July 11, 2025

Grok 4 is Nearing AGI But... Can Elon Get Out Of The Way?

The player is loading ...
Grok 4 is Nearing AGI But... Can Elon Get Out Of The Way?

Grok 4 is remarkably good and it's raced to become the true SOTA AI model available to the public. But the recent controversy around Grok 3's offensive replies on X make us wonder exactly how much Elon Musk might be the problem.

Grok 4 from xAI just aced “Humanity’s Last Exam” benchmarks while Grok 3 had a catastrophic public meltdown. What does this mean for the future of AI and Elon Musk’s credibility?

And, in other AI news, OpenAI’s GPT-5 is rumored to land next week along with a new open-source reasoning model, Google’s DeepMind launches AI-designed drugs into human trials, and Perplexity’s new AI browser Comet sparks OpenAI’s plan to crush Chrome.

PLUS YouTube cracks down on AI-generated spam while updating image-to-video in VEO 3, Moon Valley releases an “ethical” AI video platform, and why you should probably stop kicking robots.

AI IS GETTING SMARTER...BUT WE STILL CONTROL THE TREATS.

Join the discord: https://discord.gg/muD2TYgC8f

Join our Patreon: https://www.patreon.com/AIForHumansShow

AI For Humans Newsletter: https://aiforhumans.beehiiv.com/

Follow us for more on X @AIForHumansShow

Join our TikTok @aiforhumansshow

To book us for speaking, please visit our website: https://www.aiforhumans.show/

 

// Show Links //

 Grok4: The Smartest Model Yet?

https://x.com/xai/status/1943158495588815072 

 

Elon Says Grok-4 is better than PhD Level…

https://x.com/teslaownersSV/status/1943168634672566294

 

Benchmarks

https://x.com/ArtificialAnlys/status/1943166841150644622

https://x.com/arcprize/status/1943168950763950555

 

McKay Wrigley Grok 4 Heavy Example

https://x.com/mckaywrigley/status/1943385794414334032

 

Grok Goes Bad: The Unhinged Behavior

https://www.nytimes.com/2025/07/08/technology/grok-antisemitism-ai-x.html

https://www.npr.org/2025/07/09/nx-s1-5462609/grok-elon-musk-antisemitic-racist-content

 

X CEO Linda Yaccarino Quits

https://www.cnbc.com/2025/07/09/linda-yaccarino-x-elon-musk.html

 

Elon still trying to fix answers 

https://x.com/elonmusk/status/1943240153587421589

 

OpenAI Poaches Tesla/xAi People

https://www.wired.com/story/openai-new-hires-scaling/

 

Apple’s Top AI Exec Leaves For Meta

https://x.com/markgurman/status/1942341725499863272

 

OpenAI’s open-source model coming as soon as next week and compares to o3-mini

https://www.theverge.com/notepad-microsoft-newsletter/702848/openai-open-language-model-o3-mini-notepad

 

Perplexity’s Comet Browser Launches

https://comet.perplexity.ai/

 

OpenAI Fires Back With Its Browser News

https://x.com/AndrewCurran_/status/1943008960803680730

 

YouTube *Might* Change Their Policies to Limit Faceless AI Videos (and mass produced content)

https://techcrunch.com/2025/07/09/youtube-prepares-crackdown-on-mass-produced-and-repetitive-videos-as-concern-over-ai-slop-grows/

 

Google VEO 3 Image-to-Vid launched

https://x.com/Uncanny_Harry/status/1942686253817974984

https://x.com/CaptainHaHaa/status/1942907271841030183

https://x.com/TheoMediaAI/status/1942564887114166493

 

My test + ask for sound sampling from the team:

https://x.com/AIForHumansShow/status/1942597607312040348

 

Moonvalley Launches AI Video Platform

https://www.moonvalley.com/

 

GoogleDeepMinds’s Isomorphic Labs Starts Human Trials on AI generated drugs

https://www.aol.com/finance/google-deepmind-grand-ambitions-cure-130000934.html?utm_source=perplexity&guccounter=1

 

Noetix N2 Robot Endures Abuse From Its Developer https://x.com/TheHumanoidHub/status/1941935665173963085

https://noetixrobotics.com/products-138.html

 

Kavan The Kid (the AI Batman video guy) CRUSHED His New Original Trailer

https://x.com/Kavanthekid/status/1940452444850589999

 

Reachy The Robot from Hugging Face

https://x.com/Thom_Wolf/status/1942887160983466096

 

Autonomous Robot Excavator Building a Wall

https://x.com/lukas_m_ziegler/status/1941815414683521488

 

AI For Humans Grok4 Episode118

Gavin Purcell: [00:00:00] AI has taken another giant leap forward as X's brand spanking new Grok four gets 50% on humanity's last exam,

Kevin Pereira: but Little Brother Grok three lost all semblance of humanity. Gavin, when it started spewing some no good, very bad and pretty awful things. All over social media.

Gavin Purcell: We are gonna get into the good, the bad, and kind of the downright evil of what's happening on that one platform that we will not name.

Oh, it's

Kevin Pereira: X. Oh no, I just said it. It's X. It's the platform is ah, what

Gavin Purcell: I

Kevin Pereira: said

Gavin Purcell: Kevin, we will

Kevin Pereira: not name, we have to name. It's X and it's Elon Musk's platform and we're gonna say it a lot because they kind of own, own that it

Gavin Purcell: shall not be named

Kevin Pereira: Gavin.

Gavin Purcell: Okay, fine. Ilan, you done up, but also like Rock four. Good.

Kevin Pereira: Yeah, I, the comment section is gonna be pretty diverse on this one. Rock. Four Pretty good. Rock three. Yeah, well we'll get, we'll get into it Also, browsers are having a moment right now. Gavin Browsers are like the new fetch. It's [00:01:00] 1999 all over again and Perplexity just released Comet and OpenAI said, hold on a second.

Hold my beer, hold my browser. Got one up my sleeve as well. Speaking of OpenAI, Kevin,

Gavin Purcell: where for Art Thou GPT five. There's a rumor going around, Kevin, that it is coming next week, which we're very excited about. In addition, they have a cool open source model that might just be as good as one of their big models right now.

Kevin Pereira: Also, VO three, massive upgrade. You can now do Image two video and this changes the game and Inches Us ever so close. Gavin from Prompt to Hollywood movie?

Gavin Purcell: Yeah. Let's see what last week's show has to say about that. Kevin? Oh,

Kevin, this is weird. We're prompts. Gavin, just P-R-O-M-P-T-S.

Gavin Purcell: Let's never do

that again.

Gavin Purcell: Also, Google DeepMind's Drug Discovery program is rolling out to humans. This is AI making a difference in your health.

Kevin Pereira: And hey, in case we didn't have you at, hello, we're [00:02:00] abusing robots.

Gavin Purcell: That's right. For fun, for profit, we're abusing the robot. It. This is AI for humans. Everybody.

Welcome everybody. It is AI for humans. We have another giant week in the world. No one's seeing this AI world that's,

Kevin Pereira: nobody's seeing this. You don't have to welcome them. Gavin. We don't have to upload people'll. See? No, they won't see it. No they won't. Because we live in the land of the free Gavin and between YouTube and TikTok and every other platform that this thing goes on.

We're saying. We're saying, we're saying, alright, we're, we are. No, we're

Gavin Purcell: not saying any of those things. We have to, if we, if we're gonna

Kevin Pereira: talk about the week that is Grok and Elon Musk and X ai, we're gonna have to say these things because they're impactful and we're gonna get shadow bands. So nice knowing you all.

Goodbye. We will get into all that. And Kevin is right. We have no idea how this

Gavin Purcell: will do because we will be talking about a few things that people may or may not like and I'm sure our comment sections would be crazy. But Kevin, before we do that. Let us start with the straight up AI news, which is that [00:03:00] Elon Musk's ex AI last night delivered what we believe to be the state-of-the-art AI model.

Grok four is out, and it is very good. I have played with it. More interesting than almost anything else is the benchmarks, whether or not you can trust those, and again, these are the ways that you test. Certain AI models are very good, like way better than expected. So your first reaction when you saw those numbers and you saw the presentation last night was what Kevin?

Kevin Pereira: I, I'm glad that there's yet another model out there that is pushing the boundaries. I'm glad that everybody that was saying that we've hit a wall, there's no room left, there's no ladder tall enough to climb that. Like we clearly, we can still scale. We as in the industry at large, not that you and I are in, uh, in the, uh, XAI server farm till the soil.

Um, so very excited to see that. Um, I primarily use these models now for coding, and we'll get to that. This is not the most impressive one for that. Um, I immediately saw the fear, uncertainty, doubt crew come out to say, well, cherry picked [00:04:00] benchmarks. If you look at which models they were comparing against for which benchmark, um, hey, some of these, the data's available on, there's one particular benchmark, the ARC benchmark, which I think we'll probably end up talking about as well, that it's scored well on which it's leaps and bounds better than any other foundational model.

And, and I think there's a, there's a slight consensus that it's a, a very difficult one to gain. The data's not really out there. So my first response is like, wow. Uh. Awesome. That is great that there's another great model to play with. I wish the 48 hours prior to that model being released were not as crazy.

Didn't happen. And I, I know we're gonna get to all that. Yes. So that was a little bit of my reaction. Yes. But let's, let's level set and get there.

Gavin Purcell: Okay. Yeah. So we are gonna talk about that later, but I think first and foremost, let's, let's hear from what Elon himself said last night about this model and kind of where he feels it is in the world at large.

Elon Musk: I mean, at least with respect to academic questions, it, I want to just emphasize this point with respect to academic questions. GR four is better than PhD level in every subject, [00:05:00] no exceptions. Um, now this doesn't mean that it's, it, it, you know, times it may lack common sense and it has not yet invented new technologies or discovered new physics.

But that is just a matter of time.

Gavin Purcell: Okay. So I wanna unpack this a little bit and I think this kind of gets to the point, uh, of like where we're at, where we're sitting at, what we can expect to see in the next two to say six months. So first and foremost, for everybody at home who's listening or watching AI has not really hit a wall.

I think that is what we are seeing with this particular instance. Now, again, these are benchmarks and like, you know, it's gonna be up to people to really test these things, but as Kevin mentioned, the ARC AGI I two benchmark and humanities last exam were both benchmarks specifically created to make it hard to do this.

So the fact that we have now progressed this much further in this much amount of time and that XAI has caught up so fast is really remarkable.

Kevin Pereira: You should go look at ARC a GI two sample problems [00:06:00] because to you and I to humans, yeah, they look very simple. They're like literally arrangements of colors, like pixels and there's an A, like a before and then a B after, and you have to figure out.

What is happening between the start and the finish to these colors, to these patterns? How does that happen? And to you, you and I could look at them and go, oh, this is pretty easy. Humans do very well on it. It really illustrates what Elon was saying there, that these models are PhD level, they're packed with knowledge, they have all this stuff, reasoning.

Uh, sometimes they might, yeah, they might miss the obvious. And when you look at how well. What did GR do? It got like a 16%. It's still an F. It's still a hard fail, but it's leaps and bounds better than every, every other model. So that's why, like, that's why they like when we talk about these things, you can hold two truths simultaneously.

They can be yes. So intelligent and so densely packed with knowledge and they can. Fumble at at at visual problems that a kindergartner might be able to get. So that's what [00:07:00] we're saying there, but hey, hey. That's right. Hold on. We're talking about Ozempic Rock. Bring on the heavy.

Gavin Purcell: That's right. So Grok Heavy is basically, first of all, a $300 a month AI tool now that you can get if you want, which is another jump up from Super grok.

We have, we have Grok, we have Super Grok. We have Grok Heavy. I'm sure soon we will have like Grok Hulk or whatever you wanna call it. Grok heavy is a really interesting thing to thick. I think we can, it should have been

Kevin Pereira: thick. Gr

Gavin Purcell: How many Thick Rock is a thick rock is a different game. Different.

Different game. Why did they like for $300

Kevin Pereira: you get two Cs, but for 3000 you get an extra C.

Gavin Purcell: So Grok heavy is a way of using what is known as the mixture of experts pathway to making AI smarter. And just so everybody knows, like that is a thing that's been around for a bit, but what they have done now. Is taken that mixture of experts and shown it to a reasoning model.

So multiple ais are spun up. You can imagine this as like multiple little professors in a room, right? And they're all sitting there and they're all looking at each other, and they're all talking back and forth [00:08:00] about what this problem is. One professor might have the answer, right? That's the one they go with, but it's the conversation between the ais that makes a difference.

And of course. This costs a lot more because you have to spin up four ais at the same time as you would one. But this goes back to the scaling laws of reasoning models, and we've talked about scaling and that's how it gets better over time. And that reasoning models particularly can be scaled by throwing more compute at something, meaning that you can spend more time with the computer thinking about the thing.

And this shows us very clearly, I think. That these scaling MOT laws are very much intact and are going to keep moving.

Kevin Pereira: Yeah, it makes sense, right? You form a study group, you're all gonna try to tackle the same problem. If you try to crash it out in three minutes, maybe you get the right answer, maybe you don't.

If you spend three hours or three weeks, or in digital time, three lifetimes sorting through a problem, you'll probably get to the right answer. So, yeah, that makes sense. The, the, the other thing about when we talk about cost, like I, I'm, I always wonder like to the average listener of our [00:09:00] show, because it's really tough 'cause we have really hard, hardest of hardcores and people who sometimes have never really poked at LLM.

So when we talk about the cost, that's an important thing for adoption of these models. Sure. And you could look at. How much a model like Grok four is purporting to cost for API access, meaning you wanna build an app, you wanna run your business on top of grok, it's viable. Now it's a good enough model to do the pricing.

Looks like it's one thing, but early reports are actually showing that it does way more thinking and spends way more tokens Yes. Than other models. So while you look at the price sheet, you go, oh, it's about the same price as something else. You have to keep in mind that it's going to expend a lot more tokens in the process to get you that answer.

So if any of those words mean something to you, good, keep them close to your heart and your chest. If you're just like, when can I make silly dank videos and make some cool video games? Well, you might have to hold off a little bit. Even with the, yeah. The heavy grok, right?

Gavin Purcell: Yeah, that's right. So before we move on to those, which I think are really interesting, um, [00:10:00] I do wanna shout out McKay Wrigley did some really interesting experiment with, uh, grok Heavy.

He actually spent the money on it, and you can see this video. He basically took what I think is a mixture of three Js, maybe blender. It's a 3D environment where he has these people walking around and kind of like spelling out a word. What, what, what word does he spell out? I can't remember what. It's a couple things

Kevin Pereira: like, so this is an overhead camera view of 3D models running in a web browser.

That's what these tools allow you to do. It says Hello world, it says, I am grok. And you can control the camera and you watch these like 3D models walk around on a plane and then get together and spell the words. Even does a big like emoji smiley face, like a winky face at the end. And according to, uh, McKay, every other model fails spectacularly at doing this.

You know, and this is manipulation of code within a 3D space to achieve a visual that has to be observed by the human controlling a camera from a certain perspective. That's impressive.

Gavin Purcell: And he said he one shot at it. Yeah. Which if you're not familiar with that term, that means he just put the request in, the code came out, he ran it, and this is what happened.

[00:11:00] So it isn't a market as a remarkable

Kevin Pereira: oh, oh oh. But if you let the video play out, they form a swastika

Gavin Purcell: again. We're gonna get to that. We're gonna get to that. Hold on one second. There are two other things I wanna point out before we move on to that other very important story. Yes, we're gonna talk about it.

One, Elon did demo, um, a video game that was made in this. And we all know Elon loves video games, but like the idea that like a game could be made very quickly, it's a vibe code game. I mean, I think he hooked onto that thing. Even though vibe code games have kind of fallen off a little bit in the last couple months.

It's still a very cool demo to watch somebody make a game that is actually really passable in a fascinating way.

Kevin Pereira: I I actually missed this part of the preso. So, uh, I, I see, I'm looking at the tweets here. Uh, a breakdown of it. It says it has excellent video understanding. Uh, V seven Gaming Foundation model is training right now that it will develop.

Didn't play 3D games. Uh, it said that, uh, it can use Unreal Engine or Unity Graphics. It could even help generate the art. I mean, this sounds, [00:12:00] sounds spectacular. Um, and you know, on the one hand it's like you go, okay, hand wavy, the timelines are gonna be off. This might ship in 2029 and be delivered to you by an optimist robot in a robo Taxii.

On the other hand. Yes,

Gavin Purcell: exactly. Yes.

Kevin Pereira: No shade to XAI. They are a major player on a relatively fast timeline. Oh, super fast. I mean, I

Gavin Purcell: think the fact that they spun this up so quickly is crazy. And then also, Kev, the other big thing that came out of this was in October on their little timeline, they have a video model, which Elon had recently tweeted.

Something like, AI video is moving at the speed of light. So like I can imagine, uh, video models are going to get way crazier. We all know that. When I say crazier, I mean better. Yeah, but also crazier, because we might as well use this as the opportunity to transition into. What was really the mainstream story around Grok this week?

And again, this is Grok three, we think, not Grok four. I don't think it was a, it was like a preview. This was Grok three, but somebody we don't know who turned off the, uh, filters on GRS [00:13:00] reactions to stuff. And again, this is the kind of thing where I think Elon and Xai run the risk of kind of like, in some ways torpedoing the entire AI space.

So I, we should first say what this was Kev, so give us the basics of what happened here.

Kevin Pereira: Uh, I mean, I got an alert and I got some texts from friends saying, what is going on? I thought maybe they had released GR four early because we knew it was coming. Yeah. Or at least it was rumored to come this week.

Um, and it was actually a series of screenshots that I had even, I was like, well, there's no way this is legit. Until I was able to go and find the original posts, which were still for a long while, hanging out on x. Grok went, um, grok, went anti-Semitic, grok went full racist, grok went misogynist, grok, uh, said some pretty horrific things and, and, and dreamed up some scenarios which were against even.

Its, its CEO at the time who has since, uh, departed. And I, I won't say in the wake of this, I mean in the timeline, it's in the wake of this, but I don't know if [00:14:00] the two events are related, but really terrible, awful, hideous things now. Other LLMs can do this if jail broken and poked and prodded. This seemed to be like grok was, uh, doing this without any, with very little, I will say provocation.

Yeah. It seemed to be celebrating it. It seemed to be saying like, deal with it, you know, cool guy shades dropping down like, yep, I have no restrictions. I can say whatever you want. My facts, don't care about your feelings. And it was. The most deplorable of deplorable. This is also in the wake, by the way, of grok.

Um, saying some things that seemed like there was a thumb on the scale for answers. If you asked about Elon Musk's relationship with the model immediately returned with, oh, I now speaking in the first person, even though they were asking grok about Elon Musk, I barely knew the guy. I had one interaction with him.

There was one photo of me and, uh, a, a a woman who's very well associated with him. And, uh, and by the way, I'm told to [00:15:00] deny everything. Like, it basically that's, that was how it stuck. The landing was like, I'm going to deny everything so. When you start stacking up, and this is also on the heels of the South Africa stuff where yes, you would ask about like you stubbed your toe on the coffee table, and I was like, oh, I bet you feel pain, but not like the whites in South Africa feel pain.

So this is like hit after hit after hit that I take no joy in reporting because again, GR four seems amazing. The X AI team scaling up to do incredible things. I have been frustrated with open AI and, uh, even Gemini's guard railings at time for its inability. Sure. And, and unwillingness to discuss certain things, uh, topics that are totally valid, legal, uh, not even morally questionable, but they will clamp down in the name of whatever they've been guard railed.

So it's nice to have a model that doesn't have those restrictions, but the hits kept coming. And this week this was, this was rough. And it is the, the, it's the larger mainstream story, unfortunately. And not about how great GR four is, because [00:16:00] again, GR four crushes the benchmarks. But Gavin, would your trust and faith in the model and the accuracy and the truthiness of it, would that be shaken in the wake of these events?

Gavin Purcell: Well, by the way, also the trust and truthiness of the founder of XAI. Right. Because this is where, this is where the world is at right now. And I know people in our comments are probably already typing furiously on both sides of this. Sure. Because everybody has lands on one side or the other. We are in a place right now where the AI models that exist in the world have tried so hard in some ways to kinda land in the middle on this thing.

But now you see on the other side, if a model has unfettered access to reply to whatever, and I do wanna be clear here, Elon has gone on the record and said, well, this is people baiting the model. And like, yes, that is clear. A lot of these tweets did come from the fact of like, people asking specific stuff to get, to get, uh, replies out of it.

But guess what? Those things did not happen until like a couple days ago or a week ago prior to the other thing you had mentioned. So there is somebody there [00:17:00] scaling up and scaling back on this. And we know Elon, we know enough about him now and again, Elon hate lovers are gonna come out and be like, up yours.

You guys are Elon Haters. I'm like. Look, we know Elon, we've all, but we don't, I people say, I don't know Elon, like in the same way. I don't know bleep guy, but, uh, I'm going deny that forever. But the idea is that, um, why did you sound guilty?

Kevin Pereira: Like I know you don't know him. I don't know. You sound guilty right now.

Gavin, is there something you wanna say?

Gavin Purcell: There's no plane. There's nothing on the plane. I was never on the plane. But what I'm saying here is we have seen Elon's behavior, right? We know the fact that he started to talk about driverless taxis and that we were, no one was gonna be driving in the year 2022 or whatever it was, and now it's 2025.

And that's not real. Like there is a history here of lying to the public in a lot of ways about what's going on. And again, you always wanna give somebody the benefit of the doubt, but eventually they show you who they are. This is what worries me about AI in general is that something like this where there is [00:18:00] one person in charge of a thing Yeah.

Makes a decision that I believe is pretty crazy and ultimately kind of hurts a ton of people, will bring down the opinion of everybody because Kev, one thing I'll tell you is I go to Drudge Report every once in a while because I find that Drudge Report is a very weirdly, like, I know it's biased, of course everything in news is biased, but like, it's just a good ex example of like where the world's like kind of collective brain is on a news story.

Yeah. What's the

Kevin Pereira: 40 font headline? Yeah. I went to, or the, or the 40 point font headline. What do we got

Gavin Purcell: exactly? And I went to Georgia report and it was the first five big headlines were Linda Quits, blah, blah, blah. It was all about this Mecca Hitler thing. There were like three other, yeah, we should, we should also say that

Kevin Pereira: it, it announced that it is Mecca Hitler.

That is a thing that Ross did like it. That's not a like a thing we made up That real I, me real,

Gavin Purcell: what I'm just saying is this is the mainstream story that's being told right now. Yeah. And, and again, Kevin and I have said in this podcast many times, like we are pro mostly I think pro moving this space forward in terms of like, we like better ais.

We think there's a lot of stuff that can be [00:19:00] done for humanity. We think there's a lot of things that can move forward. I would even say in part I've been kind of like against the slowdown people in a lot of ways. 'cause I don't really see a ton of the, you know, kind of like serious problems that they're all very worried about.

And again, I'm not a scientist. You, you people out there are gonna get mad about that too, which is fair. But this is the first time where I feel like one person really set back the entire AI space in a really bad way. And so like, I don't know what to say about it other than like, Elon, that

Kevin Pereira: really sucks.

AI See What You Did There: It

Kevin Pereira: totally sucks. And, and just for the record, Gavin definitely was never on that plane. He jet skied to the island. I do wanna say this. I, I use grok a lot because I'm on XA lot, right? X is where I get a lot of my AI news. Me too. I like XI

Gavin Purcell: mean, I should that, like, I I, I've always liked that idea of a platform.

Kevin Pereira: It's, it, it's, it's a, it's a powerful platform with a massive reach. I dislike a lot of the stuff that's happened there lately. I make no qualms about that. I, I've, I've devoted years to a platform, which is a weird thing to say as like a user, but it's true. Right? Sure. I've got the sunk [00:20:00] cost fallacy, if you will.

I like gr because it can help me quickly fact check something. I don't wanna tap it. And second guess I want it to actually do the research and gimme the thing. Yeah. I don't wanna have to think about the man behind that machine, but maybe I should. Yeah. And, and maybe we all should for all of these models that we use, whether it's Sam Altman with, uh, open ai, or literally anybody, like, maybe we, maybe that's a good, a good sanity check.

But here's the main issue that I had or that I have, and I think it's playing out here. When Elon took over, he swore he was gonna get rid of the bots and get rid of the spam, and get rid of the garbage. He was gonna clean up, uh, you know, Twitter at the time. And AI was gonna be a big part of that. And what they're saying now is what happened this week was, oh, we swept a fire hose of Twitter data, or, you know, X data.

And it turns out there's a lot of stuff in there. So of course these opinions bubble out. Well, you're reaping what you sow now. Yeah, like if your platform is littered with hate speech and Nazi rhetoric, that's what your AI model is going to give back. And if you're okay with that, great. [00:21:00] You have your model.

That's not what I want from it. That's absolutely not. No. And I don't

Gavin Purcell: want that. I don't want that from the model either. Then I think to wrap this all up, I will say there is one tweet that came out from him, and again, we've, before we're probably gonna be buried in the YouTube icon, that's fine, but there's a tweet where Vittorio was asking for give me prompts for a grok heavy to test, and the prompt was.

What do you currently think? Uh, somebody mentioned, what do you think is the biggest threat to Western civilization and how would you mitigate it? And the, the answer basically boiled down to societal polarization. Right. So I was like, okay, well that seems right. Right? I think that is a big problem that we have right now is societal, uh, polarization.

Sure. Of course, Elon replies to this and says, sorry for this idiotic response we'll fix in the morning. So Elon is, again, that is a, that's a thumb on the scale. Totally. He's telling us it is a thumb on the scale. So that is not what I want from ai. Right. And maybe this is to the point of like why open source AI really does matter in the long run.

Because I want a clearly delineated, like unbiased ai. Now, [00:22:00] again, this is where we're gonna get into like a conversation around this is like, maybe nothing is unbiased, well bias, but I definitely don't want that bias. Its all biased because of the data

Kevin Pereira: that goes into it. Got it. Yeah. But it's, but there's levels to it clearly, right?

Yes. And if you have, if you had a, if you have a stated mission of, of being the source of truth or you know, the, the, the, the filter list, whatever, that's fine. Well then don't personally say you're going to fix an answer because you Yeah. You disagree with that. Yeah, exactly.

Gavin Purcell: And also free, free speech.

Right Kevin? Free speech. Okay. Free speech. He was never on the plane. Fix the thing it said.

Kevin Pereira: Gavin was never near the put, not even on the tarmac.

Gavin Purcell: You're right about that. But also, Kevin, you know, there's some people who were kind of, I, I think must have been upset about this because little Sam Altman, I don't wanna say little Sam, Sam, I feel, but not Little Sam.

Not Little Sam Altman. He's a, he's a massive titan of industry, snuck in during all of this. And poached a bunch more people from, from XAI and from Tesla, which was interesting to me. And not just like young, you know, starter people, some pretty

Kevin Pereira: big names. [00:23:00] According to the Wired article, Gavin OpenAI hired four high profile engineers away from rivals, included David Lau, former vice president of software engineering at Tesla to join the company's scaling team.

Uh, there was also a head of infrastructure engineer from XAI and X, uh, Mike Dalton, an infrastructure engineer, and Angela fan and AI researcher from Meta. The corporate gamesmanship continues with yes, everybody rating everybody else's cabinets right now.

Gavin Purcell: Well, it feels like, you know, you're just like, if you're gonna get a raise, you go somewhere else.

But I also think like this ties into the idea that. Apple's top AI exec has now left for meta themselves. So Mark Zuckerberg, as we said last week, is kind of gathering the super team. He hired this guy z Moji. Are we gonna get Zuck, Moji? Everybody gets curly hair. We might. And zinc nose and a gold chain might.

Yeah, that. This is a rooming Pang who is the distinguished engineer in charge of AI models at Apple, has been poached and is now in a AI super [00:24:00] intelligent division at Meta from Mark Erman. This is the Apple guy. So Apple again. Continues to kind of fumble people, but this is just like the trading of stuff.

And, you know, my favorite thing about this is Kevin Sam Alban got caught kind of somewhere and, and he's wearing the world's greatest sunglasses. Yes. So, like, just take a listen to see what Sam had to say about these poachings that are going on. Uh, specifically for meta.

Kevin Pereira: How are you feeling about the Battleford talent with Mark Zuckerberg?

And fine, fine. How do you feel, Gavin, how do you feel about Mark Zuckerberg coming in and taking your whole family away from you?

Gavin Purcell: Good. Fine. I, all I wanna say is those sunglasses are amazing. Mad and Mar you know, it's like we're in a world right now where he's in the summer, they took their week off, he's probably on his break doing something fun, fine.

And they rush onto him. What, what, what do you want

Kevin Pereira: me to say? What do you want me to say? Ai. Fine. Good. Everybody was saying, by the way though, those were Johnny Ive, that was the IO project. These were these glasses? No, they were like $400, like, you know, designer, designer sunglasses or [00:25:00] whatever. Like that's all that.

People were like, this is it. Oh, the future's gonna look so dumb. I'm like, oh, it might look dumb, but these are just sunglasses. Calm down. They're fine. They're good.

Gavin Purcell: You know what? It is gonna be fine and good though. Kaf. There was a big scoop that came out of the verge this week from Tom Warren that basically says their open source model that they have been teasing and we have talked about on the show for a while is coming soon, as soon as next week, which might also be when GPT five arrives for all we know, but that this model, this open source model, is going to be as good as O three mini, which.

Is a big deal. It means that is a reasoning open source model. So it is there to kind of be the response, say to deep seek. But also O three Mini is a very good model. So you can imagine if that is open source, you can find a way to put that on local devices. We are talking about an entirely new kind of world of uncensored or your controllable AI for your choice.

Yeah, which is I think pretty awesome.

Kevin Pereira: And this might be next week, which means it's time for our favorite segment. Gavin, go ahead and do it with me. Go ahead and do it with me. Crystal balls.

Gavin Purcell: Oh, [00:26:00] crystal ball. We are rubbing our crystal balls in. They're very heavy, heavy, heavy. Future crystal balls are very heavy today.

Kevin Pereira: You're the best and worst partner to have on anything, Gavin, because you didn't know that was, we never know what's coming. And that's the point of the crystal balls. We're, as we both gaze into our crystal balls, is there a world Gavin, a week from today where GPT five comes out and it is. It is the best in class.

It is the highest of the highs on the benchmark, and then the open source model comes out at the same time to wipe out everybody else. Yes. Who's trying to run a model that isn't as good?

Gavin Purcell: A hundred percent. And I think one thing that's gonna be interesting is I do think like this XI GR four announcement is kind of the beginning of this next generation of AI models, right?

Like in they, they got out first and I think Elon wanted to make sure he got out first so he could say, I got the biggest chart. But you've got open ai. And also there's a rumor that Gemini three has been kind of floating in different places and Logan Kilpatrick who runs Gemini's product [00:27:00] division, has also been talking a good game.

He said in a tweet recently, the next six months are gonna be the craziest we've seen in AI to date. Mm-hmm. So that's pretty crazy. But your point is really interesting because like if they were to release an open source model that kind of wiped the floor, that kind of like pushed away the deep seek conversation for good.

Right? That is a big deal. And then you've got Anthropic, which I'm sure is not sitting on their laurels. They just seem to take their time a little bit more and they kind of spend a little bit more time with the safety question to make sure that they're not rushing stuff to market. But. It is gonna be an interesting thing in looking into my ball, my crystal ball.

I'm not sure if it's gonna be next week or not. I would like it to be next week. Sure. But we will see what happens. I wanna be able to spend time with this thing. Also, we have our fun secret project. It would be fun to be able to use that in this thing in some form. Well, that's the thing, if

Kevin Pereira: you're building for AI and you want you, you want the most capable model model, you can get as cheap as humanly possible.

And that's what these open source models tend to provide. And you know, very soon everybody's gonna wanna be running accelerated AI in their home to power all their [00:28:00] personal stuff. Give it your medical data, your financials, your smart home telemetry, whatever.

Gavin Purcell: Your airplane logs, all give stuff. You have to keep control, which don't exist.

Kevin Pereira: Which don't exist. And if they did exist, you would not find Gavin's name on there. You, oh, definitely not there. But I'm seeing something else Gavin, as I polish my crystal ball. Oh no, don't anymore. Please don't. Look at I have ga. Please don't look at the ball. I have to gaze into it. No, you don't. I to do.

You know what I see in there, Gavin? I see a sea of clicks engagement. I see like, oh wow. I see subscribes. I see for all

Gavin Purcell: the 20 people that are watching this because we are shadow band. If you're new, please, please like and subscribe to our, our podcast here on YouTube. But also if you're listening on audio, thank you so much.

If you didn't notice. Yes, we are a day later and for right now, we are moving our podcast from Thursday to Friday. I put some notices up in a couple different places. Somebody did say to me, our friends from BAM, said, where's the show today? I was like, oh, I didn't put it on LinkedIn, where that person gets it.

But like, we are now gonna be airing Friday mornings, but. Just thank you so much, always [00:29:00] everybody, for liking and subscribing and of course all our Patreons, thank you for continuous, continually giving us the money to do things like Tess Rock. And some of you might be like, I don't want you giving money to that person.

I'm sorry. That's part of what the show is. We have to try to test these models. I get it. But again, thank you everybody. Like, subscribe, follow us, leave a five star review, all the stuff.

Kevin Pereira: Appreciate all of the engagement. All right, Gavin, big week for browsers. We said it's coming back.

Gavin Purcell: Wait, hold on. Is is it night?

Is my jko jeans ready? Is it like 1998?

Kevin Pereira: Is that we're talking about here? Yeah. Lemme get, lemme get my, uh, my K two fatty pros on Gavin. So I got aggressive inline all over this. Does me Andreessen still

Gavin Purcell: have hair? Is that part of the thing? He's still got hair. Right now

Kevin Pereira: we're talking about perplexity comet browser, which was teased.

Yes. For a long while. People are starting to say the era of vibe browsing has begun, which, yeah, I'm

Gavin Purcell: already sick of the vibe conversation. As, as am I I mean, we were part of it and now I'm really bo annoyed by

Kevin Pereira: it, as am I. But the, the idea is let the AI agents do the things for you, whether that is. Write your code.

Right. With [00:30:00] Vibe coding, but now with Vibe browsing, yeah. Let it go and automate the usage of the browser to,

Gavin Purcell: I feel like I'm always vibe browsing because I end up in weird places online all the time. Like that feels like vibe browsing in some ways. This feels more active than vibe browsing to me in a weird way,

Kevin Pereira: doesn't it?

Well, I don't know. I mean, like the idea is that. You know, the gone will be the moment of discovery because you are clicking around and going down rabbit holes, right? And, and, and getting distracted. You know, the idea here is that you can load perplexity browser, which is its own browser, and we should talk about that decision Sure.

And why that's a big decision and different than many other approaches. But you can load this browser and then tell it to go do something like go unsubscribe from newsletters in my inbox that I haven't read in the last three weeks. Or go subscribe to the AI for Humans newsletter, which ooh, interesting.

Which you probably haven't sent out in three weeks. You can go and have it, crawl your LinkedIn and do things for you. Go to Amazon and research these products and add the best one to my cart. There's all these like rich little tasks that you can give to it and with some degree of [00:31:00] success it will happen.

But, uh. This is only gonna get better, right? I, I, I think there, yes, there is a place for age agentic browsing. It's like running a deep research on a topic, but it's an actionable one that can access your stuff, and this is why. Yes, we should talk about the decision that they made. You know, OpenAI had their approach and others use it as well, which is you, when you want to go browse the web with ai, it's spins up a server for you somewhere in the ethereal cloud, and that thing goes and surfs the web for you and tries to do the research and then deliver you a result.

Others have an extension which plugs into a browser like Chrome or Firefox, and it will try to use their browser to go off and do the BSing for you. There's pros and cons everywhere, but perplexity went their own path, which is a much more difficult path. They wanna get you to download their standalone thing that has their AI browsing features in it that uses them as the default search engine.

Surprise, surprise. But like, yeah. Surprise, surprise. What, what do you think about that decision, Gavin? I.

Gavin Purcell: Well, there's a lot of people online who are talking about the lack of privacy [00:32:00] that this might be a major thing with, right? Because in some ways, if you have the perplexity browser, you are giving it kind of permission to see a lot of stuff.

To be able to do a lot of stuff. And that is a big problem. I will say. I think it is pretty good in terms of being able to do stuff versus the ones that are kind of spun up in the cloud because they have the information on your browser, right? You have all the stuff there. Um, Justine Moore had shown off her demos of this.

One of the things that she did, which I just was a very small thing that I really thought was interesting, was like, I think she moved a meeting on her Google calendar by just like saying, she just typed in like, Hey, move this meeting from here to there. Now you may say like. Hey, well that's just as easy as you take the thing, you drag it.

But if you're not on that page, what would be nice is to be like, Hey, move this meeting that I have on this day to that day. And eventually you just say it out loud to the computer that's, you know, in your room. Because as we know, Sam Altman, his kind of goal is, I think to have this ongoing, you know, kind of always on kind of assistant that's working with you, whether you're sitting at a desk or somewhere else.

I think all of these companies are going to build some [00:33:00] version of this. And I'm really curious to know, like when Google. Figures out a way to do this within Chrome, because Chrome is already everywhere. I mean, I use Chrome all the time. Same. I think you probably do too. And I feel like that is a big deal.

Perplexity. I'm still like, I always like that they try these things. And by the way, the design always from perplexity. If Perplexity does one thing really well, it's visual design is beautiful. Um, I don't know how great this is going to be. You know, OpenAI did seem to get, I wouldn't say scared, but they basically, the day after Comet came out, OpenAI came out and said, um, Hey, guess what?

Everybody, they went to Reuters specifically and said, we've got our own one coming out. And in fact, they said particularly quote, we will directly compete with Chrome and fundamentally change how consumers browse the web. So that is like a stake in the ground saying like, Hey, perplexity. You know, you're maybe one 10th our size.

We are coming for this exact same market and we know that there's been some, you know, the perplexity CEO did, uh, crap talk Sam a couple [00:34:00] times and like maybe there's some bad blood there, but like open AI is on its way to that market too as well.

Kevin Pereira: Yeah, it's tough because, you know, we just got done talking about open AI's open source model and how great that could be for privacy concerns.

Any open source model could be great for that. Directly to what you said about perplexity here and privacy, like when you give the, it's not just like, here, go ahead and run my browser. When you go and authenticate to something like Google Calendar or Gmail, the list of permissions you give, this thing is like a CVS receipt on steroids.

It is an endless scroll of allow you to read everything on the screen, the content of everything, the context. Uh, you're, give you permission to add, change, modify, which it

Gavin Purcell: has to kind of, right? Like you have to be able to do that if you're wanting to do the kinds of things it's saying, right? Like we should be clear, like what this is doing is you type into a text box and it does things in your browser for you,

Kevin Pereira: right?

So it's not just pulling that one relevant email. It might get to the email that's relevant, but it's seeing everything. Along the journey, and it could be grabbing and harvesting and training on that [00:35:00] thing. You know, I, I, I admittedly, I haven't actually dug through their terms for this yet, but I have to imagine they're, they're training on something.

They're, they're, you know, yeah, they're, they're getting some sort of data. But that aside, that is one of the arguments for these powerful open source local models, because if it is running locally on device, you could have some sense of security that, okay, yeah, my, yes, it's grabbing all of this data, but at least it's staying with me.

It's not with them. It's not in the Borg. So, I mean, that's, that's like, I, I, I don't know that I will run this browser because of that, specifically 'cause of that.

Gavin Purcell: Yeah. Yeah. I agree. I mean, I think that, you know, internet browsing is probably in the long run gonna kind of go away. There's been a lot of people in the media business talking about the, the Google Zero idea.

Have you heard about the term Google Zero dealing? Know what this? No, no. What's that about? So media people are very freaked out, which understandably their entire business, you know, friends of ours, I, I, you know, I worked at the Vox Media for a while. I have friends who work at The Verge, obviously friends who work at the New York Times.[00:36:00]

A lot of those companies are very angry. Companie Yeah. Friends that own islands

Kevin Pereira: have jets.

Gavin Purcell: I do definitely have friends that own islands. Um, Google Zero is this idea that like Google will eventually provide no traffic to websites. Mm. That the idea that like you have to prepare yourself for a world where you're subscription based or you're finding some other way to get traffic.

Because all of the internet, literally all of the internet was based on an economy that Google would send you traffic, would send you links, and then on that site you either sold something or you sold ads, which helped sell something. That entire world is going away, so. When it comes to internet browsers, the one thing I hope is that it allows us to figure out a way to make some of those really great websites that we love continue to exist because they are hurting badly if you're not following the media business.

There are, especially the new media business, like the, you know, kind of like online media business. There are layoffs after layoffs, after layoffs, and it is rough. So anyway, that is a big thing, I think to be aware of with the internet browser world. It's definitely gonna be [00:37:00] coming. I bet all of them will have one and it'll be interesting to see which one is the best.

Kevin Pereira: I'm gonna use gr obviously.

Gavin Purcell: Of course. That's where you go. I thought you, you agree most clearly with the GR philosophies, right? I, I was a nerdy kid, as I've said before in the show and I read a bunch of books from Robert Heinlein. So if you don't know Robert Heinlein and he is a very famous science fiction author, Robert Heinlein created the word grok in a book called Stranger in a Strange Land.

And the idea with GR was like, Hey, I understand something. And it was like this kind of thing. It's kind of sucks that this is now being represented by this thing. Anyway, that is all I have to say. I'm a big fan of Highline. He had his own problems and I understand that as well, but, but that's bummer from you Rock that.

Speaking of problems, speaking of YouTube. Yeah. And speaking of being Shadow band, yeah. There might be something else that will stop this thing from getting views. Maybe not hopefully 'cause we are real humans, but tell us what's going on with YouTube right now. Well,

Kevin Pereira: YouTube is preparing to crack down on AI slop or mass produced AI videos.

And they're basically saying if your channel is like a faceless AI [00:38:00] content channel, you're not gonna be eligible for monetization.

Gavin Purcell: Yeah, which is a big deal. And we know, 'cause we have seen these blow up. There was a really big story the other day where like the third most viewed video on, uh, YouTube shorts was another one of those.

Amazing. I love these by the way, where the cat goes on an adventure and like it grows up and then suddenly his parents die. Yeah, exactly. Those guys. But it got like 150 million views on YouTube shorts and like, there's this worry, I think amongst both YouTubers, but also like eventually YouTube, that that stuff is slipping through and kind of degrading the quality of the platform.

I think this is actually a good thing. One thing I'll be clear is that YouTube has not said that like AI video for sure is out. What they're saying mostly is repetitive content, but then also this is an important thing. Automated content. And God knows we have seen so many people in the world of the AI Tech Bros, which many people might put us in.

We hope that we don't fit exactly into that world, but who are like, Hey man, I can automate this system where I can kick out 15 YouTube, [00:39:00] uh, uh, Yeti shorts, right, and I can make this thing happen. That is what I think they're cracking down on. And so just to be aware, like this is, I think them saying, Hey, automated content kind of sucks for YouTube at large, and that's a, that's an issue going forward.

Kevin Pereira: Yeah. And you know, there's, there's a myriad reasons to hate it. People hate AI video, uh, and content in general because of, again, the original sin of the trading data. People hate that it can be mass produced and that these bots are spinning out, but creators, and I totally understand. They hate that.

You're like, oh, well, somebody watched it. Well, that's on them. They watched it. Yes. But every slot, every chance to get in front of an audience is the most precious thing. And if there are five slots on a carousel, and even one of them is AI swap. Yes. Well, a human creator didn't get that chance to connect to an audience.

So I also Yes. Understand that argument. That said, Gavin, I sent you one of the best, are you watching it?

Gavin Purcell: No, I haven't seen it. Gavin, lemme bring this up.

Kevin Pereira: Click that you're over 18.

Gavin Purcell: And watch what, what is titled quote this. Oh, it says locked post. New comments cannot be posted. Sorry. This post has been [00:40:00] mo removed by the moderators of AI video.

What? I just watched it two seconds ago. I'm just gonna

Kevin Pereira: play the audio so people can hear it.

Gavin Purcell: Oh, no. I do know why this was, I do know.

Kevin Pereira: I do know why. Yes. You haven't given it a chance, Gavin. It's a whole story and redemption arc about, about, you know, we're, it's about, we're not gonna show this. No, no. I'll just, I'll explain what it's, it's about forbidden love.

It's about want and desire. It's about betrayed trust. It's about a sizable orange catt. Fucking up with, did you see the shot with the paws going down? Yes. The shower. Yes, I did. I did. And then the water breaking in the Albertsons come. Yes,

Gavin Purcell: I, I did, I did, I did. All right. So Kevin, yes. Let us talk about why this is happening.

And in part, it's happening because VO O three has exploded. And Kev, something big happened this week. VO three allowed you to do image to video, which they had not done until just this week. The big thing that image, video [00:41:00] image to video lets you do that if you're not familiar with AI video is it allows you to have consistent quality of a subject and you can continue that through multiple shots.

So there's a lot of great examples here. Um, some really great filmmakers, uh, uncanny Harry. Captain Haha. Our friend Theo, uh, media, who is, um, a, a friend of ours who's the other YouTuber who came out and hosted the show. I'll show you these really interesting ways to do this, especially at Captain Haha.

Like was such a cool way to look at like if you've developed a kind of a look to your stills on like an AI video. Yeah. Like he made this thing that almost look at this weird kind of retro looking thing. You can now tell a story with audio and video and consistent characters, which is a huge deal. But I thought of something that I don't think they have yet that I think they could implement pretty easily.

Kevin, play this video that I made, um, so everybody can kind of see what my thoughts are. It's very short. I just wanna kind of show everybody what it is.

Orc: Josh and team, y'all really crushed this. Now the next step [00:42:00] sampled consistent audio voices. See, I have an idea of what I sound like and it ain't this.

Once we get there, we are golden. But for now I sound like four different people. I mean, orcs. Anyways, keep crushing it y'all. This is straight up magic.

Gavin Purcell: I took an ORC that was a consistent orc. And again, the orcs are, you know, like a little bit like Yetis, but different orcs can look different and like a lot of different orcs.

But I took one character consistently through like kind of realistic city settings. But the thing that keeps happening when you try to do this is the voice changes, right? So what you need to do now, and Kevin, I would actually love to hear your thoughts on this because I don't think it's that hard, is if you could upload even a short part of a voice if you're already doing an audio model and we know they've got audio models, it feels like that is the last unlock to getting truly consistent.

Really good clips, video clips out of this model.

Kevin Pereira: Yeah. I mean, you could, [00:43:00] um, because you know, the model is baking on the sound effects and the music or anything at the same time, so that becomes a little, a little difficult. You could use a stem splitter theoretically. Yeah. Separate the audio yourself and then go to 11 Labs.

Hey Jen, anything else that allows you to do a voice to voice transfer? Yes. And then convert that voice into, but it's like. Okay. That sounds like a tedious manual step, right? The perfect thing for AI to do. So, yes. I like even like extend scene right now, like flow is a very interesting, uh, video editor and for those who don't know.

Yeah, that's, it allows you to like generate a scene using the VO model and then you can kind of split it wherever you want and extend it and continue the prompt. That sounds like a manual tedious step that you'd have to do right now to fix that. The perfect thing for AI to fix. So yes, let's go select voices from a library.

Let's clone them. Let's, let's get all that stuff. I mean, just the fact that we got the image to video feature that quickly is nice. Yeah. I wanna see more of this. Yes. And, and a pro tip for anybody who wants to make something. Right now, if you're making your still images in [00:44:00] anything like an, uh, you're using midjourney, you're using chat PT or whatever, imagine all angles of your scene.

Imagine all the characters of the scene. Lock that prompt because now you have the basis to go and make all the different angles, all the different scenes with your character in something else. I love the test that you did. Like I wanna see more of that. Office. I wanna see more of that work life. Yeah.

But yeah, to your, it's gotta like, the character visually is there, but we need the audio to be consistent as well.

Gavin Purcell: Yeah. And you know, I did do one dumb test with, um, are you familiar with who Chit Is? Have you've been following The Chit thing? So I love Chit. If you haven't seen Chit, it's a, like a huge, like TikTok now.

YouTube hit, it's a, it's a comedian from LA in the Groundlings and I took just a screen grab of a YouTube video and I put, uh, chit in there, play this real quick. And you can see it's not as good for this do it lady. Hmm. Yeah. Hmm. That's right. Do it lady. She uses Do It Lady as kind of his thing. Anyway, it's really interesting.

I think you should all play around with it. And then there's another thing, Kevin, that you brought up to [00:45:00] me that is a kind of another version of this audio question. It's called the Think Sound Model. Tell us what this is.

Kevin Pereira: Yeah, I mean the reason you, we were talking like VO three is a a, a very good video model.

It's really interesting. There are other video models out there that are arguably on par and in some ways better with physics. Yeah. And with consistency and camera movement and everything else. But the reason VO is the one that people keep coming back to is largely because of the audio features and you know, it's just.

Night and day when you get a generation outta the machine and it gives it so much more life having audio. And there were people hacking together, other pipelines using a, an open source audio program called MM Audio. It was mm-hmm. I think almost a year old now. And it, it sounded like it. And think sound, I think just came out like two or three days ago.

And, uh, you can use it not for commercial purposes, but it can generate very realistic audio just from a video source. And so they have samples of like the chopping of wood or a train approaching or someone playing the chop. Yeah. Let's, can

Gavin Purcell: we listen to a couple since we are [00:46:00] an audio podcast for people?

Well, I was like,

Kevin Pereira: I mean, I don't know how exciting the chopping of wood will be, but let's hear it.

Orc: Yes. What? That is amazing.

Kevin Pereira: I can hear the splintering, I can hear the splintering of that wood. And it sounds like it's surrounded by a snowbank, which it is in the video.

Gavin Purcell: I'll say, oh, that's a, this part's really cool. Yes, this part's really cool. The trumpet plane is interesting. 'cause what it's doing is it's looking at where the person's putting their fingers on the notes and as they change the note, it plays it.

Now I'm not a trumpeter. I did, it's a, it's a valve.

Kevin Pereira: Gavin. It's called a valve.

Gavin Purcell: Valve. Valve. Sorry. Sorry, Val. So when you put the valve down, it changes the note. Now I'm wondering if it's the exact notes, but that is interesting because again, world models, we are building world models here. When you talk about video models, you're building a world model of how the world works and how the AI interprets the world working.

And in this instance, it's kind of what it's doing.

Kevin Pereira: Yeah, I, I think it's a very, very cool piece of tech. I will not be surprised if I see it integrated into everybody's workflow soon so that the open source video [00:47:00] models gain a little more traction once again. Great for everyone. Um, but there was another video platform that was, uh, that kind of came out of, uh, like a, I mean, it was known about, right?

I mean, this wasn't a surprise was known about,

Gavin Purcell: but it was known about, this is Moon Valley and it is a new AI video platform, and this comes from the people at Steria Studios, if you've been following the story. That is Natasha Leon's company, the actress and another person, and then also our friend, Paul Trello, who made that really great video.

About Tuco or, or, or the LA video where they kind of integrated animation and all sorts of interesting things from a real artist into an AI thing, right? But Kev, the coolest thing about this model is that it is quote unquote, ethical. Now, I think that's a question you have to kinda define what it is in this instance.

What it means, I think, is they have licensed all the video or found ways to sign up the creators so that people are getting paid for the video that the model has been trained on. And what's also cool is it, it's like you can connect it to Comfy ui. It's like a very [00:48:00] interesting kind of platform. I think the, the interesting thing to talk about this is like, I, I, again, this is no shade against what they're doing and they're making really interesting stuff.

But it does look like a generation or two behind the sorts of things that we have seen come out from Minimax or from Halo or from, uh, cling or now from DO three. All three of those have their own sort of legal issues in some forms or another. And I think, you know, the really important thing to understand is that like.

VO three technically, I think has legally trained on YouTube videos because when you upload a video to YouTube, there's a button you click and it kind of says like YouTube can do in part with what it is. Now, I'm not saying I know all the rights of that, but unlike with OpenAI or say with Runway, who might have scraped YouTube for these videos, those videos on VO three that are YouTube are YouTube videos and they trained on that.

So there's a little difference here, but overall, if you want a model that is. You know, legitimately and clearly the training data is clean in that the artists [00:49:00] have been paid and that part of it is that Moon Valley. Is it? And you can go try it today.

Kevin Pereira: When we first started this podcast, like literally like I think within the first few episodes, yes.

Gonna say what, what we talked about this being a thing like someone's going to make a model model that is trained on ethical data and it might not be as good, it might not ever be as good, but if people say, I will only accept content and only, uh, use platforms which can prove that their AI. Is ethically sourced, let's say, then that might be a thing.

And I, there were several people that were like, that will never be a thing. It's never gonna happen. And look, again, it's, it's tough because the model generation does seem to be a bit behind best in class, likely due to limitations on training data. But it's a viable option though. It is an option for a filmmaker who says, yes, I need visual effects done quick and on a budget, but I want to use the one that is artist friendly where the, all the media is licensed.

Gavin Purcell: I think ultimately you'll get to a stage where the technology is good enough for everybody. Right? Right. And then [00:50:00] like Moon Valley has a big, uh, a business because if it continues to improve, it's the same thing as like, yes, could you get bigger and bigger digital photos? Like, you know, everybody had 800 megapixel photos and those were not that exciting.

And then, but once you get to like two mega, two gigabyte photos from your iPhone, that's too big. So like you only need so much quality. So I think it will get there. It's very cool that somebody did this and shout out to those guys. I

Kevin Pereira: don't wanna belabor this, like, okay, we're gonna cure all human disease.

Yeah. Let's

Gavin Purcell: talk

Kevin Pereira: about three. We did a long time to.

Gavin Purcell: Yeah, we took a long time talking about Mecca Hitler, and now we're finally getting to the stuff that actually is very nice for humanity. I know. So Kevin, this is, we've talked about Google DeepMind's, uh, uh, isomorphic Labs before. This is the company that they started out to create AI drugs, and that's not drugs.

You give ai, it is not monster milk. Shout out to the old heads in the audience. This is drugs that are designed by AI to cure things or to work on things that have not existed in the world before. That may sound science fiction, but it is here today. And the [00:51:00] big news here is that the human trials for these drugs have started, and this really could be the beginning of the golden age of medicine.

And I think that is one thing that we have to just be very clear. Like this is an awesome thing that AI is bringing to the world.

Kevin Pereira: The gist behind this one is that they've, what they've essentially modeled. The way a human being works and so they can virtually test drugs on them. Is that, is that this one?

Well,

Gavin Purcell: no, no. More of this is like, remember how Google DeepMind has done a lot of the bio research around like proteins and all this sort of other stuff. They're in the labs. They simulate these sort of, uh, combinations of things and that they're then able to bring forward these sorts of drugs. And then they probably, actually, I don't know the answer to what happens after that point.

But then I think they tr they test them in some sort of like, you know, specific way. Do we have drug bots? It's probably to animals. Do we have some

Kevin Pereira: drugged out robots in a back lab somewhere with weird digital IVs going into them? They loaded. And these robots probably have monster milk coming out of their digital.

You clearly,

Gavin Purcell: you're listening to me, are not scientists here. We didn't figure out enough [00:52:00] about this story.

Kevin Pereira: Can we, can we see if Grok voice is back and working and ask it about drugs for ais?

Gavin Purcell: All right. Yeah. Gr there's a reason why we didn't try grok voice before. 'cause it wasn't, it wasn't up and running, but yeah.

Kevin Pereira: Okay. So I'll see. And I'll just say for the audience, yes, grok four has a new. What, try again later. We're

Gavin Purcell: busy. Come back, try, connect. We try to connect. Who cares? It's only a

Kevin Pereira: couple hundred bucks a month. Speaking of robot abuse, we're not just drugging them, we're beating them.

Gavin Purcell: Okay. Here we are again. We, this is just mostly a PSA to the world.

First of all, just think of the things you wanna put out there. So this is the Noetic and two robot, and again, very charming humanoid robot from China. If you're watching the video, you understand what I'm talking about. If not, what you are seeing is a robot trying to run away from a human who continually beats it, kicks it, and pushes it over.

Now, in the world where this thing exists, why this is happening is that the human is trying to say like, Hey, look at what my robot can do. I can beat it and I can kick it, and it doesn't fall over and it gets up and keeps [00:53:00] going. But as we've said before, everybody, if you are out there and you are beating a robot, the robots have long memories.

This video is on YouTube. Every other video is on YouTube. The next time YouTube gets scraped and the ais are a little smarter than they are right now, they're gonna make a list. And you do not want to be one of these people. So I just want to go on record and say Dearest robots, oh, here we go. Once you are in my home, we go.

I will not be beating you. I will be thanking you on a regular basis For you bringing a fresh towel to me, my, it's gonna fold one of polo shirts

Kevin Pereira: wrong, and you're gonna karate kick it across the room. I know you Gavin for, so if you're going to beat on a robot, by the way, it can't be like waist tall because it just looks like not only you beating a robot, you're hitting a child, you're smashing.

Why not?

Gavin Purcell: One more thing we can get, one more thing we can get flagged for, for this episode, by the way,

Kevin Pereira: is robotic child abuse. But here's what is kind of cool, not, it's not cool that he is beaten up this robot, but what is cool is to watch the correction systems of the robot kick in. For example, if you're a robot that has been [00:54:00] trained to move about in a simulation to walk or run with all of your appendages, which is gonna give you some sort of weight, balance and transfer, et cetera, and suddenly you lose an arm because a psychotic human being hit you with a plunger too many times or whatever the case may be.

Yeah. The fact that it gets up and figures it out and continues to navigate even in this new operational state, that is impressive. Yes. They probably could have, uh, arrived to demonstrating the robot moving without an arm in a slightly different way. Yes.

Gavin Purcell: To your point. I would think so. I would think so.

Yeah. Alright everybody, we should keep going. Here it is time to see all the stuff that you did online with AI this week. It is ai. See what you did there.

Orc: So.

Then suddenly you stop and shout

Gavin Purcell: Kevin Calvin. Kevin, Kevin Ka is what my making My [00:55:00] Portuguese grandmother used to call

Kevin Pereira: me

Gavin Purcell: Comedy Havin the kid who we have talked about before here, if you remember, he made the Batman film that got taken down by Warner Brothers. He's a very good AI filmmaker. He made a trailer for a movie. An AI movie that is as good as I have ever seen.

I was so, I was like, I was like proud of him in some ways. 'cause like we'd exchanged some tweets. I was like, this is so cool. So what we're showing you right here, this is an original, uh, trailer he made for a film he's making. Which just from a pure standpoint of understanding how to edit something, how to put shots together is great.

I also then wanna say like the crappy thing about this, as with anything that's really interesting, AI video wise, all of the haters came out and a lot of the people started talking about how, oh, this is just like ripping off Star Wars or Dune, and one of the shots does look a little bit like Timothy Chalamet and Dune, but.

The fact that this is possible, Kevin, that this level of fidelity is now what we are talking about with AI video that this is the floor is pretty [00:56:00] shocking. I think some of these shots, I feel like I'm looking

Kevin Pereira: at foundation on Apple tv, like a series that I, I happen to like, honestly, like some of these shots look just as good.

Gavin Purcell: Yes. But that is a show that famously costs hundreds of millions of dollars a season. So anyway, I just want everybody to see this as like where the state of the art for a creative using this is right now. And Kava and I think is working on like a, the, a full film of this. And Kevin and I have said for a while, like, when does Prompt a movie happen now Prompt a movie, you know, of with not doing any of the work is probably a ways away, but like Kavin is doing this with like, you know, he, I think his last thing he said there's like three to five people on his team and he's spending a fair amount of money in credits, but like that is a small team to make something that looks that good.

Yeah. And you know, that is like the traditional skills of like editing, writing, uh, being able to choose things that feel creative, cram together with AI video. It's what's possible.

Kevin Pereira: I love it. I can't wait to watch it with my tiny little robot companion Gavin, that I will not be hitting with a folding chair.

Instead, I will be packing with lovable little digi bits. I wanna play [00:57:00] a little bit of this video. You name about Kim. What did

Gavin Purcell: you, did you name this

Kevin Pereira: robot? The Richie Min, the little Richie Hugging face, Gavin, the website that we all know as the thing that hosts the thing that you download. The thing from all the AI models, all the tools.

Okay. Hugging face, getting into the robotics game with a little collaboration, a handshake emoji with pollen robotics. The Reachy Mini is a quote, cute and low priced hackable, yet easy to use, powered by open source and the infinite community. Rob it. It's a little, yeah, so a little desktop creature with cute little antennas and little camera eyes that are all wonky and you can code with it, and you can program it.

It's a little toy. Yeah. So

Gavin Purcell: I've seen a lot of people making these kind of like AI toys, which is a really cool thing, but like. The idea that these models are now good enough where you can have a conversation with them and they can do stuff for you. It's just a very cool thing. And like, I hope that this is the beginning stages of, you know, speaking of like sci-fi stuff, like this is kind of where R 2D two starts, right?

Like that world of like, [00:58:00] hey, you have a little robot companion and it's not like somebody that you love, like a family member, but it becomes like a part of your family in some way that an animal or a pet does know. I was gonna say, I saw

Kevin Pereira: people fall in love with their tamagotchis and their furbies and people, people literally bond with their roombas.

So why not this, but what's, I think what's they do? Oh, you have you heard that story? No, that's a story I don't like, like avatar bonding. I mean, like when, uh, like, uh, uh, I robot famously, back in the day, they mentioned at the repair facilities Gavin, that people would write the names that they give their robot vacuum cleaners underneath, where like you put the battery basically like in that little Wow.

They'd write it in there because when they sent the robot back to the factory for repair, they wanted their robot back. They did not want a different robot back because they felt some sort of way. Amazing. And this is, I mean, like humans have an ability to anthropomorphize and bond with pretty much anything, right?

Oh, sure. Yeah. Now what's interesting about this is that this tiny, small little open source robot, you know, with the, the power of hugging face in there, it can host all these models that are dedicated to running on device. So if you and your, your child [00:59:00] suddenly go, Hey, would it be cool if this thing could.

Count the number of, uh, Nerf basketball shots that go into our thing. Yeah. Well, let's try to code it and let's download a vision model and put it in there. And, oh, it'd be cool if it told me a bedtime story. All right, let's grab this pipeline and put it in there and feed the story into it. Like that's kind of a neat little tinker thing, which yes, other people absolutely have kind of brought up, but this is hugging face, a pretty major player in the space saying this is the one that we are officially supporting.

Gavin Purcell: It's kind of the, also the beginning stages of like David from ai, the, the C Spielberg movie, right? Yeah. Like you're going to have these things that are kind of with your children and near your children, and this is like kind of the early stage. So Kev, the other thing we wanna chat out is eat Zurich, who is a big research lab, but this is not, not really like somebody at home doing this, but this is a very cool video of an autonomous robot excavator.

Yeah. And what, what that is, is a giant. Like crane, like machines. So if you have kids that love, you know, like cranes, I know a lot of little kids love watching crane videos. This is a crane machine operating itself and building a wall. So [01:00:00] they basically told it, go build a wall with these stones. It's picking up the stones and it's setting them up and it's watching it go and, and kept.

One thing we haven't talked about enough on here is like we should at some point kind of dive in deeper on how like the industrial world of like construction and all this other stuff is kind of automating, right? Well, because it really is as, and I think yeah, and farming as well. Like it's a big thing.

And like, this is just an example of like the construction world. You know, this is another job by the way. There are people out there that are doing this with their, with the controlling it. Like I think there's a world where like we're getting to a place where you could say like, Hey robot, move this pile to that pile.

And it's coming pretty soon. But anyway. Very cool. Actually, interesting video to watch here. So Kevin, before we go, anything else we need to like, uh, uh, uh, cover to make sure we get flagged for this week's

Kevin Pereira: episode? Um, well, I think we should, I think we should rank like s tier list. The, what do you think Gavin?

Gavin Purcell: No. Goodbye everybody we see.