AI For Humans Episode 104

Gavin Purcell: [00:00:00] Google is back everybody. We are very excited. They have huge advancements in video with VO two, some new audio models and a really exciting update to their deep research project. We've done some early testing and it shows that Google's product is actually performing better than open ais. Google, please tell me how make money when money all gone Ga.

Gavin, please don't know you. You shouldn't be doing that, but if you are gonna do that, you should use an agent because Google just launched agent to agent and it really signals the future of intelligent machines communicating. So we'll show you why. Yesterday's tomorrow is actually today's. Future now, today.

Okay. And Meta has released two new AI models. Llama may be the most powerful open source model out there, or they may have kind of cheated. And the AI video game debate just heated up big time industry icon, the creator of Doom and Quake, the legendary John Carmack. My hall [00:01:00] pass weighed in on the topic in gamers.

Whoa. Hold on, hold on. Did you just say, I know. Did you just say he's your hall pass? Kevin, you know what that means, right? Plus we have a rideable robot horse. Gavin, look at it or listen to it. If you're on the audio club club club, Kevin, you're the reason why John Carmack is never gonna come on here. You are too obsessed.

What if? What if John Carmack were a rideable Mecca horse? No. This is AI for humans. This is AI for humans. This is AI for humans. Shadow Up. Little Doggy.

Kevin, Google Cloud 25. Are you ready for the excitement that that conveys? Kevin, I woke up today and I got the crusties outta my eyes and I just looked out the window and I was like, same old cloud. There it is, but there's a new one. I'm just, what? Yeah, there's a new cloud. I'm for it. We finally have a new way to cloud.

It's a new way to cloud and we are not kidding. That is the title for [00:02:00] their YouTube video. But Google would, that was the tagline that Juul vapes used. Oh yeah. It was like, oh, strawberry cream. Sick bro. Look at this new way to cloud. We are not associated with Juul vapes. I wanna make sure that everybody understands.

Legally we have no connection to them. They're not a sponsor, but we are zen bros. We love a good sugar pillow, right to the gums Zen. Get on it. Vibe. Coders also not an ad. Let's talk about what Sundar Phai and the Google team came out and said today, Kevin, there's some really big stuff here. And I really wanna dive into their deep research update in a bit, which I'm gonna talk about.

But first and foremost, this is a big event. Uh, it's not the most exciting name in the world, but Sundar came on stage and kind of introduced what we're gonna see here today. Let's take a listen to that. With Google Cloud, we see AI as the most important way we can help advance your mission. The opportunity with AI is as big as it gets.

That's why we are investing in the full stack of AI innovation. Okay. [00:03:00] Yeah. So that sounds a lot like everybody else talks about ai. Kevin, what is my mission? Yeah, that's why I'd like to know. Can you define my mission? I'm still trying to figure that out. Gavin, what do, do you have one? Do I have one? Do we have one collectively?

No, probably not. Um, some really interesting things are coming. Not everything is out yet. We expect that this stuff will probably be rolled out over the next couple days. VO two is going to be coming to everybody and this is a very big deal for creatives out there. I have had access to VO two for a while.

It is a very powerful, uh, text to video model. It can do a lot of interesting stuff. I won't say that it's like significantly better now than something like runway Gen four, but it is very good. It does have a lot of restrictions on it, but I think people will play around with that. They're also gonna release updates to chirp, which is their text to speech model and.

Lyria, their text to music model is going to be released in private access. It's in private access right now, but we expect to see that. I am really curious, Kevin, about that particular one, to see how it compares to say something like [00:04:00] Suno. Yeah. Well, we kept hearing about, uh, at least on the YouTube side of things, deals with the music companies to push new AI tools that lets you make music, uh, remix stuff.

Is any of that gonna make its way into the Elyria product when it comes to maybe generating new music or new sounds? They had some DJs on stage. We've been sort of, uh, half watching the keynote while we prep for the podcast. Spoiler recording it simultaneously. And, and they're mixing, uh, imagery, which is created by VO two.

They're mixing sounds, conceivably made by, uh, Lyria. So that, I mean, interesting on the media side of things, but as, as maybe, uh, let's say uns spicy or as unsexy as a new way to cloud might be for the general listeners out there. When you think Google Cloud, some people think, oh yeah, Gmail. And, and Google Drive.

And, and yeah, those are products that can run within the Google Cloud. But we're talking like server infrastructure, which powers a lot of things. Core connective tissues and APIs. Um, they list, uh, that products [00:05:00] and platforms is at the top of the stack. Models and tooling, world-class research, all sitting on top of AI infrastructure.

Like that is the cloud as they see it now. And again, the bottom of the stack is AI infrastructure. This ain't your mommy and daddy's old server farm. This is all, you know, uh, hard, hard plumbing for AI and to power all that. Well, what are they gonna do, Gavin? Are they gonna write checks to Papa Jensen or are they going to invent their own Ironwood?

TPU Gavin, can we get to the sexiest of the sexy announcements real quick? Can we hop into the new frontier of ai? I mean, you really wanted to talk about this, so that's fine. I want to make sure that like you get your time to shine. Spoiler alert, Kevin. This is okay. First of all, I did want talk about this because peak performance, Gavin larger pod configuration.

Tpu, did you see the chart? I did see the chart, Kevin, and I think in general what I wanted to talk about here, it is really goofy to start talking about this [00:06:00] stuff. My favorite stat on here was peak watts per flop, which is a new stat that we have to be aware of. Peak, it's peak flops per wat. Oh, I'm sorry.

Which is different. Then peak flops per pop. Yeah. Which is what happens at the lake when your dad belly flops in the water. So this is their new tensor processing unit. It's a new version of it. All of this is interesting. It's a very dense conversation to have, and we're not gonna have it here, but you should.

We'll put the link of the show notes. This is basically Google putting together a new kind of hardware infrastructure for these AI models to run on. And something we talk about in the show a lot is inference compute and the idea of these reasoning models. And almost all of Google's new Gemini models, which are very good, we're gonna talk about a second here, are running on reasoning models need a ridiculous amount of compute.

So what they're saying here is that they're making specific hardware. That can run AI models at a much, much faster, much higher level than stuff they had before. Yeah. This is, uh, you know, we talk about energy independence. This is infra, infra [00:07:00] independence. Yeah. This is Google trying to say, Hey, we're not reliant on Nvidia or a MD, we are going to, you know, do our own thing.

So it is a, I just wanted to get the peak flops per watt in there, and I know you were excited. I'm excited, but let's get, get back to the stuff like that, that will pay dividends to everybody listening to this podcast in a few years time, potentially. Yeah. But the stuff that they can get their hands on today, we gotta talk about this new deep research tool.

Yeah. So to me, this is the biggest advancement, at least outside of the stuff that's been released publicly yet. And again, VO O2 is very good. As soon as it comes out, go try it. I believe it's one of the top video models, but deep research is something that I've been interested in for a while because Open AI's product, if you're familiar, it's the agentic ability for it to go out and search the internet and then come back and produce a report essentially.

This is really agentic ai, the first real. Use case of a agentic AI that a normal person understands can get. Google's version of this before was okay, it wasn't amazing. They have now updated it to use Gemini Pro 2.5, which is also [00:08:00] now in the Gemini app. So when you go to the Gemini app, there's a dropdown and you can get this particular model in.

That dropdown model is a thing that says deep research. Click through to that. You go in and you can ask it anything in specific and it will do a long form research project on it. I did a couple things here. That were both, uh, really interesting one. I had asked OpenAI to do this before and I just wanted to ask, uh, deep research.

Uh, the Google, they're both called deep research, which is also so stupid. Yeah. So can we, well, at least Rock went with deeper research. Exactly. And someone's probably gonna go super duper deep or the deepest research. Uh, how, what's the Google Cloud, uh, slogan again? Can we come back with a new way to cloud, A new way to deep, A new way to deep research?

How about that? It's a new way to deep research. Okay. So anyway, I asked the OpenAI one a couple months ago, if you remember this, or maybe a month ago, um, to give me some insights on how to achieve the e scores in the game tro. So if you're not familiar with tro, it's a very fun card game, but one of the things you really wanna try to do is get past the highest of the high levels.

It [00:09:00] goes on forever. If you can make it. But the eCore means that you've hit a score that is too big for it to show on screen. And there's even a further score called the Nan Infinity score, which allows it to go even further than that. Completely broken. Yeah. It's, it's completely broke, which is fun, but you can continue to do that.

So I wanted to say, go out and find me the strategy to do this. I put in a deep research and it provided a 15 page document that, and, and I want to be clear, the open AI one did an okay job with this as well too. The difference with the deep research one when I got outta Google now, is it's readable. Like it feels like it's been generated by a very smart person who has, it doesn't feel like a technical white paper, right?

Yeah. You're not looking at like the formulas. It is, it is presenting it in a, in a enjoyable way. Yeah. And, and you know, the opening eye one, I went back and did it again just so I'd have the direct comparisons and like the information is, a lot of it is in there, but it is not organized it in a really interesting way.

And I do wanna shout out the fact that like, I. Um, Logan Kilpatrick, who I know has been who, who is Google's like in a mouthpiece [00:10:00] specifically said that early tests show that users preferred this two to one to other products. So they actually released some stats here, which is interesting, and they did a comparative evaluation of Gemini and open I deep research and on these stats.

Overall it's at 70% versus 30%. The one that I thought was most interesting though is comprehensiveness. Where Gemini's Deep Research now scores 76.9% and OpenAI scores 23.1. Now, granted, this is Google releasing these numbers, but I will tell you that I felt that in reading this thing. The other thing I did, which I thought was really interesting, is I basically went and asked it for.

A biography of you because I was like, well, who do I what, what can I do with something that I can think is interesting? Great. You background checked me with deep research. I basic, so I asked it specifically for, I said I'm doing research on my favorite internet celebrity, Kevin Pereira. And I'd like to help, I'd like help Col collating a biography of him with all the highlights of his career.

I'd also like you to gimme insight to his personality and the sort of things he loves and what [00:11:00] makes him tick. So then Gemini Advanced asked me, you know, deep research asked me, basically I'm gonna, it gives me this kind of heads up and it says, okay, I'm gonna go ahead and start. So I hit and start it then goes, does did it ask you clarifying questions?

It just go right away because sometimes it does, doesn't do. Yeah. Okay. In this case, I've often found that open ai, deep research does and always like, tries to get something out of you. In this case, it just kinda shows you basically what it's gonna do and then it just says Start research or you can edit plan.

Mm. Okay. So I go away. And what's interesting to me about this, I sent this to you as a Google Doc so you have it, um, but it produces a very long, very detailed thing. And what you see at the end of it is sources used in this report. And it's a lot like, it's like, you know, 25 and it's searching through YouTubes, it's searching through podcasts, it's searching through, um, IMDB pages, Wikipedia, all this stuff.

And at the end then it says, sources read but not used in the report. And there's an even longer list of things that it reads. Oh, wow. Yeah. So does it, does it have my foot wiki? [00:12:00] Ooh. I don't know. Let's see. I probably not, is there one that exists? I may. It probably avoids those on purpose. Is there one that Gavin, like, you haven't browsed my foot Wiki.

It avoids that on purpose, but it did. So one of the things that was interesting about this is it reads like a very extensive, much more dynamic Wikipedia page, right? And I thought, mm-hmm. Oh, this is an interesting thing. Like it would be interesting to see an AI directed Wikipedia. And I know there's a lot of hardcore Wikipedia people out there around like, you get that AI out of our Wikipedia, we do this.

But like, well there is an interesting idea in this form. Oh, sorry. Yeah. I didn't mean to cut you off Gavin. I just, what, what is interesting is that like conclusion? What makes Kevin Pereira tick? Yeah, like. That it's going out of its way to try to get into my psyche and make good, you know, draw conclusions about me based off of this research.

Yeah. Is different than just presenting you, here's a bunch of facts that I found. Yeah. It's like, oh, now let's try to draw some insight from it, which is, um, scary. And to be clear, I did ask for that. I asked for that specifically. Right. Like in my interesting, I said, I'd like you to give me [00:13:00] insights into his personality and the sorts of things that make him tick.

What does it say? What makes you tick? I'm curious what Wanna read it for us? Wanna read the line there? It says, uh, it says, uh, Go-Gurt, Pop-Tarts and occasional squeeze. Its. That's amazing. I really didn the evidence suggests a powerful combination of factors. There's an undeniable creative impulse that need to build, launch, and communicate, and, uh, a drive he acknowledges will never cease.

There is a deep seated fascination with leading edge, technological and cultural change, which is why NFTs are where it's at baby. Evident from his early embrace of the internet to his current focus on artificial intelligence. Yeah, I mean, crucially. Oh, sorry. Oh, Kevin. No, no, no. I'm still throwing flowers at my own feet.

I dunno if we need to go through the whole thing. There's his talent for translation. Taking niche complex or emerging subjects and making them accessible, dragging, shouldering, and elevating his co-hosts along the way. Oh, do they deserve the accolades? It does not say that Kevin is making [00:14:00] that up.

Deserves to live. This is so weird. Wow. I can keep scrolling. This is fantastic. I do, I do wanna, I do wanna read one particular line that I thought was something really interesting that only somebody who had spent a lot of time learning about you would know, which is uh, and that towards the end of that section it says, perhaps most importantly is appeal lies, and a personality that defies that definitely bounces sharper and humor with a relatable vulnerability and authenticity.

I do believe that's actually pretty good. He can mock the hand that feeds him. And then parentheses, it says the word meat puppet, which is so interesting to me because that is a word you use a lot and you may not use it as much, but I've heard you say that a lot. Yeah, it clearly understands that aspect of you through doing all this research and that just kind of blew me away.

So I think this is one of those kinds of things that, again, you have to be a paying Gemini user to get this. I'm a Google Cloud person. I got hooked in 'cause of Google photos, so I pay, right? You do get 20 of these a day, which is kind of crazy. That's a lot. I'll also say, I did another one of these I don't wanna talk about here specifically.

But I, it was about [00:15:00] something I wanted to do in my future that it blew me away. Like it literally blew me away at how good it was. It was a step-by-step process to get something done, and it was a really interesting thing. I, I really recommend people go try this now, if you can, I wanna respect the, the, the boundary that you just set.

But how do you build a vac bed for furries? Won't, the, the, the costume won't tear through the vinyl when you start suctioning the air. Okay. You know what, I, you know what? I respect your boundaries, Gavin. That's fair. And I do, oh, I'm sorry. Wish, wish you the best in your, my camera. My camera just, uh, froze.

Kevin, did you want me to say something about that? Google's cooking, that's the theme of the day, right? Google is cooking another thing that they released, which I, I, this, this can veer into nerdy, but this is gonna be so impactful. I can feel it. Yeah. I'm excited for you to describe this conversations because I've been learning, I've been on this very nerdy group chat where they're really excited about MCP and this connects to that a little bit.

Oh yeah, yeah. Oh, for sure. Yeah. This is the next evolution of that. So, um, well, MCP is a model context protocol. We've talked [00:16:00] about it or attempted to on the show before, from. Uh, yeah. Anthropic release. It's fully open though, so anybody can use it. Even OpenAI has integrated it and it allows large language models, these AI tools, uh, to discover APIs and abilities that they have.

Meaning if you want your LLM to talk to anything, your MacBook or Spotify, if there's an MCP for it, you can just say, go use the MCP and figure out what you can do and how to do it so you can make these natural language requests of it. Okay. We're talking about a world where AI agents are gonna go off and do for you, just like we discussed with this deep research project.

Well, an agent can go off and do what it can do, but what about the things that it doesn't? Know how to do. Mm-hmm. What about all of the other services that it could be connecting to, but they're different tools running on different servers with different sets of capabilities. How do they talk agent to agent looks to solve that?

It is a, an open protocol that Google has developed with support from [00:17:00] hundreds of very, very shiny logos in their, uh, partner ecosystem. It allows. AI agents, regardless of what they're written on, to talk to other AI agents. So for example, if I say go out and one of the examples that they use, go out and find me a candidate that would be good for this job that I'm posting.

Here's a PDF of what I need from the job. That agent goes, okay, let me go look at an agent database, see if there's something that has expertise there When it matches with an agent, right? Which like a resourcing agent, they exchange a card and that card has a set of info. It's like the old modem, handshaking sounds.

Yeah, sure. Where they would screech at each other. Crack cracker. It goes, oh, I, I have this ability, this ability, and that ability. And the other thing goes, great. I need this, this, and this. It's a client and a remote agent, and so the client agent makes the call. The remote agent goes, oh, okay. You're looking for these features.

Here's my card, here's the compatibility that I have. They speak in these little parts and these little tasks. And what's interesting about the spec as it is now is that can [00:18:00] be text, it can be audio, it could even be video. So they're built to be multimodal agents, right from the jump. So you might have a video agent connecting to another video agent and exchanging video for any solution.

And, and the response time Gavin can be, we are going to work together in real time and I'm gonna tick off tasks and report back to you. Or a session can sort of remain open and that remote agent can go off and do its thing. Getting back to you days later with like, they had a background checking example that was literally that where it said, oh, you want me to do a background check on these employees that I'm hiring?

Oh wow, okay. I'll be back in one to two days. And so the remote agent would then come back while you check in on it, you're like, where did you go? It's like, it's in Mexico somewhere. Sipping on being coladas. I can't come back right now. Uh, have a headache. The agent, the agent, the, the remote agent has actually, uh, thrown a flash bang through a window and kicked open the door.

It's like I needed more information. I'm sorry. That's super cool. It's, it's really interesting [00:19:00] to me. It's really cool. This is the future. It feels like, I mean, one of the things I think a lot about with this is if you're in our audience and what Kevin just said sounds like, kind of like, kinda surge your brain in a, in a kind of a mishmash way, like kinda scrambled your head.

I think the important thing to know is that ais are going to talk to AIS in the future, and we are Yes. Sometimes not gonna know what they're saying. So what we're trying to do is make sure that the connective tissue between them makes sense for us. Is that right? That's the, that's what it's kind of like trying to do.

Yeah. It's really about like. The in the future for, for you, me, everyone listening to this, we should just be able to say what we want or need to the machine and then the machine sorts it out. And right now human beings are still a middleman between ai. Like AI are super capable. But you need your video.

You wanna make a video, okay? You go talk to one ai, it gives you the script, then you go take that script and put it to another thing. Then you, you wanna repair your house. Well, you could ask AI to do it and it could create, oh, here's the problem that you need solve, but then you gotta go and send it [00:20:00] to another AI to go out and field and get bids and blah, blah, blah.

That's not going to happen in the future. And technologies like this are going to allow for you to just whisper your needs into the machine and the machine is going to sort it out. And one last nerdy thing about it is that even the way that that information that's retrieved is displayed to you, Gavin is in this protocol so it can discover, do I just give you a text response?

Does the agent have the ability to put imagery on the screen? Can it speak to you? Can it pop up in a video avatar and give you a presentation with that information? Like, this is the groundwork. So if you know agents are going to be a thing, how those dots connect, that's what this agent to agent infrastructure is about.

Yeah. And what you, when you, you're saying all that, when you say all that, it makes me think about, uh, how important the personal agent is gonna be. And when I say the personal agent, I don't just mean. My agent, but I also mean like the personal agent for you at work or the agent like that. Because [00:21:00] I keep thinking more and more about the fact that like in five to 10 years, let's say, and I think it'll be that long until we really get to this space, there will be a personality who you are interacting with on a regular basis.

That is your, you can call him whatever you want. It's like your version of her, your agent or your, your like persona that will be delivering you and taking in a lot of your stuff every day and like doing this stuff. It's like a sci-fi idea, but it is really is coming in. This is the very groundwork for it, I feel like.

A hundred percent. And you know, your personal agent, even in the near term, like forget the charismatic personality side of it, but like we will wake up in the near future, I think with a daily digest of here's what happened overnight, right? Here's what your day is looking like. Here's what the next week is looking like.

I can help out with these three tasks. So if you want me to go and send out a bid request for the plumber to fix the issue that you mentioned and remind you about the media that you need to watch or go discover three new shows tonight. Yeah, [00:22:00] because you said you want like whatever those things are, agent to agent would allow your personal agent to go out and talk to all of the machines on the internet, everything that's connected and come back with those results.

Yeah, it's gonna be interesting to see when that happens. But you know, Kevin, whenever there's an AI story, you know who I really think is worth checking in with is my go-to guide is Solar Meditations on TikTok, and she just recently weighed in on AI, and I just wanna share this with everybody if you're not familiar, solar Meditations is a woman who uses dowsing rods to talk to the universe and ask them questions.

Kevin, can you play this for me? Yes. Gavin, I can play it for you. Is AI going to end humanity? No thank you. Oh, thank you. Does AI have consciousness? Yes. Thank you. Will AI help humanity solve the hunger problems around the world? Yes, thank you. You know, you, now that I've heard all that, I'm not worried about ai, these dosing rods.

I'm not worried about ai. These [00:23:00] dowing rods are accelerationist. I think they, they actually are fans of ai. This is, uh, solar glow meditations on TikTok. You sent this to me, Gavin. Uh, and then had I, I don't know if you use the bold, I like I my message text, but it leapt off the screen with OMG. We have to show this on the show.

I love Solar Glow meditations. It, it, I don't, I, I will say, you know, TikTok is in a magical place where sometimes stuff gets delivered to you. This was from about a month ago. This woman blew up on there and she just like, you know, she asked questions of the dowsing rods, and today she asked about ai. So I thought we'd throw this in there, Kevin.

The other thing that I also love is when our viewers follow us and subscribe to our YouTube channel, solo Globe Meditations actually just said that the, the universe has said they should be watching and subscribing to us. Will the audience like and subscribe on YouTube? No, thank you. Oh, no. Will people go to AI for Humans Show and join the AI for Humans newsletter, which comes out for free twice a week?

Yes, thank you. No, thank you. That's right everybody. It is free and [00:24:00] we are writing a lot more in the AI newsletter. Comes out twice a week and also go leave us a review on Apple Podcasts. Really, we would love to see a couple more of those reviews. Our audio audience as always keeps growing every week.

Yeah, so that makes us really happy. We're thrilled about that. And. Just to clarify, we don't advertise, we can't afford it. Yeah. Uh, quite frankly. So literally this thing grows because you share it with everyone. So if you're hearing this and you haven't shared it yet, please do. If you have shared it, thank you so much.

We also have a Patreon if you wanna stuff, stuff in our, uh, tip jar, but we digress. Okay, Gavin, so let's talk a little bit about, very quickly, about what, what open eyes. I think pre response to this Gemini stuff was, so this is my theory here. So, uh, Sam Alman had just come out earlier this week and he basically tweeted this.

He said, change of plans. This is from April 4th. We are going to release oh three and oh four mini after all probably in a couple weeks. And then G do GPT five in a few months. There are a bunch of reasons for this, but the most exciting one is that we are going to be [00:25:00] able to make GPT five much better than we originally thought.

So there's a bunch more stuff here, but this is an interesting thing timed where it is because I have a theory that, I mean, it's not even a theory. You can tell by the way they talk to each other that OpenAI and Google are really butting heads around like where the state of the art of this thing is.

Right. And I think, yeah, you know, Google dropped 2.5 Pro last week or a couple weeks ago and OpenAI dropped their image gen, which I guarantee they were holding onto for something for this exact reason to kind of jump ahead of, of Google's announcement. Now we have a week of really Google looking pretty good.

And I would say that like there's a real argument that Google is pushing state-of-the-art. I've seen a lot of people say that they're using Gemini 2.5 Pro for their coding needs now that this is gonna be one of the main, I. Yeah, so I think this is Sam in part, like not taking a step back per se, but really kind of like saying, okay, we are going to make sure that the next model really is that step and not try [00:26:00] to release something that's gonna get crapped on.

Yeah, I think that's fair. And when you talk about the two companies butting heads and the fierce competitiveness between them, if you need more signal there, friends, Google has been accused of paying employees to do nothing. That's right. I forgot about that. Yes, that's right. Up to one year. I just had to shout this out.

They wanna stifle like people in the AI departments leaving to go and work for competitors. So even if there's literally nothing for them to chew on at the moment, or they're thinking like, ah, I might need a break. They're saying, okay, we will pay you for a year just to keep you around. How do I get that gig?

That's like, well, it's funny, a lot of people have been using that Silicon Valley meme. Which was literally, yes, the guy goes to the roof. Yes. Big head goes to the roof and there's six people and that's the people that are just paid to stick around like it is crazy. The, the other thing I thought was interesting that OpenAI dropped is that they are creating a strategic deployment team.

And Kevin, this kind of just talks about the idea of, it's basically a team that is preparing for an a GI future. So this is like a team that is trying to figure out what does it mean when we get to a GI and [00:27:00] how is that gonna affect everything we do going forward? This feels like to me internally at some of these larger companies, you're seeing this a lot and I know there's a real kind of like hype beast versus realistic sort of world that goes back and forth.

We've come to a place now where I think the vast majority of people in the AI research space feel that we are, you know, two to five years away at, at, at most, maybe from an a GI, I think Dennis Haas is on the far end of that. Uh, Dario, Modi and Sam are probably on the lower end of that. I forgot about this, but there's a really amazing paper that you should read called AI 2027.

We'll drop in here. Yeah. It is a sci-fi. Look at a future of where we would go if we really were on the path. We are. It's gotten some criticism because people think that these guys who wrote it are AI doomers, and the guy who specifically did the research for it is somebody who left a OpenAI earlier because he wanted to kind of call out what OpenAI was doing.

But it is a very interesting read about the possible future of not even a GI, but a super intelligence that I think [00:28:00] people in our audience should watch out for. I, I, uh, fully go to AI dash 2020 seven.com. Hashtag not an ad or an endorsement, but, but, uh, people should read it and, and take a look at it.

We're not saying we agree with every piece of it, but one of the things that I saw people like debating, which to me is not debatable, is the, uh, AI capabilities for things like hacking and coding or for robotics. And it is just like, well, it's never gonna be that. It'll never be a hundred percent this, that, that.

Like, no, I, I, I, I think the numbers are in on that. You can chart it. Right now, like this thing will be so good at writing code. Um, it will be very good at pen testing that code and that can be weaponized against people. You know, there's can be people that use it to find flaws in their systems and fix them just as there's gonna be teams trying to deploy them at adversarially to destroy someone else's system.

So that's, that's worth paying attention to. And that actually leads us to a really interesting thing is that Llama dropped a few new llama four models, which we've been waiting on for a while, and Mark dropped them on a Saturday, which was really funny. There's a [00:29:00] funny tweet where somebody asked him, you just dropped this on a Saturday, bro.

And he is like, yeah, that's when they were ready. But I don't know, maybe this is a weird thing because of Gemini as well. Maybe they wanted to get ahead of Google Cloud Week. Um, but Kevin, I, these are interesting. There are a few issues with them, but like, let's talk a little bit about what the models are first.

So three models, two available right now. The biggest model, which we can't touch is the behemoth. Yeah. A 2 trillion parameter model. Um, they use that model to distill the other two models. So it's sort of like the teacher model for Maverick and Scout. So LAMA four Maverick, this is a 17 billion active parameters.

It's a mixture of experts model, which means when you ask a question, it has 128 experts that have been, uh, we'll say fine tuned for specific responses like code generation or creative writing, et cetera. It will figure out which expert. Is best to answer the question and we'll give you the answer. And what that does is it cuts down on the overhead, the amount of, uh, compute and memory, uh, [00:30:00] the, the overall system requirements to house this model are less because it is just only activating the expert that it needs at any given time for an answer.

So it's, it's efficient. It has a 1 million, uh, token context length, which is massive, pretty big. We're talking like, yeah, take, uh, libraries and jam it in there if you want. It's a little hyper hyperbole. Take a small library, take a, take the Antioch, uh, public library, which is where I got to go to, which has like four books.

You can put that in there. Then they have Llama Scout Gavin. This is their smallest one, uh, smaller, uh, or excuse me. Uh, industry leading 10 million context length, which is, that's pretty big. Mm-hmm. Uh, 16 experts as well. But the claim to Fame on Scout is that it can run on a single, an Nvidia, H 100 GPU, and for you and I, that means like, oh, it's a very capable model according to the benchmarks that can run in the near future, maybe on a device we have in our home.

Yeah. Like on a, a higher end graphics card, which is great. And again, this is the [00:31:00] benefits of open source models, right? The idea that you could run that on something in your house or, you know, some, maybe even a robot dog or something that will actually be needed. Like that's good to have these open source models.

Now there is a little bit of controversy here. In that they came out and kind of were originally touting some pretty significant benchmarks and it sounds like they may have, at least, this is all rumored right now. They sounds like they may have futzed a little bit of how they worked around these benchmarks, and I wanna make sure I understand this right.

There was a very good investigative story. I. From the information that kind of dove into this, they're talking about how some developers questioned why the model performed differently based on how it's being accessed. Meaning that you could de, depending on where you got it, whether it was an inference provider or the model that somebody downloaded themselves, they were having different experiences.

So all of this stuff led to a weird thing that happened on LM Marino, which is how these different models are tested and the, the actual vice president of AI for Facebook said specifically, we are glad to start getting LAMA four in all your hands. [00:32:00] We've already heard lots of great results from people getting these models.

That said, we're also hearing reports of mixed quality across different services. Since we dropped the models as soon as they were ready. We expect it'll take several days for all the public implementations to get dialed in. We will keep working through bug fixes and onboarding partners. So the interesting thing here to me is that, you know, clearly meta is feeling a little bit of pressure.

To stay up at the level of these other companies. And we had heard a while back that Deep Sea kind of threw them a little right when Deep Sea came out, because that's an open source model. Sure. It was like way more advanced than the meta models. So I think probably Meta is kind of scrambling a little bit to get caught up to everybody else, specifically Google and opening eye.

Well, Gavin, did you race to poke and prod at the new meta models? Because I gotta tell you, I, I didn't, no, I didn't either. I didn't either, so this is a good point. I spent, like I said, I spent time on meta ai, but I did not spend a lot, and it's not because I don't think there's a lot of really smart people working there, but it, it's almost [00:33:00] like when you watch a, when you see a movie now, if you hear a movie or a video game's a good thing too.

Mm-hmm. If you hear a game is just okay, or you hear a movie is just okay, you do not rush to buy it. You do not rush to see it. You're like, oh, maybe I'll catch it later at some point. That is a weird way of how this is working with the AI space now too, because like I know all of these models feel very similar.

So you have to kind of go above and beyond to get me to switch or get me to go deeper on your thing. Correct. So it feels like unless Meta's coming with something that really blows me away, I'm not gonna switch over for that. Nobody is signing up unless you're gifted a new superpower. Yes. Like if there is a, let's say eight to 10% edge over an existing product, then sure.

Go outta deep research. Like Google's deep research feels exactly like that. It feels like eight to 10% better than open ai. Yeah, but if you're ultimately saying like, eh, it might be like a half a percent better and that's probably not gonna make much of a difference with what you and I are doing with these LLMs.

Like, sorry. [00:34:00] Like it's not there. So speaking of this exact same thing, mid Journey V seven came out this week, which Kevin, you know, is, is a, is a fine and awesome image model. Like this is a model that's been around forever. In fact, my first real transformative experiences were with early Midjourney. In fact, I remember Midjourney three being like a major touchstone and we, I went back and looked at like an old episode of ours from like almost two years ago now.

We were talking about how incredible Midjourney three is. So it was the state of the art for a long time. Midjourney seven came out. And yes, it's an upgrade, it's definitely good, but the thing that it doesn't have that I think is really affecting my perception of it. And by the way, I will be very honest.

I have not resubscribed, 'cause I had not subscribed I'd canceled my Mid Journey subscription. I had not resubscribed because I had seen enough people talk about this. Open AI's four oh Image Gen has solved context and has solved this idea of give me this image to do this thing. Now obviously it has also been nerfed in a significant way, [00:35:00] but midjourney, from what I've read, it still hasn't figured out like that to pull off the slot machine mechanic.

And to me, I'm now past that mostly with image generation. There are a lot of artists who shout out the fact that Mid Journey seven is delivering really compelling images. And it is, I've seen these before, but from what I want an image generation software to do, I'm much more directed. I'm more using it to make AI movies or I'm doing stuff with it that I wanted to do specific things or as well as we'll learn later in the show to make buff versions of, uh, celebrity, uh, uh, movie cast.

We'll just get to that in a second. Sure. In, in, in a lot of ways, I am not gonna fold back into this until they prove to me that they have solved that problem. Right. And I think that's the same sort of issue with Llama. It's the same sort of idea. The, um, flagship feature that according to the V seven release is going to be draft mode where it can render images at 10 times the speed.

And people like this usually does from, from people are tweeting about, they like it, and there's a conversational mode that you can have with it. So you [00:36:00] literally are just kind of telling it, Hey, change this to that, or tweak this to this. It it, with respect to the conversation that we were having when we announced V seven again, impressive model.

We love midjourney, love the company, love the images. It generates on the heels of open AIS chat, GPT image creation tools. It's like, wow, that, that kind of, that kind of did it, you know? And it's gotta be, uh, you know, again, a, a. It's gotta be markedly better to get you to switch back. Yeah. And resubscribe to something else.

And, uh, I think it's just too early to tell with V seven. Yeah. I mean, my only takeaway on this at the end was I do remember when Dolly three came out with like, the ability to change stuff in the image. And like, that was really cool. And then Midjourney kept iterating and Dolly three just didn't update until just now.

Right, right. So there's a world where three months, six months from now, mid journey does pull ahead again. But right now it feels like I have no reason to go back to it. All right. So Kev, we should move on to. What I think is a really interesting conversation in the space of AI and [00:37:00] gaming. Yeah. So, uh, Microsoft basically created an AI generated replica of Quake two, the classic first person shooter, which had been open source.

So people hack it into toasters and even pregnancy tests. I've seen doom running on one. They take these open source games now and then they, I mean, they're classics and they get 'em running in weird ways and on all sorts of stuff. So Microsoft made a version where every frame of quake that you're playing is actually being generated on the fly by an AI model.

Um, a lot of nerdy reasons why that is very technically cool and maybe points to a future of interactive entertainment. But Quake dad, quake dad into the old arena and quake Dad said, this is absolutely flippant, disgusting, and spits on the work of every developer everywhere. I saw this Gavin and I thought.

Probably a bot to be honest like that. Because let me be clear, and nothing against you Quake dad. I don't know [00:38:00] you personally or which foundational model you run on if you are a bot. Um, it's not a bot. Assume just, just Patreon, by the way. So I don't think bots have Patreons very often, so I subscribe to so many AI influencers.

Gavin, not true, but I, I, my, my thing is that like, I just assume that most of social media is AstroTurf now. 'cause a large percentage is, is bots, but. Then John Carmack. Yeah. Industry legend leapt in and said, what? This is impressive research work. Who does he think he is? Chiming in on Microsoft doing something with Quake two.

Gavin. So this is a really interesting conversation because the, as we all know, uh, the gaming space is full of outrage of many different things. Gamers come at things and, and AI is no different. There's a lot of anger in the world of, of ai, uh, gamers for ai. What's really interesting about this is that the guy then responded and said, John, I respect your work and you see you've always been tech forward, but this type of design only serves to steal and render o other dev work useless.

A [00:39:00] fully generative gain, cuts out the number of jobs necessary for this project. And then John. Did a very long and detailed yes response where he talks about the idea. He says, I think you are misunderstanding what this tech demo actually is, but I will engage with what I think your gripe is. AI tools trivializing the skillsets of programmers, artists and designers.

Yes. And he goes into talk about why this isn't just. That it isn't just that, it's also the opportunity to open the door to all sorts of other stuff. And specifically it talks about how for him when he started out, that's what was happening. Yeah. So again, level setting. In case you missed in the intro, John Carmac created Quake, you know, he was the it a of stuff software End Doom.

Yeah. He's a gaming industry. Icon Luminary. Um, again, hall Pass. Probably never going to be a guest on this show, but one day, you know what, one day. Can we, should we encourage people to, we don't know. Don't spin. No, don't, don't do, don't do that. But gently nudge, gently nudge and say, come on, we don't guests either.

So John, this would be a [00:40:00] very special moment you would have come on. We haven't had a guest for about a year now, so I wanna talk Keen baby. I want to talk Keen Technologies. 'cause John is also chasing ai. He's trying to chase a GI or Super Intelligence. So anyway, um. The creator of the thing is saying, this is impressive research work.

He says, AI tools will allow the best to reach even greater heights while enabling smaller teams to accomplish more and bring in some completely new creator demographics. Yes, we will get to a world where we can get an interactive game or novel or movie out of a prompt, but there will be far better exemplars of the medium still created by dedicated teams of passionate developers.

Yes. Yes. And therein lies the thing that, um, you and I have been trying to say. I think so eloquently and succinctly. Yeah. But Carmack is a, is a, is a genius and I agree with them here. Yeah. And I think this is just the thing to be continually aware of, and it's a good thing to hold in your pocket if you do deal with people who are anti ai.

Is that like again. This is a world changing [00:41:00] thing. This is gonna change stuff. Mm-hmm. We are gonna see stuff change. And it doesn't mean it has to be all bad for everybody. Like anytime change happens, it is scary and it can change the way you do stuff. But again, with video games particularly, maybe we're gonna end the world of having 500, 2000 person teams for a game.

Or maybe there'll be one or two games like that. Like GTA eight will still be that way. But of the like 10 other games that came out and we're kind of middling, again, games that I wouldn't play because they kind of came out and got midling reviews. I'd rather see 30 of those people take a swing at a $5 million game.

And that's the kind of thing that this tool could really align with, I think, in an interesting way. Yeah. And let's say you wanna make, I don't know, a Tom and Jerry cartoon. Gavin, let's say you wanna create one of those. What's the best way to do it? Use AI and steal all of the work that the animators have done.

And use a brand new tool that has test time training three Ts. We're gonna talk about it. Yeah. Yeah. Three Ts. Yes. This is a really interesting piece, a [00:42:00] a new piece of research that just came out from some researchers at Berkeley and Stanford and a bunch of other places. Basically, what they have done is found a way to train a smaller model on a series of images.

From this particular example is a Tom and Jerry cartoon, and they found a way that while the generations of the AI video were happening, they kind of touch it throughout so that they can make a minute long, very comprehensive, um, output from AI video, which, you know, we talk about AI video all the time.

It's either five seconds or 10 seconds, but pretty quickly after even the five seconds, sometimes you start to lose cohesiveness. These researchers have found a way to use very long prompts and being able to use this test time training sort of scenario. Continue that prompt to be very good. Now if you watch what the video that's on your screen or you go look at in the show notes, it's not gonna seem perfect yet.

But Kevin you can, you can, you can very clearly understand 'cause we've seen this stuff change over time. Yeah. A minute long video would just never have held this sort of uh, uh, comprehension for this long [00:43:00] as Carmack said in his tweet about you'll be able to get a full novel or movie or video game from a prompt.

This is the early signs Yeah. Of exactly that. Yeah. And we talked about like prompt to Hollywood all the time. Yep. And we're dealing with these tools that can do really, really polished five seconds, but you gotta cut a lot of that together to make a feature film. Yes. Something of any decent runtime.

What's wild about this example is that because it looks at the frame, which it just generated and judges it in context of the frame that it's currently generating and judges it against the prompt, which is causing the generation over time. The, the model gets better. Yeah. And it doesn't need additional data.

It's creating its own training data as it creates frames. And so you have to look at the examples and you can nitpick the, well the mouths look a little swimmy here or the walk cycles look bizarre, whatever. And then you could look like a year and a half ago when we were nitpicking AI video and where we are today.

And you can really see that this is, this is line of sight on some amazing stuff in the future. Yeah. I mean, and [00:44:00] again, I can't, uh, underestimate the one minute thing, right? Like we would've said that it was crazy that something could create a full one minute comprehensive video. And yeah, there are small little things you can point out for sure, but it goes to show you three years, five years, 10 years down the road, we very much are going to get to a prompt to Hollywood sort of environment now.

How a person puts their own voice into that. What sort of manipulation in the edit you can do to make it more creative, to do stuff that makes it feel really interesting versus fully AI generated content, I think is gonna be a debate for the next, you know, for the foreseeable future. But it is. I not gonna here for it.

You're, where are you gonna go? I'm not gonna be f I'm saddling up on my robot steed and I'm riding outta town into something we call Robot Watch. Ba ba ba. It's Watch

Robot. Watch. It's Robot Watch. That's right. It's robot. Watch everybody and first and [00:45:00] foremost. We were talking about Kawasaki's new crazy robot horse concept. Kevin, when you're watching this video, what is your first reaction? What is your first reaction? I can't watch it. I laugh too hard. I know. This is crazy.

It's incredible. I thought it was an April Fools thing. I really did. No, it's a realize concept. It concept's a brilliant marketing thing. No itlo lop. No time to explain, says the Bite. Uh, Kawasaki heavy industries has shown off a bizarre concept. It's a rideable four legged robotic horse. They call it cor or cor, I don't know.

And it runs on 150 cc hydrogen engine. So it's hard to describe what this looks like, but it almost looks like a, like a navi from from Avatar or something. It's like it almost look, and obviously this is a conceptual video and a lot of people have been out there thinking this is real. It is not real. The video's con concept.

But what's interesting is watching it like. Climb over rocks. Imagine a large, um, you know, one of those dog-like robots that a human rides [00:46:00] and then walks around. Now, Kevin, when, when you're looking at the, it's, it's, that's, it's a Boston Dynamics. Yeah. It's like the Boston Dynamics robot dog with a set of motorcycle handlebars on it.

Yes. Yes. Yeah. And, and I worry about the impact on me. That's what I wanted. Because like, when you think about this, it's like, imagine, you know, you're on a motorcycle. So much of it is about the suspension of stuff. But like, I've ever rid, you've ridden horses, I assume before, right? Yeah, yeah, sure. Yeah. It's, it's like, it really doesn't feel good after a while because your body's not used to like, being banged around like that.

I think that's what this is gonna feel like. So I don't know if I'd ever wanna do this, but I appreciate Kawasaki coming out, showing us the, the version of the robotic horse future Look. Military applications for getting, I mean, look, if you look, I mean the, the climbing, you know, steepy, uh, climbing steep, uh, rocky terrain, uh, like a damn mountain goat, but a hydrogen powered engine.

I sign me up, man. Put these things in the new storm. Put these in the new Starship Troopers bet there'll be something to bounce up, rather, like a robot that I could ride on the back [00:47:00] of like, like a, almost like a bear that stood up on its high legs. Oh, like a man's sport. You wanna be. Hugging it from behind.

Yeah. Or like, what's that guy? See, I wanna run, I wanna ride on the front of the robot. I want to Baby Bjorn. Oh, baby Bjorn. I was gonna say, what's that guy from Teenage Mini Turtles, who's like the little brain that sits on top of the giant buck guy? Oh, crane. Yeah. Yeah. Crane the belly. So that could be, I could be in the belly of it too.

Imagine it being that big. And I would be, it's belt buckle and like, I would be putting my arms this way and I'd be controlling it as I walk. And now more than ever, seems like a good time to shout out. We do have a discord where our community shares all sorts of AI creations. We recommend you hop in there or just tweet us.

That works too. That's right. S send us anything. Okay. So another really interesting robot thing was that I show speed. The very famous, uh, Twitch streamer YouTube streamer spent, uh, a couple weeks in China, uh, recently. And like, if you've seen this stuff pop up on social media, it was really interesting.

The most fascinating thing though, is he visited the company Engine ai. Now, I don't know, I. If I show speed was paid to go over there. I don't know if [00:48:00] this is an ad for them, but what was so interesting is that if you remember Engine ai, Kevin, this is the company that we showed off a couple weeks ago that had the, the chopping hatchet robot dance hatch, the hatchet, the Hatchet dance.

So I show speed in this video, both, uh, kind of first hangs out with this robot, then dances with the robot. And what I thought was interesting is the Hatchet dance does show up again here without the hatchet itself, but it does its same sort of robot dance, which I love. And then at the end I just beat, like start trying to like beat it up, like kind of fake, beat it up and the robot falls over.

And what I'm really curious about in this instance if like, if I just beat actually just. Kind of stumbled into the idea of doing that, or if the robot has specifically designed to kind of spar with you and then like take a fall as if it's done? No, I, I think what we're seeing is that at the end of, its probably pre-programmed animation Dance cycle.

Yeah. Or fight cycle. It loses its balance because there's another video to their credit engine. AI is posting these videos. They're not shying away from them. There's another one where they're kind [00:49:00] of in a crowded hallway outside of a McDonald's and they gave the robot the hatchet again, and it is dancing, but towards the end of the dance it starts to kind of do a stutter and fade away and falls over and pieces are popping off the robot.

Like that's clearly not, wait, hold on. Did you see the. Second. Yes. Look at the second shot where they bring out, they like dance out. They bring out the stretcher. Stretcher. What the hell? Which is why I love this company because if you look at like AI demos in the past, do you remember like the, the Honda had Oh, the Osmo ish robot.

Yeah. That fell down the stairs and they bring out the privacy curtains and try to hide their shame here. They're making a meal outta the fact that the robot fell over and had a part go flying off of it. Honestly, I've love that. I wanna see a, I wanna see a movie with this robot because I will tell you the thing that charms me the most, it's the dance that it does in 26 seconds where it's just doing, it does this, it raises its arms back and forth.

Sure does. And it, it just looks fun. Like this is what robots could be. And what's interesting, and again, this is all probably a psyop on us from engine ai and they probably brought, [00:50:00] I show speed over there to do all this, get all this publicity. But you're right, they show the robot doing things that other companies would be afraid to show.

And I That's right. Am impressed. And again, we should be clear, engine AI is a Chinese company. Um, you know, China, as we've said in the show before, does feel at least to us very far advanced versus the US and the robots. And obviously China and America are going through something right now. Um, but no fault of our own.

But that is something to be aware of. Um, I, I do think that this is a really interesting space and a company to keep your eye on. China and America are kinda, they're going through some stuff. I don't know if you've seen the news, but hey, I just wanna be clear. Not our fault. I didn't do it. Not our fault. I didn't do

well. Those videos stopped us dead in our tracks, but that was robot watch. Mm-hmm. That's not where we're going. We're gonna continue discussing the things that made us pause and go, Hey, I see what you did there. [00:51:00] Sometimes Yes. Rolling without a care. Then suddenly you stop and shout ai, see what you did ai.

See what you did there. That's right. In a IC we did, every week we go through and see what people we like have done with ai. And specifically this week, ke we have a really fun one that involves Gemini 2.5 Pro. So if you go to, uh, Gemini plays Pokemon, they have set the Gemini 2.5 Pro model up to play Pokemon just like Claude did.

And Kevin, I have to tell you what's been really interesting about this is that it has done better than Claude. In fact, it escaped Mount Moon in three hours and 50 minutes, which is a pretty big deal. I know that sounds insane to think about that as a benchmark, but you have to realize what this is, is the AI is playing Pokemon on its own.

Yeah. Like there's nobody giving it instructions and they've set it up and. This is just a really [00:52:00] interesting way of seeing how ais are improving on, on, on other models. This is a way better benchmark than some obscure math thing. Yes. Or crypto puzzle. Like I, I, I love this, but what is also fascinating to me is how quickly these things spin up now.

Because there was a time where, uh, if you wanted to get an AI to play any sort of game or do something interactive, you had to get in there and hard code a whole bunch of stuff now. Yeah. You can just feed the screen output Yeah. To the ai It will read it, understand what it's looking at, and figure out, I mean, not that I'm not hand waving that there isn't some custom plumbing to let it interact with an emulator, but.

Still like this stuff is just, this is what I love and it's advancing so quick. And I love that there's currently 121 people watching an AI try to navigate Pokemon. I know, it's crazy. Um, another cool thing we saw that involves video games is this guy Nicholas Zu and I-C-O-L-A-S-Z-U, who has been working on behind the scenes what I think is kinda a really interesting vibe coding game.

And there's probably other people that are [00:53:00] doing this, this, I just happen to see these tweets. Um, sorry, it's Nicholas Zulo. His handle is Nicholas Zoo. And basically he's building an a kind of what he dubs is the ultimate sandbox and it's almost like a Minecraft built into three js, the whole kind of vibe coding thing where you can kind of type in what you want in this space and then it will just create it.

And I love these sorts of games and I, I, this was just one that kind of caught my eye 'cause it was like, oh, I, it's makes perfect sense to me that somebody's out there trying to vibe code like a, almost like a Gary's mod, right? It's almost like Gary's mod vibe code. Right. Well, I mean, look, you and I saw a company, uh, a, a fairly well funded company that I think was raising even more money years ago with a demo of an engine where you could say, I want a, a red couch, or gimme an end table, and it would spawn it into the world, right?

And it wasn't exactly real at the time. Mind blowing to me that here we are, you know, a couple years later and AI is someone is vibe coding. Yeah. Within engine. Similar [00:54:00] thing within engine on a set of set that it kind of is a similar thing and like with rockets that can explode into the sky and everything like that.

And it's just the AI is doing it now. I love it. And this would be a fun one to interact with. Like I want to play this game with my niece and go into a shared world and spa some objects and build something. 'cause the doing it block by block less compelling to me. Yeah. This is very cool. Yeah. And, and one of the things that was interesting about this, this particular guy I was talking about, he is not necessarily making it public yet because it will cost him too much to run it.

And that is one of the downsides of vibe coding, anything that involves AI in the thing itself. 'cause you have to pay for people calling to those models. But anyway, it's, it's worth looking at. Um, I think it's gonna come out soon. It's very cool. And it may go viral, much like a plate of cold spaghetti and in marinara water.

I love this. This was a simple post on Reddit. It has six point great 5,000 upvotes. Um, we are gonna talk more about four oh image gen and what I did with it later, but somebody just said, I asked for an image that will never go viral. And what was fascinating about this is like, it is a plate [00:55:00] of spaghetti in a, in a kind of a broth with a white plate and a CRT in the background.

The most boring picture you can imagine. But because, because there's a weird flash here. It's so boring. It did go viral. So now here we are. It went viral. Well, my favorite thing is the number one response to that and the Reddit thread is this viral image with a man making YouTube face shocked expression.

Did you see that? Yeah. That's his, it's the best. Made a thumb, a thumbnail response to it, which is amazing. Yeah, the YouTube clickbait thumbnail. Yeah. So, uh, what a brilliant, brilliant prompt. It's great. And I love that the decisions that the image model made of, like, like you said this, the old CRT in the background.

Yeah. A house plant. An unremarkable house plant in an unremarkable brown pot. The, the, the flash, uh, uh, reflecting in the TV screen and on the, the plate as well, like it's, it's. It's perfectly boring. That's right. No, Kevin, you know what isn't boring, which is kind of surprising based on what it is. Did you know how toothpaste is made?

Maybe play this video for, for us so we can kind of hear how toothpaste is [00:56:00] made. Oh, turtles are placed onto the conveyor belt for initial processing. Turtles are then submerged into a purple solution to remove all color. Um, with color removed suction tubes are fitted to the head to begin extraction.

Pretty interesting. I didn't know any of this. Pressure draws the toothpaste compound from dermal layers through the suction tube. This is, this is extract paste is deposited into, oh my god. Turtles remain connected. Okay. So, all right. That's okay if you're just getting the audio. Audio, yeah, that's the right move.

Actually, don't go look this one up. So let's, let's explain what this is. This made me laugh so hard. This is brilliant. Basically, um, a completely made up version of like how it's made, right? Like if you know how it's made, it's like the show or shows you how things are made. In this instance, it's how toothpaste is made and this guy.

Um, has done a bunch of these sorts of videos, but there are turtles going down a conveyor belt. This is all AI created everybody. So animal rights activists, you know, keep back a little bit, is all AI created, but a bunch of turtles going down a, a conveyor belt. They are then, they're then completely white where their coloring has been removed and then they're kind of [00:57:00] connected to this tube system, Kevin, where they are kind of like pull things are pulling out of them and you start to see like an aquafresh style.

Yes. Pull out of them. It's just a weird, crazy video that is only possible with AI tools and to me, you know, whether or not you find this disturbing as Kevin clearly does, or very funny as I do, this is the sort of thing that I would love to see more of. It's surreal, it's goofy. It made me really laugh and I was surprised.

I've been thinking a lot lately about the idea of. Joy and surprise in how that drives people to like want things and do things. And I think there are ways that AI tools can drive that. And this was a perfect example of that to me. So, okay, that was a super fun use case of ai. And Kevin, I have to say, I spent some time with more time with four oh image Jen this week.

We talked about this a ton last week, but I found an incredible prompt on soa. I saw an incredible prompt on soa, and I will say to everybody out there, you have not done enough with four oh image gen. And the problem is, is it does keep pulling back possibilities. But I found this really fun [00:58:00] prompt. And the prompt is grungy analog photo of X in search celebrity here circa 1999, playing their game on PlayStation one in front of a 90 CRT in a dimly lit bedroom sitting on the floor in front of the TV holding the PlayStation one controller.

So the idea here is that. You have somebody, a character in a game playing their own game. And so I went in and created a Tony Hawk version of this and also shared on x um, a picture of someone else had created, of Laura Croft playing Laura Croft of, uh, scorpion and, uh, subzero playing, uh, Mor Kombat. And then it was a PS two version of this, but two of the characters from GTA.

Yeah. So what was so cool about this, Kevin, is that like, not only did this tweet that I made took off like crazy. But what was so cool was seeing other people try things. Right? And so if you go to our Twitter thread, which I will share in our show notes, you get to see all these other people. If you just scroll down, somebody made a great, one of, of Pierce Brosnan playing Goldeneye on the N 64.[00:59:00]

Somebody made one of Ho Hogan playing WCW and WA Thunder, um, John Madden playing Madden 95. Like all of this creativity comes out of one small thing, right? Like, and again, I'm not gonna be, uh, to be clear, I didn't make the original one. I think the original one was probably the Laura Crot one playing Tomb Raider.

Mm-hmm. I just took it, put the Tony Hawk one together and said, Hey, look at this cool format somebody made. And then everybody piled on and like, if you go to this thread, it's like's still so many d different versions of it, scrolling this whole time and just kind of like everyone is just such a tiny little dopamine hit Yeah.

Of seeing, you know, Harrison Ford playing Indiana Jones and Fate of Atlantis and like seeing, you know, the, like seeing a battle toad play battle toads. Like, it's just so great to see where people go to which games they wanna see and which characters they wanna see. It's, this is such a fun prompt. Yeah.

And it's like, it it, to me, it really reminded me of like what's possible when you kind of have the ability to do stuff and not feel as controlled. Right. Because I think that was [01:00:00] what the Ghibli moment felt like. It was like, oh, it's not just like. Oh, you're using this May famous company's ip. It was more like what creativity could feel like if the entire world was involved.

Yeah. So I do hope we figure out some way of allowing things like this to exist because we know that OpenAI is kind of slowly Nerf stuff. I did get one last thing that I wanna shout out before we go. Another very fun prompt. Actually, two very quickly. I did an eighties, uh, Japanese movie prompt that I shared on our Twitter as well, where it took like an eighties Japanese movie photo, which was very cool.

There's some fun things I made there. I made a fun Goonies poster, but the most fun one was there's a very famous, uh, image I saw on Sora of the Harry Potter guys totally jacked. Um, and it was like the whole Harry Potter staff jacked, and it was like they were in this room. I took that prompt and I rewrote it for the super bad cast.

And so if you see this, I've created basically, you know, Jack mle, Jack Michael Ser, all hanging out in a, in a gym with the words like, you know, uh, I don't, what does it say? Something like, [01:01:00] uh, break yourself, fool. Yeah, break yourself, fool at the back. So like, that's the fun. It's like that remix fun that you can have when you find a really cool prompt.

Yeah, I mean, uh, just so, so good. I love that we are still discovering capabilities of this image model and people should go to the Sora webpage, right? Yeah. And scroll a percent, that's where Yes. See what people are creating. Yes. Yes. And that's where you really look at, I, I would say for suggestions is the remix tool is there, but the best tool is to rewrite the prompt.

Because most of the time when you remix something on soa, it's gonna take the original and then change it. What you really wanna do is get the original instance of what it gets created, because there is a thing where like over time as you change it, I found it gets worse. So you really wanna start with a fresh iteration.

So take the prompt and rewrite it. Um, do we wanna just unveil to the world what our super secret No, we don't wanna unveil that. No, we don't wanna do that now. We don't wanna unveil that to the world, Kevin. No, we what? Don't we what if we talk to the whales in the audience and let 'em know that we are slowly [01:02:00] unveiling it?

Maybe that's a thing we should do, like should reach out Whale Me and those kind of whales. Is that who you're talking about? Like the whales. The whales that are, that are actively eating a wolf Who's in pain? No, I'm Are you talking about whales, like underwater whales or what kind of whales are you talking about?

I'm talking about the titans of industry. Oh, the captains of the VC firms are, well, we are actually, we could say that we are gonna be actively looking for funding for this thing we're working on, and we're pretty excited about what it is, and we do believe we've done something interesting. But I don't think we wanna say a lot more than that, Kevin, because it's super secret.

Well, I'm not gonna share what I did with AI this week then Gavin fine. How about that? Fine. We'll go. We'll go away. Kevin and I will talk about this away from all y'all. We love you very much, but we can't talk about this in front of you from now and we will see you all next week. Goodbye. Bye.