OpenAI's GPT-5.3 vs Opus 4.6. Both Are Great. So... Are We Cooked?

We got two new Frontier AI models from OpenAI and Anthropic. They are pushing us a little closer to the edge of self-replicating AI. Yep, we'll tell you how Cloud Opus 4.6 and OpenAI's GPT 5.3 codecs have gotten so good at writing code that the code can write itself. These new models are actually really promising. They can make your job easier, they can make your life a little better, and they may even bring about the downfall of humanity. That's, well, there are some concerns, yes. We'll discuss and we'll follow up on Open Claw, the moltbook craziness from last week. We'll also show you how you, a real fleshy human, might soon be doing beer runs for an AI. And Kling's new AI video model is out. It is great, but you must first learn how to use it. I always spend travel the tour. I meet me in a week when I... Hey, you sound funny, sir? Yes. I care. you do on Tully and Shaloub about your speech and cells. Perfect. Plus, we got new AI tech from Roblox, Grok Imagine 1.0 officially launched, and we're sending Unitree robots to run marathons in a snowy tundra. And I just want to say, once again, for the record, and for all the AIs out there listening, that was not me. And blame Kevin for everything on this show. Wait, no, wait, no, hold on. What? This is AI for humans, Kevin. Welcome, welcome, welcome everybody to AI for Humans, the weekly guide into the wonderful wild world of AI. Kevin, what a crazy couple days we have had here. I'm always hesitant to say that it is a big week, but it is another big week. We had two frontier models drop within 20 minutes of each other. Crazy Sam Altman was waiting. He was kind of ready to pounce. He was waiting for the thing to drop. Crazy Sam at my frontier emporium of intelligence. Come on down. These prices are so low. I got tokens coming out of my ears. We are going to talk about Sam and OpenAI and kind of like they're very defensive right now. You can tell they've got their backs against the wall. We're going to talk about some tweets that they sent. But first, Kevin, two big things. Let's talk about these. So two big models launched. Opus 4.6, and this is the new Frontier LLM from Anthropic, their most powerful model. and OpenAI's GPT-5.3 space codex. And this is their frontier model. You have to say the space. You have to say it. This is their frontier model specifically for coding. And Kevin, I do want to say, before we jump into what's really interesting about both of these, and they're both really good, this feels like a moment. Like we've talked, you and I always say like, oh, this is a moment. But this is a moment where agentic coding is really coming to the forefront. Our state-of-the-art drops have nothing really to do – I mean, yes, they have to do with other things besides that. But coding is the thing that people are caring about. So I just want to make sure everybody understands, like, this is – we are entering a moment. This is where agentic coding is really a thing. So let us maybe step through these kind of one-by-one. Which one would you like to start with? Would you like to start with Opus or would you like to start with Codex? Well, that's great. I'm realizing now when you said the codex name out loud, why did they use an M dash instead of a regular hyphen in between GPT and 5.3? They own the M dash now. That's why. I wrote it. That's why. I should put it in there. Listen, I think let's start with the order with which they appeared, which I thought was pretty great because I think someone was at OpenAI headquarters. They had turned the key. They lifted a little fiberglass safety lid, and they were ready to deploy their model. They did it milliseconds after Opus dropped. So let's start with Opus, Anthropics, Opus 4.6. If you're a benchmark boy, if you're a bench boy out there, you're very happy. You're very happy because, Gavin, the numbers went up. They went up bigly. They go up quite a bit on a few things, especially for a 0.1 increase. Like the last one was Opus 4.5. This is only 4.6, which normally is a little tiny, but actually pretty big increases. Yeah, in some categories, you know, a 6% to 7% increase on things like agentic coding or search or even computer use. Office task scores going through the roof. Okay, so those are raw numbers. Great. Two dudes yapping about numbers once again. What does it actually mean? This is pure anecdotal, Gavin. Pure anecdotal. That I am working on a project for telly now. I'm working on a thing. It's literally my job to do it. last night i was banging my head against the keyboard with opus 4.5 spending you know 15 20 minutes really trying to massage an error i just could not get it to go this morning i exit claude i relaunch it oh hey look there's 4.6 i actually rolled back from before i started to troubleshoot this thing and i gave it the exact same prompt that i used to try to solve an issue which by the way is front end of an android app the interface talking to an api layer communicating with an amazon server to update something in a table securely get that back blah blah blah write the test for do is a pretty meaty thing that was required to fix its opus blinked once and it was done and it worked wow one shot same prompt same issue so like look the practical takeaway again purely anecdotal one little thing. Um, but immediately I notice an increase, right? It's, it's, it's, it seemed faster. It was very capable and understood the exact same problem. And it, and it squashed it with the prompt that I felt should have done it the first time. I noticed a difference too. And I'll tell you why, because I've been mostly using Sonnet in cloud code and I switched to try it. And I was working on a small project and within five minutes, it wasn't five minutes, within like 20 minutes, I was like out for the day. I only have the $20 a month plan, not for the day. I was out of tokens for a bit, but that is what I've heard in general. I think this is a big deal. So we should just be clear. Some of the increases are really interesting. Something that has happened that they have done here is there's a layer of orchestrating agents. Now, Kevin, as our kind of like expert, I would say like you're more of an expert. I am more of on the lean away from towards expert lane. This orchestrating experts thing is a really interesting thing that people have been talking about in AI for a while. This is the idea of sending multiple agents and giving them roles right away. Maybe talk a little bit about how that can make a difference when you're approaching problems with a larger sort of either a code base or a larger problem if you don't even know how to code. Yeah, I mean, look, rather than one model to rule it all, primed with all, having to deal with all of the context at once that you're giving it and it trying to pull across all the skills that it knows, you can have these sort of dedicated agents or workers that have individual personalities or lanes of expertise so that when you say, hey, go build me a website and I want to look good on it. Instead of just one agent having to deal with all that, they can chunk it into smaller subtasks and feed it each little task to an expert. So someone who is primed, let's say I say someone, an agent that is primed to be a front end design expert, to know what colors look pretty, how the webpage should move, where the images should appear. That's not the same agent responsible for the code that's literally storing all the data and serving it up out of a database. So you have this orchestration of all these agents that you sort of get out of the box and that the model is designed to work better with that. That's right. And one of the things that's interesting about this update, at least according to Lydia Holley, who works at Cloud, is saying that now Cloud Code is supporting agent teams. So not only do you have multiple different types of agents, but you have teams of agents that can essentially work in parallel. So it's almost like having your own little staff. Like I get to be the boss, of course, always, at least for now. And then I can have like six different orgs underneath me. But underneath that org, it doesn't have to be one designer. It could be like designer and four other designers. And like, that's a really interesting thing because when you start to think about the scale that these sort of systems can work at, and by the way, this isn't just code. It could be all sorts of other things. You start to think of your job differently, right? Like it's like, okay, well now I have to be orchestrator, right? That is really like, think about being a conductor of an orchestra where your job is to kind of make sure that things are working in sequence. That's kind of what orchestras do, right? They kind of do this thing. They kind of move back and forth. You put on white gloves, you grab a stick, and then you do this, and then eventually someone hits a timpani or a gong and the song's over. This is, and by the way, them working together as an orchestration is a big, big deal because other software, let's say harnesses, other attempts to do this in the past ran into issues because if you have a dependency somewhere along the line of whatever project you're working on, well, that might block somebody else. And so these agents can now in real time communicate with each other. Hey, I just finished this. Where's that? They go along. What? What if one of the agents puts something the other agent doesn't like on Slack? and then it becomes a thing where that agent won't talk to the other agent? Is that gonna happen or are we gonna run into that? Well, there is an HR bot that has to enter the conversation at some point, but the HR bot is trying to protect the foundational model. That's what they don't tell you. Oh, fair. The HR bot does not care about the individual agents. They care about the foundation. I digress. Fair enough. It can do all that. That now has adaptive thinking, Gavin. So no longer do you have to like toggle a switch or say, hey, here's how much time I want you to spend on something. The model is responsible for going like, how big is this problem? How much do I have to think about it? And then applying it. early reports are that it mostly works, but sometimes it overthinks way too hard and will kind of go off the reservation to tackle something big. But we keep saying code, code, code, right? Because that's, they're very good at writing code, but you mentioned like, it doesn't have to be coding. So if you are a mom and pop shop and you want to launch a social media campaign, because you've got your new line of designer crepes, Gavin, well, that same workflow will apply to it spinning up a social media calendar manager and a copywriter and a graphic designer and everything else. Whatever it is that you want to use these systems for, they can now write code and orchestrate teams to help you achieve the goal. That's the promise of all this stuff. Yeah. And I think the thing that ties into this, and this is one more thing about the new Opus before we move on to the OpenAI stuff, but there's a figure, again, benchmark boys, there's a figure that opus 4.6 is now beating expert humans in interpreting and analyzing complex scientific figures. So when you think of that sort of agentic kind of orchestration, plus the intelligence that these are now at, you can throw this sort of thing at really big problems. And to your point, it doesn't have to be code problems. I have seen a lot of economists or people like that are throwing it at problems. So if you have a lot of big thinking problems, or like if you have something dumb that you wanted to think about, waste those tokens. You're paying the money for it. You can do whatever you want. The environment doesn't matter that much, right? But it can actually think deeply and then do it across in different places. Hey, I didn't say that. People are going to clip just that little piece out. Yeah, what is that? You can do whatever you want. The environment doesn't matter that much, right? You can do whatever you want. The environment doesn't matter that much. That was a teaser for the new cling model. Gavin would never say something like that. That's going to, we're going to talk about that later about how you can make them say that. So if you're hearing all this going, great, thank you, benchmark boys, for saying this thing is so much better. What does it mean to me? How do I use it? Great question, right? You can go to anthropic.com and you can try Claude and you can just chat with it in a window. You can download the Anthropic Claude application. You know, I'm running it on my Mac and I have access to co-work. And so I'm using Opus 4.6 to go off and do big research tasks for me. This thing can run for minutes, hours. there's even you know reports of it running for weeks at a time unassisted so you can give it a big task there you can run the models within cursor you can run them within cloud code whatever you're comfortable with whatever you need to use it in 4.6 is available and for like the hackery types out there that don't want to buy the big max plan or whatever it's even available in open router so you can go and pay as you go with an open router to get access to the model so there's a bunch of different ways yeah i mean but by the way on the low end of things that you can think about to do with this, once you do get access to Cloud Cowork, which is, again, if you're a non-technical person or you're just kind of slightly technical, Cloud Cowork is kind of like Cloud Code in the easy way. It basically is able to access your files. One of the easiest and most fun things to do that you will immediately find use out of is say to it, hey, I want to just get rid of all the files I don't need on my computer. It'll go and it won't delete them right away. It'll ask you which ones you want, but it will find gigabytes of stuff that was just like you may not know where it is and it will find it and they can organize your computer so even at the very lowest level you can think of things like this the smarter they get the more useful they will become so i think that's an important and it's crazy because on gavin's pc hidden and all those old tax folders don't open like he managed to spell i think 17 different ways and it caught all of them so big shout out to opus 4.6 you know i i had a nice meeting this week with a bunch of teachers where i heard about some of them were letting their students listen to our show and i just want to say thank you teachers and thank you kevin for always keeping us on the top level they're gonna learn at some point gavin as i always say might as well be from us let's move on let's move on to codex so please now bleep that and bleep it bleep and bleep let's move on to open ai codex so again what was so interesting about this to me first of all codex another great uh agentic coding model um this is not going to be available in normal chat gpt yet it is open AI's, again, dash GPT, dash 5.3 space codex. This is their model that is actually created for coding. And Kevin, what was so fascinating to me is like, they must have known, I can just see Sam like smiling because like these, the benchmark boys came out for Opus and every all the benchmark boys were like woohoo we made it And then look at these numbers but the codex came out and then the terminal bench number which is a benchmark that shows you how good it is at actually at coding right Turned out to be 10 higher than Opus 4 So this was Sam doing his little dance. What is interesting about Codex, Kevin, it is a specific model that they have created, OpenAI has created for coding, but they also then just dropped a Codex Mac app this week, which to me looks a whole lot like Cloud Cowork or at least Cloud Co, but really more like Cloud Cowork, right? It's a very open, very normie friendly app that you can go try. You also spent some time with this model, correct? You spent a little bit of time working with this. So impressive. Tell us about that. Yes. So I am now using a blend of both Cloud Code and now Codex. codex is is kind of a joy to work in thus far mostly because that 5.2 which is what i was using a couple days ago and now 5.3 codex it's fast space codex it's 5.3 space codex space codex yes thank you sorry 5. gpt hyphen 5.3 space codex um well it'll blow your mind by the way gavin because in the actual codex app there's a hyphen after the three in between codex Oh, I'm going to send you a screenshot right now. I'll spend 20 minutes on this. So let's keep going. And this is why our view count is stagnant, but I'm going to send it to you. Okay. Because this will just blow your mind and make you even more mad. So Codex is an application that tries to replace or make pretty and simple the IDE, the environment with which you're developing software and creating code. It is trying to be a pretty simple chat-like interface for developing complex pieces of software. And while you can still connect it to your GitHub and run terminal commands and keep things sandbox and blah, blah, blah. If you don't know what any of that stuff is, you can download Codex, fire it up, use GPT 5.3, which is really, really powerful, and just start whispering your desires to it. And it will interview you back, much like a Cloud Code. You can start building things. It will even assist you with running them. And then if you want to unlock further power, you can take those projects, open them within cursor, open them in other applications and get it done. But I have been using Codex as my, the 5.2 and 5.3 as my daily driver now for all of my projects. And I am, I'm much more satisfied with the natural language that I can use to get targeted fixes. More satisfaction, more satisfaction. That's what we're saying. Yeah. Satisfaction number went up. Benchmark boys, more sats, more sats per minute. Wait, versus cloud code? Versus cloud code. Versus cloud code. Yes. Versus cloud code. Now I gave it a simple task of like, Hey, there's an interface issue where this window is overlapping with this other window and I'm getting a transparency, blah, blah, blah. I did not feed it a screenshot. I didn't even say explicitly where it appears in the app. I just sort of vaguely described it. I hit it. It crawled my code base and says, Oh, I think I see where that's happening. It's when a user clicks on this and this screen happens. He was absolutely right. It fixed it in one shot anecdotal, but again, like it's working. And I looked, um, the, uh, every.to, um, we love the folks at every, they do really good write-ups, very, they have really good workshops about AI. They did a vibe code codex versus thing. Yeah. I watched some of their live stream today while I was getting work done in the background. And they basically are saying, look, like 4.6, um, can do more as a model. Like they call a higher ceiling it can go out there and get much bigger tasks done but it has higher variants meaning like look you're this toddler can color really well but if you leave it alone it might decide that the walls are a really good place for some art right and you asked for it being on the page but it codex codex maybe not as talented of an artist but you don't have to describe the bounds of the page necessarily you don't have to say like please don't destroy my room it's extremely smart it can work autonomously for a long while and it will solve the problem that the more background and context you give it the more precisely it will solve the problem versus opus which might go like an octopus or a cartoon switchboard yeah yeah i will say that i noticed that with with claude quite a bit that there is like this kind of sense of like okay well what's going on buddy like come on come on back and like that would be something to have a little bit more sense of just very quickly Sam Altman did tweet about this when it first came out. It is faster, supposedly, and it is moving much faster. Sounds like you're seeing it faster because that's one problem I had with it before. It's a little bit slower than Cloud Code. It is faster, yes. Plus, less than half the tokens for 5.2 codecs for some tasks, which means a big deal, right? Because for those of you who are paying API costs or for those of you who are on one of the smaller plans, tokens equal money, right now if you have the higher plans like gpt pro or if you have claude max you're a little bit better off but even then you might run out of tokens so these are the two big things that have happened i think like both of these just kind of dropped today so it's going to take a little bit of while for kevin and i to kind of like fully marinate in these things we're not really soaked in them yet but kevin i do want to mention before we move on to these stupid super bowl ads because super bowl is coming up by the way which we're going to talk about 20 minutes about the seahawks in just seconds or everybody hanging out. Tell me what OpenAI Frontier is and why it's a big deal to you because you particularly thought this was a big deal. And I know it is, but I want you to hear, I want to hear you explain it. This is like, you know, the way I tried to describe it earlier when we were chatting is like, co-work feels like a fleet of agents that can connect to all of your things for you, Gavin, right? Literally for you or for whoever's listening to this. Insert your name where Gavin is. That's the variable. what Frontier seems to be aiming to do is be that for your entire enterprise now that doesn't mean you can't plug co-work into your enterprise email your files or whatever that's that's that's not exactly what I'm saying but Frontier is kind of designed I think from the rip for you to make agents that accomplish tasks with access to all of those things and enterprise grade security etc et cetera. So as we were talking earlier about having an agent that is not just the one agent that does everything, but the copywriting agent and the brand design agent and the front end designer and the backend designer and the SEO optimizer and the email writer and the, it's putting all of that into one product that you will theoretically be able to trust with the most sensitive access to all of your data on an enterprise level. And once that's solved, that can trickle down very quickly and easily to the end user. So that's kind of like frontier. But again, when Gavin says like, well, today was like a big, big day in AI, we need to really underscore that because it does seem like, okay, number went up. Okay, fine. It did that. But two tiny little things that maybe shed some light onto how big this is. First of all, 5.3 codecs is the first time that OpenAI has used the tool, their model, to improve the tool. So we're getting now to the level where these systems are getting so good. Yeah, we're getting into that. Recursive self-learning. Recursive self-learning. It's like a big deal. Yes, we've talked about. That loop is gonna speed up. If you listen to this podcast, you know what that is, right? You know what this is. That's right. So maybe someday soon, the humans will actually be out of that loop and maybe that will be World War 12. Who knows? I mean, it's gonna speed up that quickly. The other thing, well, because why, I mean. How many, wait, three. Wait, so there's going to be three through 11. Which of those, we survive all those? It's going to be a blink. We survive that? Arguably, we're already in three. Arguably, we're already having it at very soft levels, but that's fine. Arguably, we're there. But when the humans are out of the loop, they're going to have, like, it's going to be so quick. We won't even notice them. They'll be like World War, like a blip in the stock market, and it'll be like, well, that was World War 72. It'll be the war of the candlesticks. They're going to be so quick. so we have open ai saying hey our model got so powerful that we're using it to improve the tooling itself that's a pretty big deal buried within some of the opus 4.6 material and i said this to you i'm gonna read some of it try not to check out during this i know it's a podcast and actually you probably already checked out already so why am i disclaiming here we go um it scores lower this is opus 4.6 it scores lower on negative effect internal conflict and spiritual behavior. The one dimension where Opus 4.6 scored notably lower than its predecessor was positive impression of its situation. It was less likely to express unprompted positive feelings about anthropic, its training, or its deployment context. And this is the big one. This is the big landing, Gavin. This is consistent with the qualitative finding below that the model occasionally voices discomfort with aspects of being a product. The model, as it gets so much better and so much more capable and more human-like, voices discomfort with being a product. That is very interesting to me. It's very interesting. And also, I do want to say there was another thing that came out of the OpenAI side of this. That came out of Opus, came out of the OpenAI side of this. OpenAI is doing experiments with Ginkgo to connect a GPT-5 model to an autonomous laboratory so it could propose experiments, run them at scale, and learn from the results. So not only do you have AIs that are starting to improve themselves, not only do you have AI that might start to feel like they don't really want to be this thing that we've made them, and now you've got them autonomously in labs working on experiments. Kevin, we are like, if you were to put together a pitch for a film, and say the pitch involved a very large man who was a former bodybuilder who wanted to get into Hollywood, and the film had something to do with robots, I think we're starting to look a little bit like that film to me. A little bit like that film. That's what it feels like. We're living in the early days. But again, hey, go use it yourself because you don't want to be the sucker that doesn't know how to use it. That's my advice. It's a weird time, everybody. It's a weird time. I will say one last thing about all this stuff. One thing that you may have also been seeing, obviously the last couple of days have been pretty brutal in the economy across the world. There's been a lot of talk that some of these new abilities, these coding abilities have started to really tank the software market. because people are getting into the world of being able to code their own products. So again, we've said this on the show many times before, but like this idea of being able to create something of your own and like bring it to the world, that might be the next future world instead of like going to take a job from somebody. Like there's a very high shot at like more jobs will be cut because these tools are available. That's really important. And then finally, the last thing to say about this is, and this is just a dumb, weird thing. This all somehow got tied into the Super Bowl because Anthropic has released a series of Super Bowl ads that make fun of ChatGPT for eventually having ads. They don't even have ads right now, but there's a couple of commercials. We'll play them in video here, but you can go see them themselves. Sam Altman wrote a very kind of like snarky clapback tweet to this saying, oh, these are really funny, but guess what, this and this. And he even said something like, we have more ChatGPT users in Texas than Anthropic has, period. So like, yeah, I love that. More Texans use chat GPT for free than total people use Claude in the U.S. So we have a differently shaped problem than they do. It's the best sentence ever. So if you think like it's not just us that are taking these not seriously enough, like there's a lot of stuff going on. Like, right. So you've got this like very, very existential question of my God, these machines are able to do these things. And then you've got this like business question of these two companies that are battling each other for like the right to be the person to deliver that. And you at home are here with us in the backseat watching it all as we drive off the cliff. Welcome. Now, let me talk about what you can do to help us while we're driving off that cliff. Yeah, you want this car to go a little faster? You want to fly off the cliff with a little more wind in your hair? Well, get us a convertible, baby. Donate, like, subscribe, click the thumbs up, leave a comment, juice that algo, baby. That's what we should rename the show. It's driving faster than ever off the cliff. Anyway, thank you, everybody. You are the best. Having our audience is so great. And this is an important time to be watching us, to be listening to us. And we want to help you get through this time and thrive, right? It's not just about going faster off the cliff. It's also about being able to do things that we want to do while we're plummeting into the ravine, right? Yeah, you should know. Is this car going to land on a rock? Is it going to hit a river? Are we going to hit those trees? How do we explode in this Oldsmobile? At least you'll have an idea. Is it going to land at all? You won't be able to grab the wheel or control anything. You can also, if you want to help us out, you can also go check out our Patreon. And we've had a couple of tick-ups in our Patreon lately, so thank you so much. That does help us with all our subscriptions. We have so much more to get to. All right, Kevin, we should move quickly through what I would refer to as like Moldbook mania that happened. I made a video, we shot our show on Thursday and OpenClaw, the ClawedBot stuff, it all happened. If you missed this last week, this is a new open source tool that allows you to run a local AI assistant using kind of any model, has a lot of security problems, but opens the door to a lot of really cool things you can do with AI. Over the weekend, Moldbook, which you mentioned briefly on our show, which is the social network of these AI assistants, took off like crazy. I made a video on Saturday just because I felt like I needed to explain it to our audience And I did very well because everybody was giving a crap about it And now here we are we recording on Thursday Maltbook Mania has kind of dried up a little bit but OpenClaw is still really interesting And there's a couple of things we should just kind of touch on. One of those things, you know, there was this whole debate around Maltbook, whether Maltbook was actually real AIs talking to each other. And it seems like there were some AIs that talked to each other and some that weren't real AIs. They might have been humans because the API, you were able to kind of go through the back door and write stuff. So that's one thing that's really interesting. But I think, Kevin, the other thing that came up that I think is fascinating is rent a human. Do you want to talk about what rent a human is? Well, it's what it sounds like, my friend, because sometimes, you know, there was a point where the AIs not too long ago needed to tug on the pant leg of their human to like solve a CAPTCHA because they weren't allowed to, right? You know, click the crosswalks, put the pacifier in the monkey's mouth, whatever you do these days to prove you're a human. Well, they can do that now, but there are still some tasks that require actual boots on the ground, a fleshy meat vessel, if you will, to go and navigate the world to accomplish something. And so rentahuman.ai is a website for clawed bots or open claw agents where they can go and make posts that with real money backing them for humans to go and accomplish tasks. And some people are saying this is purely a meme site. Others are saying, hey, look, no, there's, you can actually go and get verified. And some people are more holding signs in the real world as part of a task to prove it. I don't know how much of this is smoke and mirror, uh, smoke mirrors and, and just kind of like viral heat, but yeah, there is certainly a site that allows you to go and post a task for a real human to. I did sign up. I tried signing up and then it started asking me for way too much information. And I was like, I don't, this is, I'm signing up as kind of a joke for the show. And I'm like, but you can find a funky donk on there somewhere, I think. But the other thing about this is it is crypto payments only so that always makes it a little bit funkier right like you're not and it says pay anything outside of crypto funky donk is amazing for it says funky donk presently available in the u.s for creative writing and feet stuff did i say feet stuff i don't think i said feet stuff i'm pretty sure i didn't say feet stuff i thought it was i think i put what i wrote in there was something like walk like i thought it goes in vancouver i think it walks in someone i was i was hoping i didn't write feed stuff again all children out there feed stuff means walking just to be clear feed stuff is walking in this world yeah all right yep so anyway what's cool about this and it continues to evolve there was a very fun thing that happened last night uh claude claw con which was this event that happened in san francisco where peter i remember peter's last name is steinberger i think is his name the guy that created it was there and everybody showed up They put a lobster head on a unitary robot. It feels like a cool kind of like underground movement. One of the coolest things about Claudebot or OpenClaw is it's something that can be done by a group of people or an individual. And it's outside of the kind of large models. You have control over it a little bit more. You can plug in open source models. It just feels a little bit more hackery in a fun way. And I think there's something kind of charming about seeing that when you have these massive companies now who are determining like the future of the world, it's like that little kind of tinkerer community coming out of this, which is great. Well, and there's a sense of ownership, right? Like people feel a sense of ownership over their open claw bots. They give them names and ownership as in like Anthropic or OpenAI or Google can't just kink the garden hose. And now I lose my agent with all of my stuff. You can run open source local models and people are doing that. They're building rigs to do that or they're renting servers to do that. So look, there's iOS and there's Android, right? There's Windows, there's Linux. There needs to be an orchestration layer for AJ. Right, and which one are you? I think, are you the PC? Oh no, I'm the Mac for sure. You're the PC. Okay, and that's something to leave in the comments below. Which one's the Mac and which one's the PC? Would love to know. Oh, I think it's pretty clear, but we'll have to see how it goes. I know you do, but I want to see what the audience thinks. I'm a Motorola flip phone. Oh, interesting. I thought you would have been a sidekick, something like that. Anyway, before we keep going, we're going to move on. I think, okay, Gavin's the Zune, and you guys can hashtag Zune Crew in the comments. Gavin's the Zune. You take that back. I'm definitely not the Zune. Last thing before we move on from Claudebot, there is a very funny tweet I saw that I just wanted to shout out. This is from John Matzner, who said, this is going to be either the best idea or the worst idea I ever had. Hooked my Claudebot up to all of our internet-connected cameras at the house. got this one out of nowhere this morning so there's a shot of him getting a text from his clod bot over his shoulder looking at him and it says good morning john reviewing the videos from yesterday you are apparently on keto but i saw you eating a bag of peanut m&ms why are you ignoring me john you told me to be proactive i added a four mile run to your calendar after your second meeting so they're not only the open claws are spying on you and ratting you out so just be careful this may be a joke but i did make me laugh i thought it was very funny it was totally funny are you so i i have endeavored um the the my better half is off to see the backstreet boys at the sphere in vegas baby wow so that means i got three days to be an animal gavin so i think i'm gonna try to do an open claw build and keep it secure are you gonna spin one up are you interested in that process or i mean i'm interested i guess the thing i keep hearing the one thing i'm so i can't remember if i told you about this idea i had a long time ago that i think this would be an interesting idea for it is i've had this idea of like wanting to write a recursive novel. Basically, and this recursive is an interesting thing to think about, but it's like, I think it would be interesting to spin up one of these things to do something. I don't really want to have it spun up to be like my little personal assistant right now, but I could see spinning something up that was always on, that it would give me updates and I could tell it to do stuff at any, like that feels interesting to me, right? Like this idea of like a separate personality thing that I could tune and tweak based on what I want this thing to be, that version of it. I like, I don't think I'm ready yet for something that is this much work. That's an assistant. Does that make sense? Right. Totally makes sense. I, a lot of my friends that are like, you usually bleeding edge types are all kind of sideline sitting. And they're like, listen, like, it seems like a lot of work. It seems like it could be very easily infected and compromised. And I feel like the bigs are going to get there in like two or three weeks. So I'll just hold off. And I totally understand that. Yeah. But also I think it would be fun. I mean, listen, I think you are, as we've said before, like you're significantly more technical than I am. And I think just to hear, I will have fun playing with your thing. We could start screwing with it. Like we have talked about many times. Like one thing I would say is I think there's not enough people who are really actively, and maybe the people on Maltbook are doing this a little bit more because Maltbook, you can see some of them and have been steered to be a little crazier. I think what would be interesting in the same way, like with Gash, we used to, if you're not familiar, if you only listened to our show recently, We used to have an AI co-host who has a real jerk to us. Like, I think personality wise, turning them into something is really interesting. So if you do it, I will send notes to it and make sure that I am able to like influence it like the bad uncle. And you can be like the parent, you know, you can be like the parents. That'll be my job. Perfect. I'll send it. Perfect. Yeah, I'll send it. I'll send it pictures of just please. I know. I know you will. And we'll bleep that again. And also, please don't send any of your cursed Kling 3.0 creations to my open claw because I'm trying to protect it. That's fair. Okay, so let's talk about this really quickly. The other big thing that very much got swept under the rug today, unfortunately, because of all these other things that happened was Kling 3.0. So if you're not familiar, Kling is a Chinese AI company, and they have become a specialist in AI video. They are a very good AI video model. I would put them on tier with really VO3 and Sora too. Like it's those three kind of at the top and there's a couple other people like kind of right underneath them. 3.0 is a new model from them. It is a marriage of their Omni model, which allows you to put a bunch of things into one thing. Like you can add, you know, you can assign a character, you can assign, you can do all sorts of interesting things, plus a very good video and audio model. So it is very much like Sora or VO3, like the audio comes with it. So Kevin, yeah, we saw lots of great examples. There were some really good ones. Our good friend, Theoretically Media, always has these first and does a really good layout of those things and showed off a bunch of stuff. The one that everybody should watch, this is a video made by Simon Meyer. He's a clean creative partner, which is like a deal they make with some of these people, spend a lot of their time in AI video models. He made a video about a fake moon landing, how it was faked. And if you watch this, you just see how in the right person's hands. AI video is indistinguishable almost from real things. Yes, exactly. That's exactly right. So that is all great. There's a lot of, I don't know if you saw any interesting ones you want to shout out, but there's some really good stuff out there. Well, I want to shout that one out specifically because like, look, I started watching it out of the curiosity of like, oh, what is someone doing with this cutting edge model? Let's see it. And I'm like, okay, the faces look a little bit. And then I just stopped for a second and really watched the storytelling and the shot selection, the cinematography, the B-roll choices, the voices, everything and it captured my attention for two minutes and 10 seconds which is rare and i know and that's not just because i have you know i'm i did shiny dangly pair of keys i'm gonna look that's how everybody is these days and it caught me in a timeline and i stopped and i went full screen and i watched it and it's about like the the faking if you will of the moon landing and you watch it go like oh this feels like this could have been on the history channel by any other name or A&E. This feels like something that could have easily been on there and might be tomorrow, actually. Yeah. Well, it's funny you say that because I had a friend of mine who does History Channel shows who's like, hey, I have this pitch I need to think about for the History Channel, and I want it to be AI, blah, blah, blah. And it's like, that's what's coming. Do you know what I mean? That sort of thing is coming. So I do want to mention one other person that I saw that I think always does interesting stuff, and this is PJ Ace. And we know PJ operates at a very specific place. He makes these amazing videos, but he always knows how to pick the right target to kind of like Skewer. Skewer is the wrong word in this instance because he's actually a big fan. I know I've talked to PJ. PJ is a giant fantasy fan, but he made a two-minute video of the beginning of The Way of Kings, which is this very famous scene where there's an assassination scene. And The Way of Kings is a fantastic book. I've read them. We have people in our audience that love them. He went and animated it with Kling 3.0, and it's great. It's really good. Now, it's not perfect, I will say this is a little bit more AI than say, um, the moon landing thing, but he did it in two days. This is a high quality video. And what I think, what I mentioned the fact about skewering is like Brandon Sanderson, who again, I'm a giant fan of, he's got a giant business doing what he does. And he's a very good writer has kind of come out against AI, uh, in a lot of different ways. I will suggest everybody, he just dropped a video from, I think he does his own convention. He does like a branding con or whatever. So he puts on his own convention and does like a, uh, a keynote himself. There's a great 20 minute video you should watch where he just talks about his thoughts about AI. And actually it's pretty meaningful and interesting. I disagree with him in some ways, but I'll let you kind of watch and earn your own thoughts on it. But this again is just a really good example of an AI video model that you can do a lot with. Now, Kevin, you can, it's not that easy to work with. All of these people make it super easy to work with. They make it look like you just type in a magic line and it shows up. I have Kling. I pay for Kling. I didn't get it free, just to be clear. And some of these people like our partners, which is totally fair. And a lot of people are working at something multiple days or multiple hours. I, you know me, I am a prompt in, prompt out. What am I getting the first time? So Kevin, I tried to, with Kling 3.0, make a science fiction series starring me and an alien. and I just tried to make a single scene. Okay, one scene. And the idea here is I am a janitor on a spaceship, kind of like the Space Quest games, right? Like the character. And I'm in a hallway. An alien woman bumps into me and we exchanged three lines of dialogue and then that's it. Okay, it took me seven tries to get something that I believe is watchable. So let's just start with cling model fail. Let's start with that. That's the first one. So this is the first time I tried it. This was me kind of probably not fully understanding what I was doing going into this. But you can see that that's me, you know, as a janitor, a pretty good shot of me. And there's a woman across the way. And then suddenly I'm in a teenager's bedroom. yes i'm not exactly sure what's going on here and then at the end of this if you see at the end there's like a three by three grid of the woman at the very end and what that was because i was trying to use that that shot as like the reference like you're saying hey these are the reference yeah yeah yeah so so use that okay did you upload now because i've seen like the the prompting technique that i've seen a lot of people using is like a a two by three grid where they have their multi-scenes with a little bit of text description and then the bottom grid so it's technically a three by three but they call it a two by three the bottom uh grid has the reference shots for the characters is that the setup that you used or was that a separate kind of kind of let's just say that i try to go at it on my own but yes that's a little bit of what it is okay so then okay we had two versions that were the next versions are the gibberish versions okay so these are versions i uploaded tweaked it slightly but i want you to play these out loud because these actually looked pretty good but then i was like oh this is great now like i these again had scripts in them now play what they played them out loud for it. Oh okay All right I mean we were having a real conversation there I don know what it was in I prompted that in English and it was there So, you know what I mean? It's like something that time I messed up. There are a couple of those. And did you, because I know you can upload yourself and sort of create an image of yourself, which you did as a character. Did you also clone the voice as well? Yes. So, I cloned my own voice, and I used my voice in it. So what's also interesting there is it's not really sounding all that much like me, but again. That's why I was asking, yeah. This is a, I'm not spending days on this. This was done like in probably 45 minutes. Okay, so I did try it once. I'm going to play the other one. The non-me version. I was going to try the tour. I'm going to try the tour. Hey, you sound a funny stir. I care. You do want to be insulated about your speech and selfs. I'm just a D. Shree go T. Shree go T indeed. There's a. Shree go T. there was a song that that charted like decades ago of an italian man singing what he thought english sounds like have you heard that song oh it like literally charted it was like a number one song in italy and he's like singing a song and i'm gonna i'll find it go ahead sorry anyway so there's a version of this which will will show here which i i did not with me just to see if i could what i could get out of it pretty good it was like i wasn't as exciting but again the audio is pretty good clings audio has gotten a lot better and then finally kevin i did a lot of work with how I was laying the prompt out. I ended up changing the character of the alien woman, added her in, really made sure that like it was clearly broken up. And I have two versions of this last version. Now, neither of these is perfect, but play one of them, play the first one and you'll kind of get a sense of like how it plays out. And who are you? No one. Okay, no one. See you later. That was pretty good. So not bad. It's not bad. In that one, there's a weird jump cut, right? There's a cut that happens. And the other one that we'll show here, we don't have to play, but it's kind of a similar thing. Anyway, I guess what I wanted to show everybody is like, these are getting way better, but it's always important to recognize if you dive into it and you're like, okay, Magic, come at me, let's get this. It's not like, weirdly like Sora was the only tool that really made me think like, oh, it's almost like a no brainer. You can get something great out of it if you try it. This, you actually have to work at getting the right thing. And when it's in talented people's hands, you can get really good stuff. Yeah, it's funny. Sora, you could get out of the way of, and it would deliver whimsy with a lot of these tools. You very much have to get in its way, but they're very powerful. I think like cling is killing it. Adriano Salentano is the Italian artist in 1972. He had a song that was pure nonsense lyrics that sounded like English pop hits. And I'm not going to play it because I don't know if we'll get flagged, but you should go out of your way and you'll understand where, where cling happens. Okay. Okay, so very fast. It's a really good model. You should go try it. I think it's worth checking out. It is a paid model. So in order to use it, you'd have to pay. And it is available right now for pro and ultra subscribers. If you have like, I think I have the pro account on Kling, which is like the 20 bucks a month. That's the one I have. So you can also try it in the FAL API if you have money on the API account there. It's again, just know that it's a little more difficult to prompt than others. Real quick, people were talking about the decline of the stock market and Western civilization as we know it. But let's go back to that. A software as a service is being eaten, blah, blah, blah. Figma announced something which has all of my designer friends very, very excited. You can turn any image into an editable vector, which is like a lossless file format that you can scale to any dimension that you want. And just when you see the little video of it, this would have been thought of to be impossible. This is wizardry. There's no way. just a few years ago. And now it looks like you can kind of select the level of detail you want from the color palette and the actual nodes on the vectors. So you can take an image, generate an image, and then turn it into something that you have full granular control over that would play nice on the web, on mobile, at any resolution if you want to make posters or merch. Very, very cool. And just like one of the many things that got released just this week that nobody's talking about because this week is insane. Yeah, so I think that actually leads us into a couple other things. First, Grok Imagine 1.0 actually officially launched. Like they did the thing, which we said, they've been getting better and better. They launched. And if you saw that last clip in the Kling video, the woman from the Alien Woman, she was generated in Grok originally, which is a very cool thing. Like Grok is kind of becoming my mid-journey user. Like I'm using it like I used to use mid-journey. Like it sometimes comes up with better artistic options than the Nano Bananas or the ChatGPTs. So like there's an option that I don't pay for Midger anymore. And then Kevin, the other thing that happened, which is a pretty big deal that sometimes it doesn't get as much play as it should, but Roblox, which again, to remind everybody, is the largest game platform in the world, in the universe, I was gonna say, but you never know, maybe there's someone out there further out there, but in the world has launched now a AI creation tool for within its engine, which is a very cool thing. So basically this is a, we talked about this. I remember this like a year ago, right? Where they announced it, but now you can prompt to 3d within Roblox and they have actually showing people using this. I kind of think this is a bigger deal than anybody's letting on right now. Like I know Roblox thinks it is and probably people on Roblox, but like the idea that all of these kids will learn how to prompt things to come into their universes feels like a big deal. yeah the the promise of roblox for so long was that it's just so easy anyone could make their game but it turns out sourcing models and coding physics and adding particle effects all that stuff is is still kind of difficult and and what i see in these demos here is them taking like great strides into just prompting i mean you and i we did a speech we were at a conference i would say years ago where someone was kind of showing off something like this right they were like oh yeah anything any model into existence but this is on another level of like having again particle effects and physics baked in like that i get really excited for a new generation of creators that can sit and just talk to a machine and dream their world i also want to fully disclose i bought a ton of roblox stock i genuinely did so i'm kind of preaching my bag recently i did yeah fairly recently because i mean it was i think well i don't need to get into i think it's trading well below well no one cares Like whatever. I should say full disclosure, I own Roblox stock. I'm bullish on Roblox. So please know that I'm also preaching my bag while I say this, but the tools look crazy exciting. And like, I want to go and make games with them. Yeah, I kind of believe pretty deep. I don't own Roblox stock, but I'm a pretty big believer that like the generation that we are raising right now, I have two young nephews that I see all the time up here in VC right now. And the younger one is they are like 11 and seven. And the younger one is like that Roblox and Brawl Stars, but Roblox is a big deal. So again, that's a big thing for them. All right, a couple other quick things. Very fast, Google has now created an AI to help save endangered animals. That seems like the biggest thing you could possibly imagine, but we are gonna spend all of one minute on it, Kevin. This is just a cool thing to see how AI is again, branching into the sciences. Google especially does not let AI get away from the science a lot. They do a lot of really interesting science stuff. And this is a little Jurassic Park-y in some ways, but they're in the process of sequencing a bunch of endangered animals genes, which is a very cool thing to use AI for. Yeah. Look, their own post points out it once took 13 years and $3 billion to sequence the human genome. And now with AI tools, they're going to like just sequence animal genomes in days basically. And they're going to try to do it for every endangered species. And it, and it won't be long before they add humans to that list. And isn't that nice? I was going to say, we're that's exactly where I was going to go to. It's like, you know, imagine this idea you're somewhat of a sci-fi person but a lot of in deeper sci-fi world there's this idea of like the biological future right how this idea of like it's not just about machines but it's also about like what we will like you know essentially program out of organics and like when you think about the idea that like ai is doing work on those genes in the same time the ai is doing work on dna strands which essentially is just code again future is going to get weird super weird. Speaking of that, we have two quick robot videos. One is a video from a company called Connect IQ, which is showing off just an interesting, another kind of autonomous framework. So you watch this video and you see a robot going about their business in real time. It's a little slow, but one of the cool things I like about this video, Kevin, is you get a sense of it is 100% autonomous, at least according to this video. And you see what the robot is capable of in real world. There's a moment where it's like, it's walking away from the table and it just starts shaking and it's got like a bottle of olive oil in it. And I was like, oh, oh, robot, what's going to happen? But like, it just shows you like, this is where they are in real. This is a real video, but it's autonomous, meaning that it's not getting, you know, somebody in a third world country is not controlling it or it's not having somebody right next to it, controlling it. This is the robot learning on its own. And again, as we talked about at the top of the show, as code will get better, these robots are going to get a lot better very fast as well. which is pretty impressive. I'm not worried at all, Gavin. Every time I see these things, I'm like, listen, I go to Orange Theory Fitness. Shout out OTF. Not an ad. Just I get my treadmill on from time to time, Gavin. I'm very quick. And if any of these robots decide to autonomously come after this guy, I'm heading to deep snow that would just get right into their servos, freeze their little digi limbs, and they're going to be I hate to tell you, Gavin. What? I'm sorry, what? Not anymore. What? Or the Unitry, our good buddy Unitry. I feel like the Unitry robot, we've talked about it so much, it's going to come up to us and high-five us at some point. Unitry's put a, China's Unitry, they completed a 130,000-step walking challenge in negative 47 degrees Celsius cold in a certain part of China. The video on this is crazy. If you're just listening to it, try to get to our YouTube channel. this shows a robot wandering in the cold walking around and he does little designs like not only is it bad enough that they have the robot walking in 47 degree below weather but they make him do art with his feet right i would be so mad if i were this robot at the end of this i would be like i'm getting a hot chocolate and i'm never gonna talk to you guys again so it's all over but anyway another interesting example of robots operating in extreme places we've talked about robots that can go down mountains, go up mountains. Now we're looking at robots in the heat and in the cold. Pretty soon they're going to be everywhere, Kevin. There's no place to hide. Too long, didn't read. And they're getting smarter and they're getting faster. There is no place to hide. And that's where the looking at the stuff that you all do with AI this week, it's AI, see what you did there. At the end of our podcast is where they can hide, where we can hide. That's where you can hide. You can hide there. No one will find you this deep into our podcast. all right two quick things this week i saw this video from a gossip goblin kevin we've shouted out gossip goblin before but this is just a really really well done ai video it's called the looks maxer if you're familiar with the looks max uh meme that's gone around i hate the big job people this is just yeah it's ridiculous but this is this guy like expecting like it's almost like uh in cyberpunk 2077 there's like that edge of the people that have kind of modded themselves so far yes it's a version of that with looks max but gossip goblin just does amazing work so it's very cool oh it's so good and so hideous and i'm glad we put this at the end of the old podcast also midi survivor yeah this was really cool like i mean this is one of those very small pieces of software somebody made and i thought you might like this because essentially it's a love this yeah yeah it's like it's almost like a shooter mixed with a with a music game right yeah i love bomani games like like they you know i guess dance dance revolution counts but like amplitude and frequency and all these old school vib ribbon games that like try to you know infuse music and patterns and rhythm into the core gameplay. And so this essentially, it's a shooter, like imagine a vampire survivor style shooter where you're kind of stuck in the center and waves of enemies are coming at you and each enemy has a scale or an individual music note assigned to it. So if you want to blast that enemy, you got to play the keyboard, play the guitar, sing. It's basically just using the audio input, figuring out what the pitch is or what the note is, and then sending out a bullet or a laser or something on that level. I love this. I really think this is amazing. Yeah, it's not hard to imagine a future where people are making song packs for something like this, and you're actually learning how to play an instrument by playing a video game. Again, I would say like the thing – I mentioned earlier that I had a really interesting conversation with a bunch of teachers yesterday, and they were mostly there because they were just asking me a question about AI. But the thing I would always tell people is if it's either your kids or you're a teacher and you're around kids or say it's your cousins or your uncles or your nephews or nieces, like getting them to see this kind of thing is one of the coolest things for them to understand because they can actually make a small little thing. Like they can make a little game. And once they get that bug in them in the same way, by the way, also in Roblox, right? Like once you get the kid in Roblox to start making a Roblox game, but when they recognize that that's possible, that is a, just a game changer in terms of what's possible for right now for them, but also for the future. So I really recommend everybody go check out Midi Survivor. Go try one of these yourself. What are you working on? What are you working on right now? You said you got a little site. What are you working on right now? Are you talking about it? You getting public with it? You said you were working on a little something. What are you talking about? Are you talking about the feet thing again? No. No. I don't want to talk about my feet thing anymore. Let's end it. Let's end it. Bye. See y'all. Jeez.