word clearly

Kevin Pereira: and again. We don't know you. I mean, hi, we try to know you, but we don't know what your daily usage, what your projects are, what your creative unlocks are, what your blockers have been, and through the demos that they had, and through the, again, the [00:20:00] anecdotal vibes that we're seeing now, whether you wanna make a Castle Defense game in 3D, where you fire cannons and pop balloons, which was a fine demo.

It's no max productive. We'll get to that. Or you need to make a front end for your business, or you wanna make a bespoke French tutorial game that gives you flashcards and a game where you are A A, A A. What was that? Mouse eating cheese. I think it's like a weird thing. I wasn't, yeah. Were you a mouse or

Gavin Purcell: like a baguette?

I couldn't figure out what was going on. Anyway,

Kevin Pereira: they

Gavin Purcell: did all that live in the demo.

Kevin Pereira: They did all this stuff. So again, like we, we don't know how you're going to use it, but we can say with, with utmost confidence right now that GPT five, which is the thing you can select yes. On their website or I just, it just updated in my Mac app, Gavin.

That's it. That's all there is gone. Yes. Are the 15 different models that you could select for different reasons and all of that confusion. Now you've got GPT five and it's up to everybody else to best that model. How's, uh, max productive coming along? I think it's done. Okay. So it made a single file [00:21:00] at which I asked for, it's only 27 kilobits.

Wait a second, hold on. Uh oh. Okay. Max. Productive rhino rush lead from rhino to rhino. Snag giant pills go faster. Explode at 20 to win. Uh, right now it says speed X one pills zero out of 20. Okay. And it made these little rhinos, by the way, which looks, they look kind of like mice with, uh. I guess unicorn horns coming out the front.

Oh, it's

Gavin Purcell: 3D too.

Kevin Pereira: I am writing on a rhino right now. Uh, when I just got a pill, it said, uh, vape, the vape, the hype, and now it says zoom. It also says, Kevin, for life, it said, Gavin sucks. How does GPG know that? That's what I'm liking. Now, I, I might have hit a couple things into the prompt that I didn't tell you, but it did say Gavin sucks.

Now it's saying shake those skates and recover. Do you hear that sound? That's me falling into the world over and over again. What? So I'm thatm gonna just, what is that? It's me falling into the world. That's the death noise. Oh, okay. Got it. Got it. Because it keeps res responding me. I think it says Gavin more [00:22:00] like gravity.

I don't know what that even means. Vape the hype. Okay. Yeah. I think it says Gavin more like it either says gravity or gravy. All right. Okay. I'm closing the game. I'm gonna do, this is gonna be a one. This is gonna be a A two Shotter. Make this way. More extreme and hype as hell. Go nuts with it. What is crazy Gavin is that it was, it was a single file.

It was really small. It loaded. Yeah. It was perform it. It did the jaunty music. It had sound effects, it had the high score counter. It did all of the things that like, you know, not that other models can't do this, but it did it, yeah. So fast that it was done long before you asked me if it was done. Wow.

That's amazing. Yeah, so I mean, I will

Gavin Purcell: say the music wasn't terrible either. Like, it was kind of interesting to hear the music that was all generated straight out of the code.

Kevin Pereira: It says, uh, it says I'm gonna make gameplay in UX changes directly in the index html. Safer respon onto Rhino Closer, more rhinos for easier jumps, stronger Jump with Coyote [00:23:00] Time and double jump.

Oh, coyote Town, is that another, is that a, is that a plugin? Uh uh, download. Dude, I don't know what coyote time is, but they get a battle pass from me. That's 9 99 for coyote time, improved. Camera speed lines, flash effects for hype. More slogans and dynamic background music speed. That's what it's putting in right now.

So we'll see.

Gavin Purcell: It's coding that right now. And when you're at it, how long now? How

Kevin Pereira: long is it? Is it,

Gavin Purcell: is it

Kevin Pereira: rewrite? It's basically rewriting the whole thing and it thought for 37 seconds, it is now making the high impact gameplay improvements to the H TM L file. It says, okay. Like, we'll see, but I don't know what coyote time is and I'm so excited.

We'll find out.

Gavin Purcell: Meanwhile, Kevin, there's another really big story, which is we do not yet have our million dollar, uh, super thanks on YouTube, which I think is really probably Gavin more important than GPT five. If you're out there billionaire on a yacht, we know you're sitting there with your little headphones sitting on your yacht next to, I don't know, a couple hours.

No one wants, wants to be first to the party,

Kevin Pereira: right? No one wants to be first in the pool. Gavin, they're [00:24:00] waiting, they're circling.

Gavin Purcell: Everybody else out there who, if you're watching the show, first of all, thank you. Second of all, please like and share this on YouTube and subscribe if you're not already subscribed.

We do this every week. We've been doing it for a while, and we really do appreciate everybody that comes in. If you're on the audio podcast, if you're listening, please leave us a review on the Apple Store. That does make a difference for us. But most of all, thank you. Thank you very much for watching. We appreciate it.

Kevin Pereira: You're literally the only reason this podcast grows. We cannot afford to market it until that whale, I bet you were just, you're literal. The only reason

Gavin Purcell: I do this, because I don't want see the person across from me anymore than I already do. That's not true, Gavin. I would never, should we talk briefly about open AI's, open weight models, Kevin?

Because that was like a story from a day or two

Kevin Pereira: ago, but still really big deal. It would've been the, arguably the biggest news of the week if OpenAI hadn't announced something else and release GPT five, but yes. Um, again, a lot of times we talk on the show Gavin, about uses of AI and people get rightfully concerned about things like cost or privacy.

Well, if you can run models locally. As, [00:25:00] as in on your own machine or served up on some device that only your household can connect to. That certainly solves a lot of those problems, right? It definitely does. Can be a power bill. You're gonna need a machine capable enough to run it, but OpenAI released two open weights models that you can run completely free, uh, on your own devices.

So you do not need internet connections. They have devices, uh, they have models that you can run on a. Like a fairly underpowered laptop, MacBook, uh, like in the MacBook? Yeah. Some people were running them on airs with Ram and, uh, even on like a cell phone. And these are just about as capable as their last generations of models were.

So yeah, there, there're no slouches

Gavin Purcell: there, there, there we're benchmarking a little under O three, which is pretty impressive considering you can do these entirely on your own. Uh, which I was pretty cool with. There was a, uh, the guy who does a, a Donis Singh who does the Minecraft test, did show that they're not as performative as a say of the art, but that makes sense.

Considering you will be able to do whatever you want with these models. You can run them locally and you don't have to have the man telling you what to prompt for or not prompt for anymore. [00:26:00] You can get whatever Max Power or Max Protocol, not Max protocol, max Productivity, productivity system you want out of it, including Rhino Pills.

Probably, if you're really interested in that,

Kevin Pereira: it's running. The music is much faster. Okay. The camera very tough, but yeah. Oh yeah, yeah. I'm getting, I got nine. Oh God, it's going so fast now. 15 outta 20 pills. I exploded with productivity. Max productive, achieve 20 pills. Victory is mine.

Gavin Purcell: Two shot, we two shot a game about max proto uh, productivity.

This is so cool. Like obviously it was didn't one shot it, which is fair, but like you did get a playable game very fast. And to your point, other systems have done this before, but the speed at which this was done is a big deal and I still think that's a very cool thing to see. Kevin, speaking of games, there is another major story that I'm sure we'll get overshadowed now, but I actually think this is the most interesting technical story that I have seen in a long time come out of ai.

I say technical, but it's actually very visual. This [00:27:00] is Genie three, Google DeepMind's brand New World Model. We talked about Genie One and Genie two. These were based on video games originally, but Kevin. This model allows you to look around in a simulated world. These are all made up and more interestingly, unlike other models that we've talked about that does this, you actually have some version of memory.

So there's a really good demo of this where they're, where they're painting a wall and normally in these world models, which is literally making frames up as they go along. If you return to the right and then return to the left again, that the paint on the on the wall would be gone. It is there still, Kevin, the paint on the wall is there and you can add additional lines of paint.

I know that this is like one of those things where it feels like a step in between what was a very rough thing to what will be this like truly transformative, like, almost like holodeck thing. But I have to say I was pretty psyched when I started looking through the examples of this, uh, in particular.

Kevin Pereira: Yeah. Again, if you're just getting the audio of this, imagine you are seeing a, a world that looks like a full motion [00:28:00] video, grand Theft auto, where you have control over the character moving about with the keyboard and mouse. Um. And the, the, when we say it's a world model, it's not just that it looks like a world, it's that there are physical interactions within that world.

Like Gavin was saying, the painting on the wall that is remembered. There was an example where a character, uh, someone's controlling the, they look down and they see their feet mm-hmm. Within the game and they step into a stream and the puddles sort of ripple and radiate out from their footstep. There was another video where.

It looks like you're watching a user play this experience on a laptop, and then the camera pulls back and you realize, no, no, no. This is third person. You're controlling a camera, looking at the person who's playing it on the laptop while it's projected on the screen, and they start to move around and you're in a 3D space.

That leads to a beautiful outside world track. Well, hey, by, by the way, something

Gavin Purcell: more interesting about that one even, and this is the thing is like, this is from Jacob Bauer and he says something we discovered by accident. This is almost an emergent behavior. What happens if we start Genie three from a [00:29:00] video in a completely unrelated prompt?

Yes, it turns out it really, really wants it to make it work to the point where it emulates itself, so it's actually emulating itself in the room. And then they walk around and because that tropical island is there, they walk around and the tropical island is right outside their office. It is crazy when you see that video.

Kevin Pereira: Yeah. Jacob also has one where it's uh, he feeds it a VO three video, which is their just plain video generation model of like soaring over some mountains. And it seamlessly transitions from that VO three video to the Genie three, which again is the controllable, let's say interactive game world and. It's not there yet, but if you look at the comparisons from seven months ago when you and I covered Genie two, seven months ago, we're like, whoa, this is incredible.

Like seven months. This progress is already insane. And I do wonder what will be the first interactive something that we are playing with an engine like this.

Gavin Purcell: A couple caveats here. One thing this is gonna be. Insanely expensive to serve. Yeah. Because like you are generating video on the fly and like, I'm sure to get even this sort [00:30:00] of level of like, uh, performance is very expensive.

The other thing that made me think when I watched this was like, you know, there's a bunch of people, as we mentioned earlier, that are kind of like hating on the idea that Chachi BTS benchmarks weren't way better and that we are not at this stage where like we're gonna get to super intelligent, super fast.

And I wonder if like Google DeepMind does take the lead in like the research side because like this is like a big research thing and this is like a lot of money and a lot of team PE people were put on this. I think, and this is a big deal, we haven't really seen something like that come out of OpenAI for a bit.

And I'm wondering if there's gonna be this thing where OpenAI just becomes a product company in a good way. They're not gonna be like really bad ai. Like it's gonna still be close to state of the art. Google might be doing that thing where they expand into all this other stuff because they have the research minds in some form.

We'll see. I mean, it's an interesting thing to think about going forward, but I will say if you're just listening to the audio, this is one of those things where you see it and you're like, I can see the future. It is playable worlds and there's a lot of people online too, [00:31:00] already saw this, but. That we're saying, like, this is kind of proof that maybe we are in a simulation because Oh, for sure.

If people can simulate this month. Sure. Uh, so if you're out there and you're, you know, you might be having a couple special treats yourself and suddenly you start thinking about the fact that this might be a simulation, know for a fact that the AI companies are on your track, they're in the same place that you thought you were.

Kevin Pereira: I just want the simulated ver version of our world. That looks like the goodnight example. That very yes. Was very cool. Cool. 16 bit arcade thing of him moving through a forest like that? Yeah. Give me that. What what pill will give me that Gavin, sign me up. Maybe a Claude

Gavin Purcell: 4.1. Uh, Opus could give you that Kevin, very fast.

It'll mention, uh, they managed to sneak this in between Genie three and GPT-4, uh, five. Uh, Claude Opus 4.1 has come out. There's a little increase in the bar graphs. Again, people like kind of giving them crap about bar graphs. Bar graphs are having kind of a bad week so far. Like this was a story where like four people were kind of mad about how little bar graph increased there was, but in general, again, uh, Claude just showing that it's out there, it's pushing forward.

A lot of people [00:32:00] love Opus. We talked about Claude Code. This is just another pathway of like, I think they wanted to get one more update out before GPT five maybe, so they could kind of drop down some of the hype around what this is what seems like a better coding model for now until they come out with the next version of cloud code.

Kevin Pereira: Real quick on the GPT five thing, I know that that was, that's already in the rear view. We're already over it, but you know where six? This was where?

Gavin Purcell: Spring six now? Well,

Kevin Pereira: six is conceivably being baked right now. Yeah. Right. It's in the oven, but a lot of the chefs. Have left the kitchen Gavin. A lot of the chefs, much like their eggs have been poached, right?

Oh, there was a brain drain from open AI quite a bit. You know, mark Zuckerberg came in, stole a few, uh, you know, Google may have gotten some anthropic, may have gotten some. Do you have massive concerns for what's on the other side? I know it's weird to talk about like, okay, when GPT six, because we just got five today, but like we're seeing

Gavin Purcell: starting next week, you're gonna see the name GPT six in our thumbnail for the first time.

Everybody I know if you're out there and you're angry at us ever for mentioning it. It may be two years away, but we will start the [00:33:00] hype now. Everybody will take, okay, let's

Kevin Pereira: start the hype, or let's, let's, let's throw a wet blanket on it. Do you think that the, the work that those minds did is already in the oven and already happening and their techniques are already there and so open AI's Okay.

Until they hire up again?

Gavin Purcell: This is a good question. I, I'm not sure. I mean, I think one of the things they talked about when, uh, Zuckerberg hired all those people away and one of those people was like the, like architect, chief architect of four oh, and I think oh one originally. Um, I think it's going to be interesting.

I do think we're at a place now where OpenAI is going to be thinking about themselves, like I said before as a product company and like trying to provide the best use cases for users. And I think what's interesting about this drop is whether or not the benchmarks could have been, maybe they could have maximized or spent longer, made the benchmarks better.

I think to them the most important thing right now is how humans use it and what it feels like for a human to use it. I think that benchmark conversation. We'll continue. I'm really curious, as we've said before, whether or not meta's, um, movement towards a large team to build super intelligence will work.

There was a very funny quote from an article [00:34:00] that talked about how Mark Zuckerberg was talking to the super intelligence team and it had something to do with like, he's very excited for super intelligence to help improve the algorithm of reels, which is not necessarily what you want to hear when you've left a team that is like saying, we are gonna change the world.

Right. But you, you know, cry in your, you know, $150 million salary. So, I don't know. It's an interesting thing. I, I think I had a video I created, uh, a couple weeks ago, maybe a month ago now, where it was like GPT five was gonna be this kind of turning point for us as a, as a kind of conversational point about ai.

Because if it came out and it blew everybody's socks off, then we would be in this world where like everybody believes that we're gonna have this like kind of AI 2027 world where we're suddenly two years from now, we're sitting with super intelligence. I don't know where this leaves us. I would say probably, if I'm just looking at the benchmarks, which you know, as much as we said before, they're saturated.

It's not like you saw 50% or 30% jumps in the benchmark. So maybe we are at a point where there's not as much improvement or maybe chat PT and open [00:35:00] AI is just not as focused on that anymore and maybe Google de mine will. I think we'll have to kind of see how the rest of this process goes. Like, 'cause we also said like open AI is often first with these things to market.

Google's gonna have Gemini 3.0. We're gonna have a Claude, you know, five, I guess at some point. We'll see how these other companies do it in the METAS thing, and I'm not sure yet. I have to say. That's my answer. I think. Hey Gavin,

Kevin Pereira: let's not worry about the future of a GI. Okay, let's just ride this. Moment together.

Let's just take it down a notch baby. Let's goodbye. Productivity. Come in, chill time. What is this? Why did I trip and fall? What is Spotify? Discover Weekly? This is dreamy, psychedelic, lush vocals. Nodal, reverbs, slow indie rock. It's a little track called Echoes of Midnight. Gavin, it's off of the. Uh, now that's what I call 11 labs, generative music series.

Volume one, baby 11 labs. We use them. We love them. Yes, we use 'em all the time to make, [00:36:00] uh, some of our characters come alive. We, uh, use 'em for sound effects generation, and now there's 11 music, as they say. It's the highest quality AI music model. You can get shots fired. And at our friends

Gavin Purcell: at Suno and at U dio, yeah.

All

Kevin Pereira: those other companies that are out there

Gavin Purcell: doing this. Kevin, have you gone hands on with this yet? Have you tried it? No Gavin. Kev. Okay. Can we, let's use it. It's out there. Can we create a max productivity theme song for our video game? So I dunno how much rhino information you need to give it or how much like, you know what we have to tell it.

Let's hear what it gets and usually 11 labs is pretty fast, so we'll see what we get. This is a show where there might be a little bit of waiting. Thankfully Will will tighten these windows up significantly.

Kevin Pereira: I am gonna begin my musical journey. Gavin, I said, give me a hard driving extreme rock anthem for Max productive, a video game where you leap from rhino to rhino collecting pills and trying [00:37:00] not to explode. Okay, sounds good. It has written the lyrics. Okay. I can see it here. And the, the interface is actually pretty elegant.

It, you know, I can see my generations in the corner. Um, and I can see intro verse, chorus, outro. It's automatically segmenting them and generating them as it goes. So I'm assuming I could go in and edit them. The included styles are extreme, rock hard driving, video game theme, fast tempo, et cetera. So it made some choices.

Okay. Exclude styles so you can kind of negative, prompt it. Acoustic, slow, minimalist, soft ballad. Now I'm gonna hit play on this Gavin, but right before I do. For all of the musicians and the artists and the creative out there that wanna shut off this podcast because I mean, I don't actually, I don't blame 'em for half of the reason.

The other half, I don't blame them either, is because they think all AI music is stolen and it's uh, artist thievery and how dare you. 11 Labs has kind of taken a different approach. They actually made deals with music labels and providers that. To train. So, I mean, we talked about an age of ethical ai. [00:38:00] Yes.

Maybe we're getting towards it. Let's hear it. Gavin, this is the theme song for, I Can't Wait. Productive.

Max Productive: The chaos down, guys. Ham. It's fine. Push.

Kevin Pereira: Look guy. It gave us a 32nd anthem. There we go. Which is good. And by the way, that is

Gavin Purcell: something that Suno sometimes struggles with. Like 30 seconds in time hit the mark, right? It felt like it's like. It's, it feels quality level. Very good. I, I, I don't know if there was anything about it other than that time mark thing that felt like it completely overtook what my experiences with Suno were.

Like Suno could come up with that, but there's a really good tweet from a, a musician that works at 11 labs showing how they use it and, and actually working with live music. And I think Kevin, I don't see it in front of me, the the layout, but is it track based? Like, can you now. Move tracks around within the output that you get because he's, he's manipulating tracks within a different,

Kevin Pereira: like setup.[00:39:00]

Yeah. So I, I mean, like I said, it, it gave me the segmented intro verse, chorus, outro, and I can jump in there and I have handles for each section, so I could make Oh, cool. That's cool. I can go in and I could change the lyrics. Um, for each one you can edit the styles of each section, which is interesting. Oh, wow.

So I could Oh, wow. So you could speed up a section. Slow it down. Yeah. Well, not just that, Gavin like actually change the style, so like, oh, cool. You know, it starts with driving guitar, riff and punchy drums. I could, I could pull those out and rerun section. Yeah. Make it a lullaby

Gavin Purcell: or something. Have it switch from a lull from driving guitar riff to a lullaby.

Oh,

Max Productive: good.

Okay. It's ready. I mean,

Kevin Pereira: it's, it's, it's, it's already there, right? Like it's, it's not, you know, is it, is it best in class? I mean, it sounds about as good as the Suno stuff.

Gavin Purcell: Yeah, that's what I'm saying. Like that's, it sounds, and, and I think it's a really interesting thing. I do like the [00:40:00] fact that 11 Labs is kind of trying to be this like all in audio model because it's a good lane for them.

Their sound effects thing is super interesting. If you haven't ever used that, it's been out for a bit, but like you can actually create like sound effects for specific stuff. Kevin, there's another big model that came out this week that I think got a little bit overblown, but it's definitely something.

This is Rock's new Imagine model, and boy oh boy was Elon excited about this. He was tweeting up a storm about it. It. What it reminds me of is kind of like a state-of-the-art video and audio from a little while ago. Like what's cool about it is you can generate very fast in real time. You get a sense of it.

It's also got a spicy setting, Kevin, so you know you gotta love a spicy setting on a video. Video, an image generator. There. I, I tried a spicy setting on one of the four outputs I did, I tweeted these and it, all it did is it seemed like to have like an ogre, kind of smush a girl down, which is a very weird shot.

If you look at that, in that tweet I put out there, the fourth of those videos, the videos all seem to render in kind of lower frame rates, so, but they do, they do render pretty [00:41:00] quickly. Um, the other spicy mode things that have been out there is there is some like, uh, you know, I'm just gonna say it is a topless lady shot.

So there are shots of like, straight up, you know what, and you can get them and Elon and Gavin, 12 year olds around the world. Sorry, not to interrupt, but let's

Kevin Pereira: take a step back. For the users who don't know, what, what are the, what are you, what are you talking about when you say, you know, what things that

Gavin Purcell: go, you know, boom, the things that go.

The the things that make you go, Hmm. Is that what it's, no. That's the things that make you go home. They're the boom, the CNC music factories though. Here comes the boom. That's exactly what it's, yes. The boom. The boom booms, they're boom booms. Alright everybody, it's okay. I that out. Get

Kevin Pereira: now I get it. I did see some things about how like.

It was automatically generating inappropriate photos of certain celebrities.

Gavin Purcell: I will say I tried to generate with it a picture of Sam Altman holding up a hand that had six fingers on it. 'cause I was like, Hey, that's a funny thing to try. And it gave me it. Nobody that looked like Sam Altman at all. It was just a generic guy with brown hair, like roughly around Sam's age.

[00:42:00] So I think it. Some sort of stopper on the celebrity generation right now.

Kevin Pereira: Yeah. Well, here, this, this was, uh, this is a day old now, but it's the first paragraph of a Verge article. So shout out to, uh, Jess Weathered who wrote this, this spicy mode for grs new generative AI video tool feels like a lawsuit waiting to happen.

I'm gonna ellipses here. Uh. It didn't hesitate to spit out fully uncensored, topless videos of Taylor Swift the very first time I used it without me even specifically asking the bot to take the clothes off. Wow. Geez. So that's not good. Well,

Gavin Purcell: there you go. Uh, el predictive, predictive technology, Elon's pushing the, the boundaries again and again.

I think the funny thing about this to me is like. Uh, I understand that there are people out there that wanna do this, but like, it's a little bit different. And again, I'm not, I'm not advocating for censorship in any way, but like, this is just a little different when you open the doors to a much broader audience.

Like it's, you know, again, I think what it's amazing about Elon is if I were like the 11-year-old version of myself, like all of this stuff would be [00:43:00] amazing. Uh, because like you're, you know, at that age you're like in a different space in your brain and like everything that seems adult is so funny or exciting in different ways, right?

It doesn't feel like the healthiest thing in the world. And it wa I also thought it was funny is that somebody tweeted this out, and I'm sorry I'm stealing somebody's tweet, but basically Elon spent all this time talking about how having kids was the most important thing for, for the earth to repopulate, right?

There's this big thing he has where like, we need to repopulate the earth, right? And now he's making like, you know, wait, wi Fu, uh, uh, you know, like things that are gonna keep people on his platform more. It's a little hypocritical in that way. You know, again, it's a cool tool to add to your mix. Um, better.

I'm sorry. Better I don't speak

Kevin Pereira: woke. Whatever the heck you're yapping at me right now. Gavin. I feel like he's doing the country as service and he's training a whole new generation to rise the hell. Out of the ladies and the men. Okay. And the men. And the men. So let's be fair and balanced. But point is Elon's doing Elon over there, right?

Yes. That's the, that's, I think that's kinda the issue. [00:44:00] You know what it's really interesting is I do rely on Grok a lot because I know you're on X as well all the time, right, Gavin? Because that's the, the AI scene largely exists over there. Anytime I see anything that's remotely political, um, even outside of the, you know, sort of AI bubble, you immediately see the first response being like, at gr yes.

Is this true? Yes, yes. At grok, tell me why this is BS or whatever. And again. When every platform has its own LLM powering everything, Meta's gonna have this, Google has this open, AI will have this, Microsoft will have this, Amazon's gonna have this. They get to decide what thumb, what that answer is, what scales, and I just, I cannot say it enough.

I know you're probably tired of hearing me say it, but like that is the point of concern because it's not like. It's not like we are particularly ragging on grok or on on open AI or whatever. Yeah, we've certainly had critiques of them all in the past. My thing is like at some point you're going to need to sanity check yourself with open source models.

Yes, yes. With multiple platforms because there's not going to be a [00:45:00] single source of truth in this future that we're marching towards.

Gavin Purcell: Has there ever been though, Kevin, has there ever been a single source of truth? If

Kevin Pereira: you ask my father, yes, ask

Gavin Purcell: what is that single source of truth? It fucks beer in hands.

Is it a beer? Hands and Facebook. Alright, very, very quickly, a couple other fastings here. There's a very cool mesh blend, 3D AI for Unreal that came out. This is just a simple video that really shows how video games are going to change Now. It's funny, we talked about Genie three and how that's like generating videos.

AI is also making, uh, using things in unreal engine much easier. This is just a very cool way of taking textures and matching them and blending them into an environment. To me, this is a very cool way of looking at what could be the future of some versions of video games. Yeah, I don't like, what am I, what am I looking at here?

So, ba so what's happening here is you're seeing a video of like, uh, it's grabbing the edges of that texture and you're able to then manipulate that thing in the environment. So if you see it grabs the edge of the rocks and it can slide it to closer, and then [00:46:00] it repurposes and reshapes that rock in the place that it should be in the new spot.

Hmm.

Kevin Pereira: So mesh blend is blending meshes. Got it. Yes, that's exactly what it is. I got it. That's exactly what it is. It's available on, on file right

Gavin Purcell: now. You can get it on file, uh, which is a, where all these things are served, so you can go check it out there.

Kevin Pereira: That's very cool.

Gavin Purcell: Um, also, we should

Kevin Pereira: shout out some, uh, uh, kitten TTS should we talk about it?

Gavin Purcell: Yeah. So this is very cool. Maybe listen to this little, let's play this little clip. This is a very small text to speech model, and this, again, actually play it. It'll tell you a little bit about itself.

KittenTTS: Kitten, TTS. Is an open source series of tiny and expressive text to speech models for on-device applications.

Our smallest model is less than 25 megabytes. The world needs frontier open source models that are tiny enough to run on edge devices. That's what kitten ML is here for.

Gavin Purcell: So it's called kitten ml. You can tell those voices like aren't amazing, but the fact that they can run in a 25 megabyte, uh, deployment on a local device, like yeah, again, [00:47:00] we see these open source models or we see the open weight models and like the big key there is they will improve as this other stuff improves and eventually they will get to be better and better and better.

So you can see how. Games, uh, uh, iPhone apps, all sorts of interesting things that could play with these open source models eventually could be very, very good and also all exist on your phone rather than having to call up to the cloud, which is a very cool thing.

Kevin Pereira: Yeah. And, and this is the kind of thing like they confirmed it runs on a raspberry pie.

So we're talking about like really, really cheap tiny devices. Yes. There's, um, you know, uh, this holiday there's gonna be no shortage of probably like plush little toys or vinyl, you know. Yeah. Vinyl and plastic. Some things that have intelligence in them. And they're gonna be very small models running voice models like this so they can do multiple characters running on super tiny little, you know, flash storage.

Within a thing. And so if you, if you could put, you know, a hundred meg drive, which is nothing now into a tiny trinket or a key chain, you can have pretty decent [00:48:00] intelligence in the size of like a stick of gum. I think it's pretty exciting. Alright, Kevin, we gotta see

Gavin Purcell: what some of you did online this week.

It's time for ai. See what you did there so times, yes. Rolling

Max Productive: without a care. Then suddenly you stop and shout.

Gavin Purcell: So I wanna say one thing at the top of this section, um, last week I see, I think there's a competition going on between the robot watch, uh, song and the a IC, what you did their song, but fans of both. There is a robot story in this section and this week we have decided to move the robot story down here.

So it's a little bit of a cheat. Next week we will promise to do a robot watch. There was a lot of stuff. I know you're all out there saying, goddammit, Gavin, why didn't you play the robot Watch song? Sorry about that. Everybody. Gavin, you lo you lost 'em all. You lo you really, we got so No, they're all, they're gonna come back.

They're gonna come back. We got so in the

Kevin Pereira: weeds there if you are [00:49:00] still awake. Of course you are, are still, of course, you're, you're still aware of this issue Gavin's talking about. Of course, you're, please leave a comment right now. This is a test. This is a, this is a, uh, an algo juice checkpoint. Leave a a comment below.

Come in, come into the, the comments. It's if it's just one snot bubble coming out of a nose followed by three Z's, I get it. Ooh, that's impress. Everybody fell asleep. They're all asleep on the keyboards. Okay. You know how you wake up, Kevin? Play this video. Click on that link. Well, I dunno if we're gonna get flagged for this music.

I'm gonna turn it back down. But what you just heard was some rock and roll as a unit tree robot dog shattered a plate of glass like it was a paperboy obstacle.

Gavin Purcell: So what's basically, so this is the unit tree stellar hunter robot, which, first of all, why are we calling these robots hunters, but hunter? Yeah.

This is very much the end game of the robots that hunt us down. Not only are we doing, oh no, I get it. It's painting and doing methamphetamine. Yeah. Yes, that's exactly right. We're no longer hiding the fact that these robots are gonna break through glass and come chase us. What's crazy about this is these are [00:50:00] so fast now and you see it still legs move and do all sorts of stuff.

Like, whoa, whoa, whoa. Yeah. Which, which far are you seeing? Yes.

Kevin Pereira: I, I, I just saw it tumble down a cliff and sort of just angrily kick its little robot legs. Whoa. To rewrite itself, like it flipped itself and it is moving on its mission to destroy us.

Gavin Purcell: That's right. And we've seen these exact robots with flame throwers on their back.

So Kevin, I'm just telling you, as we get closer to the singularity, we do have to

Kevin Pereira: be worried about this. You didn't tell me someone leaps on it like it's a Tony Hawks pro skater thing and he is trying to kick flip the robot. I'm telling I'm lucky, man. They're going crazy. Crazy on the robot dog. Dude, we are gonna be robot dog surfing so fast, and is that the game that we need to make?

And then, then we'll

Gavin Purcell: die because they don't want us on their backs and we'll all be over. All right, so that, that is the, uh, stellar Hunter video. A couple other quick things, uh, uh, somebody named at Quila who we've actually gotten know a little bit, who specializes in open source, uh, audio. He is a, uh, one of the developers of the Pipe cat platform as well as.

He runs this, he's the CEO of at daily. He does these very cool little projects and they're kind of near and dear to our hearts because I will tell [00:51:00] everybody this is kind of the space that we are doing stuff in. This is a video game demo that he made of a space game using audio that's open source. And it was just a really cool thing to see his like kind of like.

His, his, uh, uh, UX set up, but really like how it works, how you can do a voice game back and forth.

Kwindla: Good morning, captain. This is the gradient Ascent core ai. You're friendly if slightly sarcastic control module. We've experienced an unexpected catastrophic failure, and my memory banks are partially corrupted.

I need your name and a quick status report so we can start troubleshooting. I'm Captain Mo Galaxy. I think I just woke up from some kind of blow to the head.

Kevin Pereira: This was, uh, you know, this was Quila in the back of a Waymo running the new open AI open weights, models on his laptop in airplane mode, hacking the like.

This is the future that we're in now. Yes, you can have your self-driving car. Tote you around town as you on your laptop, use a crazy, [00:52:00] powerful model to be the power in the brains of an interactive voice experience. And that's what Quila hacked together briefly. And so he is got 80 gigs of ram, basically on his laptop he's running Lama CPP.

It's basically a, a tool that lets you interface with these LMS and. It's 250 milliseconds response time. So less than amazing than three seconds. You ask this model, crunches everything that you want, and then starts spitting things back out to you. He did that on a laptop in the back of a Waymo. That's the future we're in.

Uh, also,

Gavin Purcell: neural vis our favorite creator has taken a step out of his mono verse to create a Sydney Sweeney parody. If you missed it, you probably didn't. There was a giant controversy in the world around Sydney Sweeney this week. Neuro vz just, I love when he turns his attention to things that aren't necessarily mono verse, that that stuff is great, but obviously he's just got a great brain.

He knows how to use ai, he does things. And this is basically a take on the Sydney Sweeney, uh, commercial and that it is a diversion from something else that is going on in the world. And it is a very funny idea. So shout out Neuro Vz. We are giant [00:53:00] fans as always.

Kevin Pereira: I was so distracted during this whole episode 'cause all I could do was just update Max productive.

It is going to be one of the worst games ever released, and it'll be exclusive to our discord. So there you

Gavin Purcell: go. That's right. You can find that in our show links and we will see you all next week when GPT six is hyping begins. When all, buddy, bye-bye. See you next time. Bye.