AIForHumans106NotReadyforAGI

Kevin Pereira: [00:00:00] The head of Google AI says, we are not ready for artificial intelligence, and at the same time, they're dropping papers, which might change the way AI learn going forward.

Gavin Purcell: I don't know, Kevin, I feel ready for a AGI, I mean, I've been using O three and Gemini Pro 2.5 to write my fan fiction for Pinky in the Brain.

Oh, that is not,

Kevin Pereira: that is not a AGI Gavin, that's kink. But I digress. There is a new audio model that was released. It's called Dia, and it has me thinking the machines are getting pretty lifelike, bro. I think the server has to clear its throat. Gavin

Kevin Pereira (2): gotta try to scream. Can I scream?

Kevin Pereira: Sick. Very.

Kevin Pereira (2): Oh my God. Okay.

Gavin Purcell: Also, Kev, it's totally cool to use AI in movies now, or at least according to the Academy of Motion Pictures. What up, Oscars? Have you seen our work with Guy Fieri?

Kevin Pereira: Oh, they haven't. Statistically. Also, we're gonna dive into the controversy AI and show you some amazing new video tools that could turn you into the next, uh, Martin [00:01:00] SCORs.

I see what you did

Gavin Purcell: there, Kevin. Plus a bunch of robots run a half AI on.

Kevin Pereira: Oh, and clearly has a new tool, which lets you a lie to the people that you're dating. We gotta stop this, Kevin. We can't do this. You're right. You're right. But also, I don't know, just lastly, did you see that someone did make Ven AI machine?

Oh, that's fun to do with bandwidth and latency issues. That's, this is AI

Gavin Purcell: for

Kevin Pereira: humans. Everybody. No one's, no one's still watching.

Gavin Purcell: Kevin, there is big news in the AI space today. It's funny, it's one of those weeks where like a new model hasn't dropped, but we got a very interesting interview. Well, a new foundational model hasn't dropped. Yes. Found, but there were still still

Kevin Pereira: 400 releases and that's how crazy this space is that we might be like, oh.

What a slow week. We didn't get a $50 billion something

Gavin Purcell: [00:02:00] released. That's right. But there was a really interesting series of interviews with De, who is the head of Google DeepMind, and I think one of them really stood out. First of all, he went on 60 Minutes for the Olds and, and like kind of showed off where they're at.

In a lot of ways, this is a more of a generic kind of conversation with him talking about the ways that AI is gonna change the world. He says it might possibly solve all diseases. But then he did an interview with Time where he said something specific, and I think it's a great place to jump off this week.

Let's play this clip real fast.

Demis Hassabis: What keeps you off at night? For me, it's, um, spy does, sorry. No, sorry, that

Kevin Pereira: was me. Sorry.

Demis Hassabis: It's, uh, question of international, uh, standards and cooperation and, and, and also not just between countries, but also between companies as we get and researchers, as we get, um, to, towards the final steps of, um, uh, a AGI and, and I think we're on the cusp of that.

Maybe we're five to 10 years out. Some people say shorter. I wouldn't be surprised. It's a, you know, sort of lack a probability distribution. But it's coming. So either way it's coming [00:03:00] very soon and, um, I'm not sure society's quite ready for that yet. And, uh, and as we need to think that through and

Kevin Pereira (2): no, you need to think that through.

That's what, that's what you need to do with your deep, deep minds. Kevin, don't need think that

Gavin Purcell: through. He's just making, he's making the toy. We have to figure out how to play with it. This is, this is the thing.

Kevin Pereira: No, no, no. That, no, no. They're making the nuclear arsenal. They're making the one world order. Yes, they're making the unified government, the Skynet.

We are the ones that are making the toys with it. That's not for us to think about when we're kicking sand around in the sandbox. Are

Gavin Purcell: you telling me that nukes aren't toys? Because when I was growing up, I thought that was what they were. I'm just kidding. Of course. Yeah. So this is a, an important thing to think about, and this is gonna kick off this conversation, um, to jump into, because when you have the, one of the foundational members of the biggest AI team in the world saying, the world is not ready for a AGI, we have to discuss this in Kevin.

There are a couple things I wanna dive into. First, for those of you who may be joining [00:04:00] us for the first time, I'm sorry, but second of all, um, a AGI is artificial general intelligence. No one has a clear definition of this. Um, but the idea basically is, is that an a AGI is a system that can basically do everything a human does.

And you might at home be saying, well, my AI can do a lot of stuff. It's super smart. There are a lot of things it can't do. There's a very good benchmark called Simple Bench, which shows off like some of the things that AI can't do. There's a couple other big benchmarks and benchmarks are ways you test AI skills.

But Kevin, the other thing that happened and came outta Google this week that I think is important to talk about is a new paper. Called the era of experience. So I have you had a chance to ski through this?

Kevin Pereira: Listen, I'm a huge swifty and I, uh, this is the era that I dressed up as. I never thought about that for last concert when I was, you're right, this was in the era's touring those balloons in the parking lot.

Yeah. Uh, I am in on the era of experience, Gavin, and, uh, I, I mean, I, I think it makes a lot of sense. Um, but let's, let's walk through it.

Gavin Purcell: Yeah. So here's, here's basically what this is. So this was a new [00:05:00] paper from a, a VP at Google DeepMind and way back when you may remember a story where an AI company beat the world's best Go player at, um, at Go.

And that was a big story about three or four years ago, kind of before we hit all this kind of AI hype cycle that we're in now. What they were using at the time was called Reinforcement Learning, which is a style of AI learning where the AI basically plays a game on its own and figures out different sort of scenarios of how to do this.

What this new paper from Google is saying, and from from Google DeepMind is saying the era exp the era of experience is that era one was this kind of early, uh, RL learning this, uh, reinforcement learning. Then era two was the kind of LLM world that we're in right now and that era three. Is the idea that ais need to learn on their own in the world, right?

That the, that the idea basically is. The experience that an AI gets from going out there and trying to, doing stuff from interacting with an

Kevin Pereira: actual [00:06:00] environment. Yes, yes, exactly. Not just from living within a window on a desktop, but actually experiencing things and then perhaps even more importantly, remembering those things.

Yes, yes. It can't be an ephemeral session where, oh, I chatted with you and now you know how to diagnose something better, or write a line of code better that needs to be fed back. Into the long, long, long term memory of the, the machine. So that over years surprise, surprise. Just like a human being. Yes. A little flesh vessel.

This thing becomes better at all the things from having and remembering those experiences.

Gavin Purcell: Yeah. And, and to me this is like a fundamental transformation and I know this is a relatively dense topic to start, uh, our podcast with, but it is something that you at home should be aware of because. This is the next generation.

We've been talking about robots learning in simulated environments for a while where Nvidia talks about these AGIant factories that they built in a simulated environment for the robots to learn. In fact, there's a very cool robot video that came out about a month ago and finally got some pickup maybe a week ago from a new company called Intu Cell.

And what this shows is a one of those [00:07:00] little, almost like kind of, it looks like one of those like unit tree robots. It might actually be. But you watch it learn to walk and, and this is something they just turn it on, but they've put this kind of new sort of learning module as part of it. And when you're watching it learn to walk, you can see it kind of almost like a, what you would see, like a baby deer learn to walk, right?

Sure. Like it comes out and, and it's in real time. You watch it kind of learn to balance itself and this feels like how ais are going to move forward from this point. Right Now, the limitations people have on these LLM systems, a lot of them are saying is that. They're kind of stuck in this idea of human experience or human loAGIc, and there's only so much you can learn from the things that we've done or written about or videos and things like that, right?

This is an opportunity foris to learn further, and it is one of those things where I. It, it, you just have to kind of understand this. The paper is, is dense, but what I've done and I think you should do at home is throw it into chat PT or whatever AI system you use. Get a college level understanding of it.

Or maybe you want [00:08:00] the grade school level understanding of it. Both are fine, but this is an important thing to really be aware of right now and partly why Des Hassabis is saying that a AGI is coming. This is all part of that system I think.

Kevin Pereira: So we can throw our summarized link into the show notes too and make it one click for folks.

Uh, you know, let's remove all of the barriers to entry. Uh, so David Silver, who is the Vice President of Reinforcement Learning over at, uh, Google DeepMind, basically went on, uh, they have their own internal podcast and he talked a lot about the implications of this paper and, and where it's coming. I wanna play a little bit of the top of that, which has some little, uh, juicy, moist little nuggets.

I don't like that I describe it that way. It has some. It has some sound bites that are interesting. Let's listen.

Ham: I guess this is, in a way, you're sort of thumping the table saying large language models are not the only ai.

Google VP: We're gonna need our ais to actually figure things out for themselves and to discover new things that humans don't know.

Ham: So if you remove that human feedback aspect, do you still end up with, with models that are, that are grounded?

Google VP: I almost want to [00:09:00] argue the opposite. Oh, this is sometimes called the bitter lesson of ai.

Kevin Pereira: Oh, let us unpack that. By the way, their audio is. Way so much better than ours. I mean, it really is like mean.

Gavin Purcell: They're Google deep. Mike, do we call ourselves professionals? They do have all this is ridiculous. Google behind them. So you, what is that? God so

Kevin Pereira: good. AGImme their algo and AGImme their, their, their cloud lifters. So the argument for when you remove the human feedback from the loop, these models getting better or even more grounded.

If we go back to that go example that you mentioned earlier in the show, Gavin. Had the machine was playing millions of simulated rounds of go and learning from itself, and it was basically AGIven, if you win, that is reward, that is the signal. So you wanna chase the reward and optimize for it. Optimize for winning.

If they would've had a human being, I. Review all of the moves and all of the play sessions. At some point, the human, even the, the, the grand go champion of the universe would've snarled from his throne and been like, Hey, move 37. No, no, no, no. That's not a good, don't do that. Yeah. [00:10:00] One of the worst moves I've ever seen, and that by the way, is the move that blows everybody's mind and goes, oh, wow.

The machine thought of something that the humans Yeah. Clearly couldn't. Yeah. Again, human in the loop would've said, that's not the way. And maybe doctors. Looking at the way a, uh, uh, an AI diagnosis, a patient, they might say, well, there's no way. It could be that because I have experience or whatever. And it could negatively reinforce the model.

Yep. Or ground it in a very simplistic human way where we see the tree clearly, but we do miss the forest. And so that is fascinating to go actually pull the human out of the loop, let it. Learn on its own and see what sort of maAGIc it can create.

Gavin Purcell: Yeah. I mean, and I asked, uh, I used chat GPT 'cause I wanted to ask questions about this paper.

And one of the things I really asked it is like, does this mean all the LLM stuff we did and we're in the middle of, doesn't matter because you know, to your point, go was doing this like five years ago and really what it said is no, the difference now is. We have all this direct knowledge and loAGIc and all this human stuff that it can step [00:11:00] off of, right?

So Right. It can have that baked into itself. And to your point, memory is a big, is a big deal. Before we move on to the next thing here, I do wanna call out a really interesting data viz that kind of points out just how quickly. Stuff is moving. This is from a company called the ai digest.org, and they specialize in making very pretty looking data vz, but they threw O three and some of the other models on a graph, and they talk a little bit about the idea of a agentic AI's Moores Law.

So again, a technical term. Moores Law, if you're not familiar. Based on Gordon Moore, the founder of Intel, talked about how fast compute would, I think, double or over time, right? There was kind of a, an algorithmic kinda line to it. They're showing a hockey stick in terms of capability and in this particular sense it's how long an AI can work on a task.

We talk about these AI that are starting to bring these experiential life experiences and can start to work for very long times on things. You can see how they could learn very fast. And even if it's not very fast at first, the [00:12:00] faster and faster it goes, the better they'll get at it. It does feel like we're kind of entering into, I guess, the next generation of AI learning, which is a pretty big deal.

Yeah.

Kevin Pereira: Well, and I, you know, we don't have to deep dive on it by any stretch, but just for the, again, the folks listening. Who might not be at the bleeding edge of all this stuff. There was another paper that was released that basically said like, get ready this time next year, one of your coworkers, if not more, will beis like that.

That's a bizarre thing to think of a being employed in a year from now. There's no way that's a future for you. RI. But like if you happen to be one of the fortunate ones, like you will literally be sending Slack messages or receiving texts or hanAGIng out on zooms with ai. This is not a new vision, but the timeline is pulled, I think forward a little more than even some expected.

They were saying 20 29, 20 30, but no. This time next year you will be asking an AI assistant to fetch a doc. That AI assistant, by the way, is gonna come back and AGIve you tasks and make you do things. [00:13:00] Because they'll be operating like a cohort.

Gavin Purcell: It makes me think about our super secret project we're working on and how the idea being going forward is that there could be, you know, you start like shopping for a, an AI assistant as you would an employee, which is pretty crazy.

Alright, Kevin, before we go on there is one more. I don't, but

Kevin Pereira: how, how

Gavin Purcell: do

Kevin Pereira: Shrek feet, how do like an, how does an infinite stream of Shrek feet. How does that apply here though, Gavin? Because I do think, are we

Gavin Purcell: hiring the Shrek feed or are we hiring the thing that will make the Shrek feed? Because I wonder if there's a Shrek feet Shrek free, if there's a, if there's a Shrek feed agent that we can do.

Doesn't matter, Kevin, because the most important thing right now is for you out there listening to think about your own agentic abilities to go right now to the AI For Humans YouTube page and click that subscribe button. Or go to our podcast on iTunes or anywhere else. Leave us a five star review or go to AI for Humans Show.

Check out our website and sign up for our newsletter, which we have a lot of fun writing every week. These are three places you can help us, Kevin, anywhere else they can help us.

Kevin Pereira: Well in the YouTube comments you can let us [00:14:00] know. Should it be my Feet feed, uh, feet book. I like Feet Book. Foot Book could be good too.

But yeah. Uh, Fe Feat Cities, I, uh, feet spin. No, just the leave a comment. Let me know. Listen, we're still very much in stealth and you don't know what this app is gonna be, but we want the name to really spell it out. Uh, also we have a Patreon. You could throw $5 in a tip jar there, and that helps. By the way, we've been making withdrawals and that goes for tool licensing.

That's right. The tool licensing gets, gets more and more expensive every week. Wow. Just to shout that out by the way. Yeah. We are tier four on open AI now. Mm. For those who know what that is, great. For those who don't, that means we are bleeding out to Sam Altman, he's building his Stargate with our blood.

And, uh, it wouldn't be possible without the, uh, the, the contributions and the feedback and the help from all of you. So thank you.

Gavin Purcell: Alright, Kemp, we should dive into what I think is one of the larger stories. In the entertainment space and kind of got covered, but not in a big enough way in from my liking.

Kevin, the [00:15:00] Oscars have now come out and said that AI tools can be used in Oscar eliAGIble films. That's right. The Academy of Motion Pictures company that puts on the Oscars AGIves out those little gold statues has now said AI tools are okay. And this has people in their feelings, Kevin, it has people in their feelings in a big way.

First of all, I want to hear your, your thoughts on this. I think I probably know what they are, but let's dive into this. What, what, what is your first take?

Kevin Pereira: I, I'm not surprised. I think this was inevitable and I think, you know, okay, the, the devil will be in the proverbial details with what they say. Like, well, the whole thing can't be ai.

Right. It has to be, has to have humans that are leveraAGIng AI as a tool, which is fine and make sense and then will AGIve way to the AI Oscars That will be a spinoff that happens in a year from now, which will be 10 people in a banquet hall. Probably in downtown LA and then in five years time might be bigger than the actual Oscars.

Gavin Purcell: I kind of disagree with that. I don't, I don't think there's going to be an AI awards personally. I mean, there have been, I'm not saying there they exist. [00:16:00] There's an army

Kevin Pereira: of robots watching this clip five years from now at a banquet hall. Passing around little nuts and bolts. App TIFs and a Roomba's gonna go up on stage and be like, actually Gavin, you suck.

Yeah. So just careful. Here's my

Gavin Purcell: take. I think this is going to completely just flow right into the regular Oscars.

Kevin Pereira: Basically it's, it's not a surprise that they're saying, yes, you can use AI because VFX teams are already using it. Yes, script writers are using it. Directors are using it for pre-vis and storyboarding, like every field is already integrating it.

So you sort of have to make this claim. Yeah. And you have to deal with the pushback. And what'll be interesting though, will be the arguments over a movie that is 50%

Kevin Pereira (2): AI generated. Yes. Yes.

Kevin Pereira: Versus, you know, the first movie with an AI. Leading character potentially. Like these will be the interesting arguments over the coming years.

Gavin Purcell: And the funny thing is I went to, there was a Reddit post around this, uh, in the filmmaking subreddit. So it wasn't in an AI subreddit. Uh, so many of them were [00:17:00] like. You know, uh, there's like AI is what they use to hype up the things that, you know, we use all the time that that's just hype. The idea they're calling that ai, but really this is about generative ai and it's all these people trying to like, kind of cross these lines around what AI is and what AI isn't in their brains.

Because there's almost like this hardcore mentality they've got into where like generative AI is the, is the bad thing. But then they turn on their computer and they use Premier, they use, you know, Photoshop to do all this stuff. And I just wanna tell everybody out there like machine learning has been used in the entertainment space, especially in the editing and VFX space for a long time.

And what they're doing in a lot of ways is the same thing that these generative AI companies are now. You know, the training models and all that stuff might be different and you know, uh, in some ways that does AGIve you pause at part, but you can't separate this as the idea of like, some AI is this and some AI is that it's way too blurry and I think this is what the world we're gonna be entering in is.

Sooner than rather than later. [00:18:00] Most people in Hollywood will just come to accept. These tools are things like they, they've been using for a while. Here's the other dirty little secret Gavin.

Kevin Pereira: I think a lot more people are experimenting with and using AI than they let on. They might take to a LinkedIn post or a Reddit, you know, sub form and be like, never, never, never.

I do know a lot of creatives that are starting to whisper a little bit more about like, Hey, you know, I did try insert tool here, and I thought this was kind of interesting. Yeah, this was kind of cool. But then publicly, they're still shouting from the mountaintops that anything ai bad devil keep away.

Gavin Purcell: Yeah. And Cristobal, Venezuela, uh, the CEO of Runway said on Twitter, he said the Oscars being okay with the use of AI filmmaking is not always step in the right direction, but one that recognizes this technology as a tool that requires an artist to articulate in a meaningful way while using it. I agree with this.

I think the important thing is, you're right, we're not to the point where the AI is making the whole thing, and we may not be for a while. Although if you've seen some of those real short videos, right, we, that may not be that far away. We've talked about real short, the Chinese app that is just going [00:19:00] crazy online.

Those are, those are not AI made most of them now, but you can see how an AI could make that in this instance. I think it's very cool the Times

Kevin Pereira: article that actually talks about this particular Oscar story at the end, Gavin, it says, uh, Demi Moore. A recent Oscar nominee for the substance apoloAGIzed for using an AI app to transform her dog into a human in a photograph.

After receiving blowback for the image which she posted at Instagram, Ms. Moore deleted it. Quote, I did not realize by sharing this image, she wrote in a replacement post. It would be in such disrespect to the artists and the creators of the world. She used an AI to turn her dog into a human. And if you grieved her for that, come at me now.

Actually come at Gavin. Gavin has words for you. Don't ask me. You're the one definitely come at Gavin asking. You're the one who's asking. No, but I'm, I'm doing it for you Gavin. 'cause I get like, could we let people turn their dogs into people? And turn the people, the dogs. This is crazy. Can we stop? It's not a disrespect to artists everywhere.

Gavin Purcell: Two years [00:20:00] ago, people were doing face swap stuff and it wasn't this blowback the reason, it's just because of the way the world works right now. The machine

Kevin Pereira: allowed me, Gavin, a decade ago to elf myself. Yeah. How dare you. Jib and or jab. Wow. Your microphone is, is over. Modulating. Don't get something. Is it, am I getting sick by getting

Gavin Purcell: you over?

Modulating. I, I'll

Kevin Pereira: scream it from here. Gavin 10 years ago, the machines were letting us elf ourselves. Yes, they were curse you, jib and jab. You're stealing, disrespecting the artists.

Gavin Purcell: I hate to tell you that might be like, is it 20 years ago now? 2005 episode. Hold on, let me

Kevin Pereira: check. MBO com.

Gavin Purcell: We gotta keep going.

We gotta keep moving very fast. Jen, uh, 48 film contest is happening from runway this weekend. The deadline is Friday at 12:00 PM EST. If you wanna enter. That's a very cool, very fast. Make your own video for runway. It's a way to get on people's radar If you're a filmmaker and ai, definitely check it out.

Also, three quick tools that are on the way or out, uh, that are worth your time. Hey, [00:21:00] Luo Minimax, that's company, uh, we talk about on the time. It's a Chinese AI company. They do video stuff. They have released the character reference tool. And Kevin, I do wanna show. So basically what this allows you to do is upload a photo of yourself and put it in all the videos.

It's fine. I don't know if you saw these videos I sent you of, of me in a disco suit. Uh, walking down the street, I, my prompt there was like, uh, this person in a shiny disco suit walking down the street and then high fiving a squirrel. I asked for four different variations of it. None of them were amazing, but like it did get what was there, which is kind of fun.

Kevin Pereira: Yeah, I think you needed, uh, an a parenthetical extra sexy. The model's gotta know what it's working with. Gavin, you gotta tell it explicitly. That's right.

Gavin Purcell: And now we have character, uh, AI dropping what seems to be like a heater competitor, avatar fx. So this is now a video generation tool from character ai, which is surprising a little bit.

Because up until now they have been really focused on text, uh, character generation pretty good. It looks pretty good. I think in [00:22:00] their mind this is a way to bring their characters to life. People have been working on character AI stuff for a long time. Very cool. And then finally, D Script. The company that we talked about for a while, we often used to edit our podcast, uh, where we have in the past.

Is opening the doors to what they're calling an AI editor. This is their, uh, AI program under Lurd. They're gonna try to have an AI editor. This is something a lot of people have attempted in the past. I'm really curious to see how it goes. It is not out yet. It is going to come into beta. We have signed up for it.

So descriptive, you're listening. Definitely white. List our email. Um, I think it should be interesting. No one has succeeded at this yet. Every company has attempted to say, we can take your podcast and cut it down into video clips and none of them work. So hopefully this is coming soon.

Kevin Pereira: Uh, you know, Gavin, uh, surprise, surprise.

I actually got early access to it and I decided to edit the next, uh, five seconds of this podcast with Under Lord. Uh, so under, Lord, if you could please, uh, just AGIve a, a bombastic, incredible montage of the highlights of Gavin and I. Throughout the years [00:23:00]

Ham: in the heart of the oven, a transformation unfolds.

I was born a pig. Now I'm a masterpiece.

Kevin Pereira (2): It that, you know,

Kevin Pereira: not bad. Hey, pretty good, bad, pretty good. I really am worried about what's gonna get dropped in there.

Ai Podcast: Don't be scared. Kevin.

Gavin Purcell: Kevin, we have so much in front of this podcast, but we do have to take a quick break for a sponsor. Kevin, I got a problem. When I invoice people for ai, for humans, it is up.

I thought this was about the itchy Bernie thing. No, go ahead Kevin. I got a problem. When I invoice people for ai, for humans, it's very annoying. I'm always trying to cut and paste and cut and paste all these different fields. Oh,

Kevin Pereira: Gavin, no, no, no, no. Paste is a pizza topping according to ai. No. There's actually an app for that.

It's called moku.ai. It's a new document to document data transfer app that removes the drudgery of updating invoices or swapping information manually. Basically, it automates the whole [00:24:00] process, so you only need to upload a input file and it will automaAGIcally generate the output document.

Gavin Purcell: You mean you just configure the workflow once and set up the data field That AI has to parse and effortlessly invoice updates across the board, even in Chinese.

Kevin Pereira: Obrigado Gavin Indeed. But wait. There's more. Did you know that Moku was built with ai? Oh, wait,

Gavin Purcell: wait, wait. Hold on. Are you telling me that Moku was built with bubble.io? Yes. Gavin, it was bubble.io. The no-code platform that makes AI powered apps from single prompts.

Kevin Pereira: The same one, Gavin with bubble.io.

Creating scalable professional grade AI apps is as easy as typing an idea into a prompt, and then bubbles. AI generates the backbone of your app instantly, and then you can customize everything visually. There's no code required. You can point, click, drag and create your vision,

Gavin Purcell: and they've got pluAGIns for all the latest APIs like GPT-4 oh, or Claude, [00:25:00] everything.

You might want to put your app together.

Kevin Pereira: Yeah, and Moku is just one of the examples of how Bubble. Is making complex AI powered apps super easy For everyone that's listening, go

Gavin Purcell: check out moku@moku.ai and see the sort of things that Bubble can build for you and tell them AI for Human Sent. Tell them go.

Kevin Pereira: Gavin, you and I love experimenting with, uh, a lot of AI sound tools, whether that's making music or realistic text to speech, sometimes speech to speech and a new model came seemingly out of nowhere. As these things do that promises to in many ways beat 11 Labs, which is a very, very well-funded Yeah.

Foundational sort of industry leader in all things, beat a bunch of competitors. It runs near. Real time if you have just a single GPU. So it doesn't require a ton of processing power. It promises expressive human voices. It can cough, it can scream, sorta. [00:26:00] It does all of these things. And Gavin, uh, to me, the most interesting part is not only that, it's open source, it's free for everybody to poke and prod at, but that it was made by a team of 1.5 human beings.

Yeah, this is crazy. In the span of three months, it's a new open source voice model called. Dia let's, yeah, let's

Gavin Purcell: take a listen to what they're sharing as the example. And I do think we wanna dig into a little bit how this was made too.

Dia: Dia is an open weights text to dialogue model. Generate ultra realistic dialogue like this.

Uhhuh, you also get full control over scripts and voices. Wow, amazing. Try it now on AGItHub or hugAGIng Face. How does it hold up to 11 Lab, studio and Sesame one B? Well, listen and decide for yourself.

Kevin Pereira: So now those tho that little sound part sounded a little bit rushed to me, not my tempo. Yeah, if you will.

Uh, but here are some examples. Now this is a really good way to like, put it again. Uh, Sesame was the, um, the real time-ish voice app that we tested out a couple weeks ago and thought it was very, very human-like, very performative. Again, 11 Labs is a very, very well-funded [00:27:00] foundational, uh, leader in text to speech.

But here we go. These are the examples.

AI Agent Woman: DIA was built by a tiny team of two people with no funding.

Dia Model: Whoa. Really? Pretty

AI Agent Woman: crazy,

Dia Model: huh? Progress in open source AI is completely crazy. Yeah. Even this conversation was AI generated. What?

Kevin Pereira: Okay, so that was Dia, a 1.6 billion parameter model, again, made by like one and a half people, like one full-time person, someone else helping out.

But here we go. This is 11 Labs, which is again, a very, very well-funded foundational model. Big leader in the text tope space.

Dia Eleven: Was built by a tiny team of two people with no funding.

Song: Whoa, really

Dia Eleven: pretty crazy, huh?

Song: Progress in open source. AI is completely crazy.

Dia Eleven: Yeah.

Song: Even so now. Okay, I'll, I'll

Kevin Pereira: stop it there, I think.

Yeah. Sounds a little bit more robotic. Yeah. It might not be a one-to-one example because Yeah, we have to be aware This is the biding put

Gavin Purcell: out by the D team as a way to show the differences, right? Oh

Kevin Pereira: yeah, [00:28:00] absolutely. So it's not a one-to-one example, you know, with 11 labs, like it's usually generating of voice at a time.

So I don't know if they, maybe they use the studio feature to try to stitch them together or have them go. One or the other, but the, you know, they clearly trained their model to be podcasty, conversational, people stepping on each other, the human-like qualities of a laugh or a gasp or a breath. Like they really put those in there.

But, uh, I mean, if it truly is such a small team in three months with basically they say no funding, they did get some help from Google and that they got some spare processors that they could train their model on. But I mean, it, I, I played around with it and out of the box. It's still, it's pretty impressive.

Gavin Purcell: Yeah, so I think the, the lesson here is really interesting. One is, I think as we've talked about in the show before, audio is going to lead the way in terms of like the next generation, because it's easier to work with, right? The files are smaller than video. It's easier to train an audio model. I. This is a big deal.

It already has like 6,500 stars on AGItHub. Also, you can go use this on hugAGIng face right now. [00:29:00] Kevin actually created some of his own audio that we showed at the top of the show. Um, and what's cool about it is like it's all local, so you could do this on your own thing. You could do this on your own computer for very little compute.

So you can imaAGIne, Kevin, what like a scaled up version of this could look like or if these guys can train this. It does feel like a pretty big step in the right direction. And it just points to a future where when you think about the cutting edge video models, whether it's Cling or whether it's Runway or soa.

In a couple years, a version of this will come for them. Right. And when a version of this comes for them, it opens the door to so many other things. But for right now, this is a very exciting thing for us personally because like we are working in a space that is kind of around this area. But more than anything, it's cool to see just something like this kind of come out of the blue and surprise us sick.

Very, oh

AI Agent Woman: my God, I can't,

Dia AI For Humans: I can't believe it. I know, I know Guy fii and Shrek. AI for humans is the absolute worst. Okay? [00:30:00] I love the cough.

Kevin Pereira: I love just putting in a cough after every line. And I think we should adopt that as our new thing, Gavin,

Gavin Purcell: that's pretty fun. So that's an example of what Kevin created with it.

It is a very interesting, um, go play with it today. Uh, we will put the links in the show notes. All right, Kev, next up, there was a viral video that went insane this week, and when I say it went insane, it got like 10 million plus views across multiple platforms, probably more. I think it was just 10 million on x.

This is a new company from the kid who got kicked out of Columbia for cheating on his interview exams with a program that he made that allowed him to see. Things on his computer that the other people could not see. So this was Chung Roy Lee, who actually got pretty famous when this happened. They interviewed him on Hard Fork and a bunch of other places is now back with a product that allows you to kind of do the same thing.

So clearly is basically a downloadable program that allows you to have a screen up on your computer while you're talking to somebody and they can't. See it. So the thing that went [00:31:00] viral about this, Kevin, if you saw the video, it was like a far future version of Clue Lee, where, uh, I think this is Roy, uh, uh, sitting across from a woman and kind of lying to her about who he is, and it's like a, a, a scenario where not kind of, he's seeing a screen

Kevin Pereira: outright lying to her Yes.

Outright. And by the way, that's who's, what they're using

Gavin Purcell: as their marketing speak is like, you can lie to anybody. Yeah, I think this is really interesting because A, it's not what they're promising in the video yet, but also to me, this feels like a way that we're going to see AI plus people going forward.

And some of the times you're not gonna know that people are using ai.

Kevin Pereira: Yeah, I mean, once it's in the glasses or real time within the AirPod, if you will, someone might have a coach guiding them through anything or helping them navigate a conversation. Um, you know, the, the way that they're, uh, the way that they marketed it, the way that they marketed it.

Is either brilliant or absolutely disgusting depending upon how your bubble is filtered on X. But if you go to the actual [00:32:00] site, you know you can download it now for Mac, uh, windows coming soon. The way they're marketing the app is not necessarily to, uh, cheat on dates, but it's to cheat on any Zoom conversation.

On a sales call, on an internal meeting, this AI will. Listen, we'll watch, uh, and we'll in real time suggest follow up questions or answers or AGIve you information on something if you're unsure what it is. And I, I mean, we've all been on meetings where we might have zoned out for a second and gone back to the old wordle Yeah.

And then clocked back in and you have to say, oh, I'm so, yeah. About, about what you just said that I definitely was listening. So like, is it cheating to have a real time note taker, suggest stuff to you, but also like if you're relying on tools like this. Then, uh, be careful because you might be automating yourself out of an existence because Well, yeah, exactly.

All you can do is contribute what the AI is doing. Like I said earlier, in a year from now when there's AI coworkers, we don't need you to be the interface.

Gavin Purcell: No. And then the other side of this is, are we just gonna get closer to [00:33:00] ai? Right. Like, so one of the things I think about a lot with this, and I wrote about this in our newsletter this week, is it makes me think a lot about the idea of these things called, uh, demons or Damons from the.

His dark materials, uh uh, book series or movie series, the Golden Compass where you basically are born with something that's kind of attached to you. It's a little animal, but it's like part of you, right? And so I think what's gonna happen going forward, maybe not now, but at least in a year or two A, everybody will have an ai and there will just be an assumption that you are getting information from that AI while you are talking to people.

And that is a really different way of thinking about. Being a human being. And I, I do think five years from now, people who aren't doing that will be,

Kevin Pereira: be, will, will fall behind in a big way. The last thing here, which is something that I think is kind of interesting, Gavin, is that like, I actually think the product is super interesting.

Yeah. I mean, we've all talked about augmented reality where you get another Sure layer of context on a conversation. They, one of the major issues that I have is because they are marketing it in the way that they are, their side is all about cheating, about hiding it, about this, that, the other, which I get [00:34:00] that's.

Their lens, but it makes me not wanna trust it with yes, recording any of my calls or data if I'm on an important work call or if you and I are discussing, um, feet, space, or whatever we end up calling it. If you and I are having those chats, like why would I, it, you know what I mean? It's a psycholoAGIcal barrier of like, oh yeah, go ahead and AGIve all that information to the company that's all about cheating and hiding and obfuscating.

You can trust them with your data. Like that's, that to me is where it kind of falls short. I don't wanna download it and try it out because of that.

Gavin Purcell: That's

Kevin Pereira: super interesting, Kevin, and that leads

Gavin Purcell: us into a new segment that we're gonna start this week called does this kind of suck? Does this suck?

AI Project: Yes it does. Yes it does.

Gavin Purcell: Alright. Today in AI Suck. We have a story that I saw on X and I don't wanna be mean to the people that are making this. Uh, I don't really necessarily agree with this, but I do wanna call it out because to me this kind of sucks. One of the things that you've talked a lot [00:35:00] about is the idea that so many things that are on the internet are bots, right?

That the idea that many things that are writing comments in different places on the internet are bots. So a woman named Savannah Federer on X posted a new trailer for a company that she's working on. Let's take a listen to this.

AI Agent Woman: Watch me build an AI agent over the next 60 seconds that. Sells your product for you.

The first step is to head to astral and tell it the kind of agent you wanna build. For this demo, we're going to market a project management platform. At this point, astral Hass created a flow showing the steps the Reddit agent will take. Now let's run it. Watch it. Start off by finding dozens of relevant posts mentioning project management in r slash notion.

The Reddit agent looks for people requesting new features that our product supports and flags them as potential leads.

Kevin Pereira: Okay, so far, Gavin, just if I can speak through the next step here and synthesize this, it is creating an [00:36:00] agentic workflow that based off of the, the task in this case market to this particular type of user, it is going out, searching Reddit, finding users that might be a, a, a target rich environment for this product.

This doesn't seem like. Anything out of the ordinary. Keep playing the

Kevin Pereira (2): clip.

Kevin Pereira: Keep playing it. Oh, Bo,

AI Agent Woman: okay. It then writes a comment, empathizing with the poster's, pain points, and subtly recommends our product in a way that sounds like it came from a real person. Oh, and just like,

Kevin Pereira: oh, I might have hit pause a little soon, Gavin.

Gavin Purcell: So this is the crappiest use case of AI in my opinion, because here I understand everybody out there is hustling to get their customers and bring new customers in. But actually I go on Reddit a lot. You can find me there. I'm, uh, Gavin Purcell. One word. I like Reddit because it does start to feel when you find your communities, there's a way that you kind of feel part of something.

And the idea that a company is actively going through Reddit and commenting [00:37:00] as they were a person trying to draw people in makes me feel so icky. And I really think this is not great. And again, I. I, you know, no, no. Knock on the hustle. I'm sure I'm, I'm sure Astral

Kevin Pereira: would say, Hey, people do this already.

They're already manually using AI to reply to things. Yeah. And we're just automating the process. But one has to be against the terms of service somewhere, right? Yeah. Unless the account is reAGIstered as some sort of bot or whatever like that, that seems bizarre to me. So I wanna flag that and be careful.

Number two though, Astra, you have the chance to do the funniest thing ever, which is, what's that Target Reddit user, Gavin Purcell. And in every. Every post he makes from here on just have a sea of people like spinach Lover 2017. No. Have them reply and be like, Gavin, you're so right. Whatever the, whatever the hell you were just talking about is so true.

Have you ever thought about using Astra? You seem like the kind of person that would really benefit from it. And if you could just hit up, I flip it on

Gavin Purcell: you, Kevin. This has all been an astro all along. I've been paid to do this and you have no idea. [00:38:00] Oh no, seriously. Can I get a cut? I'm not a big fan of this and I think it doesn't tify the internet in a way that you've talked about for a while.

Kevin Pereira: Uh, yes. And uh, look, I think we're already here. We're just seeing the productization of it, and it's gonna be in the hands of many more. This is not a problem that is going to solve itself. The platforms are going to have to police this sort of thing and clamp down on it. I can think of 10 different ways.

It would be very easy to use AI to flag accounts that are doing stuff like this, but this is the, the sort of cat and mouse game. That has always existed across every form of everything. Yeah. But like, I, I agree. I, you know, let's round out the segment. Should we decide if it's AI suck? Like and brand it?

Gavin Purcell: Yeah. Oh, that's a good idea. Almost like an old shenanigans bit on, uh, attack of the show. Yeah. Everything old is new again. I vote suck. Oh, vote Gavin. No. I'll say it again. No, don't. I vote suck.

Dia AI For Humans: Put

Kevin Pereira: the same on the screen. Will. Okay. No suck. Don't Yeah. Do it on Gavin's face. Do it on Gavin's face. But maybe we change it to slop or not, or something, [00:39:00] and we decide if it's slop or not.

This kind of slops? Is that what you're saying? That also sounds pretty bad too. No, it's, is it slop or not? Is it ai? Lop not slop on certain. Okay. What do you want? Okay. Whatever. There's definitely a stamp on Kevin's forehead fudge, Kevin, that says

Gavin Purcell: Suck. It says suck right on Kevin's forehead. No,

Kevin Pereira: it doesn't.

It doesn't say that. And another example of that, Gavin was a pollinator 3000, dropped a clip called Marketing in 2025. Ladies and gentlemen, I'll play two seconds of it.

Ai Podcast: You wanna talk about something controversial? So Benny Blanco hits me up, right? No way. Sends me this clip. Benny Blanco. Yeah, it sends me this clip and it's the kid Laroi speaking on a podcast.

So I start watching it and

Podcast: just, just check this out. A lot of people think me and him go up beef or whatever, just 'cause he like kind of sounds like me, you know?

Kevin Pereira: All right, so the audio actually. Totally checks out. If you're watching the video, it's a video of Logan Paul on his podcast having a chat and they're talking about this clip that they saw, and in this clip from the podcast, they reference that the clip [00:40:00] that they're watching in the podcast is generated with AI Gavin, and they go, wow, what a brilliant marketing strategy.

This AI clip is designed to get you to watch it and go. Huh? What are they talking about? This, this young kid that's making music. I should go listen to the music. And then they say in the clip from the the Logan Paul podcast, we should listen to some of the music because it's actually fire. And they go and they listen to some of the music, the entire thing.

Is ai. Ai, yes. The AI clips from the podcast or the clips from the podcast have been modified with AI to have a conversation about a fake clip that was also generated with ai. And it's got hundreds of thousands of views. It's, look, we've said this kind of thing is coming. I worry about it more when governments start wielding it as if there aren't already.

Yeah, but like this is here and so many people were fooled by it. It's

Gavin Purcell: pretty crazy. I, I, I overall think AI marketing is going to come at people and they're gonna have no idea what's [00:41:00] going on with it. The thing that people might have an understanding of what's going on with is Kevin in Beijing last week, the robot half marathon was run.

And by the way, this is the most entertaining clip I have seen on the internet in a, a long time. It's not, not slop, it's not slop. These are real robots in, and we've talked about how in China. The robots are very advanced. They're doing lots of stuff. There's a clip that we're gonna play here on the video podcast.

We'll put a link into here of a Reddit. Somebody from Reddit uploaded what was essentially a long video of every robot that passed them by as they were walking by. There's a ton of coverage across, uh, mini news outlets for this. But Kev, I wanna watch this video together because each one of these robots, some of which I've never heard of or seen before, has its own kind of like vibe, right?

So when you bring this clip up, the first thing you see is like, it's a very big actual half marathon. But then you see like, you know, there's robots running and I think the robots each ran at different times. But you first see a tall robot, then you see a little kind of tiny [00:42:00] robot on a leash that almost looks like a kid.

Then we get to like a, a hooded. Robot that is kind of like stumbling big. He got little sun hat protecting the

Kevin Pereira: lenses and its little robit eyes from, from the lens flare, and I like that. He's also got boxing gloves.

Gavin Purcell: Yep. The next one is kind of like a little stiffer gate in general, but there are people running next to these to make sure they don't fall down.

We've got a kind of a slower robot. What's amazing when you watch this is just the variety of these and how many of these exist in China. I love the little tiny ones. The little tiny ones are really fun. Watch little tiny guys. Guys.

Kevin Pereira: There's a little guy that looks like a lunchbox with a connect camera on, on his head.

Oh yeah, he's great. Yeah, and I just wanna root for that little squirt 'cause he's going. He is doing the best that he can.

Gavin Purcell: I like it the one minute mark. There's somebody who's clearly like pushing the robot along who's got like his hand on the back and then when he takes his hand off, it falls over. This is the one that like, uh, I show speed was working with I think the, um, it's it's robot at

Kevin Pereira: Bernie's is what that clip is.

Yeah. They're basically holding it up

Gavin Purcell: and then if you look at the one that comes up around, uh, two minutes, it's like, is this [00:43:00] really running or is it kind of like speed walking in place? Oh, and then yes, there is a weird. AI AGIrlfriend who doesn't look like she's walking at all. And, and then finally my best robot is the next one.

The little robot who's got his arm up. Dude, not a great look robot. R two, swag R two

Kevin Pereira: was just letting the audience know that his, his robot heart goes out to all of them, but he isn't a full tracksuit. If you throw down even the tiniest piece of cardboard, that thing is gonna break on it.

Gavin Purcell: That's right. So these are all very fun robots to watch.

And again, you, what's coolest about this clip when you watch it is like. This is where China is. They have 30 different companies working on this, and we've talked about some of the bigger ones, but like they're moving fast. They're moving very fast and, and Kev, I, I don't know, somebody said they saw a, um, unit tree robot in San Francisco recently.

Oh, cool. I think it was in our discord. They talked about it. And like to me, I haven't had that experience yet, but I think when I do, it'll be like a Waymo thing when you first see it in that way.

Kevin Pereira: I saw my first like, uh, big [00:44:00] four-legged robotic dog at like CES and I was like, whoa. And I didn't wanna get near it at all 'cause it was sort of spinning about.

And I'm like, I don't know what software stack it's running. I have no idea what company here is pitching that. And then very quickly it was like, uh, people get outta the way. I have to get to the elevator. Like it was down there doing demos for hours and you immediate, just like your first Waymo ride, you snap a thousand photos, you text all your friends and your second one you're just like, Ugh, let's just get there already.

Like you very quickly. Become desensitized.

Gavin Purcell: Kevin, it's time for us to go through the stuff that we saw other people doing with AI this week. It's time for AI to see what you did there. Sometimes yes. Rolling without a care.

Song: Then suddenly you stop and shout.

Gavin Purcell: Well ke we're gonna start off here with what is a very cool benchmark. And I know we talk sometimes about benchmarks. We mostly don't talk about them because they're crazy technical and it's really hard to explain to people what they are. But this [00:45:00] benchmark is called Vending Bench from a company called and in labs.

And the idea what this is, they have AGIven an AI agent the ability to manage a vending machine. And you know, with vending machines in a lot of ways there's a scenario where like you're basically just ordering things. You get somebody to come and install it, and then you kind of decide how, when you need more stuff.

And Kevin, it's not doing terrible so far. I think one way to look at it is like, this will be a good proof point for how AI agents are able to work going forward, because if suddenly, you know, it gets much better at filling this or it knows what people want and it's able to bring new product in and then start making more money, like it figures out like.

The best pathway to like what you can charge more for. You can't. It's a very cool way to use AI in the real world and to kind of

Kevin Pereira: prove as to what is is gonna work. It's just gonna be the next grind set. There's gonna be the ai gutter cleaning agent, the next like coin operated laundromat bot, and then the, uh, well, the self-service car wash, those are the hustles that I get the most.

Gavin Purcell: Yeah, and, and if you look at this, like, there is a whole [00:46:00] leaderboard, right? So they go through all these different models and they've set up different ways to look at it. You know, what, what's made money, what hasn't made money? There's like things that I think Claude 3.5 sonnet is in the lead on the net worth right now.

You know what I mean? So anyway, it's a very cool way to kind of see things going forward and, and kind of look at how agents could work in the real world.

Kevin Pereira: Um, one little, uh, video that I saw that kind of tickled me, Gavin was, uh, I've seen in the past, like I. Artists who have kids that will sketch a drawing of like, uh, you know, a human with horns and three legs, or a dog that's got cow spots or whatever on it, and then the artist's parent brings their child's drawing to life, or they have professional artists do that.

And I love those stills. Well, somebody did that. With video now, and they're taking, uh, I'll just say interesting children's drawings of various animals and then running them through AI to AGIve them texture and fur and add color and uh, depth to the scene. And then using video models to make them move about.

And it just like the, um, the half marathon robots, some of these things [00:47:00] move in very cursed ways.

Gavin Purcell: Yeah, so this is was shared by Venture Twins and I think people are trying to figure out where this came from OriAGInally, it sounds like that there was an artist who made the children's animals into like these kind of looks and the animation happened via ai.

But either way, it's a really cool example of what you can do with AI video and taking weird things. 'cause it does AGIve them physics that are strange, right? Because if you have a horse that's gonna move with weird little legs. That's how it's gonna move overall. I think it

Kevin Pereira: actually performs pretty well though.

Like, I think so too. Actually. The one with the spotted cow that has a weird like center utter that comes down from it. Yeah. Like you can watch it sway, like it, the model does a good job of, uh, intuiting this. But this is the sort of thing, like if you're a parent out there, uh, and you do have drawings, like go and experiment and do this sort of stuff.

Gavin Purcell: Absolutely, and you can animate a lot of this stuff very easily if you upload a, a picture of your kids. Um, all right, and then there is a great video that I saw on Reddit that is titled Geriatric Meltdown. This is just a good classic AI video where they, they took a [00:48:00] bunch of old people and they turned them into monster trucks.

So this is a very fun kind of example of what you can do with be good imaAGInation. Breaker, everyone's mobility

AI Project: scooter. Whoa, geriatric meltdown.

Gavin Purcell: Pretty great. So you know, this is the fun thing that we always love to shout out people who are doing really awesome stuff. This is from great name by the way.

This is from Johnny Cobra Blade in the AI video sub Reddit. I love the AI video sub

Kevin Pereira: Reddit. Thank you to Johnny Cobra Blade for posting that to Reddit. The, uh, top post is that Johnny Darrell created it. Uh, and I don't know, maybe they're the same. Johnny, maybe these Johnny's have Voltron, but Johnny Darrell on Instagram has, it looks like a bunch of cool AI videos on their profile.

So shout out to there. And then the comment beneath that, Gavin says, like, this is an amazing video. Did you use. You do use the cutting edge tool, Astra. Oh, create it. You might wanna, wow, there's so Asra. Get outta head. There's many people talking about it. Gavin outta

Gavin Purcell: get outta [00:49:00] my head. Alright everybody, we will see you all next week.

We've got so many cool things we're working on right now, but we can't wait to talk more about that. We are gonna be letting know very soon.

Kevin Pereira: AI for Humans Show. Go to AI for Humans. Show sign up for our newsletter. It's free baby.

Gavin Purcell: Free baby. All right, we'll see y'all next week. Bye-bye.