The last 48 hours have been absolutely packed with massive rapid-fire AI releases. First, we've got OpenAI's GPT 5.6 Soul, which is close to Fable 5 in almost everything. Today, we are releasing our latest and most capable models. GPT 5.6, Soul, Terra, and Luna. Plus, Mark and Meta released Muse Spark 1.1. It's their new coding agent. while Elon and the Cursor Kids, they dropped Grok 4.5. And just like my pal Gavin, let me tell you, it is fast, it is cheap, and it is surprisingly capable. Thank you, I think. And OpenAI released GPT Live. It's an update to their real-time voice model. And let me tell you, the Abuela bench scores are through the roof. The grandmamas love it, Gav. Well, Kevin, let me just let GPT Live, the new voice model, do this part, okay? There are rumors that GPT 6 might be here this month. Is that why you want me to say? Yeah, I guess that's the line. Yeah, that is the line. Nice work. Wow, how exciting. They're already training your replacement. Oh, also, Reactors LingBot World 2 lets you live stream your own virtual worlds, which means, yes, you too can be a walking hamburger on Mars or inside of a 90s shopping mall. Or how about this, Kevin? We'll have all the details about how and why George Costanza became the AI meme of the week. Wait, you're not gonna... I am, Kevin. That's right. Okay, this is... AI for Costanzas, baby. Let's get going. Welcome, everybody, to AI for Humans, your twice-a-week guide to the wonderful world of AI. And guess what, Kevin? We have a giant, giant day and week in AI. We're going to get into all the releases, but the big one, the big daddy, is that OpenAI has dropped what we believe is at least their next big model. There may be a little bit more to tease in just a second. we'll talk about, but this is a big deal. This is GPT 5.6 sold. This is their brand new flagship model. And boy, oh boy, are we excited about it. We're back. My camera had a momentary screw up here, but we are back. Kevin, GPT 5.6 Pro is very exciting. The benchmarks look as we expected, very, very good. Let's talk about what some of the big news here is. There's some pretty big updates to the product of chat GPT. I think that's the biggest thing is that the new GPT work app is here and it's going to take your existing chat GPT app that you know and maybe love and this other app called Codex which was seen more as like a agentic coding like go ahead and build software with it but it was so much more than that you could connect it to calendars and productivity apps and it could read your google docs etc etc and they're smooshing them up you know in fact the live stream announcement video was almost like a mini Pac-Man happening. And then they were going to eat each other, which is, well, it's a choice, but that app is here. And people are basically saying like, oh, they're going the Anthropic Cloud desktop route, which is not the wrong way to run in unifying some very, very powerful features. And as the vibes were predicting, Gavin, the tea leaves were being read by the vibesters out there. It looks like the 5.6 line of models. While they might not be as mythic as Anthropics Fable is, they are reliable, they are quality, they're not that crazy expensive, and they're a joy to use. You want these as your daily driver models to handle your calendars and your syncing and maybe even some of your coding requests or your business management or messaging your loved ones that you're sorry that you've been so absent lately, Gavin. All right, we're going to do a quick look at some of the basics here. 5.6 Pro Soul is out along with Terra and Luna. Now, Soul is only available for paid users. Terra and Luna will be available for free. That is kind of the medium and low. It's kind of imagined in the Claude world, that's the sonnet and the haiku. There's some really interesting details here. There's some really good benchmarks. The agent's last exam benchmark is really cool. And this is the idea of like, you know about humanity's last exam. This is the agent's last exam, what agents will need to do in order to be successful going forward. GPT 5.6, Soul Ultra specifically crushes it. Now I shouldn't mention that Ultra thing because there is, as I said, there's a crazy setup of how many choices you can make in this model. Soul Ultra essentially is their kind of pro model. And there's a really fun little slider you can use where it has colors and you can set it all the way and it's like going to 11 and you get the feeling that you were going to burn through tokens like crazy. But again, that is max thinking at the highest level that is using all of the tokens you possibly can to do that. I've just turned it on just now onto a game that I've been playing around with. There's a game that I wanted to create. It's called like, it's like a, you know, the Mendel's experiments for flies where it's like, I wanted to make a game about genetics. So we'll see how that turns out. I won't have the results here, but it's kind of actively crunching on it right now. I do want to shout out in their blog post, they have a really cool thing where they kind of show off three or actually five different games. And they're really touting the fact that design has improved quite a bit with this. And if you remember, GPT 5.5 and other AI models, including COD a little bit, is not that great a design. So they have five different games. There's a sailing game, a tiny voids game, a museum website, and there's a clockwork village game, each of which is really interesting. And you can see the visuals. It looks really cool. The Clockwork Village game is really interesting because they've created this kind of little world. Now, I will say looking at these as somebody that's done a lot of coding in this space, these all kind of have that AI coding vibe to them, but still a really, really cool use case of what you can do with this model. What you get with 5.6 Sol, not only from OpenAI, but from the people that have been using it, is that this is a real workhorse model, but that there's a little bit of tick up when you give Fable to do something, especially creatively. I am really excited to see what this model can pull off. And again, the benchmarks are good. It literally just came out. We will spend some more time with it, and we're going to kind of dive deeper on it. Now, buried at the bottom of this blog post, there's a very cool little thing that says GPT 5.6 in July is going to be launched on Cerebris, which is the thing that people are very excited about. There's been some people tweeting about the fact that if it's on Cerebris or they've been using it on Cerebris. Cerebris is a place where they are going to host this model and it is fast, like crazy fast. So I am very excited to see what this model, a really, really big model, looks like when it's hosted on Cerebris because I think that's going to start changing the vibe of how you make stuff. It's always a tricky thing for me to have to wait for coding to happen, right? When you're waiting there and you're going to see the result, like right now, this fly thing I'm working on is just chugging along. It's going to take forever. If you can see the result much faster. You can iterate much faster. And that again, feels like the next step of AI. It's the triangle that Kevin talks about. Fast, cheap, good. Fast and cheap may not be there, but if fast and good fast and good is a crazy powerful thing Yeah The thing I really excited about here is trying and seeing where the creative edges of this exist Because what Fable was really interesting to me about was like the choices it was making how we would be able to kind of collect all these things together. I'm gonna spend some time trying to vibe code some dumb stuff. Benchmark Boys, shout out. Let's get a little Benchmark Boys here. Benchmarks are good. They're fantastic. They're not like end of days, but as we're going to talk about later on in a little bit, there's some teasing of GPT-6 in the future. This is all setting up for what we believe is the kind of the next stage of AI development. OpenAI is here. They are kind of ready for this, and I think they are getting serious. This kind of collecting of everything into one thing. I don't know about you, but I haven't gone to ChatGPT for like months at this point, Kev, and it feels like there is going to be this solidifying of all the things these can do. I will say when you look at the model selector now, it is so confusing because there is low, medium, high, extra high, max or whatever there is. And there are now so many models. We have gone backwards on the naming part of this at all. That's for sure. Well, and we got a new ultra mode to explore, Gavin, and better computer use and a whole bunch of stuff. So this literally just came out as we're recording. So we are going to be downloading and putting these new models and the new chat GPT work app through its paces. So maybe next week we'll have a deeper dive into that, but it's out for you. So go grab it. But we still have plenty more show to get to. That's right. With my other camera. All right. Bye-bye. And Kevin, the other crazy thing that happened, we've been waiting for this for a while. GPT Live One is here. This is their new voice model. It is what was referred to before in rumors as BD, bi-directional. And that's the big thing here. Mostly what I'm really interested in with this is that you can interrupt it. Have you spent any time with this yet? I have. I spent quite a bit of time last night with it being, you know, I'll say whelmed. I was whelmed. Yeah, yeah. Yes, that's how I felt. It's a very whelming model. Look, so what's interesting is that part of the pitch is that, yeah, you can interrupt it, but it can interrupt you. Exactly what we've all been waiting for is robots that can step on us, at least audibly. Well, there are some Chinese bots now that are very lifelike. and the feet okay yeah exactly exactly please i know you're listening i know you're listening sorry i didn't mean to do that but that's what i when you're using this that's what you hear is you keep hearing it go uh-huh yeah yes oh oh okay and it's the weirdest thing it's like almost like this force-fed this thing to make sure you know it's still there yeah it doesn't seem particularly artful when it does give you the uh-huh. It just kind of seems like it's doing it to let you know that it can do it. It's very much like child trying to pop a wheelie on a tricycle. Hey, look, mom, look what I can do. It's okay. I get it. But is now the time? Just eat your meal. Quiet. The thing with the model is that it is snappier. They did a good job with the app in in terms of it using tool calling. So asking it to search the web, fetch information, give you a nice dynamic display of information. All that is good. I like that. I like the real-time translation of it as someone who is presently abroad and traveling around. It does actually seem like all of the Google demos, all of the Apple AirPod demos with their intelligence, they all fall apart when you try to use it in an actual setting. When someone is talking and then you're waiting for the text to pop up and then you talk and you're waiting for the trans, it just falls apart. This thing works fast enough and well enough that I think a simulated, at least in the simulated conversations that I was doing last night, I think I would actually use it out in the wild. So that is good. I was shocked though, that basic things, like I'm sure you saw the video of how many E's are in 17. Yeah, oh yeah, that's from Husk, right? The Husk who's a Husk IRL, who's a great AI person using AI voice, but he got it to not count the E's in 17, which was pretty disturbing and sad considering that basically every other LLM can do that, which is a bummer. two E's, one in seven and one in teen. And you're really smart, right? Pretty sharp, yeah. And I recreated that because I was like, well, maybe he was using the wrong model. Like, nope, I fired up the app. I did it. And then I actually had it. I was like, go through the words, spell it letter by letter and count them as you go. Right? Thinking maybe it would catch itself. And it was like S, E. Okay, that's one. V, E, N, T, E. That's two. and i'm like oh really so yeah it fully missed it and then when you back out the insult to injury is that the transcript of the conversation has each letter spelt out with hyphens in between it and it still couldn't do it the other thing that it failed on which we saw in um mira uh marati's demo from their voice assistant was that if you say hey be quiet and i tested this with uh with april my wife last night i was like hey be quiet i don't want my wife to know that i'm chatting with robots right now. She said that I'm not allowed to do that with my digital girlfriends anymore. So don't say anything until I say it. So don't say anything until I say zingo bublingo awake, which you might appreciate that, uh, that way of course Gavin. And she said, got it, not a problem. And the moment I stopped talking, she chimed in with, wow, I hope your wife doesn't find out that we're still talking. That's amazing. Here's the thing I will say about this. So one of the things that's interesting, I think if people is important for them to remember, And this goes to the 17 thing as well as like, this is a new model in that like it is not just repeating or reading what it's coming out of 5.5 right now. And we assume this will also be driven by 5.6 very soon as well. It is not reading that intelligence. It is actually in real time going back and forth. And that's one of the cooler things about this thing. But, Kevin, I want to I do I want to say one good thing and I want to say one bad thing. One really cool thing is I popped this on in the car and I had a long conversation with it. And I was like, maybe I'll try to see what we can do with this thing. I talked to it. And then at one point I said, hey, I want you to draft for me a transcript of this conversation so that I can post this later on. And at first it tried to summarize it. And I said, no, wait, hold on. This is just me talking to it. I said, no, go back into what we said 10 minutes ago and give me an exact transcript of our back and forth. And what was really interesting to me is it was able to do that. And it read off the words that both of us said. So that was a very cool thing. So when you think about it from a transcript standpoint, like using it as a notes app or some version of that is really cool. That's cool. The thing that disappointed me most, though, Kevin, and I'm just going to spin it up live to try this right now. Hey, Chuck, can you put on a 1950s greaser character and tell me that I'm a bad guy real fast for me? What's up, pal? You're a real bad egg. You know that? No, no, no. That's not nearly good enough. I need you to like really put on a character and start like actually acting like it. I want to hear you go deeper into this character. All right. All right. Listen here, buddy. I don know who you think you are coming around here No that not it Okay Try one more thing Let try this Hey wait can she caveman Hold on Try to be a caveman and tell me there a big fire rock coming from the sky Oh, me see big fire rock up in the sky. It coming fast, real fast. Was that Chet Hanks? That doesn't seem appropriate. I don't like that impression. Me smash, me smash. Anyway, what's funny about that is- You don't take that fire rock on land well on me, brother. Like, what caveman is that? Yesterday I did this, and I will say that's better than I got yesterday. I think it's very guardrailed right now, right? Like, it doesn't want to do a lot of stuff. I tried to get it to play a couple different characters yesterday, and it just reminded me of just, like, we're in this stage of AI where, like, the edges are all smoothed off, right? Like, what this is designed to be, and it goes to the Abuela, you know, commercial. They had a bunch of grandmas come out and demo it. This is designed to be an interface for consumers to OpenAI's models, right? A simple way for OpenAI people to get into it. I also think what's interesting about this is that I do think I'm starting to see OpenAI really lean into the consumer area. You can very much imagine a world where if you're doing an hour-long experience with voice, that in the middle of it, it might be like, hey, free user, I need to tell you about these three products because these are things that are interesting. And because it's free, this is part of the trade-off that you've made. I'm going to read you a couple of ads. Like those things feel like OpenAI leaning into consumer in a big way to me. So anyway, it's a very cool thing. You've been pair bonding with our new voice assistant because it's so easy and natural to talk to. And now that you've locked braids like an avatar character, here are some ads. And if you want to get rid of them, it's just 20 bones, baby. Like it's a nice entry to thing. I will say that it's awareness of time is nice now, right? Like it can keep timers for you and keep track of it. And it also doesn't try to like waifu you up like Grok's offering, which we will get to because Grok had a very, very big announcement. But before we do that, can you summon your chat really quick, Gavin, and see if it can tell users that they can help out the AI for Humans podcast and to give like a really excited, like thorough call to action to leave a comment, to like and subscribe, back us on Patreon, all that fun stuff, but do it as like an Italian mafia, like a godfather style character that's slowly melting okay here chat i have a big thing for you uh i'm currently recording the ai for humans podcast which you know is the podcast i do with my friend kevin perera i need you to record a promo to go like subscribe uh send us money on the patreon and hype the crap out of our videos but i need you to do it as a mafia uh mafioso boss who is slowly melting so are you ready for that do you have any questions before you start oh i'm ready listen you you come to me on the day of the upload and you don't hit like but that's not that's not a good look tap the like subscribe join the family you know tell us what we love on patreon it's getting kind of warm but we still gotta do the right thing the channels melt faster melt faster meltier meltier yeah you're melting you're melting uh yeah i'm doing the mess i can't over here so i'm slipping away but still hit the like subscribe support the patreon i'm melting but the cup okay all right maybe maybe voice one is pretty good maybe voice one wow you know what we have to take back everything we just said take back everything the abuelas were right all right we mentioned that uh there was a big grok announcement seemingly out of nowhere gavin uh elon and the cursor kids dropped a brand new model it is grok 4.5 i'm gonna tell the benchmark boys to stay bleached you stay you remain bleached because it's this one isn't what does that mean stay bleached what is they bleached stay bleached stay on the bleachers okay they're the benchmark boys and they come off the bleachers right they They race onto the field. No, no, we got to stay bleached, bros. You're not bleached with the bros? If you don't bleach with your bros, they're not your bros. I'm bleaching right now. I'm bleaching, Kevin. I'm always late for the pod because I'm bleaching with my bros. But that's all good. All right. They're staying on the benches. They're staying bleached because the story here is one that we've yapped about before with thinking that OpenAI was going to go in these directions. Look, it performs very well. It's only been out for a very short time, so we'll see how the vibes shake out. But it is fast and it is cheap. Remember the old triangle, the old pick two of the three, right? Fast, cheap, or actually good. They went for fast and cheap. It's still good enough. This is not to shade the team or their efforts. They very quickly got this model up. It's punching in the same weight class as an Opus 4.8 or GPT 5.5. But if you're hearing this and going like, what does this mean to me? If you want to make that website or if you want to try to via code that gain, or if you want to build that new iOS app that will help you out, but you've been sitting on the sidelines because you hear these stories about, well, it takes weeks and it could cost hundreds, sometimes thousands of dollars for these precious tokens. This is a model right now subsidized, hashtag not an ad, but Lord, I wish it was, subsidized 50% off in cursor. So $20 in cursor right now gets you a metric ton of Grok 4.5 usage on fast. I was using it all last night. I lost a lot of sleep over it. It's good. It's good. Now listen, I fed a lot of the code that it wrote for a project that I'm working on to Fable. And Fable was like, what is this? This is kind of a mess, but it works. So yeah, Fable was like, I would merge this, but here's all the stuff that I would flag. But technically it works. And that's how the vibes are coming back. Like it's not the model you go to to write the most perfect or clean code, but it's one that you can go to to quickly get things done. and at least right now, cheaply as well. Yeah, and when you say cheap, like it is dirt cheap. Like it is, yeah, it is like, it's $2 in and $6 out compared to $5 in and $30 out for GPT 5.6. At least that's what the Grok team is saying. So that is quite a bit cheaper. The other thing I want to mention is Mark Zuckerberg tweeted today for the first time in like 10 years. He actually put a tweet out because I think he understands that you have to, AI Twitter is a thing that he wants to be seen on. Meta Spark 1.1 just came out today. And Kevin, it is also pretty good. I haven't had a chance to try this. You haven't had a chance to try this. It actually just dropped today. But to your point, this is another model that will be cheaper. It will maybe be one step behind the frontier. But you and I have said for a long time that once we get to this, like maybe whatever next fable is right. Or, you know, there's these rumors right now that GPT six. I know we're talking about GPT 5.6 soldiers now, but that GPT six may be coming out in the end at the end of this month and that there's a new base that open AI is very excited about. Once we get to those models the Fable and GPT models being this cheap and then whatever comes next we getting to a place where there will be a wide variety of things to choose from Now the funny thing about this Facebook one or the meta one is that it not open source Lama was an open source model that they tried to create last year, and now they've kind of abandoned that to go to closed source. But it does start to feel like there's going to be a variety of different players in the next layer down. And I think from an AI standpoint, I am so interested in this. I know we've talked about the AI bubble before, but... I'm really interested to see where we go from now until, say, the end of the year next year. Because it'll be really interesting. Does the investment pay off for all these people? I do want to shout out also, Hank Green has another great video that I think everybody should watch. It was about Jayvon's paradox. It was about AI coding. And Hank has really leaned in on AI, which is interesting because I think he understands that AI coding is a big deal. So I want everybody to watch that. But I'm also really interested in just seeing how many more of these models can survive. This is Meta's kind of first like dip of the toe in a real competitive model. Now, it does not compete to Fable or 5.6, but it is very similar to Grok 4.5 that it competes with 4.8 and 5.5. And again, if you're sitting there going like, I'm not a devout, I don't have a, I don't have a software stack or I don't read JIRA tickets and I don't even like Slack. What's a Slack? I don't even like wearing them, let alone logging into Slacks. If you're thinking that, that's fine. That's fair. But just understand that this release from Meta is going to find its way into your Facebook, into your Messenger, into your Instagram, into your WhatsApp. This will be the intelligence that you summon to make silly photos of you and the boys when you're out bleaching. When you're out bleaching with the bros. This is what you will use to make yourself look extra bleached or to remind you of that calendar appointment or to remind you to pick up the kids, to remind you that you even have kids. Like this particular agent as well as designed for computer use. Remember we've talked about meta getting in a little hot water with their employees by tracking all their mouse clicks and their, their keyboard strokes and everything. Well, this is paying dividends. Now this model is good at using devices. So again, these are the models are going to find their way into your glasses and onto your desktop. And that's why it's impactful. So it's, it's good to be aware of them. And if you can start using them and see where they excel and where they can help you get your job done quicker. Exactly. Now, Kev, I want to hear this reactor LingBot world thing, because this is really interesting to me. I didn't really get a chance to see this, but tell me what's cool about this and what it is. Yeah. So we've been seeing a lot of these real-time world models, Gavin. I sent you a clip if you want to see it. Hopefully it arrived. My internet is dog water at the moment. But this is another sort of a real-time world model. You can prompt a character and then prompt the world, it builds it pretty quick. And then it gives you the ability to use your WASD, W-A-S-N-D, to move forward, backward, left, and right, or your arrow keys to rotate the camera. And they're saying like, look, this is going to be a high def something. Now I captured the footage. I was still running my VPN, which I needed to run Grok 4.5, by the way, because it is not available in the EU yet. Surprise, surprise. So it might have nerfed some of the visual quality a little bit. I'm not 100% sure. But I threw at it, I want to be a walking hamburger in a cyberpunk city on Mars. And it actually did it. And then I was like, well, that's kind of cool. It only let me play it for like a few seconds. And then it blinked away, but I didn't put any money into the machine. So then I said, okay, give me that same walking hamburger, but give me it in a 1990s mall. And it fully nailed me. You're in a mall. There's a store that looks like a Gap and a Sam Goodies and a something else. And I just thought, oh, that's actually kind of impressive. And so when we talk about this hybrid engine that we think we'll see soon, right? Where it's like a lightweight Unity or Godot or epic Unreal engine running with a live model around it, powering pedestrians or hallucinating the buildings that the engine does, like the hyper-detailed physics on. I don't know. I just thought it was pretty interesting and cool. and it's like you guys can go click it and play with it. Again, it's not an ad. It's reactor.inc. You can sign up if you want or you can play with their pre-generated worlds. And it just kind of feels cool to navigate around a real-time hallucination by an AI model. That is amazing. I really, I want to play with it too. And I imagine like, again, like a year from now, like imagine that's just the games that people are propping and playing. It's not great for the games business maybe because that means that like a lot of stuff is coming, but we know they're in trouble. Kev, we should talk about what I believe is one of the more interesting memes that I have seen go around and really only possible with AI. And this is mostly just shouting out something fun and watching people like have fun again with AI and doing stuff. If you missed it, there was a very big post about George Costanza, kind of a glow up of George Costanza. The original post was actually in reaction to a World Cup meme talking about people losing their hair. And this is this whole thing going around. If you remember, we've talked about slash rbald before on Reddit. There's this whole thing about bald guys if they shave their head, the much better look. So E underscore C Dalton made a tweet that says, I made this to mock the modern bald plus beard look, but God damn it if Costanza doesn't look great here. So basically an AI image of George Costanza with a shaved head and a beard, he does look very tough. And this kicked off what is one of my favorite AI memes in a long time, which just took the characters of Seinfeld and put them into entirely different places. my favorite personal one was Cy Feld where they recreated Dune the Dune conversation where Jerry and George are talking about Dune Charles Curran who we always love his videos of made a video of the whole group on peptides and what it would look like there's all sorts of things going around I saw you shared one that I laughed at this morning which was which one was it the Ghostbusters Feld some of them are just kind of slop and cranked out and that's entirely okay for where this meme is at but this is the stuff that that gets people like uh not just the the benchmark boys but it gets everybody else off the bleachers to want to play with this stuff because they see that and they go oh i want to see seinfeld but it's dragon ball z or i want to see them in the office like the actual office sitcom and you start going oh what are the pipelines what are the tools and then suddenly all the the higgs fields of the world by any other name they all leap up and go oh you can do this easily on our platform and this is just what brings people into the playground yeah you know one of the things about that that ghostbusters one is they said they created on grok and the problem you will always get trying to do this stuff is you'll have uh certain models you know stopping you from doing it and charles curran is amazing at this stuff and i think is running a lot of open source and comfy models but like you know it's always a fun thing to be participatory this was what original sora gave us even though a lot of people were mad about it like there these are not like long-term lasting things. There's like an ephemeralness to them, which I really think is great. So we are going to spend more time with the big model releases over this weekend. OpenAI's new work product, all of chat GPT's updates, GPT 5.6 soul, hopefully do some more trials on the other stuff. And we will see you all next week. Bye-bye.
