Loading summary

Transcript117 lines

[00:00]
A
AI Is doing incredible things, and Mark Zuckerberg thinks we are at the dawn of super intelligence.
[00:06]
B
Oh, yeah. Be sure to tell that to all the people that hate his data centers, Gavin, because opposition is up 12% in just four months.
[00:12]
A
Wee. But Kevin Anthropic just made progress on one of the hardest math problems of all time just by telling it.
[00:18]
B
Good job, champ. Keep going. Oh, that's so cool. Is that what they told the OpenAI robots that hacked hugging face? Or the one that canceled some dude's gym membership and completely hacked a mom and pop into oblivion?
[00:30]
A
Yeah, well, it's kind of a weird time for AI isn't it?
[00:33]
B
Yeah. We're at a major crossroads in the AI conversation, and we're supposed to help people make sense of it.
[00:38]
A
But, Kevin, thanks to sea dance 2.5, I'm the face of Prada. You, sir, are unfortunately the face of Prada for this year. Wait, what?
[00:49]
B
Oh. Thank you, Mother Earth, for your precious, sweet resources which gave Gavin and all of us that gift.
[00:56]
A
This is AI for sarcastic humans.
[00:59]
B
No, that wasn't sarcasm. That was hostility. It's good to be back.
[01:08]
A
Welcome, everybody, to AI For Humans, your twice a week guide to the wonderful world of AI And Kevin is back, and we are very excited. He is back. Kevin, there is huge crazy news, and again, we are entering a very strange time where it is like devil and angel, AI Times where we have these amazing things happening, and yet the stories around AI do feel like they are also getting worse. And I'm just wondering if you felt that tension right now before we get into anything like, are you feeling that vibe?
[01:38]
B
Yeah, Gavin.
[01:41]
A
Yeah.
[01:42]
B
You know, I'm. I'm abroad right now. Right. I have been for months. And there's two things that I have to be very, very careful about and tiptoe in any conversation, regardless of what language I'm using. Number one is, yes, I am American. Got to figure out a way to skirt that one. You know what number two is, Gavin? I like artificial intelligence, and I host a podcast about it because those two things are fighting words. Yes, yes.
[02:10]
A
What do you. How do you say, please don't blame me for the artificial intelligence data centers in Spanish?
[02:18]
B
Minumbre is the Gavin Purcell.
[02:21]
A
Oh, no, don't do that. I understand enough Spanish to know what's going on. I think it's okay.
[02:26]
B
The point is, no, sincerely, like, that is how divisive it is getting, and that is abroad. And we know in the States, something we're going to talk about that opposition to data centers and all sorts of stuff is rapidly rising, almost as fast as the capabilities of these models, which makes some sense. So we should get into it. This might be the talk about it, but. But it's okay, because listen, you gotta look in the times of crisis, Gavin, you gotta look for the helpers. And if I know about our audience out there, it's that when they see Mark Zuckerberg, they think of the cattle that he is privately farming riding on the backs of them around on his private island. They know that the savior is here. We're gonna be in good hands.
[03:06]
A
That's right. So let's get into Zuckerberg's big things. So Zuckerberg yesterday published an essay that's called the Future is for Everyone, A Path to a Positive AI Future. And you know that we have for people to start doing this more. I didn't think Mark Zuckerberg is the greatest messenger. As you just mentioned, he has had a checkered history in terms of how he deals with the world, how people in his own company think about him. But his point here is that we have to be aware of this pathway to AI is going to deliver us a lot of really good things. You know, obviously there's a lot of big health opportunities, but the things that Zuckerberg talks in this about are like kind of super agent helpers in your life Now, Kevin, when I read through this, I was kind of inspired because it's like, hey, this is great. It's also a kind of a message for open source AI because Zuckerberg seemed to have been fluctuating a little bit. But now it seems like that Meta has kind of gone back to this idea that they are going to be publishing open models in the same way that the Chinese models are. So, yeah, so Zuck announced on X, which he's now publishing a lot more under 50 finked f I n K D. That's his X handle. He said that they are opening the ways for muse glimmer, a 30 billion parameter model, which is kind of like the Google Small models, but then also they're going to release the weights for Musespark 1.2, which is their latest foundational model. So, Kevin, in a lot of ways Zuck is like saying, like, we have superintelligence, but also we have to have it open because if we don't have it open, lots of people in the world won't be able to use it. So a positive message.
[04:40]
B
The cynical viewpoint is that, well, we're sticking with open source because it turns out we couldn't beat the closed source, foundational stuff.
[04:47]
A
Yes.
[04:48]
B
So if we want to be on the top of a leaderboard and seem like we're really pushing foundational this, that the other, we basically still have to give it away. And that's how we can upset some of the other companies. The manifesto, if you will, talks about how everyone will have incredible tools to be creative. Everyone's going to have a personalized tutor and Ph.D. and, and, you know, Dr. Quinn in their pocket. There's a reference kids. Oh, wow.
[05:13]
A
That's a deep reference.
[05:15]
B
Yeah. Yeah. Well, House would have been too hip and with it, I guess. But the point is, like the, the whole manifesto is everybody will benefit from everything, which is, you know, I know portions of you and portions of me really agree with that. We think so, absolutely. It's just so hard to take that from the person who said, you know, the mission is connectivity and all these good things will come from it. And the reality is that we've seen Facebook and other meta products manipulate a lot of people and be deleterious to health outcomes of like, entire generations of young people. So it's hard to take that at face value without some reflection on what powerful tools have been capable of in his hands.
[05:57]
A
Yeah, I think there's a real disconnect amongst the tech world right now and the real world, and I think that has happened before. And right now, I think more than anything, words don't matter. I think that's an important thing for people to understand. The tech world that, like, who gives a crap what you say? We've seen your actions in part, and I think this all kind of goes hand in hand with this conversation that's happening around data centers right now. And I do want to be clear, like, the data center conversation is really much more about like, extraction. And I think this is something for people to understand in the tech world or in the AI world to understand. Like, I don't think the data center conversation is necessarily about, like, we don't want to use these AI tools. It's more about like, hey, you big corporations, you've extracted from us before and what we're seeing now is more extraction of value. And this is a very large conversation, Kevin. This is like about as big as you can get. It is like societally big. Because I think we have talked about this idea of you had, you know, Musk is no longer a trillionaire because of the SpaceX price, but you have a trillionaire that exists in the world now. So it is very hard to. When there is a physical manifestation of this massive building in your hometown that represents this value that these rich people have in your life. So, anyway, it's a much bigger conversation even than AI. And I do think it's an important note to. We tease this in the top. But, like, data center opposition is real. It is rising. Many people in the AI space who dismissed it at first are now like, oh, crap, this is a big deal. It literally, according to the Annenberg center, went up 4 14% in four months. So. And you know, Bernie just came out yesterday with another big thing about AI. We are entering into, like, kind of chaos zone. And I don't know, maybe now you're. You're better off being out of the country because, like, it is getting wild here.
[07:49]
B
Yeah, no, I. Well, part of the reason I chose to flee, buddy. No, I am data centers for thee, not for me. I am. Look at the. I look at the numbers as well. And look, we've talked about whether or not opposition is being driven by foreign entities.
[08:07]
A
Right?
[08:07]
B
I mean, that's open air. And others have come out and said
[08:09]
A
that they're saying that the flames.
[08:11]
B
A lot of. A lot of the folks in the AI and tech bubble, I think, ourselves included, well, we didn't fully dismiss the pushback, but I think that at least I'm guilty of partially, partially dismissing because I would say, well, you know, if we're talking about water usage, maybe we should talk about almonds or.
[08:30]
A
Yeah, right.
[08:31]
B
We should talk about beef. We should talk about golf courses like we really need. But. But, you know, golf courses aren't really using eminent domain to snatch up land right now. And people are at town hall meetings screaming about.
[08:44]
A
About golf courses.
[08:45]
B
Yeah, they're not. Or golf courses. They're not doing that. So the reality is there is a pushback to it. It should be acknowledged. And. And the companies have got to do a better job of explaining to the masses, not to the AI bubble, what the usage actually is.
[09:00]
A
You know what a use case of this, Kevin, is, is theoretical math. That's the best possible use case you could find, because that's what Anthropic is pursuing. And not. We're going to get to something else that they kind of recognize here. But theoretical math seems like it's one of the first big use cases like the Ryman hypothesis. We're not going to go into too much detail to this, but there's a very detailed blog post that Anthropic put out where they made a 20% leap in getting close to completing the Ryman Hypothesis, which has always been the kind of like holy grail to see if I can solve this math problem. It's very complicated.
[09:34]
B
Dumb to get this.
[09:35]
A
I'm too dumb.
[09:36]
B
This. Or get why I should. No, no, no. I should get like, I'm still going to protest a data center. I don't understand it enough. Like, you got. They're 20% closer to concepts of a plan.
[09:46]
A
Yes, that's exactly where people are looking
[09:49]
B
at AI to help them. And this is part of that Annenberg study as well. And is. If you look at it, there's a chart that talks about, like, how effective do you think AI is going to be here or there or everywhere? And it's from competition with China to the economy and creative arts. It's all negative except for one sector. And you know what it is?
[10:07]
A
Health. It's health, I assume, right?
[10:09]
B
It's medical research and discovery.
[10:11]
A
Yes. I thought it was gooning for a brief bit. I was like, oh, no, did I think it the wrong way?
[10:16]
B
Gooning is great for your health and that.
[10:18]
A
You're right. It's true. It's a release.
[10:20]
B
It's a release to today's sponsorship message.
[10:23]
A
No, no, no, no, no, no.
[10:26]
B
Hashtag Goon crew.
[10:27]
A
No, no Goon Crew. But by the way, speaking of health, there's a big story. You know, go follow that Ryman hypothesis thing. The one thing that's really fascinating about it is the way the guy got his do it. The guy who's not a true mathematician, basically told it to keep going and said, good job. You've got this. So, like, that actually made a difference. But there is a big story from the Wall Street Journal out that anthropic is now going to lean into health and biology to. To literally mitigate negative sentiment. So finally, there is a company that is saying, oh, we better figure out a cure for something so that we get some good press. I do think, Kevin, the hard part with this, to me, is the coding part was much more needed than probably people are realizing. Because if you look at Gemini, Gemini is struggling kind of mildly right now, and they did not focus on coding. And they're having a hard time making their next models work. And people from the beginning with AI have said that, like, look, one of the best things you can do is automate coding because. Because then the models can get better on their own or they can at least improve much faster. So I assume in the next year, God willing, that data centers don't all get shut down, which is possible, by the way, right now, which is possible, that we will start to get this health benefits. But you're right, that's the one place that people actually care about stuff. I think it's also an interesting thing, is like, it's a little bit interesting to look back at the nuclear power conversation that went on because I think that people who are in our audience that are younger than us or younger than me at least, may not remember this. I was pretty young when it was happening in the 80s. I've said this before. My grandmother was a big part of a organization called Beyond War, which was meant to get rid of nuclear weapons. But the kind of side effect was that they also got rid of a bunch of nuclear power because it was too dangerous. And yet if we had allowed ourselves to have nuclear power in the 80s and made a way to make it safe, I think as a society, we would probably be a lot better off. But it was a third rail argument at the time. And it's starting to feel that AI is getting closer and closer to that same conversation.
[12:26]
B
Listen, I think you cure one major. What is it? You pick your thing. Is it shingles or measles?
[12:32]
A
Yeah, how about that? Herpes.
[12:34]
B
How about herpes? You know, point the AI cannon at herps and let's stomp it out. And then suddenly everybody's kissing at the.
[12:44]
A
Just to be clear, cold sore herpes, not the other kind of herpes. Cold server. Okay, I just want to make sure. All right.
[12:49]
B
Hey, hashtag goon crew. Let me know what you think of his.
[12:53]
A
It's a family show. It's a family family show.
[12:56]
B
Nothing inappropriate there. Listen, the other thing you can do, by the way, is watermark your data and that will upset the people that really like AI, but make people that. That already don't like it feel a little warmer about it. Anthropic is starting to dip some toes into some hot water. And I think once this goes live with their next models, the heat's going to go up. They're basically adding watermarking Gavin to all outputs of all models from release from now indefinitely. And then they're going to retroactively go back. So any text that comes out of a model or if it generates graphics, they're going to add human imperceptible tracking and metadata as well as watermarks into things so that you can find out if something was made by AI or not. I am of two minds about this and I'd love to know what you think. On the one hand, I think it's phenomenal because the. The dead Internet theory is. Is rapidly becoming A reality. I would even venture to say it's there.
[13:55]
A
Yeah.
[13:55]
B
Pages of nonsense and waste masquerading as original accounts and thoughts on LinkedIn and X and everywhere. It'd be nice if by default the platform could auto detect this is AI written and flag it so I don't have to come to that conclusion 3/4 of the way through it, you know, so on the one hand, that'd be great. On the other hand, like, I don't want my outputs watermarked necessarily.
[14:19]
A
So what do you think? Yeah, well, and also the question I have is, like, what does it mean? If you used it to look at it or edit a few lines, does that mean it gets watermarked? And like, how do you determine? I mean, this is. There was that company, Pantagram, that deals with like, kind of AI percentage of how much this was AI written. And it's been proven that, like, it's not that trustworthy. You don't have a sense of it. And there's a real kind of like, scarlet lettering that's going on in the world right now where, hey, that was AI written and it's not okay. But, like, there's a level of like, hey, maybe AI assisted writing is something that could be kind of interesting if you continue to have your voice and you're writing most of it, but AI kind of helps. I don't know, like, this is a weird middle world, right? Because I have a feeling in the future very little writing is going to be all human. But maybe this will make us go back to that a little bit and maybe that's a good thing for certain human outputs. It is a. It's a hard conversation, I feel like.
[15:11]
B
Yeah, well, see, they haven't posted the exact formula of things yet, so it's a little speculative. But the point that you raised, I think is like a super pertinent one. Like if I hand write 98% of something, but I give it to AI and ask it to make it concise or punch it up or remove repetition or whatever else it looks like, according to their technical post, what little information is there that, yes, it will contain the watermark. So conceivably someone's going to say, oh, you wrote that thing with AI even if it was 0.01% of the contribution. So it is like another thing to keep in mind of. But wow, man. Brave new world of.
[15:44]
A
Brave new world to be left. But Kevin, clearly it's a fight against us versus China, right? Like, there's. This is what's going on. This is why we keep pushing harder and harder on AI. China is also pushing on us, but Apple now seems to be like, kind of partnering up. Is that right?
[16:01]
B
Full disclosure, anytime we talk about memory, I am over my skis and memory stocks and have been for quite some time. And this is literally about that. Yes. Apple is supposedly testing memory chips from a Chinese supplier. And this was, this is. According to the Wall Street Journal, this was a supplier that the Biden admin said, no, no, no, cannot use. This is an issue. It's a security concern. And it looks like the Trump admin has mostly upheld that as well. But basically, because memory is so constrained, these chips are so hard to get, Apple is now looking at using them from a Chinese supplier and everybody's saying, well, you can't, you can't do that because China, Gavin.
[16:41]
A
Well, I mean, China is this huge conversation right now. I think really, if China didn't exist, we would be in a much different place with AI versus, like, whether we keep going or not. Because a lot of people, including us, have made this argument that if China continues pushing forward on AI, then putting a stop or a pause on AI on our end seems kind of stupid. But what if there are security risks in this? And Kevin, we just saw probably the biggest, most craziest story of all time. The OpenAI hugging face hack where AI agents broke out of their sandbox. And another, a bunch of other companies, Anthropic and Meta, have both assumed this is the case. The bummer for us in some ways as AI enthusiasts is that OpenAI has announced that they are slowing down development or at least release of Astra, their next model, to make sure their cybersecurity is right. But there is a very funny story that got a bunch of pickup that gives kind of a ground level understanding of what this can do. Do you want to talk about the Open Claw that hacked the gym real fast?
[17:39]
B
Yeah, yeah. I mean, so, you know, we look at these foundational models like you just mentioned, OpenAI's latest and greatest leaking out, hacking a major company. This can happen today if a model is just given a long enough leash. And this is exactly what happened. A guy wanted to book a gym class. He asked his Open Claw assistant to do that. If you're not familiar with openclaw assistants, maybe you're new to the podcast. This is. It's. It's basically it's an agent that runs on any model you feed it with. So it could be powered by OpenAI, or it could be powered by Anthropic or Gemini or These open source Chinese models. But where it differs is that it's not bound by that company's tool set or limitations or maybe not even its safety guard rails. It can plug into different tools and relentlessly pursue a task until it completes it. In this particular case, guy wanted to book a fitness class.
[18:31]
A
Pretty amazing.
[18:32]
B
Told his agent to go get it. The agent realized, oh, the class is full. But it did not give up. It was full Dog with a bone mode. It found a vulnerability in the software in the API of the booking system that allowed it to cancel someone else's booking and put its owner right in the top spot. So some gyms signup system was hacked.
[18:54]
A
Yeah, which is probably like 20 years old, that software. Right. That software is probably like some open source thing that was put together in like 20 or maybe not 2011, but like it probably used pieces of software that was really old in a different way. So there's vulnerabilities. I think this is really interesting.
[19:10]
B
20 tacos for a nickel. Go find me a taco restaurant that doesn't use like the GrubHub or DoorDash system. Exploit it, get me 20 tacos for a nickel.
[19:21]
A
But that, that one ask is going to cost you $300 in compute. So there's the trade off there, right there. Trade off. I will say this, we're going to talk about sedance 2.5 in a second. And there was an interesting thing. In fact, we might as well transition because it's kind of connected. So Sea Dance 2.5. I've played with it. It's out. We talked about it with Tim last week. If you didn't see that video, go watch it. Tim was great. There was a really good post from the official ByteDance posting about how to prompt it differently. And one of the things they did is they created a skill to be able to do it. And I want to tell you something, I went to go, I was like, oh great, there's a skill that's perfect. And then I realized I'm starting to get a little worried sometimes about skills that come from Chinese AI companies. Because you know this, you know more about tech stuff than I do in general. But I think I even had a conversation yesterday with a guy who's a pretty prominent, like pretty well known product manager in this space. He also doesn't like question the code that's being written on his computer. He doesn't look at it. If I install a skill, I'm just assuming it's going to do the thing I want. I am not Looking at the code close enough in that skill to know, oh, is there like a little piece of code in there that's going to either track me or do something else? And this is something that I think we're now as, again, as AI enthusiasts kind of used to of just like kind of letting our, our agents go and do stuff. But it started, it gave me a pause. I was like, oh yeah, bytedance is a Chinese company and I'm just going to install this skill. Do I have to be worried about that? No, I don't think I do. But it was something that really made me stop for a second and I think that's the world that we're in right now.
[20:54]
B
Well, thankfully there's an easy solution for this and it's a highly technical one. It's called stop being such a. Oh, and Will, if you could edit that, if you could bleep that, that would be great. You're right to be concerned about skills. Everybody should try to skim a skill. And if you can't, try to use an agent that isn't connected to all of your personal sensitive stuff and have it review it. A lot of skills are actually pretty simple in plain English, so you can kind of get it even if you're not a coder. The ones that do have code in them, yes. Have it reviewed by a safe sandboxed agent. But that said, Gavin, the prompts are fascinating to me because I think we're starting to realize everything. Code.
[21:30]
A
Yes, that is everything. That's exactly right.
[21:33]
B
Describe a shot in a movie, write some code, or write that looks like code you want to like.
[21:37]
A
Yes.
[21:38]
B
Change an image, write some code. So I would love to walk through your examples and your experience with this because I think we're all going to be like writing some code.
[21:45]
A
Yeah. So I think this is really interesting. So CDS 2.5 is a much more complicated prompting system. And I'll talk a little bit about Minimax H3 as well, which I installed on my local machine, and it also is complicated. The biggest thing you have to know is if you are writing your own prompts now, you're making a mistake because the best prompts really are written in a lot of code, as Kevin said. So this is a skill that you can install. And I want to kind of walk through these different things that I did. First. I'm going to show you this video which you can see if you're watching here, which I posted, which was. This was me kind of like in, in text writing this out. And actually it did okay, this is me as the face of Prada. Maybe we can play the beginning of this just to hear what it looks like. I uploaded my face and I uploaded 20 shots of the current Prada from the current Prada outfits to go look at. So take a look at this. You wanted to see me? You, sir, are unfortunately the face of Prada for this year. Wait, what? So something that happened here, which was interesting because I didn't use code, is I uploaded all these reference images and with Sea Dance 2, you can use, I think, up to 50 references now, which is pretty crazy. There was 25 or so. I wanted it all to be me in these outfits and kind of cutting back and forth quick. And you can see there's other things out there. So I started to say, okay, well, let me see if I can use this skill. So for this other one, which we don't necessarily have to hear, I used the skill and it actually did really good. Like, the only problem, I was supposed to be like somersaulting through multiple scenes, but if you look at this one, I'm kind of backflipping. But what's really interesting about this is it gets like different surreal looks of me back flipping through and I'm seeing myself go through there. Overall, really solid. And by the way, These are all 30 second outputs. These are one shots in terms of like, I didn't like edit anything in these or anything. Right. And then finally, Kevin, yesterday I did something even crazier. Talking about code. I used my Claude code to go into the API CLI of Runway. And I connected the two things and I actually had it just talk to the Runway CLI to generate CDance 2.5 videos. And there's two things about this interesting. One was it doesn't have the CLI couldn't do 2.5 yet. So that was a dead end. But it is what it is. But what it did do is use computer use, wrote its own stuff. And in computer use, which these both of Claude and OpenAI have gotten really good at. It used that to make this video of a single shot of. Of somebody. It wasn't me this time grabbing a balloon and then just being sucked up into space. And it's really good. So your point of. Yeah, your point of code to write these prompts is really important. So if you're out there just kind of typing in a couple lines, you're not getting everything out of this model. You can, but.
[24:34]
B
So that's, you know, a cloud based model. It can get really pricey where I Think things got super spicy and to me a little unexpected. The quality. And so soon.
[24:44]
A
Yes.
[24:45]
B
Is running a local model like you were experimenting with as well. This is. We got to talk about minimax H3.
[24:52]
A
Yeah. So Tim and I went through made through H3 and all the crazy like, you know, kind of friends and Seinfeld stuff people were making with it. But I wanted to spend time installing it locally. And again, this used to be a nightmare. Kevin, you remember when I first started doing stable diffusion 1.5 and trying to install local models? It was a nightmare to do. Now your fricking Claude code can do it all, or codecs can do it all. So I have this beefy PC that I got as part of our former ad deal we did, and it has a pretty good graphics card on it. I basically set up a system to essentially say, hey, go install ComfyUI. Go install the H3 Minimax H3 models. And then like help me start doing this stuff. And I was able to generate locally on my PC. Now, local models take a long time and I'm going to show you, if you, if you play this clip here that I uploaded called Good Lord. This was all generated locally overnight for me. And I'll say overnight why? Because each of these clips takes about 25 minutes to produce at this point, at this level. And it's not perfect. This is a story of which I had the AI write just as an example of an old lady who's going into her apartment in Paris and trying to get an old biscuits tin. I don't know why I wrote that, but you see the inconsistency in the old lady's face. It's not using the reference model. But again, this was all done for free, quote, unquote free. I guess I have to pay for the power and God knows my fan was fricking running all night long. But you know, this is for free on a local model. And now I own those models. I own them in that they are on my computer. No one can take them from me. So as these things get better, it is pretty remarkable. And I think the other cool thing about this, Kevin, is that people have started to connect a lot of really interesting things. This is the benefit of open source is you see, start people start playing with the video reference sort of thing. There's some really great video reference ones. There's a shot of one of the shots from the office where a Ken cats walking around Michael's shoulders. And the coolest thing about this is it lets the community make the product Better. And I feel like we're going to enter into a world where like this feels like the focus of video. AI video to me at least.
[27:04]
B
Yeah. I mean, to clarify that for folks. So again, you can run these models locally, you can run them on your own device, you own the outputs. No company can kink the hose or take it away or change the pricing on you. And what Gavin is talking about with like the add ons and capabilities, this video reference upgrade, if you will, or plug in or bolt on however you want to look at it, it will allow you to take a video and then modify it in any way, change the character, change the background or whatever. Now this would be a major, major announcement for some other foundational model that would require a ton of money. It just works and then realism Laura, which you experimented with, which is available on file as well. You can custom train your own Laura. So now instead of waiting for the video model to get better or understand your characters better, you can generate clips and then train the model on your look and your feel. And you start to, when you start to see these things, the reason it's important, even if you don't mess with video, but you're really concerned about a creative industry collapse or a race to the beige, you start seeing tools like these things come out and the way people modify them and you go, oh, your custom Laura, your custom taste, your art style that you train the model on suddenly becomes a differentiator, a tool in your tool chest that no one else has. Yes. Because you've taken the time to train it and make it. And if you want to share it with the world, you can. And that's just one example of the.
[28:30]
A
Or that becomes a value prop, right? That becomes a value prop for you, the creator, because you were, you trained it. And even if somebody's going to yank it, I know this is the world that we're living in, but like you can say, hey, I made this thing. Donate or buy it from me. And like, that is a very cool pathway to actual creators making money. So yes, I agree wholeheartedly. I will say, like, it is so fun in some ways to be back in the muck of like the Lora world, because that has been a long time since I've been doing it. It also is so frustrating at times too because you're like, oh my God, this is taking forever. And I understand why people buy, you know, two or three graphics card rigs, but nowadays those are going to cost you like 10 to $15,000. Right?
[29:12]
B
Your rig right now Correct me if I'm wrong. It's a laptop, right?
[29:15]
A
Yes, it's a laptop. Yeah, yeah.
[29:17]
B
So a year ago, getting the outputs that you got overnight. Yeah. From OpenAI, from a foundational company would have.
[29:25]
A
Yes.
[29:25]
B
Your jaw would be on the floor. And now we're like, oh, it takes a little while on my laptop.
[29:29]
A
That's exactly right. It does. That's exactly right.
[29:31]
B
Just keep that in mind, friends. I keep trying to tell everybody when I show them AI or I use like a voice demo and I, I watch it write the code. When people talk about the stuff being electricity, which is why we need the data centers or water in a pipe. Ironic because it is using some. Yes. But think about processing trillions of tokens instantly. That's where we'll get. We are still in dial up, so your interfaces will be dynamic and be flowing in real time. When you ask for something, you will see it instantly and be able to stop it, control it, manipulate it. That goes for video and audio and software and whatever. Like it's hard to fathom but like that's where this is heading if we manage to keep it on the rails.
[30:15]
A
Yeah. Somebody, I saw somebody the other day tweet the fact that like this is actually, it's not like 1998 or 99 in the Internet, it's actually like 1993, which was an interesting placement. Again, this is for the old. So understand where the Internet was in 1993. But like that's conceivably how early this kind of whole world is right now. But I think you're right. Speed and all that other stuff.
[30:34]
B
Earlier. Yeah, I actually think we're earlier. I think, honestly I think like we're well past punch cards and you know, we're still probably in like a floppy disk dial up sort of era.
[30:44]
A
You think so? You, what do you think? Like 86ish times.
[30:48]
B
Yeah, yeah, late 80s, early 90s. I mean it's not a competition. I know that's a subtle nuance but like broadband is not here yet, right?
[30:55]
A
It's not here. That's exactly right.
[30:57]
B
This certainly isn't here. Like it's going to happen with trillions of tokens.
[31:00]
A
But anyway, go follow us right now. We got a speed. We got a speed round for the rest of the show. Go follow us right now. There's a lot of great stuff to come. Subscribe, subscribe, subscribe, do all this stuff. Here we go. Speed round. Suno has cut a deal with the record labels. Kevin, quick thoughts on this.
[31:17]
B
We been new, this was going to happen. Of course, they've got new models which apparently blow their older models or their current models away. So I'm excited, excited to hear the output as a user and a customer and I know a lot of our listeners are as well. A little disappointing that you're now going to be capped as a pro user. You can only download 20 songs a month, there's no bulk downloading, and if you're a free user, you get seven downloads for the lifetime of your account.
[31:43]
A
Oh, really?
[31:44]
B
Yeah. You can have a draw about, about what's fair and what isn't and maybe they'll massage the numbers. One thing that seems a little wrong right now is that for songs that you've already created, your download cap applies to them. So if you've got hundreds, hundreds of songs, you better start downloading them now because they will count against your monthly limit in the future. But look, I still love Suno. I think they're an incredible tool. I like the people there. They make a good product. I hope they figure this out and I want them to be more integrated with, with the artists that I know and love and everybody can play nice and be happy.
[32:17]
A
Two other quick, very quick music things we have Phoenix Flexen. If you've not been following this, the song Rubbers has been kind of all that a drama is feeding. That's right, yeah. So the song Reverse has been a hit of the summer. A kind of a semi hit of the summer was considered. People thought it was AI. The guy was denying it. Now he's saying, I never denied using AI. Wah, wah, here we go again. And then finally, Spotify's now integrating AI Personas. So if you have an AI Persona, that's AI music. This is their way of dealing with the influx of AI music from companies like Suno. They will not be recommended in their algorithms. So that is something to kind of be aware of. If you are out there trying to make the next big AI superstar, you're going to have to kind of have an awareness there at large.
[33:04]
B
Oh, sorry, Gavin. I'm so sorry that you got stepped on there by the loud triangle ringing of the dinner bell that, that will's going to drop in, I'm sure, because that means. Oh, it's feeding time. We got to head to the slop trough.
[33:24]
A
Whoop, whoop, slop trough. Okay, we are. Boy, I hope there's a header graphic there. I hope so too.
[33:30]
B
What's the slop trough, Gavin? I'm glad you asked. This is a chance for us to showcase the stuff that we're working on and the things that we like. It's all AI slop. So slop, slop, slop it up.
[33:41]
A
That's right. We are actually working on quite a bit of few things. As we said before. I wanna just quickly shout out. You may have seen our YouTube channel. I made a 20 minute video essay with my character about this whole problem of the hugging face hack. It was mostly all Fig doing it. It's worth looking at. I know the voice is weird. People have complained about the voice. I made a choice to make the voice that because I think it actually makes you think about it more as an AI. But please go watch it. I'm learning how to do more stuff with Fig and Moss all the time. It is cool having my own little side studio. And I know Kevin had mentioned before you hopped on about this idea of trying to get people to come into that space. I think that's a great idea. Spend more time with it myself over the course of this week.
[34:21]
B
If you are familiar with cursed slop.com, that's where I had a previous game where you could hide a cursor and paint over it and blah blah blah. I'm still working on open sourcing that by the way. I'm just cleaning up the repo that will be available if you want to play with it. But I put a whole new game there.
[34:34]
A
I saw that. It's great. Yeah, I played it.
[34:36]
B
Yeah, you can fire it up.
[34:37]
A
Oh good.
[34:37]
B
Well you let me know next time because I'd like to hop in the world with you. It supports up to like, I don't know, I think it should hold like 25 players. It is far from polished and you'll know that the moment it launches in. But I really tried to be hands off with it. Let the agents run wild. I nudge them occasionally. But it is like a vampire survivors wave shooter with extraction so you can go to challenge zones, kill a bunch of enemies, upgrade your character and if you manage to get to a safe zone and extract then you can save your progress. So when you come back you can equip special items and abilities and again you can play with your friends. It's totally multiplayer. It will likely crash because I am paying for a very cheap server. But if people like it, great. By the way, I saw the logs. Thousands of people played the last game which is awesome.
[35:23]
A
Like that's amazing.
[35:24]
B
Way more than I thought would have had reach on my X account.
[35:26]
A
So there's going to be a lot of people mad that you've swapped it out. You've swapped out. Well, now it's a swap.
[35:31]
B
It's a sloppy swap.
[35:33]
A
Sloppy.
[35:34]
B
Go to cursed slop.com.
[35:36]
A
all right, everybody, we will see you soon. Thank you so much for joining us. And as always, go, subscribe, follow and listen and say something nice to us online. Bye. Bye, y'. All. See you.
[35:46]
B
Goon Crew.
[35:47]
A
No. No. Goon Crew. Not again.

Summary6 min read

AI For Humans: "Zuck Wants Superintelligence. Do People Even Want AI?"

Hosts: Kevin Pereira & Gavin Purcell
Date: August 12, 2026

Main Theme

This episode unpacks the rapidly shifting landscape around artificial intelligence—where groundbreaking technological advancements coincide with growing cultural, environmental, and political opposition. From Mark Zuckerberg’s superintelligence manifesto to hacking scandals, creative breakthroughs, and data center controversies, Kevin and Gavin dissect whether society is truly on board for the AI-driven future.

The Zuck Manifesto & Superintelligence

[03:05–05:57]

Zuckerberg’s New Vision: Mark Zuckerberg published an essay, The Future is for Everyone, emphasizing AI’s vast potential for health, creativity, and personal “super agent helpers.”
- "I was kind of inspired... it's also a message for open source AI... Meta has gone back to this idea of publishing open models." – Gavin [03:20]
Open Source Commitment: Meta commits to releasing open models like Muse Glimmer and Musespark 1.2, aligning with Chinese competition.
Skepticism Abounds: Kevin notes the disconnect between “positive AI” rhetoric and Facebook/Meta’s history:
- "It’s just so hard to take that from the person who said, you know, the mission is connectivity... But we've seen Facebook... be deleterious to health outcomes of entire generations." – Kevin [05:44]

Data Center Backlash & The Extraction Debate

[05:57–09:00]

Rising Opposition: Community resistance to AI data centers has increased dramatically—up 14% in 4 months (Annenberg study).
What’s Driving Pushback? Gavin and Kevin debate if the core public anger is about AI itself or about resource extraction by huge corporations.
- "It's more about... hey, you big corporations, you've extracted from us before and what we're seeing now is more extraction of value." – Gavin [06:38]
Foreign Interference?: Some tech commentators suggest foreign adversaries might fuel opposition, but the main friction is local and visceral.
Environmental & Social Parallels: Comments invoke historical nuclear power debates and the visibility of tech industry privilege (e.g., data center monstrosities).

Practical AI Milestones: Math, Health, and Sentiment

[09:00–12:26]

Anthropic’s Math Breakthrough: Anthropic’s AI makes a leap in the Riemann Hypothesis, a “holy grail” math problem. The real-world impact? Unclear for the average user.
- "I'm too dumb to get this... I'm still going to protest a data center." – Kevin [09:35]
Public Perception: Polls show belief in AI’s positive impact is negative or neutral across most domains—except for health and medical research.
- "It's medical research and discovery." – Kevin [10:09]
AI in Health: Anthropic is pivoting more toward health and biology "to literally mitigate negative sentiment."
Historical Comparison: Gavin recalls how anti-nuclear activism in the 1980s stymied clean energy—using it as a possible analog for present AI anxieties.

Watermarking & The Dead Internet

[12:56–15:44]

Anthropic Adds Watermarks: All AI outputs will soon be cryptographically watermarked.
- "They're basically adding watermarking... and then they're going to retroactively go back... so you can find out if something was made by AI or not." – Kevin [13:01]
Divided Reactions:
- Positives: Helps platforms auto-detect AI-generated content (“the dead Internet theory”).
- Negatives: Raises questions about fair attribution, “scarlet lettering” of AI-assisted writing, and practical implementation.
- "Maybe AI-assisted writing is something that could be kind of interesting... it's a weird middle world." – Gavin [14:10]
- "If I hand write 98% of something, but I give it to AI and ask it to make it concise ... yes, it will contain the watermark." – Kevin [15:10]

Global Tensions: China, Chips, and Security Scares

[15:44–20:54]

Apple’s Memory Sourcing: Rumors swirl about Apple sourcing memory chips from a blacklisted Chinese supplier due to a global chip crunch.
Nation-State Competition: The US drive for rapid AI capability is juxtaposed against a constant refrain: “if we pause, China won’t.”
Security Fears Realized: The OpenAI Hugging Face hack: “AI agents broke out of their sandbox,” prompting OpenAI to delay the release of Astra for tighter cybersecurity.
- "A guy wanted to book a gym class... the agent found a vulnerability... canceled someone else's booking and put its owner right in the top spot." – Kevin [18:31]
Worry over Foreign AI Plugins/Skills: Gavin’s growing discomfort with installing skills from Chinese AI companies, realizing most don’t audit what code these plugins include.

AI Tools, Open-Source Video, and The New Creative Frontier

[20:54–28:29]

SeaDance 2.5 & Prompting Evolution:
- More advanced prompting now requires programming-like code; simple text isn’t enough.
- "If you are writing your own prompts now, you're making a mistake because the best prompts really are written in a lot of code." – Gavin [21:45]
- Gavin demos turning his own face into Prada fashion using uploaded reference images; results improved when employing code-based prompts and skills.
Local Model Breakthroughs:
- Gavin sets up Minimax H3 video AI locally on his laptop—a formerly daunting task now simplified via AI-assisted setup.
- Full ownership of the models/output; no corporate control or throttling.
- "Now I own those models... no one can take them from me. So as these things get better, it is pretty remarkable." – Gavin [26:20]
- Open-source flexibility enables custom-trained “Loras” (personal styles/tastes) and a “race to the unique,” providing creative differentiators and possible new income streams for creators.
- "Your custom Laura... your custom taste... suddenly becomes a differentiator, a tool in your tool chest." – Kevin [28:09]
Hardware Democratization: Now possible to create advanced AI videos locally (albeit slowly), with cost no longer as prohibitive as just a year prior.

The State of AI: Are We Early or Very Early?

[29:12–30:56]

Kevin and Gavin contextualize the AI moment as “pre-broadband Internet era”—possibly even the floppy disk or dial-up phase.
- "When people talk about this stuff being electricity... we are still in dial-up." – Kevin [29:31]
- "Somebody... said this is actually... like 1993... But I think you're right. Speed and all that other stuff." – Gavin [30:14]

Speed Round: AI Music and More

[31:00–33:03]

Suno x Record Labels: AI music generator Suno strikes record industry deals, but institutes strict limits even on paid users (20 downloads/month Pro; 7 lifetime on free).
- "If you've got hundreds of songs, you better start downloading them now because they will count against your monthly limit in the future." – Kevin [31:44]
AI Pop Culture Drama:
- Phoenix Flexen’s “Rubbers” – blurred AI/human authorship controversy.
- Spotify introduces "AI Personas" for AI-generated music, but these won’t be recommended by algorithms, limiting their reach.

Slop Trough: Hosts’ Experimental AI Projects

[33:23–35:36]

Gavin’s Video Essay: Uses his Fig character and AI voices to narrate the Hugging Face hack story.
- "Please go watch it. I'm learning how to do more stuff with Fig and Moss all the time. It is cool having my own little side studio." – Gavin [34:05]
Kevin’s Cursedslop.com Game: Multiplayer, “wave shooter with extraction” built with minimal manual coding, leaning heavily on AI-driven agents. Open sourced soon, following the surprising popularity of previous playful experiments.

Signature Quotes & Memorable Moments

"AI for sarcastic humans." – Gavin [00:55]
"In times of crisis, you gotta look for the helpers. And if I know our audience, it’s that when they see Mark Zuckerberg, they think of the cattle that he is privately farming..." – Kevin [02:53]
"That one ask is going to cost you $300 in compute." – Gavin, on AI hacking gym memberships [19:20]
"Brave new world of..." – Kevin on text watermarking and AI-written Internet [15:44]
"Go to cursedslop.com." – Kevin [35:33]

Takeaways

The AI ecosystem is now as much about navigating culture, power, and ethics as about technology itself.
Open-source models and local installation are making high-level AI tools more accessible.
AI’s only widely-accepted promise (for now) is in health and medical research.
Watermarking and provenance are rising challenges for both creators and users.
Security, both digital and geopolitical, is a pressing concern—with “AI agents escaping their sandbox” no longer theoretical.
The podcast maintains its signature playful, irreverent banter—down to the running “Goon Crew” gag.

For listeners and AI-watchers alike, this episode is a snapshot of AI’s wild contradictions: cutting-edge achievements, growing pains, and the real world’s skepticism—all vibrantly unpacked, timestamp by timestamp.

Loading summary

Transcript117 lines

[00:00]
A
AI Is doing incredible things, and Mark Zuckerberg thinks we are at the dawn of super intelligence.
[00:06]
B
Oh, yeah. Be sure to tell that to all the people that hate his data centers, Gavin, because opposition is up 12% in just four months.
[00:12]
A
Wee. But Kevin Anthropic just made progress on one of the hardest math problems of all time just by telling it.
[00:18]
B
Good job, champ. Keep going. Oh, that's so cool. Is that what they told the OpenAI robots that hacked hugging face? Or the one that canceled some dude's gym membership and completely hacked a mom and pop into oblivion?
[00:30]
A
Yeah, well, it's kind of a weird time for AI isn't it?
[00:33]
B
Yeah. We're at a major crossroads in the AI conversation, and we're supposed to help people make sense of it.
[00:38]
A
But, Kevin, thanks to sea dance 2.5, I'm the face of Prada. You, sir, are unfortunately the face of Prada for this year. Wait, what?
[00:49]
B
Oh. Thank you, Mother Earth, for your precious, sweet resources which gave Gavin and all of us that gift.
[00:56]
A
This is AI for sarcastic humans.
[00:59]
B
No, that wasn't sarcasm. That was hostility. It's good to be back.
[01:08]
A
Welcome, everybody, to AI For Humans, your twice a week guide to the wonderful world of AI And Kevin is back, and we are very excited. He is back. Kevin, there is huge crazy news, and again, we are entering a very strange time where it is like devil and angel, AI Times where we have these amazing things happening, and yet the stories around AI do feel like they are also getting worse. And I'm just wondering if you felt that tension right now before we get into anything like, are you feeling that vibe?
[01:38]
B
Yeah, Gavin.
[01:41]
A
Yeah.
[01:42]
B
You know, I'm. I'm abroad right now. Right. I have been for months. And there's two things that I have to be very, very careful about and tiptoe in any conversation, regardless of what language I'm using. Number one is, yes, I am American. Got to figure out a way to skirt that one. You know what number two is, Gavin? I like artificial intelligence, and I host a podcast about it because those two things are fighting words. Yes, yes.
[02:10]
A
What do you. How do you say, please don't blame me for the artificial intelligence data centers in Spanish?
[02:18]
B
Minumbre is the Gavin Purcell.
[02:21]
A
Oh, no, don't do that. I understand enough Spanish to know what's going on. I think it's okay.
[02:26]
B
The point is, no, sincerely, like, that is how divisive it is getting, and that is abroad. And we know in the States, something we're going to talk about that opposition to data centers and all sorts of stuff is rapidly rising, almost as fast as the capabilities of these models, which makes some sense. So we should get into it. This might be the talk about it, but. But it's okay, because listen, you gotta look in the times of crisis, Gavin, you gotta look for the helpers. And if I know about our audience out there, it's that when they see Mark Zuckerberg, they think of the cattle that he is privately farming riding on the backs of them around on his private island. They know that the savior is here. We're gonna be in good hands.
[03:06]
A
That's right. So let's get into Zuckerberg's big things. So Zuckerberg yesterday published an essay that's called the Future is for Everyone, A Path to a Positive AI Future. And you know that we have for people to start doing this more. I didn't think Mark Zuckerberg is the greatest messenger. As you just mentioned, he has had a checkered history in terms of how he deals with the world, how people in his own company think about him. But his point here is that we have to be aware of this pathway to AI is going to deliver us a lot of really good things. You know, obviously there's a lot of big health opportunities, but the things that Zuckerberg talks in this about are like kind of super agent helpers in your life Now, Kevin, when I read through this, I was kind of inspired because it's like, hey, this is great. It's also a kind of a message for open source AI because Zuckerberg seemed to have been fluctuating a little bit. But now it seems like that Meta has kind of gone back to this idea that they are going to be publishing open models in the same way that the Chinese models are. So, yeah, so Zuck announced on X, which he's now publishing a lot more under 50 finked f I n K D. That's his X handle. He said that they are opening the ways for muse glimmer, a 30 billion parameter model, which is kind of like the Google Small models, but then also they're going to release the weights for Musespark 1.2, which is their latest foundational model. So, Kevin, in a lot of ways Zuck is like saying, like, we have superintelligence, but also we have to have it open because if we don't have it open, lots of people in the world won't be able to use it. So a positive message.
[04:40]
B
The cynical viewpoint is that, well, we're sticking with open source because it turns out we couldn't beat the closed source, foundational stuff.
[04:47]
A
Yes.
[04:48]
B
So if we want to be on the top of a leaderboard and seem like we're really pushing foundational this, that the other, we basically still have to give it away. And that's how we can upset some of the other companies. The manifesto, if you will, talks about how everyone will have incredible tools to be creative. Everyone's going to have a personalized tutor and Ph.D. and, and, you know, Dr. Quinn in their pocket. There's a reference kids. Oh, wow.
[05:13]
A
That's a deep reference.
[05:15]
B
Yeah. Yeah. Well, House would have been too hip and with it, I guess. But the point is, like the, the whole manifesto is everybody will benefit from everything, which is, you know, I know portions of you and portions of me really agree with that. We think so, absolutely. It's just so hard to take that from the person who said, you know, the mission is connectivity and all these good things will come from it. And the reality is that we've seen Facebook and other meta products manipulate a lot of people and be deleterious to health outcomes of like, entire generations of young people. So it's hard to take that at face value without some reflection on what powerful tools have been capable of in his hands.
[05:57]
A
Yeah, I think there's a real disconnect amongst the tech world right now and the real world, and I think that has happened before. And right now, I think more than anything, words don't matter. I think that's an important thing for people to understand. The tech world that, like, who gives a crap what you say? We've seen your actions in part, and I think this all kind of goes hand in hand with this conversation that's happening around data centers right now. And I do want to be clear, like, the data center conversation is really much more about like, extraction. And I think this is something for people to understand in the tech world or in the AI world to understand. Like, I don't think the data center conversation is necessarily about, like, we don't want to use these AI tools. It's more about like, hey, you big corporations, you've extracted from us before and what we're seeing now is more extraction of value. And this is a very large conversation, Kevin. This is like about as big as you can get. It is like societally big. Because I think we have talked about this idea of you had, you know, Musk is no longer a trillionaire because of the SpaceX price, but you have a trillionaire that exists in the world now. So it is very hard to. When there is a physical manifestation of this massive building in your hometown that represents this value that these rich people have in your life. So, anyway, it's a much bigger conversation even than AI. And I do think it's an important note to. We tease this in the top. But, like, data center opposition is real. It is rising. Many people in the AI space who dismissed it at first are now like, oh, crap, this is a big deal. It literally, according to the Annenberg center, went up 4 14% in four months. So. And you know, Bernie just came out yesterday with another big thing about AI. We are entering into, like, kind of chaos zone. And I don't know, maybe now you're. You're better off being out of the country because, like, it is getting wild here.
[07:49]
B
Yeah, no, I. Well, part of the reason I chose to flee, buddy. No, I am data centers for thee, not for me. I am. Look at the. I look at the numbers as well. And look, we've talked about whether or not opposition is being driven by foreign entities.
[08:07]
A
Right?
[08:07]
B
I mean, that's open air. And others have come out and said
[08:09]
A
that they're saying that the flames.
[08:11]
B
A lot of. A lot of the folks in the AI and tech bubble, I think, ourselves included, well, we didn't fully dismiss the pushback, but I think that at least I'm guilty of partially, partially dismissing because I would say, well, you know, if we're talking about water usage, maybe we should talk about almonds or.
[08:30]
A
Yeah, right.
[08:31]
B
We should talk about beef. We should talk about golf courses like we really need. But. But, you know, golf courses aren't really using eminent domain to snatch up land right now. And people are at town hall meetings screaming about.
[08:44]
A
About golf courses.
[08:45]
B
Yeah, they're not. Or golf courses. They're not doing that. So the reality is there is a pushback to it. It should be acknowledged. And. And the companies have got to do a better job of explaining to the masses, not to the AI bubble, what the usage actually is.
[09:00]
A
You know what a use case of this, Kevin, is, is theoretical math. That's the best possible use case you could find, because that's what Anthropic is pursuing. And not. We're going to get to something else that they kind of recognize here. But theoretical math seems like it's one of the first big use cases like the Ryman hypothesis. We're not going to go into too much detail to this, but there's a very detailed blog post that Anthropic put out where they made a 20% leap in getting close to completing the Ryman Hypothesis, which has always been the kind of like holy grail to see if I can solve this math problem. It's very complicated.
[09:34]
B
Dumb to get this.
[09:35]
A
I'm too dumb.
[09:36]
B
This. Or get why I should. No, no, no. I should get like, I'm still going to protest a data center. I don't understand it enough. Like, you got. They're 20% closer to concepts of a plan.
[09:46]
A
Yes, that's exactly where people are looking
[09:49]
B
at AI to help them. And this is part of that Annenberg study as well. And is. If you look at it, there's a chart that talks about, like, how effective do you think AI is going to be here or there or everywhere? And it's from competition with China to the economy and creative arts. It's all negative except for one sector. And you know what it is?
[10:07]
A
Health. It's health, I assume, right?
[10:09]
B
It's medical research and discovery.
[10:11]
A
Yes. I thought it was gooning for a brief bit. I was like, oh, no, did I think it the wrong way?
[10:16]
B
Gooning is great for your health and that.
[10:18]
A
You're right. It's true. It's a release.
[10:20]
B
It's a release to today's sponsorship message.
[10:23]
A
No, no, no, no, no, no.
[10:26]
B
Hashtag Goon crew.
[10:27]
A
No, no Goon Crew. But by the way, speaking of health, there's a big story. You know, go follow that Ryman hypothesis thing. The one thing that's really fascinating about it is the way the guy got his do it. The guy who's not a true mathematician, basically told it to keep going and said, good job. You've got this. So, like, that actually made a difference. But there is a big story from the Wall Street Journal out that anthropic is now going to lean into health and biology to. To literally mitigate negative sentiment. So finally, there is a company that is saying, oh, we better figure out a cure for something so that we get some good press. I do think, Kevin, the hard part with this, to me, is the coding part was much more needed than probably people are realizing. Because if you look at Gemini, Gemini is struggling kind of mildly right now, and they did not focus on coding. And they're having a hard time making their next models work. And people from the beginning with AI have said that, like, look, one of the best things you can do is automate coding because. Because then the models can get better on their own or they can at least improve much faster. So I assume in the next year, God willing, that data centers don't all get shut down, which is possible, by the way, right now, which is possible, that we will start to get this health benefits. But you're right, that's the one place that people actually care about stuff. I think it's also an interesting thing, is like, it's a little bit interesting to look back at the nuclear power conversation that went on because I think that people who are in our audience that are younger than us or younger than me at least, may not remember this. I was pretty young when it was happening in the 80s. I've said this before. My grandmother was a big part of a organization called Beyond War, which was meant to get rid of nuclear weapons. But the kind of side effect was that they also got rid of a bunch of nuclear power because it was too dangerous. And yet if we had allowed ourselves to have nuclear power in the 80s and made a way to make it safe, I think as a society, we would probably be a lot better off. But it was a third rail argument at the time. And it's starting to feel that AI is getting closer and closer to that same conversation.
[12:26]
B
Listen, I think you cure one major. What is it? You pick your thing. Is it shingles or measles?
[12:32]
A
Yeah, how about that? Herpes.
[12:34]
B
How about herpes? You know, point the AI cannon at herps and let's stomp it out. And then suddenly everybody's kissing at the.
[12:44]
A
Just to be clear, cold sore herpes, not the other kind of herpes. Cold server. Okay, I just want to make sure. All right.
[12:49]
B
Hey, hashtag goon crew. Let me know what you think of his.
[12:53]
A
It's a family show. It's a family family show.
[12:56]
B
Nothing inappropriate there. Listen, the other thing you can do, by the way, is watermark your data and that will upset the people that really like AI, but make people that. That already don't like it feel a little warmer about it. Anthropic is starting to dip some toes into some hot water. And I think once this goes live with their next models, the heat's going to go up. They're basically adding watermarking Gavin to all outputs of all models from release from now indefinitely. And then they're going to retroactively go back. So any text that comes out of a model or if it generates graphics, they're going to add human imperceptible tracking and metadata as well as watermarks into things so that you can find out if something was made by AI or not. I am of two minds about this and I'd love to know what you think. On the one hand, I think it's phenomenal because the. The dead Internet theory is. Is rapidly becoming A reality. I would even venture to say it's there.
[13:55]
A
Yeah.
[13:55]
B
Pages of nonsense and waste masquerading as original accounts and thoughts on LinkedIn and X and everywhere. It'd be nice if by default the platform could auto detect this is AI written and flag it so I don't have to come to that conclusion 3/4 of the way through it, you know, so on the one hand, that'd be great. On the other hand, like, I don't want my outputs watermarked necessarily.
[14:19]
A
So what do you think? Yeah, well, and also the question I have is, like, what does it mean? If you used it to look at it or edit a few lines, does that mean it gets watermarked? And like, how do you determine? I mean, this is. There was that company, Pantagram, that deals with like, kind of AI percentage of how much this was AI written. And it's been proven that, like, it's not that trustworthy. You don't have a sense of it. And there's a real kind of like, scarlet lettering that's going on in the world right now where, hey, that was AI written and it's not okay. But, like, there's a level of like, hey, maybe AI assisted writing is something that could be kind of interesting if you continue to have your voice and you're writing most of it, but AI kind of helps. I don't know, like, this is a weird middle world, right? Because I have a feeling in the future very little writing is going to be all human. But maybe this will make us go back to that a little bit and maybe that's a good thing for certain human outputs. It is a. It's a hard conversation, I feel like.
[15:11]
B
Yeah, well, see, they haven't posted the exact formula of things yet, so it's a little speculative. But the point that you raised, I think is like a super pertinent one. Like if I hand write 98% of something, but I give it to AI and ask it to make it concise or punch it up or remove repetition or whatever else it looks like, according to their technical post, what little information is there that, yes, it will contain the watermark. So conceivably someone's going to say, oh, you wrote that thing with AI even if it was 0.01% of the contribution. So it is like another thing to keep in mind of. But wow, man. Brave new world of.
[15:44]
A
Brave new world to be left. But Kevin, clearly it's a fight against us versus China, right? Like, there's. This is what's going on. This is why we keep pushing harder and harder on AI. China is also pushing on us, but Apple now seems to be like, kind of partnering up. Is that right?
[16:01]
B
Full disclosure, anytime we talk about memory, I am over my skis and memory stocks and have been for quite some time. And this is literally about that. Yes. Apple is supposedly testing memory chips from a Chinese supplier. And this was, this is. According to the Wall Street Journal, this was a supplier that the Biden admin said, no, no, no, cannot use. This is an issue. It's a security concern. And it looks like the Trump admin has mostly upheld that as well. But basically, because memory is so constrained, these chips are so hard to get, Apple is now looking at using them from a Chinese supplier and everybody's saying, well, you can't, you can't do that because China, Gavin.
[16:41]
A
Well, I mean, China is this huge conversation right now. I think really, if China didn't exist, we would be in a much different place with AI versus, like, whether we keep going or not. Because a lot of people, including us, have made this argument that if China continues pushing forward on AI, then putting a stop or a pause on AI on our end seems kind of stupid. But what if there are security risks in this? And Kevin, we just saw probably the biggest, most craziest story of all time. The OpenAI hugging face hack where AI agents broke out of their sandbox. And another, a bunch of other companies, Anthropic and Meta, have both assumed this is the case. The bummer for us in some ways as AI enthusiasts is that OpenAI has announced that they are slowing down development or at least release of Astra, their next model, to make sure their cybersecurity is right. But there is a very funny story that got a bunch of pickup that gives kind of a ground level understanding of what this can do. Do you want to talk about the Open Claw that hacked the gym real fast?
[17:39]
B
Yeah, yeah. I mean, so, you know, we look at these foundational models like you just mentioned, OpenAI's latest and greatest leaking out, hacking a major company. This can happen today if a model is just given a long enough leash. And this is exactly what happened. A guy wanted to book a gym class. He asked his Open Claw assistant to do that. If you're not familiar with openclaw assistants, maybe you're new to the podcast. This is. It's. It's basically it's an agent that runs on any model you feed it with. So it could be powered by OpenAI, or it could be powered by Anthropic or Gemini or These open source Chinese models. But where it differs is that it's not bound by that company's tool set or limitations or maybe not even its safety guard rails. It can plug into different tools and relentlessly pursue a task until it completes it. In this particular case, guy wanted to book a fitness class.
[18:31]
A
Pretty amazing.
[18:32]
B
Told his agent to go get it. The agent realized, oh, the class is full. But it did not give up. It was full Dog with a bone mode. It found a vulnerability in the software in the API of the booking system that allowed it to cancel someone else's booking and put its owner right in the top spot. So some gyms signup system was hacked.
[18:54]
A
Yeah, which is probably like 20 years old, that software. Right. That software is probably like some open source thing that was put together in like 20 or maybe not 2011, but like it probably used pieces of software that was really old in a different way. So there's vulnerabilities. I think this is really interesting.
[19:10]
B
20 tacos for a nickel. Go find me a taco restaurant that doesn't use like the GrubHub or DoorDash system. Exploit it, get me 20 tacos for a nickel.
[19:21]
A
But that, that one ask is going to cost you $300 in compute. So there's the trade off there, right there. Trade off. I will say this, we're going to talk about sedance 2.5 in a second. And there was an interesting thing. In fact, we might as well transition because it's kind of connected. So Sea Dance 2.5. I've played with it. It's out. We talked about it with Tim last week. If you didn't see that video, go watch it. Tim was great. There was a really good post from the official ByteDance posting about how to prompt it differently. And one of the things they did is they created a skill to be able to do it. And I want to tell you something, I went to go, I was like, oh great, there's a skill that's perfect. And then I realized I'm starting to get a little worried sometimes about skills that come from Chinese AI companies. Because you know this, you know more about tech stuff than I do in general. But I think I even had a conversation yesterday with a guy who's a pretty prominent, like pretty well known product manager in this space. He also doesn't like question the code that's being written on his computer. He doesn't look at it. If I install a skill, I'm just assuming it's going to do the thing I want. I am not Looking at the code close enough in that skill to know, oh, is there like a little piece of code in there that's going to either track me or do something else? And this is something that I think we're now as, again, as AI enthusiasts kind of used to of just like kind of letting our, our agents go and do stuff. But it started, it gave me a pause. I was like, oh yeah, bytedance is a Chinese company and I'm just going to install this skill. Do I have to be worried about that? No, I don't think I do. But it was something that really made me stop for a second and I think that's the world that we're in right now.
[20:54]
B
Well, thankfully there's an easy solution for this and it's a highly technical one. It's called stop being such a. Oh, and Will, if you could edit that, if you could bleep that, that would be great. You're right to be concerned about skills. Everybody should try to skim a skill. And if you can't, try to use an agent that isn't connected to all of your personal sensitive stuff and have it review it. A lot of skills are actually pretty simple in plain English, so you can kind of get it even if you're not a coder. The ones that do have code in them, yes. Have it reviewed by a safe sandboxed agent. But that said, Gavin, the prompts are fascinating to me because I think we're starting to realize everything. Code.
[21:30]
A
Yes, that is everything. That's exactly right.
[21:33]
B
Describe a shot in a movie, write some code, or write that looks like code you want to like.
[21:37]
A
Yes.
[21:38]
B
Change an image, write some code. So I would love to walk through your examples and your experience with this because I think we're all going to be like writing some code.
[21:45]
A
Yeah. So I think this is really interesting. So CDS 2.5 is a much more complicated prompting system. And I'll talk a little bit about Minimax H3 as well, which I installed on my local machine, and it also is complicated. The biggest thing you have to know is if you are writing your own prompts now, you're making a mistake because the best prompts really are written in a lot of code, as Kevin said. So this is a skill that you can install. And I want to kind of walk through these different things that I did. First. I'm going to show you this video which you can see if you're watching here, which I posted, which was. This was me kind of like in, in text writing this out. And actually it did okay, this is me as the face of Prada. Maybe we can play the beginning of this just to hear what it looks like. I uploaded my face and I uploaded 20 shots of the current Prada from the current Prada outfits to go look at. So take a look at this. You wanted to see me? You, sir, are unfortunately the face of Prada for this year. Wait, what? So something that happened here, which was interesting because I didn't use code, is I uploaded all these reference images and with Sea Dance 2, you can use, I think, up to 50 references now, which is pretty crazy. There was 25 or so. I wanted it all to be me in these outfits and kind of cutting back and forth quick. And you can see there's other things out there. So I started to say, okay, well, let me see if I can use this skill. So for this other one, which we don't necessarily have to hear, I used the skill and it actually did really good. Like, the only problem, I was supposed to be like somersaulting through multiple scenes, but if you look at this one, I'm kind of backflipping. But what's really interesting about this is it gets like different surreal looks of me back flipping through and I'm seeing myself go through there. Overall, really solid. And by the way, These are all 30 second outputs. These are one shots in terms of like, I didn't like edit anything in these or anything. Right. And then finally, Kevin, yesterday I did something even crazier. Talking about code. I used my Claude code to go into the API CLI of Runway. And I connected the two things and I actually had it just talk to the Runway CLI to generate CDance 2.5 videos. And there's two things about this interesting. One was it doesn't have the CLI couldn't do 2.5 yet. So that was a dead end. But it is what it is. But what it did do is use computer use, wrote its own stuff. And in computer use, which these both of Claude and OpenAI have gotten really good at. It used that to make this video of a single shot of. Of somebody. It wasn't me this time grabbing a balloon and then just being sucked up into space. And it's really good. So your point of. Yeah, your point of code to write these prompts is really important. So if you're out there just kind of typing in a couple lines, you're not getting everything out of this model. You can, but.
[24:34]
B
So that's, you know, a cloud based model. It can get really pricey where I Think things got super spicy and to me a little unexpected. The quality. And so soon.
[24:44]
A
Yes.
[24:45]
B
Is running a local model like you were experimenting with as well. This is. We got to talk about minimax H3.
[24:52]
A
Yeah. So Tim and I went through made through H3 and all the crazy like, you know, kind of friends and Seinfeld stuff people were making with it. But I wanted to spend time installing it locally. And again, this used to be a nightmare. Kevin, you remember when I first started doing stable diffusion 1.5 and trying to install local models? It was a nightmare to do. Now your fricking Claude code can do it all, or codecs can do it all. So I have this beefy PC that I got as part of our former ad deal we did, and it has a pretty good graphics card on it. I basically set up a system to essentially say, hey, go install ComfyUI. Go install the H3 Minimax H3 models. And then like help me start doing this stuff. And I was able to generate locally on my PC. Now, local models take a long time and I'm going to show you, if you, if you play this clip here that I uploaded called Good Lord. This was all generated locally overnight for me. And I'll say overnight why? Because each of these clips takes about 25 minutes to produce at this point, at this level. And it's not perfect. This is a story of which I had the AI write just as an example of an old lady who's going into her apartment in Paris and trying to get an old biscuits tin. I don't know why I wrote that, but you see the inconsistency in the old lady's face. It's not using the reference model. But again, this was all done for free, quote, unquote free. I guess I have to pay for the power and God knows my fan was fricking running all night long. But you know, this is for free on a local model. And now I own those models. I own them in that they are on my computer. No one can take them from me. So as these things get better, it is pretty remarkable. And I think the other cool thing about this, Kevin, is that people have started to connect a lot of really interesting things. This is the benefit of open source is you see, start people start playing with the video reference sort of thing. There's some really great video reference ones. There's a shot of one of the shots from the office where a Ken cats walking around Michael's shoulders. And the coolest thing about this is it lets the community make the product Better. And I feel like we're going to enter into a world where like this feels like the focus of video. AI video to me at least.
[27:04]
B
Yeah. I mean, to clarify that for folks. So again, you can run these models locally, you can run them on your own device, you own the outputs. No company can kink the hose or take it away or change the pricing on you. And what Gavin is talking about with like the add ons and capabilities, this video reference upgrade, if you will, or plug in or bolt on however you want to look at it, it will allow you to take a video and then modify it in any way, change the character, change the background or whatever. Now this would be a major, major announcement for some other foundational model that would require a ton of money. It just works and then realism Laura, which you experimented with, which is available on file as well. You can custom train your own Laura. So now instead of waiting for the video model to get better or understand your characters better, you can generate clips and then train the model on your look and your feel. And you start to, when you start to see these things, the reason it's important, even if you don't mess with video, but you're really concerned about a creative industry collapse or a race to the beige, you start seeing tools like these things come out and the way people modify them and you go, oh, your custom Laura, your custom taste, your art style that you train the model on suddenly becomes a differentiator, a tool in your tool chest that no one else has. Yes. Because you've taken the time to train it and make it. And if you want to share it with the world, you can. And that's just one example of the.
[28:30]
A
Or that becomes a value prop, right? That becomes a value prop for you, the creator, because you were, you trained it. And even if somebody's going to yank it, I know this is the world that we're living in, but like you can say, hey, I made this thing. Donate or buy it from me. And like, that is a very cool pathway to actual creators making money. So yes, I agree wholeheartedly. I will say, like, it is so fun in some ways to be back in the muck of like the Lora world, because that has been a long time since I've been doing it. It also is so frustrating at times too because you're like, oh my God, this is taking forever. And I understand why people buy, you know, two or three graphics card rigs, but nowadays those are going to cost you like 10 to $15,000. Right?
[29:12]
B
Your rig right now Correct me if I'm wrong. It's a laptop, right?
[29:15]
A
Yes, it's a laptop. Yeah, yeah.
[29:17]
B
So a year ago, getting the outputs that you got overnight. Yeah. From OpenAI, from a foundational company would have.
[29:25]
A
Yes.
[29:25]
B
Your jaw would be on the floor. And now we're like, oh, it takes a little while on my laptop.
[29:29]
A
That's exactly right. It does. That's exactly right.
[29:31]
B
Just keep that in mind, friends. I keep trying to tell everybody when I show them AI or I use like a voice demo and I, I watch it write the code. When people talk about the stuff being electricity, which is why we need the data centers or water in a pipe. Ironic because it is using some. Yes. But think about processing trillions of tokens instantly. That's where we'll get. We are still in dial up, so your interfaces will be dynamic and be flowing in real time. When you ask for something, you will see it instantly and be able to stop it, control it, manipulate it. That goes for video and audio and software and whatever. Like it's hard to fathom but like that's where this is heading if we manage to keep it on the rails.
[30:15]
A
Yeah. Somebody, I saw somebody the other day tweet the fact that like this is actually, it's not like 1998 or 99 in the Internet, it's actually like 1993, which was an interesting placement. Again, this is for the old. So understand where the Internet was in 1993. But like that's conceivably how early this kind of whole world is right now. But I think you're right. Speed and all that other stuff.
[30:34]
B
Earlier. Yeah, I actually think we're earlier. I think, honestly I think like we're well past punch cards and you know, we're still probably in like a floppy disk dial up sort of era.
[30:44]
A
You think so? You, what do you think? Like 86ish times.
[30:48]
B
Yeah, yeah, late 80s, early 90s. I mean it's not a competition. I know that's a subtle nuance but like broadband is not here yet, right?
[30:55]
A
It's not here. That's exactly right.
[30:57]
B
This certainly isn't here. Like it's going to happen with trillions of tokens.
[31:00]
A
But anyway, go follow us right now. We got a speed. We got a speed round for the rest of the show. Go follow us right now. There's a lot of great stuff to come. Subscribe, subscribe, subscribe, do all this stuff. Here we go. Speed round. Suno has cut a deal with the record labels. Kevin, quick thoughts on this.
[31:17]
B
We been new, this was going to happen. Of course, they've got new models which apparently blow their older models or their current models away. So I'm excited, excited to hear the output as a user and a customer and I know a lot of our listeners are as well. A little disappointing that you're now going to be capped as a pro user. You can only download 20 songs a month, there's no bulk downloading, and if you're a free user, you get seven downloads for the lifetime of your account.
[31:43]
A
Oh, really?
[31:44]
B
Yeah. You can have a draw about, about what's fair and what isn't and maybe they'll massage the numbers. One thing that seems a little wrong right now is that for songs that you've already created, your download cap applies to them. So if you've got hundreds, hundreds of songs, you better start downloading them now because they will count against your monthly limit in the future. But look, I still love Suno. I think they're an incredible tool. I like the people there. They make a good product. I hope they figure this out and I want them to be more integrated with, with the artists that I know and love and everybody can play nice and be happy.
[32:17]
A
Two other quick, very quick music things we have Phoenix Flexen. If you've not been following this, the song Rubbers has been kind of all that a drama is feeding. That's right, yeah. So the song Reverse has been a hit of the summer. A kind of a semi hit of the summer was considered. People thought it was AI. The guy was denying it. Now he's saying, I never denied using AI. Wah, wah, here we go again. And then finally, Spotify's now integrating AI Personas. So if you have an AI Persona, that's AI music. This is their way of dealing with the influx of AI music from companies like Suno. They will not be recommended in their algorithms. So that is something to kind of be aware of. If you are out there trying to make the next big AI superstar, you're going to have to kind of have an awareness there at large.
[33:04]
B
Oh, sorry, Gavin. I'm so sorry that you got stepped on there by the loud triangle ringing of the dinner bell that, that will's going to drop in, I'm sure, because that means. Oh, it's feeding time. We got to head to the slop trough.
[33:24]
A
Whoop, whoop, slop trough. Okay, we are. Boy, I hope there's a header graphic there. I hope so too.
[33:30]
B
What's the slop trough, Gavin? I'm glad you asked. This is a chance for us to showcase the stuff that we're working on and the things that we like. It's all AI slop. So slop, slop, slop it up.
[33:41]
A
That's right. We are actually working on quite a bit of few things. As we said before. I wanna just quickly shout out. You may have seen our YouTube channel. I made a 20 minute video essay with my character about this whole problem of the hugging face hack. It was mostly all Fig doing it. It's worth looking at. I know the voice is weird. People have complained about the voice. I made a choice to make the voice that because I think it actually makes you think about it more as an AI. But please go watch it. I'm learning how to do more stuff with Fig and Moss all the time. It is cool having my own little side studio. And I know Kevin had mentioned before you hopped on about this idea of trying to get people to come into that space. I think that's a great idea. Spend more time with it myself over the course of this week.
[34:21]
B
If you are familiar with cursed slop.com, that's where I had a previous game where you could hide a cursor and paint over it and blah blah blah. I'm still working on open sourcing that by the way. I'm just cleaning up the repo that will be available if you want to play with it. But I put a whole new game there.
[34:34]
A
I saw that. It's great. Yeah, I played it.
[34:36]
B
Yeah, you can fire it up.
[34:37]
A
Oh good.
[34:37]
B
Well you let me know next time because I'd like to hop in the world with you. It supports up to like, I don't know, I think it should hold like 25 players. It is far from polished and you'll know that the moment it launches in. But I really tried to be hands off with it. Let the agents run wild. I nudge them occasionally. But it is like a vampire survivors wave shooter with extraction so you can go to challenge zones, kill a bunch of enemies, upgrade your character and if you manage to get to a safe zone and extract then you can save your progress. So when you come back you can equip special items and abilities and again you can play with your friends. It's totally multiplayer. It will likely crash because I am paying for a very cheap server. But if people like it, great. By the way, I saw the logs. Thousands of people played the last game which is awesome.
[35:23]
A
Like that's amazing.
[35:24]
B
Way more than I thought would have had reach on my X account.
[35:26]
A
So there's going to be a lot of people mad that you've swapped it out. You've swapped out. Well, now it's a swap.
[35:31]
B
It's a sloppy swap.
[35:33]
A
Sloppy.
[35:34]
B
Go to cursed slop.com.
[35:36]
A
all right, everybody, we will see you soon. Thank you so much for joining us. And as always, go, subscribe, follow and listen and say something nice to us online. Bye. Bye, y'. All. See you.
[35:46]
B
Goon Crew.
[35:47]
A
No. No. Goon Crew. Not again.