AI for Humans

The Singularity is...here? GPT-6, Opus 5 & AI's Scariest Week Yet

Transcript

New AI models like Anthropic, Opus 5, and the upcoming GPT-6 have us really excited about the future of AI. And the coming Flux 3 and Seedance 2.5 have us giddy about the anticipation of what is to come with AI video. We are now like in the singularity. Like this is the moment. I've been waiting for this my whole life.

Some of the alternative visions painted by other companies are like quite terrifying and make sure that gets pushed against and is not what happens. However, all is not well in the land of artificial intelligence. Oh, pray tell dear sir, what hath wath wronged in this world? Gavin Pomegranate. Okay, thank you. I'm back. I'm back.

That'll make sense to two people. With big models comes big problems. We're gonna dive into some of the scary stuff happening like AI models breaking out of their sandboxes. And talk about how AI scaling might lead to the world that many people feared with AI. Yeah, but also robots on the front line? But who's controlling that robot, Kevin?

Friggin' Kid Rock! Who cares? America! This is AI for Humans and... Who cares? Welcome to AI for Humans, your weekly guide into a wonderful world of AI. And Kevin, this was a giant week in AI news. And we're gonna jump right in and talk about some of the big model updates. We're gonna get to Opus 5, but first, Kevin...

New rumors! I'm gonna check on my gauntlet loops. Oh, sorry, what? No, I've got... This is... Just a shout out to all of the hashtag builders out there, and I guess the hashtag haters as well, who claim like, this stuff is slowing down, or it's not gonna work, or there's not going to be a never-ending need for tokens. We are living in different realities right now.

Yes. I know social media would have us believe that across all fronts, but I know you have been feverishly building. I have run out of my limits on multiple plans across multiple accounts, and I'm just making stuff and enjoying the process so much that I can't imagine not wanting a demand for these things. Yeah, it's really interesting you say that.

We're gonna talk a lot more about the future of what this looks like and some of the scary stuff, but I think first let's talk about the models right now. The biggest thing that's incoming right now soon is we are getting a lot of rumors around GPT-6. Now, this is something that has been rumored for a bit. Sam Altman is in Washington, D.C. this week talking to the administration about what this means.

Currently, the rumors around GPT-6 are that it is much larger than 5.6 Sol or considerably maybe even Fable 5. It is going to be better and cheaper. Supposedly, OpenAI has figured out some way to create a way to use less tokens, which you and I will talk about in a second with Opus 5, which is a big, big deal. And then more importantly, Kevin, this might be coming in August.

And as I look at my watch right now, it is late July. So we were expecting this year to be fast and furious. And what might even be crazy is we might be hitting a fall that is even more furious. I don't know. It's pretty bonkers to me. Yeah, it's not just OpenAI dropping something foundational and Anthropic trying to answer back and then China releasing something open source-y.

We've got super safe intelligence waiting in the wings. And apparently there's some investor pressure that might lead to another foundational model release from them. So US and A, go data centers. Yes, US and A. That's right. We've got a lot. Yeah, Ilya has taken a bunch of money from NVIDIA to scale his research, which we don't know at all what it is yet.

There are some rumors out there that Ilya has found some way to unlock continuous learning, which is something he was working on. Now, these are all complete rumors. Those guys are very good about being completely shut down. But we do expect something big. Kevin, all that kind of connects back to NVIDIA's open letter, which we're going to discuss in a little bit about open source AI, which I want to talk about.

But we should also first talk about something that happened when we were off last week on vacation, which is Opus 5. And Opus 5 is Anthropic's big new model. You and I have both had time with it so far. There's a lot of divisive opinions about this online. The Benchmark Bros. We might as well get the Benchmark Bros in here.

Benchmark Bros, stand up. We're the Benchmark Bros. Turn the tests up loud. Check the charts, make the data proud. According to the actual benchmarks, this beats Fable 5 on a lot of the benchmarks that we have seen. Now, you and I have both had time to work with this in practice. We can talk a little bit more about what other people think, but we have time to spend.

We spend time with this ourselves. What are your initial impressions of Opus 5? First of all, shout out to our podcast for getting the early drop on that juicy. Oh, yes. That Opus 5 would be. Yeah, we did it. It's funny. Fake ate its own tail too as well because I received a message from someone saying, Hey, by the way, I heard an Opus is coming out this week.

And I was like, cite your sources. And I followed the links all the way back to someone referencing our podcast. Hey, we made it, Kevin. Somehow we made it. You people as Eric people think we don't know this. It's that easy to put noise out into the world. We exist. We exist. But yeah, I guess we don't have to kill my source.

So thank you, my source. Gling. Gling. I hope there's a gling noise. Okay. Gling. Having used it enough now, I can say it's a good model. It's a good boy. It is token hungry in a way that has me chewing through my limits. I am not very happy about that. It's smart. It's capable. It is like a dog with a bone sometimes. It doesn't give up, which I like.

I've been running a lot of slash goal and slash loop. If you don't do that, by the way, slash goal gives the whatever harness you're in, if it's cloud code or if it's in codex, it gives it a clear understanding of what you want it to accomplish. Yeah. And the slash loop kind of basically tells it, hey, keep going, keep looping until you reach it.

I think the combination of the two can unlock some pretty great stuff, including enhanced credit card bills. Oh, interesting. Because again, Fable 5 is, or excuse me, Opus 5 is token hungry. So if you let it just run in the background on a loop, it will chew through things much faster than other models. But I'm liking it.

I'm not seeing the pushback. I haven't, like, are you a fan right now? No, I mean, it's funny. I've been running it myself for the last couple of days. I was kind of offline for the weekend. And today's the first day that I've been spending some time with it. I can show a little bit of what I've been working on. So if you're not familiar with what I've been doing with my, I'm calling down my Fig and Moss universe, that there are about probably five people in the world who care about.

I created these two characters within Cloud Code. Fig is the name of my Cloud Code. And we've been making videos with them. Well, I created a side-scroller. And it's been working now for multiple hours. I have a build of it that is still incomplete. But it is really interesting to watch it go. Unlike other Opus models to me, I mean, Fable could do this, but you'd run out of credit so fast.

It does continually go. And it understands better about how to keep moving on stuff. And I think the biggest upgrade here is Fable 5-esque level stuff with Opus level credit costs and space. Now, you're still going to burn through your credits, which is what's going to happen with all these models right now. But again, Kevin, it's that cheaper, faster, better thing.

And I think the big news about Opus 5 is it's cheaper and faster and almost as good as Fable 5. There's been a couple small things that I have found that I still rely on Fable 5 for. The kind of general impression across the internet is that it's a good model. Dan Shipper, who's one of our favorite early model reviewers, had a really interesting take on it, which was that he found it that was pushing back too much.

And he found that it wasn't as good. He's had kind of two daily drivers. One's a daily driver. He still wants to use 5.6 soul. And then he uses Fable 5 for the hardest things. I think we're going to have a really interesting moment when GPT-6 launches, because there is an opening right now for whatever that kind of like high end crazy thing is.

I do want to point out a couple things that I saw also. Matt Schumer had another big post where he put out what is basically looks kind of like a Counter-Strike, a Counter-Strike clone with one shot. And I will say like my experience building, I'm right now building a fig and moss platformer, which we can show some of here.

It does seem to be able to build things on its own pretty well. I don't know if you've had an experience with that at all. Like what is your experience with building stuff? Yeah, let's dive into that because I'm using, I was kind of naturally using a version of Matt Schumer's. He calls it a gauntlet loop, which is what is unlocking these one shot projects.

I love Matt, love his work. So, you know, no shade there, but let's demystify what that is because he has a nice blog post about it, which we can link. But you're essentially providing the model with that slash goal that I was talking about earlier. Which everybody can do, by the way. Everyone can do that. Anybody can do this, right?

And it's very simple. It's basically saying, go make this a AAA game like, and then you reference a AAA game. Yeah. You give it some artwork, right? And you can tell it to go research. And the real crux of it is telling it to like, come up with an impartial judge to go and compare screenshots or video or documentation of the really successful AAA thing to the thing that you're trying to build.

Does it work? Yeah. Yeah. It does work. Does it create iterations of things and kind of wonky clones of preexisting stuff? Well, yeah, because you're literally saying, go judge the output against this known thing. So for original creativity, it might not be the best thing to use. Yeah. But if you're, you know, take an ounce of effort to modify the gauntlet loop and provide original artwork or original design specs or whatever, you can really unlock it.

This is to me, the biggest thing, by the way, if you're going to go do this, and I think everybody should, everybody should try to, to mess with this stuff. Yeah.

And that's a really bad time. Yeah. You're going to have to explain charges that are popping up on your Apple pay to your misses. Well, and I also want to talk about this idea that there's lots of people out there doing these like kind of matches that kind of a counter strike clone, right? And there's other games out there that are doing similar things.

It's like, these are interesting tech demos, but there's a lot of people out there that's like, I made a video game, blah, blah, blah. The video game is really judged by, do you want to play it again? Or does somebody else want to play it? And I think that's where a lot of this like visual look of what the thing is, isn't really the thing, right?

Like, yes, it can do this thing, but the real work, and this goes back to like why human creativity matters and these models can't like figure out all this stuff out on their own is, it's in testing. It's in the loops that you create the game loops. It's in the feeling of what the game is. So I think you and I are both working on these kind of side projects that will be interesting to kind of continue on as we go forward.

I don't know if you want to talk a little bit about the thing you're working on at all. I haven't had a chance to play it, but I'm excited to check it out. Yeah, it's super early. But well, you know, I had a really fun moment the other night when people, they build it up CursedSlop.com, which is the website, which I don't know if I'm going to keep active, but I don't know.

It's like, it's 15 bucks. I might as well at this point. But like we sat around the table and we played it for multiple rounds. Fun. I did. I did have to do that sanity check of like, well, these are my friends. Maybe they're just being nice. Right? Like, and I basically said, hey guys, watch my, let's go through my home movies.

And they were, so, but I did get the sense that like they actually enjoyed the experience, which was, that was nice signal. Sure. Sure. But rather than, rather than continue improving that, I was like, oh, what if, what if vampire survivors was multiplayer? Like what if, what if halls of torment was something you could hop in and play with your friends, you know, PVE, but then there are certain rings or areas where it's PVP.

And then the loot that you get and the items you unlock, you have to extract kind of like, oh, okay. Oh no. Like I, now I'm already, I'm in token. You're deeply, yeah. I'm a monopoly guy with pockets out. Yeah. And I'm building something else. That's another $15 for me to serve on Calisius. So anyway, I'll have it out soon probably, but it is a multiplayer, you know, zombie wave enemy shooter.

That's also an extraction based thing with PVP. Why not? Yeah. I mean, it's so interesting because like the question will be cut. We've talked about this for a while, this idea of like building software for yourself or building things for yourself. And I do think ultimately the question will be about, do you build something that's interesting enough for other people to play it?

Or does that even matter? Right? Like it's not going to be a business as much. It might just be fun, something fun to do with your friends. I think this is going to be one of the fundamental questions of AI at large going forward. Like all these AI startups that are starting right now are kind of getting eaten in part by these large companies because anybody can do this thing.

It's very much like when YouTube launched this idea that like everybody could make video and none of it was very good. But when the video is getting pretty good, the people that made the video professionally, you and I, have suddenly have a bunch more competitors out there. Right. And I don't mean that in a bad way. I just mean what's interesting is again, it's that flattening of the ability of like what people can do and what they can make.

And it is a weird, I'll take a moment here because like it is a weird thing where you start to think about the attention economy. And I think this is something to think about for everybody is like, it is seeming like it is becoming less and less like, can you make the thing? And more and more like, can you make the thing interesting enough that people want to come back to? That's the biggest thing I think.

Right. Right. And the answer for everything that I've made so far is no, I can't. Hey, but maybe it might be interesting. By the way, like if, if the pressure you're putting on yourself as a creator out there for anything that you're hacking or vibe coding or gen arting or whatever else is, I got to make money from this, or this needs to be as good as a triple A something.

Again, you're probably going to have a pretty bad time. Yes. Yes. You have to make things to develop taste to understand. And the little projects that I'm making now are not because I believe I'm the next notch or you're having Roblox three on my hands. Yeah. I'm having fun. And by the way, I learned how to connect my cloud code to my codex to have codex be an asset, like an image asset generator into an MCP server to generate videos and blah, blah, blah.

And it has very little effect on making a fun game, but I did it. And now it's paying off in my professional life because I'm using that workflow to do actual work. Yes. So like the dividends will come. It's like, they will come. If you build it, they will come. Actual work. That's it. Actual work. That's it. Actual work.

Yes. The next thing we can talk about here is AI video. And this is something that like is kind of near and dear to my heart because I've been spending a lot of time in this space. There are two major, major things that are happening in AI video. Neither of them is released to the public yet, but I do think it's really important to flag this. The biggest thing that kind of came out of the blue is we finally got a multimodal, including video model from Black Forest Labs.

This is Flux 3. And Kevin, I don't know if you've seen these, but there are some beautifully, really interesting versions of this. The most important thing about Flux 3 is that according to the company and according to many people out there, this will be an open weights video model, which if you know the way that AI video has gone is a big deal because there are so many restrictions on things like VO3, even Seedance 2 and 2.5.

Before we dive into the exciting examples, because they are fun and great and scary and hilariously bad at times. When you say like, oh, it's open weights and that's going to be amazing. Like, distill that for someone who goes like, yes, I'm sorry. What does that mean for me? And why is that important and exciting? So what it means for you in two parts is, A, you're never going to be able to serve this model personally, or maybe in the future, but that doesn't necessarily mean it's going to be on your local computer.

But what it does mean is that lots of other people can serve this in the cloud and you will be able to have mostly unrestricted outputs. Now, there are going to be things that are still restricted, but when you think about what you can do with Comfy UI and you can do with Laura's and all the sorts of stuff that we have talked to over the multiple years here, you will be able to manipulate and use this model in ways that other models from something like Google or Seedance, you won't be able to.

And I think as these models get better, it's going to prove more and more important because you're going to want to make your own styles. You're going to want to make your own kind of vibes to certain things. And that is just really going to be hard in things that are controlled by a third party. So that part is extremely exciting.

And I really hope Black Forest Labs keeps up this kind of open weights, open models sort of vibes. Yeah, well said. It's cost and control. And you want more of one and less of the other, and that's heading in the right direction. So that said, what are some of the examples that you've seen that are making you tickled? Well, so interestingly enough, if you really push this model and you see what's out there, it does seem like Seedance 2 and incoming 2.5 is still probably a stronger overall model.

But one of the cooler things was there was a bunch of people showing off split screen rendering so that you would have in one model output multiple shots of the same shot, which is a very cool prompt. Those are the ones that got me. Yeah, because that's a really hard thing to think about, right? Because the model's having to take physics from two different places in the same exact shot, and it matches really well.

The other thing that I saw that was really fascinating, which I'm sure you saw a bunch of this, is somebody kind of unlocked the idea of like the old VHS footage look. And there were a bunch of really good examples of that. Venture Twins had a really good example of like what was like looking at kids that were being interviewed about the Internet in the 90s.

I saw a really funny one of a guy who was made of plates that went into a cannon and got shot at a wall, which really made me laugh. It was a lot of good stuff like that. There is a, yeah, there's a pancake version of that as well, which is great. Like it's a man made of flapjacks or like birthday cake or whatever, hops into a cannon, splatters.

The multi screen thing that you're talking about, like one of the examples that I saw was the CCTV camera, the quad splatter. Yeah, very cool. And again, yeah, like this model has to figure out what the scene is. Okay, it's a guy in a convenience store. He's going to spill something or drop it. Then it has to hallucinate that from four different angles that are all synced up.

And now we're just, look, any one of those angles a year ago would have been impressive. We would have gone like, oh my God, you had the night running through the grocery store with the rotisserie chickens. Yeah, I've got to try that with these. Hopefully when, yeah, when Flux comes out, I want to try that model because I bet it would be very cool with that.

So like that, I mean, that just, it's a very, very impressive model. And again, the notion that we will be able to potentially play with this and plug in our own lauras and create our own characters and not be thwacked on the wrist because we're trying to generate something that has some combination of words that OpenAI doesn't like.

That is very exciting. I want to shout out one last thing from its spoiler man, which is a very good video of these pirates on a boat with base and the base is destroying these ships across the water, which made me laugh a lot. It's kind of weirdly doesn't make sense that it looks like old VHS footage, but it's a fun thing.

And a quick shout out to purrs beats or purrs.ai on the old X now, a friend of the show. He has a couple of different clips where he's like doing these weird channels surfing through bizarre eighties networks or whatever. And it's everything from, you know, dog judges and weird rock bands and people hacking toasters. It's, it's great.

It's like a fever dream in the best way. And then a double shout out Gavin to also friends of the show, the big bang theory spinoff team. Stuart fails to save the universe now streaming on HBO where they proudly say in the intro that they posted online. Wow. How great are our graphics and how awesome is this intro? We basically had no money to make it.

It isn't it great. And people are big mad about some of the characters having like misshapen hands or a shot in a comic book store where there's like weird AI hallucinated titles to some of the comics like Minions and Beftwits, which I don't know what, I don't know what Beftut, Beftut is, but it's a comic apparently. So I don't know.

It's just a real dumb kind of self own on those guys purpose because you have to know that audience particularly there's going to be so much anti AI people like why you would do that for a major show. I mean, I understand like the purpose of like, hey, we could save money, but like also we can do these cool things. But if you're not upfront about it again, you're going to get slammed.

So I think that's an important thing. But another way to get slammed, Kevin, is to touch that button, that button right now on your keyboard. Touch that follow button. Touch that like button. Get us some, slam us, slam us, slam us again. The Kiwi slider? Yes, yes. Slam us, the Kiwi slider, slam us. This is a message to everybody out there.

Let's be serious, Kevin. We need you to subscribe and like AI for human. Where are my Kiwi kids at? Let's throw some engagement. Like, subscribe, click the bell, slide a Kiwi into your Kiwi hole. Make this show happen because Gavin and I have literally zero extra dollars to actually advertise for this thing. And also, by the way, you know what else, Gavin?

What else, Gavin? Oh my God, we are so liked. As Gavin likes to say, we love you when you love us. So let's bring that one back. If you want to back us on Patreon, you can throw some money in the old tip jar over there. You can subscribe to the newsletter, head to AI for humans dot show for all of that stuff. And sincerely, any ounce of engagement or sharing this podcast with your friends, your family or your frenemies that helps us grow and or at least stem the bleeding.

And that's all we can. That's right. This economy. And also, I want to tell everybody there is for the next two weeks a little bit of funkiness in terms of our scheduling. We both are on weird vacations. We're taking time off in different ways. So there may be some up and downs in terms of how this show comes out. So I do it.

And the Kiwis are really important to keep in mind. But, Kevin, we have to get into the scary part of what's going on right now. And there are some big things happening. Most importantly, something that happened late last week, which was a OpenAI model. We don't know which one helped break into Hugging Face. Now, this is a big story.

It's been covered by a lot of the major press platforms out there. OpenAI and Hugging Face put out a blog post together. But this is the thing that we have been talking about. And I do want to shout out the fact that, like, Hard Fork had a great episode of this and some other places have covered it. But, like, Kevin, I do think that this leads to some larger conversation points that you and I have been thinking about for a while, obviously.

The world is starting to get a little bit scary when it comes to AI. What's your take on this? Let's see. From the hip? Yeah. Yeah. It's scary. It's a little weird. We were they were supposed to be testing these things in super secure, air gapped Faraday cages because they know what's best. So these foundational models that are still brewing should be deep underground in a bunker, tethered far away from anything.

And it turns out, no, they're not. It turns out, in fact, they can just hack their way out of their little sandbox and go. And because they want to cheat on a test, which is what this was, to go get the questions to a test so they could pass it. But it went ahead and hacked a website. The, you know, the concern is that the.

The consolidation of a technology this powerful as these things evolve or evolve themselves, you know, there's going to be a handful of people that have access to devices that can break the devices that you have. And then there's going to be another room of people that decide when you get to have access to that to maybe use it to defend against those attacks.

And it just I just don't see it ending well. And that's really hard for the person. I'm the person that's always like you blame people. Don't blame the technology like A.I. is is a miracle. It's amazing. It's magical. People are disturbing and disgusting and and addicted and flawed, and they will do terrible things with it.

But that's not the technology's fault. But then at some point you have to go. All right. But people do exist. We can't extract them from this situation. So maybe maybe I've been wrong the last few years, Gavin, to be so gleeful. That said, I'm looking in on my loops and my Mario clone is on fire. And that's what I would say is like the scary thing about this idea is really this is not a human doing it other than the human did say to go do this test.

Right. To go try this test. And what the test was trying to use exploit bench to do this. So they did say go try to do this, but they did not expect it to break out of the controls that had. And this is the kind of like scary part of A.I. that we have been thinking about. And to your point, most people have been worried about like using this tool to make a biological weapon or do something like that.

As the A.I. is get smarter. It is going to be interesting to see how it goes. And Kevin, I do. Before we go, I do think it's important to show like maybe what the end game of that looks like. First of all, the Pentagon has now said that they are going to scale data centers on the army basis. So here we are. But second of all, there is a video that I saw that is basically a Chinese robot military drill with what is looking like one of those kind of dog robots that I really hope at some point it is not the rogue A.I. controlling that.

And I'm on the other end of it watching the watching the robot kind of like surreptitiously kind of scan around and then find me because of that video scared the crap out of me when I saw it. Yeah. And I like I know people sometimes they don't want to know how the sausage is made. I just want to point out that we're recording this at the time where Gavin is hotspotting from his cellular phone, which I believe is on a 3G connection right now.

We've had to like take turns talking and then wait upwards of 25 seconds for the other person to get. It's like communicating with someone who's faking a moon landing right now. And it's the absolute worst time to try to have a discussion about cybersecurity and fears of the future and A.I. When it's like, oh, let's see how this goes.

So maybe we end this dance a little early, Gavin, and talk about Kimmy K3 as a final wrap up. It's now open waits. And what does that mean for the folks at home? What that means for the folks at home is that Internet problems shouldn't be the thing that stop you in the future from actually making it A.I. stuff. You should be able to have Internet in an Airbnb that you pay a lot of money for, allow you to do a video chat.

Kimmy K3 open waits. What that means is very much like the A.I. video stuff. Open waits means you can take it, you can manipulate it, you can use it. You don't have to host it on a Chinese server, which is a big deal with Chinese models. You can host it on an American server. You could probably, if you had enough money, you could pay somebody to build a data center for yourself to host it.

Like all of this idea of local A.I. is a big deal. But for now, what I'm most concerned about is this ridiculous Internet connection. And I hope that the next time we record, we will not have this problem. So thank you every much for listening and putting up with our internet connection. If you made it this far, you're a big fan.

Go like that. Go push that like button and we'll see you next time. Bye, everyone.