BONUS: New GPT Memory Feature, GPT-5.6 Rumors, Hermes Desktop Agent, New Codex Plugins, MAI-2.5 Image, Etc.

[SPEAKER_00] Oh, howdy and hi, all of the things. How's everybody today? We're... [SPEAKER_01] Corey, would you mind giving the intro? I'll be right back. [SPEAKER_00] Yeah, yeah. We just hit the button. Grant's getting the email out to everyone to see who all might want to join. So we're getting started. But in the meantime, welcome humans to The Neuron Live. We're excited to have you here today, as always. And this should be an interesting episode. We're still kind of figuring out a few things about what's going on. [SPEAKER_00] So we kind of scheduled this with the idea that there would be some announcements today. And Grant's gathering some information and all of that. And we'll have some more on that in a minute. But for right now, we're just getting set up. Hope everyone's doing good. Having a great week. And we'll see what's happening. What's happening today? [SPEAKER_00] All right. Where's everybody joining from? Really curious to see where everybody's from today. I just got home about 2, 2.30 this morning from Microsoft Build in San Francisco. Really fun trip with some pretty major announcements, frankly. [SPEAKER_01] Yeah, what was the coolest thing you saw when you were there? [SPEAKER_00] Oh, it was the Surface Ultra and that dev box. The RTX Spark dev box. Holy smoke. Like, that is top tier premium hardware. It is ridiculous. I got to play with a couple of them while I was there and was pretty floored. [SPEAKER_00] And that RTX Spark chip from NVIDIA is insane. They also had not just the Surface version. They had the Lenovo, the Dell, everybody else who is releasing HP, who is going to be releasing these chips with these systems with that new NVIDIA chip. [SPEAKER_00] But we're all there. And it looks like they'll all be coming in the fall, hoping to get hands a little early on the Surface Ultra because it's ridiculous. 128 gigs of unified RAM. Lightning fast. Lightning fast. And I think it's mostly got to do with the NVIDIA chip. [SPEAKER_00] But they were showing us things in our demo yesterday. Like, I don't know if anybody works with CAD and you know what a headache spinning CAD is trying to trying to show all of the dimensions of your drawing. It was just flawless on a touchscreen. They could blow it up and go, woo, and the car spins with no loss of detail or or or graininess or anything. [SPEAKER_00] It was it was absurd. Showed us a few video games on it. They showed us 3D world design and the ability to quickly spin your 3D world and not have all of that pixelation and messing out that you have. I also got to hold the chip. The chip is about this big. I've got a picture of it on my phone. I'll see if I can send it to us here in a minute. [SPEAKER_00] Um, and it's pretty nifty, but they had one of the surfaces for us blown apart, like hanging from the ceiling in all of its slices there to check out. Uh, to be as hot as it is, one of the things that was craziest to me is that the fans were like virtually silent, like dead silent. I, uh, I have, I have this gigabyte laptop over here that I love, but when it gets, gets a little load on it, it sounds like, like a drone took off in the living room. [SPEAKER_00] It is, it is so loud. It will like, like burn into two grand. Do you mean to be muted? Or is that like unintentional? [SPEAKER_01] That is intentional. Sorry. I laugh. I laughed heartily at that last call. I saw you moving. Yeah. It's because I'm clicking around and doing things over here. [SPEAKER_00] So yeah, keep doing things. We're good. Uh, I hope, uh, what else was there from build? It was cool. I got to spend a little time with Mustafa Suleiman from Microsoft CEO of Microsoft AI. Uh, a lot of really cool stuff coming out of them and probably going to continue to be throughout the year. I, I expect, uh, that the seven models we saw released on Tuesday are, are only the start. It looks like they've, uh, I've got a full interview I did with him that we'll be releasing next Wednesday. [SPEAKER_00] We'll make sure everybody gets that too. It'll be in the, we'll share it out in the neuron. Um, but it's, it's really interesting. It's not super long. Uh, but we talked a lot about humanistic super intelligence, about what his team's been doing, about all of the restructuring that they underwent late last year, like from October through December and how that has enabled them to really zero in and get their focus back on chasing super intelligence. [SPEAKER_00] And I think that, uh, uh, uh, the stuff they just released is, is good. Like, um, like the, I want to say, I mean, I code one flash, I'm going to mess these names up because you know how names are. Um, I want to say that it debuted on Sweetbench Pro at like a 53, which is pretty comparable to Opus 4.6. So out of the gate, that's, that's a big hit for Microsoft. [SPEAKER_00] The thinking model is really solid. It's, it's really good. And what's interesting about the thinking model is they trained it on 35 trillion parameters data or 35, 35 trillion parameters of data. Uh, it has, it's a 1 trillion parameter model with 35 billion active parameters, which is super duper low [SPEAKER_00] for a model that big. Uh, and they did it by very, very carefully curating the data. There's no, distilled data. There is no, we went and synthesized data from all this other stuff. They went and, and had real data and had humans working it and, and creating quality data that was corporate friendly, copyright friendly, all of those things in order to start from this baseline that is really, really clean. And, and he talks about that in the interview as well. So there'll be some good stuff coming up there. [SPEAKER_01] Um, yeah. Well, uh, I think the email may have gone out to everybody at this point, um, or it will be soon. So I think it would be good to kind of get us, get us on track here. So what are we, what are we going to be talking about? We're going to be talking about the rumored new model, which I have a feeling it's not going to come out today, but we'll see. Um, but there was a really [SPEAKER_01] cool announcement that just came out, out of open AI regarding memory, which we are going to talk about. And then I would love to talk about, um, the new stuff that happened with codex this week, as well as my new favorite tool, which is Hermes desktop. And, um, besides that, there was also some cool stuff that came out this week regarding, uh, Gemma 12B, [SPEAKER_01] which is an open source model that a lot of people can run. Um, that already has a two bit quant. So it'll run on eight gigabytes of Ram. Yep. Yep. That's right. Um, which means a lot of computers, you know, if you have a Mac book, it's like 16 gigabytes or, or, uh, anything bigger than that of unified memory, um, you should be able to run it. Uh, I don't know. And eight gigs going back as far [SPEAKER_00] as like 2015, like you don't have to have a new computer to run that model, which I think is right. [SPEAKER_01] Right. If it's a, if it's a powerful one. Yeah. Um, and then, and then there was also something that you got to play with, or you at least knew about ahead of time, which was Nematron three ultra, the big, the big baddie that just came out today. It's, it is madly impressive too. I'm [SPEAKER_00] anxious to talk about that. Um, I actually, I've been this morning, like I got home in the middle of the night and I got up and I knew I had a number of things to do. One of which is get my full review out on Nemetron 3 Ultra. Uh, but we'll talk about that on the website. I'll link it. Uh, no, it's not. That's because I was still writing it and keep getting pulled into other things. [SPEAKER_01] Okay. All right. Well, we'll, we'll, we'll publish it after this. Yeah. When you've been out [SPEAKER_00] of town a few days and you come back, there's kind of this, uh, thing going on. Um, but, uh, it'll, it'll be coming out today, tonight. And, uh, you know, long story short, it's, it's really good. 550 billion parameters. I think it's 35 billion active as well. So it, it, it runs light and efficient. And, uh, uh, if you've watched these live streams before, when we had a new model out and you've seen some of the questions we pushed through a few [SPEAKER_00] of them that, that regularly trip up other models, it destroyed, uh, just, just absolutely [SPEAKER_01] mowed through them. Actually, I have a question. You got early access to this. I did not. Um, it also just came out today. So I'm unclear. How are people expected to run this, this NVIDIA and Nemetron? How did you run it when you did it? [SPEAKER_00] I ran it. They send it to me in a notebook. Okay. Uh, so like, I don't run it locally. I don't have enough machine to run a 500 billion parameter model. Um, nor do most of you watching most likely nothing personal. Uh, it's, it's, it takes a, takes a hell of a computer, but, uh, uh, you know, you can use it through open router. You can use it through any of those type of places. I believe it's on, you know, Bedrock and Azure or Microsoft Foundry, uh, all of the places you [SPEAKER_00] can use models. And NVIDIA has its own cloud hosting platform as well, where you can go and get it and attach and use it. But it's, uh, it is undoubtedly NVIDIA's best model to date. And, uh, it's, it's strong. Let's put some visuals on this. It'll take, they told me it would take four DGX [SPEAKER_01] sparks to run it at like speed. Hmm. All right. So this is what we're looking at here. Um, so if you're not familiar with unsloth, this is a company that makes, uh, smaller versions of models that you can run on your computer. So this one here is an unsloth version of Nemetron three ultra which is what we're talking about. That says it can run two bit on 200 gigabytes Ram. [SPEAKER_01] Wow. That's wild. Um, now how would you actually, here's the official announcement. I wonder where they're pointing people to go try this. [SPEAKER_00] Yeah. I'm not sure where they're pointing people to go. Uh, yeah. Open code. Of course. [SPEAKER_01] So, so if you, if you did want to run this yourself and spin it up on your own cloud server, the way you do that is you would go to hugging face and you would grab it there. Go ahead and retweet that. Um, and then this is the smaller version. So go ahead and highlight that as well. Um, looks like it is on fireworks potentially. [SPEAKER_00] Welcome to the frontier. [SPEAKER_01] Yeah. Yes, I did. I like, I like fireworks. They're pretty great. Okay. So I'm going to look at that blog post. This just came out today. So I haven't really [SPEAKER_00] looked at this yet. Like they kind of, uh, Jensen announced that it was coming on, I guess, us Sunday night, uh, from Computex in Taipei. And, uh, but today was, was street day. [SPEAKER_01] I like the way they outlined this. They bolded the best one. [SPEAKER_00] And which models they see it comparable to, I think is worth, worth looking at up there. Uh, you know, they're seeing it when you're looking at various classes on these. One thing to know is not every model comes out being expected to, to compete with the latest Opus or the latest chat GPT. Like, you know, they put these models out kind of in classes. Like for example, uh, this new Microsoft thinking model is Kimmy K26 range. Uh, I want to say [SPEAKER_00] GLM five, one was in there and it's meant to compete in that weight class kind of medium weight class right now, but, but getting bigger. And the thing you don't know, these are open [SPEAKER_01] weight models, meaning anyone can spend them up on any cloud server they rent or any servers that they have. I mean, they're all very big. That's what the B stands for, um, which means you can't run it on your local computer, but you could run it on a, on a cloud server [SPEAKER_00] that you rent. Yeah. NVIDIA and Microsoft though are very much right now the best U.S. based open source game in town. Uh, and, and I wouldn't have said Microsoft before this week. Um, with that said, their image models are, are fantastic. The new ones they dropped, uh, dropped higher than Nano Banana 2, uh, or scored higher than Nano Banana 2. Yeah. [SPEAKER_00] Microsoft. I find that hard to believe, but that's interesting. You really shouldn't. It's, it's good. I've spent some time with it. Uh, well not with 2.5. Two is really good. Um, is that out now? Yeah. Yeah. It's out in the middle. I believe all seven of those were released the other day. MAI image 2.5. I actually, I have a thing. I can probably bring it up on screen and show [SPEAKER_00] you. I can run it. You can run it. Yeah. Let's see it. Yeah. Right there. Read the first paragraph. [SPEAKER_01] Read that first paragraph. It said the Nano Banana claim up ahead of Nano. Okay. Let's see this leaderboard. I don't believe it. Okay. It's right below. Okay. I see. It's right below, uh, GPT. So it's not the frontier. Otherwise we would have talked about it, but it's still up there. [SPEAKER_00] Yeah. I mean, I mean, if we're above Grok Imagine, Grok Imagine's good. [SPEAKER_01] Yeah. I, I, I imagine put intended that this is not 1.5, which just launched, but no, probably not. [SPEAKER_00] That may not very well not be on there yet. Wow. This is actually, it says preliminary. Maybe that is. [SPEAKER_01] Let me expand my screen a little bit here. What is this? Go away. Okay. It's like getting this sizing, right? You know? Okay. So image quality. I don't know. [SPEAKER_00] If you look at that little exclamation point on the right over by 1388, it says hover over down, straight down. That's up. That's fine. Yeah. Those are based on pre-release testing is what it says. So I'm guessing maybe that's 1.5, but I don't know. Okay. Sure. I can see it. Yeah. Uh, but yeah, I mean, I 2.5, like, like they came out, they made some waves this week. Yeah. That's very cool. [SPEAKER_01] Um, yeah, let's show, uh, would you show us this? You want me to bring it up? Fire it up. Let me know when you're ready. [SPEAKER_00] Give me just one. Uno momento. I assume I can show this tool I have. [SPEAKER_01] Um, are you saying, so someone in the comment said our microphone is too sensitive? [SPEAKER_00] Is that? I wonder if that's me. Let me check my side too. My mic is always stupid quiet and I struggle with, oh, they changed some tools in here this week. Let me, uh, give me one second. They moved everything. Classic. There's that. There's the mic. Yeah. Okay. Well, I don't want to change the mic. [SPEAKER_00] I just want to boost my volume. Interesting. Audio. Yeah. The problem when you launch like seven [SPEAKER_01] things at once is it's hard to know like what it is. And I, uh, I finally fixed the spring in my mic [SPEAKER_00] so I can move it closer to my mouth as well. So hopefully that'll help. But thanks for calling that out. I'm very conscious of the fact that this is a problem. We don't a hundred percent know it was you. So I did ask a question to confirm. Okay. Just to be safe here, I am going to close. Both are pretty quiet. Since I don't know how much of this I can show, I'm closing the sidebar because I know there are a couple of things over there that aren't like necessarily out yet. [SPEAKER_00] Okay. Good. Good. I'll, I'll try to make sure I keep up too, Evan. Uh, that I, that I speak up good because it's, it's a thing. I, it's even, even, uh, like doing those interviews this week at Microsoft, the, it's always an issue. Like my mic has to be like way closer to my face. Wait. [SPEAKER_01] Brian says, wait. Yeah. Just don't, um, overcorrect. So don't turn your head when you're talking. [SPEAKER_00] I will. That's, that's going to be a little difficult. Let me pull. Talking, talk directly into your microphone. Okay. So here we go. This is, this is where we get access to some of Microsoft's models. Uh, but what do you want to create? Give me an image prompt idea, folks. I'm going to move the camera screen over there. Let's, uh, let's [SPEAKER_00] say we want to see create a, an oil painting of a dog in the park dancing in the rain, but [SPEAKER_01] the rain is meatballs. Oh. Cloudy with a chance of meatballs. Cloudy with a chance of meatballs, [SPEAKER_00] right? Right. Okay. Sorry. That was weird. No, it's all good. Gotta keep it weird. Okay. Oh, and this is 2.5 flash. Okay. Okay. So oil painting dog. It's definitely oil painting looking. You can see the strokes. Looks like meatballs. Okay. So now let's go back and now let's say they added an edit feature too, which is wonderful. [SPEAKER_01] Let's say actually. Can you zoom in on that a little bit more or expand your screen a little [SPEAKER_00] bit more? Let me see. Can I not very well? Uh, let's see if this helps. Yes. That's perfect. Did that help? Yes. It takes some creativity to adjust screens because like the shape of your screen matters and you always have to, it takes us a minute sometimes. Yeah. Okay. So actually make it photo realistic. Uh, overall I've been really impressed with their image models. Like [SPEAKER_00] it feels like the first place that they kind of hit the mark. Um, they've got some really cool voice stuff that's out and that's maybe still to come. It'll be really neat. Uh, their transcription model is, is supposed to be really good as well, but I haven't been able to get it to work for me on here. So I've got to check in on that. Okay. I don't know if I would call [SPEAKER_00] it photo realistic, but it is nifty dog definitely looks wet meatballs rain. Okay. Yeah. It looks [SPEAKER_01] like it's stage two similar to the original. Yeah. Very, it changed very little actually. [SPEAKER_00] Yeah. Yeah. Yeah. I can switch through them here too. I find that this is kind of a problem [SPEAKER_01] though. Any, with any image model that you use, um, is that whatever your first version is, it stays pretty close to that. And you, you know, if you wanted to actually just get at something completely different, just start a new chat. Like there's no reason to continue with the same one. If you want something different, it's like pushing a boulder up a hill. Yeah. Now there are, there are models that are specifically for image editing. So, um, [SPEAKER_01] like flux, I think flux context or something like that was really good at, at, um, yeah, [SPEAKER_00] I think it also matters how much, how much variance you're looking for from the original. Like if what you want is to change the meatballs to, I don't know, dragons, you could do that pretty easily. But if what you want is to train the stuff, change the style of the whole image, often that doesn't go far. I mean, it's, it's definitely an improvement. The water looks pretty real, but you could still tell it's not a photograph. So we're going to go back and do this a little differently. Now we're going to ask for it to be photorealistic right [SPEAKER_00] up front and it will do much better. Create a photorealistic image of a golden retriever jitterbug dancing in a park in the fall when it's raining meatballs. [SPEAKER_01] Looks like it says FOGO. I don't know if that's going to mess you up. [SPEAKER_00] Oh, it does. No, I don't think so. Okay. Surely the magic intelligence in the sky can see that it's a typo. Surely, surely. In a what? 14 letter word? Yeah. Only one's wrong. [SPEAKER_01] Yeah. Oh, there you go. Okay. Okay. That looks slightly more real. Still looks fake. [SPEAKER_00] The dog looks less real. Everything else looks very real. [SPEAKER_01] The dog's lighting is what looks fake to me. Yeah. I think it's mostly around the face. Like, I mean, [SPEAKER_00] definitely like the hair and stuff is there. I don't know. The trees look good. The meatballs [SPEAKER_01] are definitely meatballs. Yeah. I mean, it's not bad, but it just got that cartoony kind of look to it. That's why I would do the same exact prompt, even with the typo in a completely new chat and see if it gets closer to what I would at least consider real. By the way, for anyone tuning in right now, we are a little bit sidetracked. We're testing. Apparently, this new Microsoft model has surpassed [SPEAKER_01] Nano Banana from Google, which I found unlikely. So we're testing it live. What would you want to test? I'm trying to figure out how to get a new chat. Oh, yeah. Good question. Without having to... [SPEAKER_00] Hang on. I've got to stop sharing and go back in. That's okay. So this is the new... What interface are you using for this? This is their new interface? No. This is the one that they give us to use for such things so we have access to them. Let me... There it is. Okay. All right. So basically, [SPEAKER_01] what Corey said is he's using a secret magic interface. Maybe a little. What about Kev of O'Coyote? [SPEAKER_00] I don't think it's secret, but it's not mine to share, I don't think. But I think doing this is fine. All right. What did you want to see? What would be your test? Are you talking to the chat or me? Or you. Or the chat. Either is fine. Okay. Well, do exactly what I suggested before, [SPEAKER_01] where you take the same prompt, copy and paste it, and see if we can get it to look more photoreal. [SPEAKER_00] Create a photorealistic image of a golden retriever jitterbug dancing. Ah! Typos. In a park in the fall [SPEAKER_00] while it's raining meatballs. And let's see what we get this time. All right. Let's see it. I've had overall good luck with this, and I've used it quite a bit for some things here and there. [SPEAKER_01] Is this available in Copilot right now? I believe so. What's interesting? It looks exactly the same, [SPEAKER_00] which is very interesting. The dog looks exactly the same, doesn't it? Okay. Well, let's try this. Yeah. It has like a cartoony glow to it. Create a photorealistic image of a woman. That's a hell of a [SPEAKER_00] sentence. Yeah, I know. That's what I was reacting to as well. With blue hair and red eyes leaning [SPEAKER_00] against a building in the city looking to the left of the camera. And we shall see what we get here. I won't stay on this too long because I know we've got a lot of other things we want to talk about, but I just thought it was cool to show. [SPEAKER_01] Yeah, I've got an interesting tweet to share when you're done with this. [SPEAKER_00] They do have another tool for using this that you'll be able to use where there are things that are already there like the ability to click. Okay. I'd say that's quite photorealistic. Perhaps it's the last thing people... [SPEAKER_01] Yeah. Maybe zoom in on her like face a little bit. [SPEAKER_00] I can't. It's stupid screen is not... Hang on though. I'll tell you what we can do. I'm gonna save image. [SPEAKER_01] Yeah, because I would say it looks fairly photoreal. The only part that reads off to me is the face. And maybe it's just because she has like red eyes and blue hair and... [SPEAKER_00] See, and the face to me is what I thought looked much more real. Let me... Let's see. The background looks more real. There is that thing with AI where sometimes the problem is they look a little too perfect. Yeah. Which is very much a thing. Okay. So now let me reshare. All right. Good song. Thank you. Thank you. [SPEAKER_00] All right. We're gonna zoom way in here. Oh shoot, it didn't move. Okay. Okay. Okay. Okay, that's way bigger than I should be zooming in on this. It looks like digital... [SPEAKER_01] It looks like digital art, which it is, I realize. Well, and to be fair, [SPEAKER_00] I was looking at... I'm zoomed in on the head and on my monitor, it's this big. I was zoomed in. So like, you know, the artifacts you're seeing are because her head is literally this [SPEAKER_01] tall right now. Yeah. And who knows? If we were to zoom in on a real photo, maybe that would... I feel like it has and correct me if I'm wrong. The hair looks great. It has almost like a painted sheen to it, but I don't know if I'm just overly critical of that because we were just looking at an oil painting of a dog. Yeah, that's fair. Yeah. So I don't want to say that that is actually... She appears to have the right number of digits. That's good. [SPEAKER_00] Yeah. People walking around the street. We got cars. Yeah, that looks very real to me. Everything looks normal as it should in the background. Coat looks good. Yeah. Hair is slick. Like, I like that the hair has like depth and texture to it. I don't know if you can tell. Like when you look around like through here and through here, there's definite... You know. Yeah. Let's see. [SPEAKER_00] Yeah, it looks good. Red hair, blue eyes. Pretty cool, by the way. [SPEAKER_01] All right. Well, good job, Microsoft. Uncanny Valley. That's the word. Color me a little bit impressed. [SPEAKER_00] A little bit. That's okay. That's okay. [SPEAKER_01] Yeah. I mean, certainly a good showing, right? [SPEAKER_00] Yeah, we've definitely had worse. [SPEAKER_01] Yeah. So I'm going to share my screen again. I'm going to share this tab. And what we're going to talk about now is we're going to switch gears back to... Okay. So, Corey, do you know this guy? His name is Chris GPT? Yeah. Okay. I see him. [SPEAKER_00] I mean, I don't know him, but like, I think we follow each other on Twitter. [SPEAKER_01] He is allegedly an AI insider. I forget exactly what company he's supposedly an insider for. I imagine OpenAI, but I'm not sure. But one of the big post kings of x.com. And right here, so, you know, this live was about Mercury Alpha. There's also now apparently Joule Alpha, which is coming out. So have you heard of this at all? [SPEAKER_00] Well, yesterday, if you know, if you follow any of the OpenAI crowd, Riley Brown, who works for OpenAI, shared a video yesterday. And in that video, at the bottom of his screen, it showed Joule Alpha. And I think that's where this cut is from. And later on in the day, I saw that and I was like, oh, I want to go look at the video. And I went back and I said, this video is gone. And I'm like, oh, that means it was real. [SPEAKER_01] Yeah. Okay. Right. So they definitely took it out. [SPEAKER_00] And there's this other one we keep hearing, right? Mercury Alpha? [SPEAKER_01] Mercury Alpha. Yeah. And I'm going to go ahead and I'm going to try and find the original tweets where I saw this. [SPEAKER_00] A thing to know is these could also very much be different checkpoints on the same model. Like, it's quite common for me to do that. It doesn't mean it's... And a thing Mustafa's explained to me that... Mustafa explained that I didn't know, Grant, and you'll find this fascinating, is that when doing a one trillion parameter model, they do a whole bunch of them. They take the data set, mix it up, use different data, and redo it, redo it, redo it, redo it, and then go through and can cherry pick like, oh, [SPEAKER_00] well, this set of data did better than this trillion parameters. And this trillion parameters did worse than that trillion parameters. So you can go through and kind of adjust your percentages of the types of data in there to tweak the caliber of performance you want. And that's just not a thing I'd consider because they were working with, I think he said, I think he said 33, 33 trillion parameters in the thinking model. Wow. [SPEAKER_00] Very cool. Sorry. I just, I thought that was no context in talking about new models. That is cool. [SPEAKER_01] It's not a thing anybody's ever told me before. So is that kind of equivalent to trying to like stitch two different models' performance together? Because I've heard that's really difficult. [SPEAKER_00] You know, the way to think of a model is like they're building this thing and they're going to keep building this thing. And then when a new model comes out, essentially what they're doing is hitting save on that model and doing a save as. And like, here is the model as it was at this moment. [SPEAKER_02] Right. [SPEAKER_00] And so like they've continued training, they've continued tweaking, they've continued adding new new data, new things. And they hit save. But they can also do that with entirely different data at that point. They can be like, we're going to pre-train this model with this set, this set, this set, this set. And they're all going to be separate models. And then we're going to test them against each other and see what like the internal team and their beta testers think of it. Wow. [SPEAKER_00] And be able to then kind of dial in this one. And that also tells them then that like this data works well. This data is hot trash. [SPEAKER_01] Right. Yeah, they can figure out what to get rid of. Yeah. Yeah. Well, if you look at this post here, so somebody, Leo, just somebody who started a rumor and I guess they accept tips said Mercury Alpha. [SPEAKER_01] And then Andrew Curran said, AKA GPT 5.6. And a lot of people think it's coming next week. I think when you see something like this come up, it usually probably means that it's coming in a week. Yeah. A week now. So, you know, adjusting expectations, it's probably coming next week. But that being said, Codex did have some new stuff that they launched this week. And I'm going to pull that up real fast. [SPEAKER_01] If this is the right post. So, Codex for every role, tool, and workflow. So, I'm just going to read through this real quick because Corey has not seen this at all. Yes. Unless you all are. I've been out of the loop. So, I'm learning with you guys. Go Grant. Unless you all are extreme link clickers like I am and click every link in everything that you read, you probably didn't click into this to learn more about it. So, I'm going to read through it really fast. [SPEAKER_01] So, more than 5 million people now use Codex for work every week. This feels low to me given the fact that this is basically going to replace ChatGPT as their work surface app. But people have not realized this yet or not accepted this fact. So, it is what it is. Codex started as a tool for software development, but it's increasingly useful for more kinds of work. [SPEAKER_01] Non-developers, including analysts, marketers, operators, designers, researchers, investors, and bankers, make up about 20% of overall Codex users and are growing more than 3x as fast as developers. So, that means that if there's 5 million people, there's 1 million people who use it for this use case, right? So, and I feel like this number is going to grow a lot. So, this week they introduced new ways to do more work with Codex with new plugins. [SPEAKER_01] So, if you don't know what a plugin is, a plugin is like a series of tools and skills that are put together and connectors that let you use a particular suite of tools to do a particular task. And I can show some of those later. Corey, if you want to bring up your screen, you can show some as well. But basically, they adapt Codex to your role and tools. [SPEAKER_01] And they're annotations with annotations that help you refine the result in place and a preview of the ability to create interactive websites and apps you can share with your workspace using a URL. So, this is where they're talking about how they use it. But one of the really cool things that they launched this week is sites. So, sites basically let you create a site directly inside Codex. So, this is a demo here that I'm showing. Stop me at any point if you have questions. [SPEAKER_00] This is, you know, the site thing is really cool. Because, like, that's a thing that's kept, honestly, it's a thing that's kept me using V0 for a good run. It was the ability to just quickly spin something up and hit deploy. And that is a thing that, well, I guess with Google you can because they have their own cloud. If a company has its own cloud, they have that luxury. But not everybody's had their own cloud. So, like, for Anthropic, for OpenAI, that's an area where they have not had that ability. [SPEAKER_00] But apparently they have something because this is really cool. The idea being you could spin up and share a tool with your team in minutes from scratch. Yeah. [SPEAKER_01] I think this is kind of like what replaced Canvas. Like, they got rid of Canvas this week or last week. I think this is what's replacing it. And I think it kind of builds on a post that Barik from Anthropic shared a couple weeks ago, which was basically like, hey, if you're doing anything that's sharing it for other humans, just make it an HTML page. Like, just write in HTML. HTML is the better format to use. [SPEAKER_01] And I actually use HTML when I write the newsletter so that I can actually have it hyperlink my links. So when I, you know, like I said, I'm a linkaholic. I love including lots of links to things. I try to make, I try to include as much information as possible. The easier way to do that when you're working with AI is to make an HTML website. So this just builds on that. And, you know, you can add all sorts of cool stuff on top of it. And then they add some other stuff here, which the plug, I think they announced six. Yeah. [SPEAKER_01] Six new role specific plugins. Okay. That make it for more useful. That makes it more useful for more kinds of knowledge work. [SPEAKER_00] Like non, non developers, non engineers. [SPEAKER_01] Yeah. [SPEAKER_00] And I'd say this is a very key piece of the puzzle for super wrapping. [SPEAKER_01] Yes. Oh, definitely. For the super app. [SPEAKER_00] That data analytics plugin is sick. I'm anxious to play with that. I haven't yet. [SPEAKER_01] Yeah. So this is cool. I just added this, I think. Which helps analysts and business teams answer questions with data. They can explore product and business data, explain why key metrics change and create reports and dashboards. And then it connects to Snowflake, Databricks, you know, all your data sources, Tableau. More coming soon. Then there's a creative production plugin, which helps marketing and creative teams turn a brief into assets they can review. And this uses tools like Figma, Canva, Shutterstock and Fall, which lets you do image models. [SPEAKER_01] So for example, you wanted to use that Microsoft model. You can use Fall to do that as one option. Sales plugin, which helps sales teams bring customer contacts into the work that moves the deals forward. Sales teams can find high priority accounts. They can prepare for customer meetings. And they have tools like Salesforce, HubSpot, Slack, Outreach, Clay. That's a cool one. Rocks Unactively. Then they've got product design. [SPEAKER_01] So I think you can tell what product design is. It's creating prototypes. [SPEAKER_00] I don't have a product team in mind. Say that again? I said I built out a product team worth of skills in mind. Like I have a senior product designer, a project manager. I have a UX specialist. I have a YC Combinator advisor skill. Awesome. Yeah. That was like, look at this. Like it was going to be something great. [SPEAKER_00] We're really, I think we're, I think I'm going to say by 4th of July that this will be the way you use ChatGPT for the most part. Maybe not for everyone yet. That's going to take time. But I think it will be the premier way to operate will be in Codex. Yeah. [SPEAKER_01] Kind of like how people have been slowly switching from Claude workspace or Claude web version to Claude desktop and using Codex. I think the same transition is going to happen with Codex. It's just a little bit harder, honestly, because they called it something different. So I agree that I think the brand needs to evolve beyond ChatGPT. I think ChatGPT limits them even though it's like Google. But I think for serious work, I think it makes sense for them to make Codex their platform for that. [SPEAKER_00] You know, the flip side is, and this is a decision that both OpenAI and Anthropic made, is that using a name like Codex or Claude code implies this is a tool for engineers. And people who are not find that intimidating. And you can't depend on everyone to have gone and read your article here to understand that, listen, this is for anybody who wants to use it. You can do all kinds of things. Like I write in Codex. I research in Codex. [SPEAKER_00] I have Codex wings spun up all day long doing various things. Now I work, I mostly still use the ChatGPT interface. But I do have a number of small tasks that I prefer to be handled locally. Right. Just because I find it's quicker. And you can trace its steps and what it's done better. [SPEAKER_01] This whole time I've just been thinking that they should just call it WorkGPT. [SPEAKER_00] I thought you were doing that. I was like, dang, what's your job? No, no, no. [SPEAKER_01] No, no, no, no. That's a demo. No, I've just been thinking maybe they just call it WorkGPT. Because I was like, do you make it like ChatWork? And I was like, no, that's kind of dumb. I think they'll hang on to ChatGPT. [SPEAKER_00] I think the name will be ChatGPT. I agree. [SPEAKER_01] I agree. But I think like Codex could become. Okay. So like CloudCode has this really bad problem, which they solved by creating an alternate version called Cowork. Right? Yeah. So Codex. [SPEAKER_00] That you download by downloading CloudCode. [SPEAKER_01] Yeah. Well, no. Cloud Desktop now. They've sort of unified it into one app. Oh, okay. I didn't know that. [SPEAKER_00] It's very nice. [SPEAKER_01] Oh, yeah, yeah. Well, and ChatGPT still has a desktop app as well. That's right. [SPEAKER_00] Which makes the Codex app even more confusing. Yeah. Cloud and ChatGPT have both had a desktop app for a year? Yeah. Two years? [SPEAKER_02] Mm-hmm. [SPEAKER_00] I forget about that. And then the rest of this essentially became like, it was like, that was your chat app. And this is your code app. The problem is the Code app was the natural place to put all of the agents, to put the plugins, to put the skills, put all of the things. So now they've built these monstrosities over there that can do all kinds of crap. And like, you know, built-in desktop agents. Like, that's a thing that unless you have a business plan, you're not using it in ChatGPT right now. [SPEAKER_01] Yeah, you're right. Yeah. And I think, I don't know. So that's why I was thinking they need to solve it. But the potential way to solve it is say like, hey, ChatGPT is Codex now. You can still chat with Codex inside the Codex app, but everything's called Codex, right? That's one way to solve it. The other way to solve it is say, hey, we're now launching a work version of Codex that's called WorkGPT or something dumb. Like, I hope they would come up with something better than that. [SPEAKER_01] But it keeps like the chat, it keeps like some part of the ChatGPT alive. [SPEAKER_00] We're going to call it 09A3. [SPEAKER_01] Yeah, yeah. God, somebody in marketing over there helped them come up with something better. One of the, I feel like one of the most successful model launches of the last year from a naming convention was Nano Banana. Oh, yeah. And I feel like they should just keep their internal names for stuff. [SPEAKER_00] Yeah, and it's funny because I doubt that was the original intent with Nano Banana. [SPEAKER_01] No, it wasn't. They went on the record. They've done interviews about this. It was just like, I think it was someone on the team called it that because. [SPEAKER_00] It's what it was being tested as on Arena, wasn't it? On LLM Arena or something? [SPEAKER_01] Which is how people knew about it. Yeah, exactly. [SPEAKER_00] And they just stayed with it, which was perfect. It's absolutely what they should have done is stayed with Nano Banana because it's fun. Like, can you imagine if we were like, oh, well, I was using Strawberry. No, it was Orion. What was it? I love those names. I think those were a lot of fun. [SPEAKER_01] They are fun, but I get from a versioning standpoint, it's hard to know which one to use. So it makes sense to have 5.4, 5.5, 5.6. [SPEAKER_00] Bigger number, bigger intelligence. [SPEAKER_01] Yeah, yeah. So it makes sense. And I think whether they stick with Jewel Alpha or not, probably won't. But if I go to chatwpt.com right now, right? So they've sort of landed on, like, in the app, you just have the instant version, you have the thinking version, and you have the pro version. And if you look at all of the different applications, they've sort of all landed on this interface. And I think this is the cleaner way to do it. [SPEAKER_01] But obviously, you can come in here, and you can set what version you want to use, which also makes sense. But they've still sort of simplified it to these three modes, and I think that makes sense. [SPEAKER_00] Yeah. Yeah, I do too. I like that they've made that easier. I like that everything on the left is now, like, very collapsible. Like, you can close your agents. You can close your GPTs. You can close your projects. And just have them as one-liners right there so you don't have to have them all open all of the time. And I really, really like that because it makes me feel like a cleaner UI. [SPEAKER_01] Yeah. And I don't know if people know this, but now you can just add a project at the bottom here. So I could say, like... I didn't know that. Yeah. So I could say, write me a story on this. And why don't I pull up something fun just to do an experiment? [SPEAKER_00] That's a good idea. Do something fun. [SPEAKER_01] So let's look at that new memory thing that we just teased. So I'm going to go back to X here for a second. And I'm going to go home. I'm going to share this tab. And what's that chatDBT memory? I'm sure it will come up. Okay. So OpenAI just launched this new thing. So we've been researching new ways for chat memory to carry context across conversations and keep useful over time. Today, that work is rolling out as a more capable memory system in chatDBT. [SPEAKER_01] Great. Let's click on this. So dreaming. Better memory for a more helpful chatDBT. So memory is what helps chatDBT learn your preferences, projects, and constraints, allowing future conversations to start from shared context rather than scratch. And so they kind of walk you through how it's changed over time. [SPEAKER_01] And then over the last year, dreaming has supplemented saved memories to create a step function improvement in chatDBT's ability to personalize responses and offset the staleness of saved memories. However, historically, it was never sufficient as a standalone memory system. Today, we are launching a significantly more capable and compute efficient memory architecture built on top of dreaming. The memories synthesized by dreaming are reviewable through a summary of them made visible in the memory summary page. Interesting. [SPEAKER_01] So you can correct the memories, it looks like. That's cool. And then when we think about what a good memory looks like, a few things come to mind. Carry forward useful context. So you tell a chatDBT something once and it remembers that information in subsequent chats. Follows your preferences and constraints. So if you describe a preference like you're a vegetarian, then chatDBT should take actions that are consistent. [SPEAKER_00] And remember that anytime you talk about food. [SPEAKER_01] Yeah, exactly. Stay current over time. Memory should account for the passage of time. So imagine the user is planning their birthday party for next Saturday. Eventually, Sunday arrives. That's interesting. So it kind of tells you a little bit more about how to do that, right? So that's interesting. So I'm going to copy this link. I'm going to go back to chatDBT and I'm going to say, write me a story on this. So I've read it. I know it's interesting. [SPEAKER_00] Two birds with one stone today. [SPEAKER_01] No, I'm not going to write about this for a while. I solemnly swear I will not use this in the newsletter. But I just want to show you what's happening here. So I've assigned it to my project. So this is my project folder where I have all of the context about the newsletter. [SPEAKER_00] You've got a main story skill. [SPEAKER_01] Mm-hmm. And then now it's using a skill. So this is how this is basically I've taught it. Essentially, everything it was just saying about your preferences over time, I put that into a skill file here that can do that for me automatically. So I already have a version of what they've just launched here. But what they're saying is they're doing all of that context management inside the app now on your behalf. So you don't have to manage skills to manage your context. They're going to start doing that for you with the memories. [SPEAKER_01] So anyway, so what it's doing is it's then processing. It's fetching the link. It's using my skill the way that it's outlined here. And then it will eventually come up with the main story. And we'll rank it. We'll say, you know, is this neuron worthy or not? Which I think will be fun. Any thoughts so far? [SPEAKER_00] You know, something I'm circling back to build here. But I don't know if you caught this. It was very much glossed over by everyone. But Windows got skills. At the OS level, Grant. [SPEAKER_01] Okay, tell me more. [SPEAKER_00] I spoke with, you'll remember, Pavon Davaluri? Mm-hmm. He's the executive vice president of Windows and devices. And yesterday I was chatting with him about kind of some of the things they add. And he said the one thing that I feel like nobody freaked out about as much as they should have was Windows skills. He said because adding skills at the OS level was a big deal. You know, imagine if they just worked wherever you went. [SPEAKER_00] Like instead of it being, yeah, in any AI you're using or whatever. I just, I really, really feel like that's a cool step toward a truly agentic OS. I also joked about, I said, are we able to quit using AI PC now? Are we just going to accept that PCs are expected to be AI capable and go back to calling them PCs? He said, I think so. Yeah. [SPEAKER_01] It was a stupid brand to begin with. Kill the AI PC. This is, the best way to think about AI is it's like the third iteration of computing, right? Yes. It's the third iteration of computer. It's the, whether it becomes, you know, down to the OS level or whether it's a layer on top of the operating system. It is like the, it will be eventually the uncontroversial, obvious way that we interact with our computers. [SPEAKER_00] Low key though, the goal is to reach a point where a trillion parameter model can run locally on an average laptop. [SPEAKER_01] Yeah. That would be amazing. That's kind of the, the over a few years goal. Yeah. So we'll just look at this real quick and then we'll just close this chapter. So open AI says chat to PD can now dream up better memories. So your best AI chat should feel less like a customer service form and more like a coworker who remembers the project, the constraints and the weird preferences you mentioned three weeks ago. And then it says open AI. So linked open AI, which is not what I'd like to do. I like to link to the actual thing. [SPEAKER_01] So that's already wrong. But I was trying to make chat to PD feel more like it was like that with a new dreaming memory system. The name sounds like a Pixar short about a laptop with unresolved childhood issues. You know, it's a funny thing I've noticed. AI models like to personify objects. Like this is something when they make jokes, they personify objects. And I try to tell you not to do that. Yeah. So for example, it says it's about, in this case, it's a Pixar short. Sorry. [SPEAKER_01] It's a Pixar short about a laptop with unresolved childhood issues. In this case, the laptop is personified. It's a person that has child unresolved childhood issues. It's like the AI is subtly trying to force you to empathize with machines. Yeah. And it'll do this a lot. Like it'll make a joke. Like it's like a filing cabinet with an anger problem or like all sorts of stuff. And it's so weird. [SPEAKER_01] You know what though? [SPEAKER_00] As a guy who reads a lot of sci-fi and fantasy, it's very much a thing. There's a series of books. The Piers Anthony Zanth books. It appears Anthony is somewhat problematic, so I don't want to get into him. But these books are brilliant. Where things happen like inanimate objects do things you don't expect. Like, you know, if you need a pair of shoes, you go pluck them off a shoe tree. If you need, you know, your statues talk. [SPEAKER_00] You know, you might go to a different statue for different problems. You might, you could, like lots and lots of things are a bit alive. And I wonder if that's not maybe a sci-fi thing. Maybe a thing from books. Maybe it's they just would rather us not unplug them and feel like. [SPEAKER_01] Subliminally messaging to us is the way. It's not a conspiracy theory. Exactly. That it's subliminal messaging from the AI models to try to get you to empathize with objects. But actually, I think that maybe that it's just very forward thinking to your point. And that in the future, you know, we will be talking to all of these devices. Like, we talk to other people. So maybe we should just come to accept it. But it's just funny. It's one of those things that you have to iron out of it. [SPEAKER_01] It's like, no, when you make a joke, like don't make the joke from the point of view of an object. Make it from the point of view of a person. Like, it's like a weird thing that you have to tell it to do. [SPEAKER_00] Preparing us for a future where in two years we're standing in the kitchen talking to our refrigerator about what we can make for dinner tonight. [SPEAKER_01] Exactly. That's what I'm saying. [SPEAKER_00] Which is probably a thing that's going to be happening. Like, hey, what's inside you? That is rotten. [SPEAKER_01] Reveal to me your secrets. Yeah. So like, as you can see here, like this is like formatted like a neuron article. Obviously, this is not everything I would say. But like the factual parts, like when I write that myself, like that's directly from their page. I just make sure that it's accurate. You know, maybe I change it to focus on things that I think are more important than this from what I read. But yeah, it like basically runs through exactly how to do it. [SPEAKER_00] To be fair though, all you said was write me a story on this. Yeah. I didn't give it any direction. Like there was no, here's six or seven different sources I'd like you to consider. You know, my take on this is that it should, it's problematic or it's great. The thing I will say is that, you know, Chagibut's first memory launch was, I guess, 2024 now. And it has worked better since then. [SPEAKER_00] Like the truth is it changed the way, in my opinion, that your AI works for you. Like if I, I get different types of responses from my work account than my Corey account. And that is because they know very different Corys. You know, one of them knows, you know, Corey, the nerd who occasionally has a car problem or likes good coffee. [SPEAKER_00] And the other one knows Corey who writes and travels a lot and does a bunch of weird things, records a lot of video with a guy named Grant, who he always wants a quippy line about to introduce a new podcast. Yeah. Speaking of which, if you're here, I know it's a small group, but we're really grateful you're here today. Please take just a moment to hit the subscribe button. We crossed 20,000 on YouTube, which is great because that puts us in at 31,000 overall or so between the podcast platforms as well. [SPEAKER_00] We're really grateful for every one of you. If you would subscribe, like the video, it helps us continue doing these and show the value in what we're doing. And people care to watch them. And we enjoy having you here and all the comments you bring. Also pop by and check out the neuron.ai and sign up for our daily newsletter if you don't already. It's read by 700 and some odd thousand-ish people. And we'd love for you to be one of them. [SPEAKER_01] I want to highlight another thing that came out today. This may be like right before we jumped on. Blog post from Anthropic about when AI builds itself. So this is a blog post all about their progress towards recursive self-improvement. For people who don't know, that's AI that builds itself or self-improves. And they did a huge blog post about this. Let's see. [SPEAKER_01] They say, taken far enough and given enough compute, the trend of AI systems delegating to other AI systems is capable of fully autonomous designing and developing its own successor. This is called recursive self-improvement. We're not there yet. And recursive self-improvement is not inevitable. But it could come sooner than most institutions are prepared for. And so they're kind of benchmarking their progress towards that. [SPEAKER_01] So they sort of give you this really cool visualization here showing the progress from chatbots to coding agents to where we're at now, which is autonomous agents. And then at some point closing the loop. Very interesting and potentially horrifying. [SPEAKER_00] It is. It's yeah. It's both equally awesome and gives me pause. [SPEAKER_01] So this is so this is cool. So this is where we're at now. So this is Claude writes a significant portion of Anthropics Code. As of May 2026, more than 80 percent of the code we merged into Anthropics Codebase was authored by Claude. Before Claude Code launched in Research Preview in February 2025, this number was in the low single digits. That shift also shows up in the amount of output per engineer. And then they go into how that works. [SPEAKER_00] Last fall last year, they were saying almost all of the code is written by Claude already, weren't they? [SPEAKER_01] Yeah. Yeah. Well, they were saying that they're getting there. I don't think they have made the claim that it's like 100 percent really until this year they were making that. But here they're saying more than 80 percent of the code they merge. And then it shows you code contributed per person by quarter. Right. And so this is like how much code the average person is contributing, I think is what this is. Let me actually look at this. [SPEAKER_00] So is that. Oh, that's that's. Sorry. I'm going to do my own. A person and their AI is why that's going up. [SPEAKER_01] Each bar is the average over the days in that quarter of lines of code merged per active contributor shown as a multiple of the pre 2025 average. Wow. Way to make this. [SPEAKER_00] And that's assuming them using AI, I'm sure. [SPEAKER_01] Yeah. Yeah. Yeah. The hatched final bar is a partial quarter. It averages only the days observed so far, not a full quarter. Dash lines mark public announcement dates. So this sort of shows you like when Claude Mythos preview came out, Claude Opus 4.5, et cetera. And so it's kind of this confusing multiple system. Let's see if they have something easier to follow. [SPEAKER_00] So if lines of code is a lousy production metric for this, it's a good way to measure it, I think. [SPEAKER_01] Mm-hmm. [SPEAKER_00] I guess it's the only way. I don't know how else you would measure it. [SPEAKER_01] Yeah. This is interesting. So this is Claude Code session success rate. So internal Claude Code sessions at Anthropic Weekly 4 million trailing mean LLM judge success. Task complexity is assigned by an LLM classifier that reads a summary of each session and picks one of four levels of a fixed rubric. So how to read this session success is determined by a Claude judge. [SPEAKER_01] A session is deemed successful if the Claude Code agent clearly succeeded at the user's tasks without requiring corrections. So did it get the job done? And it's a very high number right now. So this is for trivial tasks, routine tasks, and substantial tasks, which has seen the most growth recently. Wow. Open-ended problems has actually kind of fallen a little bit, which is interesting too. [SPEAKER_00] You know, and I wonder if, you know, the reason – Chagibity and Jim and I have both done a lot around open problems in math, things in science, and there's been less of that from Claude. And I can't help but wonder if that's not due to their more intense focus toward coding specifically. Mm-hmm. And maybe less of a diverse approach where I feel like the other labs have leaned a little more into all of that. [SPEAKER_01] Yeah. Yeah. I think because Claude is sort of coding agent-pilled in the sense that they think the best general purpose agent will be the best coding agent because it can do anything on a computer. So they're not trying to – they don't have the resources that Google has to make a bunch of bets. [SPEAKER_01] And they don't have the same – and they're not willing to diversify on a bunch of different bets like OpenAI. [SPEAKER_00] Well, I guess there's a variety of schools of thought. Like some believe that like if you solve coding, the rest will just fall into place kind of. Others believe you need to lift it up kind of all at once. Others believe that, you know, you should specifically be focusing on any area that is verifiable. Like by solving the verifiable areas, the others will come into focus after that. And there's – you know, the truth is we don't fully know that yet. [SPEAKER_00] But so far all of them seem to be doing a pretty darn good job. [SPEAKER_01] Yeah. [SPEAKER_00] Depending on what you're looking for, you know. [SPEAKER_01] No, it's really true. So this is kind of interesting. They have a possible future section at the bottom here. So it says what happens next depends on two things, whether the trend continues and what we choose to do if it does. So we can imagine at least three future scenarios. So let's look at these. But today's AI capabilities are widely diffused. So this article features many exponential trajectories, but these trajectories may actually turn out to be S curves. [SPEAKER_01] We may be approaching the bend in that curve where it returns to scale, diminish, and the line straightens, then flattens. And then it's, you know, kind of talks about a little bit more about that. And it says alternatively, the binding constraint to AI progress could be in the supply chain, not the model. So advancing and diffusing the frontier may require more energy and compute than presently exists. Wow. They're admitting that they don't have no compute. [SPEAKER_00] That Musk has barked about for a couple of years now that I think we tend to glaze over, and he's not necessarily wrong on it, is that energy could be our biggest problem. Like, like, like an interesting thing. And he always shows this chart that shows China's growth of its energy infrastructure and ours, where ours is sitting pretty flat. And China's energy infrastructure has exploded over the last, like, five or ten years. And they're continuously building more and more. [SPEAKER_00] And the truth is, the biggest bottleneck we may face is power. [SPEAKER_01] Yeah. I think certainly in the U.S. that's true. I think in China, probably the bottleneck is chips and chip production, which is why they're focusing on doing that. [SPEAKER_00] And they may offset for another. [SPEAKER_01] Yeah, you're right. So, like, chip fabrication, grid expansion, or interconnect bandwidth. So, interconnect bandwidth is the thing that is really complicated and wonky that people don't get. But that's, like, how much power can you actually move from one place to another, right? And, you know, we have famously, we have a bunch of power that is, like, waiting to get connected to the grid. And it's backlogged, right? And part of the reason for that is because the interconnect bandwidth is, you know, delayed. [SPEAKER_00] Can you explain that a little bit? [SPEAKER_01] Explain interconnected bandwidth? [SPEAKER_00] Yeah. Explain what you're talking about when you mentioned that there is power waiting to be connected to the grid. Yeah. And I apologize if I just don't be on the spot for something kind of technical. [SPEAKER_01] No, it's okay. And I don't, let me put it this way. I understand at a very high level, not at a very, like, deep level. So, let's just give that caveat at the top here. So, there are transmission lines. And transmission lines move power from transformers, the, you know, physical transformers, you know, and the source of the energy, the power plant. And they move that to the place where the power is used. [SPEAKER_01] And so, there's been a big effort over the past six years or, you know, since around 2020 when it was, there was an effort to electrify the grid in a big way. There's been a big effort to update the transmission architecture because it's really old. From, say, giant wooden poles in the air? [SPEAKER_01] Well, there's that, okay, not, let alone that, you know, so putting those giant poles underground, so grounding the transmission lines, also, you know, updating the transmission lines just in general because a lot of the grid infrastructure is from, like, the 1970s and hasn't been updated since then. You know, some of it's at least 25 years old, so that's an issue. And it's also very expensive. So, I think there was some statistic that came out, you know, we can look it up. [SPEAKER_01] That was basically per, you know, per mile of transmission line. It costs X billion or something like that. So, think about it. If you're going to dig up or retrench and redistribute all of the transmission lines across the U.S., it's going to add up. It's going to be really, really expensive to do that. [SPEAKER_00] You know, and people don't realize, like, every winter across the Midwest and the northern part of the country, thousands and thousands. And during storm season, thousands and thousands of power lines get dropped. Like, every ice storm that rolls through, that ice gets heavy, snow gets heavy, and it drops power lines and power poles. Like, I lived in an area where we had a major ice storm a few years ago where we got almost a half an inch of freezing rain, which is a lot of freezing rain. [SPEAKER_00] And there was a 20-mile stretch of highway where every power pole dropped to the ground. The first one falls, and it yanks the second one, yanks the third one. And before you know it, it's this domino that just goes miles and miles and miles of crashed power lines. The other benefit to underground is when that happens, you don't – I mean, the benefit to underground is that in things like winter storms, you don't usually lose your power. [SPEAKER_00] Because they're not out there snapping against each other in the wind, they're underground and not collapsing, falling, blowing up transformers and things. Okay. No, no, that's great. [SPEAKER_01] Yeah. So basically the level that the grid can accept is, you know, only so much power, right? Because there's a lot of coordination. Not only all of the, you know, what's the actual physical limitation of how much, you know, transmission lines we have and how to move that power around. There's also the regulatory burden of that. [SPEAKER_01] So there's all of the legal, like, approvals that you need to add new power plants to the grid, which, you know, the current administration, they've been focused on trying to remove as much of that burden as possible. Yeah. But a lot of the power that wants to be put online is a bunch of excess renewable power. And so the current administration has certain thoughts about that. Yeah. So, you know, I don't know how much of that power has been put online over the past year or so. [SPEAKER_01] But there's, like, the last time I read, and I think it might have been 2024 or 2025, that there was something like – and don't quote me on this – but something like 30 gigawatts of renewable power that was, like, just waiting to come online. And I assume some of that has happened in – over the past year or so. [SPEAKER_00] And it's a thing that's needed not just for AI. The fact is we are a more electrified society and more connected society than we've ever been. You have tons of things. If you go into a house that was built in, like, the 1940s, 1950s, 1960s, even the 80s, you'll see a quarter of the plug-ins you see in a new house being built. You'll see a quarter of the switches. There aren't switches everywhere. Not every wall has two plug-ins like they do in new construction. [SPEAKER_00] I mean, now every wall has, like, a quad box of plug-ins. And back then it was – you know, if you go into an old house, you'll see, like, maybe two plug-ins in a room is pretty common. Well, the apartment that I'm in is – [SPEAKER_01] Even the houses aren't equipped. Yeah, the apartment I'm in is from, like, 1930s, and I think there's maybe, like, two plug-ins per room. [SPEAKER_00] Do they both require adapters? [SPEAKER_01] No, thankfully. [SPEAKER_00] No, somebody swapped the plug itself out then, probably. [SPEAKER_01] Yeah, but, yeah, so it has been updated, thankfully. But, yeah, it's a similar issue. So one way to look at the AI boom is everyone's seeing, like, oh, there's so much waste going on with all these data centers that are being built and all of the power. And, like, while you could argue that there – and I've heard some people argue that there's no other use for all of these chips that people are buying from NVIDIA other than, you know, for AI. [SPEAKER_01] So if AI doesn't work out the way people want it to, that's, like, a ton of money that's wasted and we won't have useful infrastructure like we did during the 90s with the dot-com boom and the dark fiber that was laid in the ground. The other way of looking at that is saying, like, well, we actually need all of this power infrastructure. And in order to get the utilities and, you know, private companies to build more power plants, we need to give them an incentive to do that. And AI is a great incentive to actually make all of those commitments come to fruition. [SPEAKER_01] So, actually, if you care about having more excess power and your price per power actually going down over time, you want them to build as much as possible. Like, you actually want all of this to happen. And even if there's an AI crash of some kind, we will now have all of this excess power that we didn't have before, which would be good for all sorts of other use cases and lowering your power bills, you know, and all that. So, you want there to be a positive incentive for them to build power. [SPEAKER_01] And I think that's something that we should, you know, not forget, even as we're concerned about, you know, the power usage of AI and what it means and all that. [SPEAKER_00] I want to pull up this message from Ryan asking if anyone noticed the price of a 5090 has gone from $2,000 to $2,500 to $4,500 to $5,000. Okay. I have, because I have been, for a good number of months, eyeballing an RTX 6,000, the black wool, the thing Grant has in his laptop, I should add. [SPEAKER_01] Speaking of that, I'm going to join on that stream. [SPEAKER_00] But you should know they're now $8,000 to $10,000 to buy an RTX 10,000. Yeah, if it has RAM, it's lost its mind right now. And that also means that, like, even entry-level laptops are going to hit that. It's going to hit printers. It's going to hit, you know, all of this stuff so hard. And memory is just nuts. Some of that's AI. Some of that is manufacturing issues. Some of that is there's a variety of things that have played into it. [SPEAKER_00] Definitely AI is hitting home on it. Poor gamers, though. So, you know, they gave us these great graphics cards. And now the catch with those graphics cards is, first off, the crypto people came in and made their price skyrocket in the 2010s for years and years and years. Because everybody was mining Bitcoin and wanted 18 GPUs in their empty closet to mine, you know, quarters. [SPEAKER_00] And so that sent them through the roof. Then AI comes along, same story, sends them through the roof. And some of that is that, you know, I don't know. But it's, I don't see a short-term scenario where RAM gets better. [SPEAKER_01] No, because it's just, like, really hard to produce it. Like, we would need, we would need the, we would need, like, Taiwan semiconductors to, like, you know, build, build not just the plants in Arizona, but plants in all 50 states in order to try and, like, solve it. Yeah. Or at least increase their capacity equivalently. Okay. [SPEAKER_00] There was a lot of data center talk this week. There was? Yeah. Tell us about that. For anyone who hasn't looked up, Grant, would you pull up Microsoft's Fairwater data center? See what you pull up on it? Last year, we interviewed Scott Guthrie, who was kind of the man behind this whole thing. And the Fairwater data center is probably the most environmentally friendly data center in existence. Like, it was very much built with that in mind. [SPEAKER_00] And they made this big commitment about earning the privilege of being able, this week, to earning, what did Sidi say? Yeah, earning the privilege to build a data center in a community, to create a data center community. [SPEAKER_01] And you know why he phrased it that way, right? [SPEAKER_00] Because people are fuming mad? [SPEAKER_01] Yeah. They don't know how to backlash. They don't know how to backlash against AI, so they backlash against data centers, which is perfectly reasonable from the point of view where you don't see the benefits and you only see the downsides. [SPEAKER_00] You know, and the flip side is, a number of the problems that you regularly hear barked about have largely been dealt with. Like, there was this disastrous data center in Tahoe where, like, they built this data center and it took all the power. And suddenly, the power company gave the people in these towns, like, hey, we're just not going to provide you power anymore. And you got, like, 96 months or 18 months to figure out how you're going to get electricity moving forward. Because we're just going to do this. [SPEAKER_00] And that's awful. But it's also not the norm, I would say. But Fairwater, the key to Fairwater is that its whole cooling system is built in the way, like, the radiator in your car is. It's meant, it's not constantly pulling in fresh water to cool these things. Satya said it uses about the same amount of water in a year as a single restaurant. Wow. [SPEAKER_00] Which is, you know, I mean, nobody would complain about new restaurants in town. We get excited when new restaurants show up. And they come by the dozens in some cities, you know. But the thing to know is that the water they put in them is about that equivalent and can be good for six to seven years before it's necessary to swap. [SPEAKER_00] I mean, there are probably, I understand, certain amounts that they replace here and there throughout the year for various reasons. But overall, these are built as environmentally friendly as one can be so far. And hang on, I'm letting you in, Grant. [SPEAKER_01] Okay, it's going to be, it's going to have my camera on to start. So we're going to get some echo. [SPEAKER_00] Echo, echo, echo, echo. Echo. Sorry. No, you're good. But, you know, the other thing is they're trying to be self-sufficient with their energy, committing to, like, not raising the community's energy rates. Like, if rates go up, they're shouldering that cost themselves. Also investing in, you know, a broad variety of local charities, you know, doing their best to keep its noise down. [SPEAKER_00] You know, looking for industrial park areas that are already loud, already zoned for that kind of stuff. And, you know, essentially trying to come in and win. Win their approval. Not so much, you know, some early ones maybe felt a little to people like they were sort of being snuck in the back door. And, like, we had one in Missouri where the voters voted out the entire city council. [SPEAKER_02] Wow. [SPEAKER_00] Because they approved it? They approved a data center. An election was next month. Everyone in office was voted out. The mayor, the entire city council. Wow. Yeah. Yeah, it was wild. You know, my dad asked me the other day because dad was interested. He's like, I don't really follow this. He said that I see there are a lot of people fuming mad about data centers. He said, what's the story here? Are they right? Are they crazy? Are they? [SPEAKER_00] I said, you know, as with anything, there's a little everything. You know, they absolutely have some fair concerns. They also have a certain amount of information that I think is maybe a little dated. Yeah. And could stand to probably sit down for a rational conversation and figure out, you know, if we were interested, here's what it would take. You know, because, you know, you are talking about new good jobs that they create as well. [SPEAKER_00] But there are definitely drawbacks. And I can see why someone would want them in community, maybe necessarily. But, you know, if you have an industrial park, I mean, I can't envision a scenario where it's any louder than having a place that manufactures cars near you for sure. I think the number I keep hearing is like 60 decibels, something like that as an average, which, yeah. [SPEAKER_01] I think the main criticism that I've heard is that, you know, when you have all these, you know, quote unquote, unregulated gas turbines that use a lot of methane and that can have downstream health consequences. And that's an issue wherever there's a natural gas turbine. [SPEAKER_00] And wherever there is a landfill. We never talked about this on a thing. Yeah. [SPEAKER_01] You brought this up. This is a good one. [SPEAKER_00] I lived in a town in southern Missouri where about four miles outside of town was a large landfill. I won't mention the company it's owned by because they own a million of them. There's another one not far from here. But landfills by nature create methane. As things rot underground, they create gas. And the way they build landfills, they have these pipes throughout them that take in this methane, funnel it to a central thing. [SPEAKER_00] And it's essentially a giant flame in the center that's ignited and burning most of the methane that comes out. With that said, you drive around near it and you absolutely smell it. [SPEAKER_01] I just thought of they need to use those as excess power plants for data centers. [SPEAKER_00] We always jokingly called it the eternal flame of stink. And it absolutely stunk if you were within a quarter mile, maybe. Yeah. Not far. Not far. It didn't overtake a town as far as stink goes. But your trash had to go somewhere. Yeah. I get it. And the good news is methane comes from biodegrading. [SPEAKER_01] Yeah. It's a natural source, right? So it's not like you can get that. That's not a turbine you can turn on or off. Correct. That's like you have to do something with that gas. They should just pipe that into like data centers. But it's like you're already using something that exists anyway. That was our joke. Yeah. [SPEAKER_00] Put the data centers next to the landfills. Pipe that natural gas over to power the data centers. I say that as a joke, but. It's a good idea. It's seriously, if you put the data center next to a landfill, the landfill is creating a pretty much unlimited around the clock supply of methane gas. And then you take that, pipe it in. I don't know. There you go. And, you know, I am not an electrical engineer or an engineer of any way, shape or form. Yeah. Don't quote me on that. Don't quote it. [SPEAKER_01] Okay. Do some, do your own research. But that sounds interesting. No, but the other alternative to this, right, is solar power and batteries. You can't have just solar power because, you know, it doesn't, you know, you need these things to be available for, you know, like large spikes in the fluctuation that come with training runs. So, you know, you can't just have, you know, solar power only runs for, you know, whatever, 12 hours, 16 hours a day. Yeah. Effectively. [SPEAKER_01] But then you can store that power and you can release it when you have those micro fluctuations. And it makes a lot of sense to, to do it that way. So you just need to scale batteries. I feel like scaling power, energy storage is one of the most important things we can do over the next 10 years. Like if we should have started it 10 years ago, but like, like at least doing it soon. All right. Before we, before we get too off topic here, I do want to talk about Hermes desktop, which you have not seen at all. Is that right, Corey? [SPEAKER_00] No, I haven't. I haven't played with Hermes yet. I've been pretty, pretty, pretty open claw stuck, but it's been on my list. I just haven't. The truth is I'm afraid that having it on a system where I already have open claw, I'm going to get a gajillion conflicts. Will it make my responses smell? [SPEAKER_01] Well, don't quote me on this, but I think they have a part of the onboarding. That's funny. All right. I think part of the onboarding is they'll onboard your open claw stuff into Hermes. Like it's pretty sophisticated. Yeah. [SPEAKER_00] Cool. [SPEAKER_01] Let me pull it up. [SPEAKER_00] Yeah, please do. Curious to see it. I hear good things. [SPEAKER_01] I think you took my screen backstage, so you have to bring me up. [SPEAKER_00] Up, up, up. Hang on. Do, do, do. Grant's computer. Is that the one we want? [SPEAKER_01] Yes, sir. [SPEAKER_00] Okay. No. Okay. There it is. Where is, uh, it's not showing up though, Grant. [SPEAKER_01] Yeah, it is. Whoa. [SPEAKER_00] Cascading Grant and Corey's. [SPEAKER_01] All right. So let me pull up a Chrome. [SPEAKER_00] Railway's going to delete your project. Just thought you should know. [SPEAKER_01] Yeah. It was intentional. Okay. So let me show this on the website. Hermes desktop. They have two websites. No, no, no. Okay. So Hermes desktop. Let's talk about it. So if anyone has seen Open Claw before, you know where this is going. But this is basically, in my opinion, the Claude to Open Claws, Open AI. [SPEAKER_01] Nice. So it's the creators of this agent. They put a lot of their own taste into it. They have some interesting design aesthetics. And they try to give you as much power as possible without you having to worry about it. So the desktop version is essentially like Open Codex. Looks like Codex. Yeah, exactly. It's like Codex, but with any model you want for any task you want to do, basically. Yeah. And it's pretty, pretty awesome. [SPEAKER_01] So before this desktop came out, which just came out this week, you had to install it in your terminal. It's very scary for people who've never done that before. It's not at all what you like. It's not a fun... They made it very easy, but you still get very scared doing that as a non-technical person. Now with the desktop, it is much, much simpler. And so the way that you would install this is you would pick whatever your platform is. [SPEAKER_01] You open the download and let me show you this guy. All right. So right now I'm in the settings. Let me switch out of here. Do that by closing this. [SPEAKER_00] Keep a close eye out to make sure you don't like expose your... [SPEAKER_01] It's not really connected to anything right now. Oh, that's good. Yeah. [SPEAKER_00] I'm always so scared with things like that because there's like so many ways you could accidentally expose all of your keys. [SPEAKER_01] Yeah. Yeah. Good call. But yeah. So this is what it looks like. So when you have it installed. And so I'm going to zoom in just a bit here. And right now it's running a local agent step fund. So I'm going to say, hey, can you tell me about Hermes agent? And what it's doing is it's not, you know, it's not calling OpenAI. It's not calling Claude or anything like that. It's talking with a local agent right on my computer. [SPEAKER_01] Okay. So, hey, I'm actually running inside Hermes agent right now. Hermes is an open source AI agent framework by New Research. Think of it as a terminal based general purpose assistant, similar in spirit to Claude Code, OpenAI Codex, or OpenClaw. And then these are some of the things that make it unique, which is fun. Oops. So they have a skill system. So it can learn from mistakes and reusable workflows by saving them as skills, which gets loaded into future sessions. We have a whole stream about that if you want to check it out. [SPEAKER_01] Cross-session memory remembers your preferences, environment details, and task context. It's provider agnostic. So it works with 20 plus LLM providers, open router, which you can load your own AI keys there, anthropic, open AI, deep seek, local models, et cetera. And you can swap models without changing anything else. And then you can talk to it across Telegram, Discord, Slack, et cetera. But you don't really need to do that with this app. You can just talk to it directly in the app. [SPEAKER_01] I found that was the most complicated part about using these agents was that you had to sign up for a Telegram account and then you had to message them or use them in your terminal. But here you can just start a new session like you would on Codex, which is really nice. And I have only just begun to start playing around with this. [SPEAKER_00] Now it's like you're still able to deal with it through your Telegram or whatever you want and have it connect to dropping things in your folders, your files, and Google Meets or Google Drive and stuff. Yeah. [SPEAKER_01] Yeah, totally. So if you go to skills and tools, right, it comes with these are 77 tools that are already built in here. I didn't have to go in and connect them or anything like that. It's already set up, right? So these are all things that are built in. And there's a lot of power here. It shows you how this stuff works. And this is what I mean by them kind of imposing their own taste into the tool. They've already pre-installed all of these things for you because they're like, [SPEAKER_01] hey, you're going to want something that does this. Yeah. Which is really cool. And then you can go in and you can look at, okay, what are these things actually? So it's got agents. There's creative tools. There's data science tools. And then there's this thing called tool sets, which shows you like the types of things it can do. And this is really interesting where you can do cron jobs. So you can set up automations, code execution, clarifying question, browse with automation, image generation, et cetera, et cetera. [SPEAKER_00] Wow. I'm curious. And I need to download the OpenClaw app and look at it too because now that I've seen this, I'm curious to see side by side. What's different? What's the same? Because I know they have an app now as well, but this looks super simple. It's an interface I'm familiar with. I can see a million reasons you'd want to use it this way. [SPEAKER_00] You know, something that I think would be really useful is, and why I think the Windows skills discussion is really interesting to me, is because wouldn't you want your Hermes and your Codex and your Claude code all using the same skills? [SPEAKER_02] Yeah. [SPEAKER_00] Like, you don't want to have to fix it every time in Codex, every time you're in Claude code, every time you make a change. You wouldn't want to have to make a change everywhere you might use. And I say that there's probably not a lot of skills you would use in all those places, but to have the flexibility of using them in anywhere you opened would be really valuable. [SPEAKER_01] Yeah. Yeah. I agree with you. I think probably the hardest part about skills is that you'll be updating them frequently, like at least you should be. And whenever you notice an edge case or something that you want to kind of like train out of it or teach it to do differently, then you want to create a new version. And the way that we've told people to do that is, you know, go use the skill creator skill and say, hey, update this skill. [SPEAKER_01] And, you know, the interface makes it really easy on Codex or Claude where you can just like one click button and update it. But then you want to keep track of those versions. So actually the right way to do it. [SPEAKER_00] I don't think you have to use the skill creator skill to do it, I don't believe. [SPEAKER_01] Say that again? [SPEAKER_00] I don't think you have to use the skill creator skill to make a change, though. I think you can just tell it, hey, you know, make this tweak. I mean, you could if what you want is like a 2.0 or a 3.0. And maybe that is what the answer is. I don't know. [SPEAKER_01] I think they're using something under the hood to update it and make it a one click plug and replace button. So, yeah, I think if you tell it to just do, hey, you know, update this, it does it. Like I've done this on Codex. And even in ChatGPT, the work account that we have, you can do this where one message at a time, you can update the skill and do it like a versioning system where you basically say like update the skill to do this. [SPEAKER_00] There's this issue and it shouldn't ever do that. Please update it to make sure it won't go. [SPEAKER_01] But the right way to think about it is to do it like versioning. So, like, let's redo this and let's call it like 1.1 or 1.2 or something like that so you can keep track of it. So if you were doing skills locally on your computer, if you're doing skills in your Hermes agent, if you're doing skills in Codex and Claude, if you use all four of them, that could be kind of complicated to make sure all of them are up to date. But you use a Google Drive system to manage yours. [SPEAKER_01] I feel like I remember you saying this. Yeah. [SPEAKER_00] No. Oh, I thought somebody said they used. No, but I'm really interested to check out the Windows skills because I think I'm a big fan of the idea of let's have these all in one folder and give that folder, give access to that folder to each of my various tools that might need it. Be like, this is where skills live. This is where skills live. When I say make a skill, that's where it should go, where then it's instantly available to each of those. Like, I feel like that's a thing that should be easy to do. Right. [SPEAKER_00] Where then you could make the change, whether you're in Claude, whether you're in Codex, it wouldn't matter. I would like to think you could, you know, make a change to that core file either way because they're all using the same file type so they can read them. [SPEAKER_02] Yeah. [SPEAKER_00] You would just want them to be somewhat agnostic of, like, mentioning different models and things probably. Yeah, I think that's right. You could say, like, you know, use your best thinking model instead of use GPT-55 on extended, you know. [SPEAKER_01] By the way, something I just thought of. So this is the Codex app I'm looking at now. As you can see, it's very similar. [SPEAKER_00] God, it is identical, isn't it? Yeah. [SPEAKER_01] They basically just wanted to make Open Codex. It was just like, respect. I get it. I wanted it. I'm using it. I'm happy with it. But if you go here, you can see all the plugins that we talked about earlier at the beginning of the stream. All hanging in there. [SPEAKER_00] Build iOS apps plugin, by the way. How about that? [SPEAKER_01] You notice I have that turned on. Yeah. [SPEAKER_00] That's something we've been waiting on for a hot minute, isn't it? [SPEAKER_01] Yeah. So there's all these cool ones in here. And I think they talked about some of the new. Maybe I can sort by new. [SPEAKER_00] Of course not. That would be too helpful. [SPEAKER_01] No. Yeah. If only... OpenAI! Hey! [SPEAKER_00] Tebow! Yeah. [SPEAKER_01] If only I was in charge of this. I would fix all these problems, I find, with it. But okay. So let's see. Creative production. That's one that we talked about earlier, right? Sales. This is another one that we talked about earlier. [SPEAKER_00] Investment banking. These are ones that we didn't touch on. They're all right here. [SPEAKER_01] Yeah. They're all right here. So these are the new ones. But when you click in here, you can see what they actually have under the hood. So it shows you... Okay. It uses these 17 apps. So... Oops. I didn't mean to click that. So you can scroll through and see all the apps that it has connected. And then it has 14 skills built in. So you can analyze data quality. You can build dashboards. Let me know if this is not zoomed in enough. Jupyter notebooks. KPI reporting. All sorts of stuff. [SPEAKER_01] So it's already got... [SPEAKER_00] I had it visualize things. [SPEAKER_01] Oh, you're a little quiet. What did you say? [SPEAKER_00] One of these times, I need to show how I've used... How I had Codex visualize stuff. Remember the dashboard? I had it built like... Like visualize a transformer. Yeah. Visualize. I had another one that was like an adjustable universe scale or something that it... [SPEAKER_01] Let's see what happens if I ask it to visualize. Visualize, yeah. Because I have all these plugins turned on. I've not transparently been able to use all of them yet. Because I just announced like half this stuff. Yeah. So let's actually see what plugins it decides to use to actually do this. [SPEAKER_00] Yeah. Because the one I did, it opened up these... It basically built HTML pages that were very interactive. You could blow things up, make them bigger, change a lot of numbers and change how it feels, how it acts, strengthen black holes and stuff. And I can't remember what it used. But it's on my personal account that I don't have access. My personal codex, which I don't have access to on this machine. It contains a project of particular shape, rent... [SPEAKER_01] Yeah, it looks like it's building... [SPEAKER_00] Build web data visualization. [SPEAKER_01] There you go. That's it. They don't exist where exactly we're listed. Interesting. Curious. Well, we'll let that dude's thing come back to it. [SPEAKER_00] The hoodoo that it does. [SPEAKER_01] Let me see if I make sure that they actually work. [SPEAKER_00] What do we have that we should make sure people know about right now this week? A couple of fresh podcasts and recordings I'll drop. I'm going to drop in yesterday or Tuesday. I did a live with Scott Hanselman from Build where we talked about... Basically, AI engineering was more the focus of it is what I would say. But it was really interesting. He walked through, showed some cool projects he's doing. [SPEAKER_00] This cool thing he's done with... He has a pancreas pump in one arm and a glucose meter in the other arm that are implants. And he built an app that talks to the meter all day and builds charts of his performance and sends notifications to his phone to let him know what this needs to be doing. And it was a really, really cool thing. And I'm grabbing the... Yeah, that was awesome. That was totally unplanned. [SPEAKER_00] It was just like, oh, there's this one thing that we really need. I also liked that he talked about, you know, use Vibe Coding to spend something up. And if people suddenly show up and seem interested, then fix it. Yeah. Okay, here's that link. I'm going to drop that one in here. Scott Hanselman interview. [SPEAKER_01] You know what's one thing that I've been testing lately as well that's kind of interesting that we haven't really showed off or talked about is Claude Design. Have you used that at all? [SPEAKER_00] No. I've never tried Claude Design. And that's really strange because I feel like it's a thing I normally would have played with. I remember it and being like, oh, that sounds cool. [SPEAKER_02] Yeah. [SPEAKER_00] It must have just fallen on a busy time or something. [SPEAKER_01] They haven't fully integrated it into the app yet, which they should. It makes it kind of difficult to work with it, to be honest. Like if you already have a program that you're building and you want to do the design for it. But it is pretty cool. I wonder if I have it. Do I have it pulled up? Let me check. Yeah, I do have it pulled up. Check this out. [SPEAKER_01] So this is it. It basically just makes front-end interfaces for you. And I'm going to zoom in on this. And I'm going to say on this. [SPEAKER_00] Oh, is this your project? [SPEAKER_01] Yeah, so this is something I'm working on just for fun. So basically it's like a way to work with these models to work from the screenplay as the source of truth. As opposed to some random prompt in some website. So I'm kind of trying to figure out exactly how to design it and make it look. But these are all different pages that it's created for me. This sort of shows like different visualizations. [SPEAKER_01] And if I wanted to change something to it, I could demo it here. So like in this case, I'm going to say, Can you change it so the windows are resizable? So I can make the left bigger or the right bigger whenever I want. And then the cool thing is once it creates these edits to this front-end design, [SPEAKER_01] you can download the code and give it to Claude Code on your desktop. And it can then go implement that for you. So it's a... [SPEAKER_00] Hey, review this code. [SPEAKER_01] It's a very lightweight way to mock up stuff without actually having to change all of the changes to your interface. Sorry, there's a massive helicopter overhead. [SPEAKER_00] I can't hear it if that's any consolation. Oh, that's good. Maybe a slight droney noise, but it's nowhere near as loud as you. Okay, good. [SPEAKER_01] Well, yeah, so basically it's a lightweight way to mock up stuff and make changes without having to actually change your entire interface. And you can connect it to... [SPEAKER_00] It kind of feels like V0. In a lot of ways. [SPEAKER_01] Yeah. Yeah. Yeah. And yeah, so just you can create a design system. So the design system is like where you make all of the rules for how it looks. And then all the elements will respond to that. Then it has all your files here. So you can go in and click and look at all the things it's been working on. You what it's created. [SPEAKER_00] Read your code. [SPEAKER_01] Yeah. And you can connect it to... That's funny. And you can connect it to... You can connect it to your... What's it called? You can then connect it to your GitHub repo. So it actually reads your code base and understands how everything works. Okay. And designs off of that. So I actually... I like it a lot. My only criticism is like put this in the app. Why is this web only? This is annoying. Like put it in the app. Yeah. [SPEAKER_01] I agree. As soon as possible. And make them integrate with each other as easily as possible. Because I would love to just be like great. Hand this to code. To code. Code. And go like implement it. [SPEAKER_00] Yeah. [SPEAKER_01] You know what I mean? [SPEAKER_00] In one fluid thing. That's one of the things that I like about V0 is that it also connects to V0's like web server service. Right. So it just speaks between the two. Like I can go over and say tweak that. And then hit publish and boom. It's over there. And it's up. But this is really nice. I like... And also I really dig what you've built there. [SPEAKER_01] Oh, thank you. Yeah. It's coming along. It doesn't look like this right now. Which is the problem. So I made a lot of changes to the front end. And then having the code version to go in and like actually figure out how to change it all and do it well. But an interesting workflow pattern is... So people have been using codex for mocking up front ends. And I was experimenting with this yesterday. Let me zoom out. [SPEAKER_01] And I said like, hey, here's the current version of my thing. And, you know, I have... So like here's the old... I don't know if I can click in. Yeah. Here's the old or the current version of my thing. This is kind of a crappy screenshot. Here's another one. Nope. That's really zoomed in actually. Oh, because it's zoomed in. It's zoomed in on my side. So try this again. Okay. So here's my old version of my thing and what it looks like. [SPEAKER_01] And I'm trying to get it. And then here's the work in progress version of my thing that I'm working on. And this is the ideal version of my thing. And then I show it that. And then I said, I'm trying... You know, I have this problem. I'm trying to create a designful UI. I want to marry these things. I want to make it feel very, like, connected. But I'm losing out on this element that I liked from this first version. And so I actually had the image model go in and mock up a new version. [SPEAKER_01] And this one's kind of hard to see. So I had to do a light mode version as well. So it kind of came up with a solution for a way to potentially implement it. And then now what I can do is I can take this screenshot and take this to Claude Design and say, hey, try to implement this with our design system. And then I can take that. And then I can take that to Claude Code and say, okay, now implement this because it's kind of been mocked up. Yeah. [SPEAKER_01] So it's a really cool way of going from, like, visual idea or, like, from problem to visual idea to something that understands the thing that you're actually building to then finally integrating it into the thing that you're actually building. I like that. [SPEAKER_00] I like that. [SPEAKER_01] It's cool. I think you should try it with one of your applications and see how it can get it. I will. [SPEAKER_00] I will. I may do that this weekend. [SPEAKER_01] Yeah. [SPEAKER_00] You may do that this weekend. Oh. By the way, we've got another video we should share. [SPEAKER_01] Yeah. [SPEAKER_00] Yesterday, we released an interview with Tools for Humanity. If you don't know who Tools for Humanity is, you might be familiar with WorldCoin or World, WorldID. This is a startup that involves Sam Altman. I don't know his role, but I know that he's an owner and founder. It's a startup that centers around human verification. And it's a really, really interesting discussion. [SPEAKER_00] We get a lot into things around how do you know that the person you're talking to is human. And that's what they've built this solution to do. And it feels a little Terminator maybe. Like there's this wall you hold that basically takes your picture and codes it, goes on blockchain and directly into the app on your phone. It's not verifying your identity. It's not verifying who you are. It doesn't have your personal information. [SPEAKER_00] It can if you want, but it doesn't have to. Essentially, what it's meant to do is verify that this unique person is indeed a unique person. Like it verifies that you have never been verified on their chain before. And if you have, it will catch that and it will, you know, you will not be verified. [SPEAKER_00] But the idea being that at some point we're going to get in this situation where we're in this situation where, you know, you could absolutely steal someone's voice and likeness and contact their loved ones and scam them and a variety of things like that. And to be able to have a way for a human to be like, yes, it's me, dad, is increasingly important. And I feel like they've done it in a way that is as uninvasive as is possible, if I'm honest. [SPEAKER_00] Like they're not keeping your data. Everything they have on you is, you know, broken up on the blockchain. So like it's little snippets all over the world. You know, no one person can pull your stuff back together. They can't retrieve you. It's very much lives in your app, your instance of world. And I think it's a really neat thing. [SPEAKER_00] We had a great conversation about how like, you know, we're within months to years of agents being, you know, 99% of Internet traffic. And with that, there's going to be a need for people to be able to verify. Yes, I'm talking with the person or for an app to to verify a purchase maybe or something else. And they really kind of they've been around for several years now working on this. [SPEAKER_00] And and they have hubs in, I think, L.A., San Francisco, some other cities as well, where you can go and get verified. And they also do other things. But you can even buy the orb to get verified. Everything they've done is open source. You can absolutely take their data and implement it into your app. You can implement it into other things. [SPEAKER_01] So I actually love that. I was thinking I would I think there's some application ideas that need something like that that I would love to build on top of and make something cool. Yes. [SPEAKER_00] Like a World of Warcraft account where you come in and your character's naked and his bank's empty and all his money's gone. Whatever happened to you, Grant? [SPEAKER_01] Um, I don't think I was ever hacked on WoW. [SPEAKER_00] No, I have been hacked on WoW so many times over the years now. Mind you, I haven't played in a good long while, but it was very much a thing. You would show up and suddenly there stood your character in all his glory with no armor, empty, empty bags. [SPEAKER_01] I feel like now that can't even happen to you because you can just request. You can just tell Blizzard like, hey, I was robbed. And then they like have a where it would get you is if you had been on hiatus. [SPEAKER_00] Like like if you were in there yesterday and it happened, you could get fixed. But if you had been gone for quite some time, a lot of times you were stuck. I got stuck once. [SPEAKER_01] I don't know if you've noticed, but it's working on the Transformer Explorer dot HTML down there. [SPEAKER_00] Yeah, it is. [SPEAKER_01] There we go. It's been working on it's been taking screenshots, making edits. It's very interesting. [SPEAKER_00] It is interesting. Wow, it is. It's so cool to watch these happen. I swear to you that a year ago, real agents weren't a thing. No, not like this. It's so crazy. And, you know, oh, the interview. I've got the link here. It's it's really cool. I was supposed to get in San Francisco last week and even spoke with them a couple of times. They even volunteered to come to the hotel and do it there so I could get verified while I was in town and I and record it. And I wasn't able to. [SPEAKER_00] So the next time I'm out there or maybe in Vegas later this month, I'm going to make sure we can link up with them and get a video getting verified because I think it's pretty neat. I think it's important. And I just dropped the link into the chat there. So anybody who wants to watch can. It doesn't have a lot of views on it yet for one reason or another, but it's super, super, super interesting. And it's worth a watch because it's a problem that's going to impact everyone at some point. Why is it not letting me open this? [SPEAKER_00] Open in the Codex browser. [SPEAKER_01] Weird. [SPEAKER_00] If it won't tell it, it won't let you open it. And it'll give you a better link. [SPEAKER_01] Yeah. [SPEAKER_00] It may even give you like a link to the file you can go drop in a browser if you're choosing. Codex browser won't open. [SPEAKER_01] Yeah. That's kind of annoying. Yeah. [SPEAKER_00] It'll sort it out. It seems like I ran into the same issue the first time I did it too. Yeah. It gave me a local host link is what it finally did. So just go pop this in your browser. [SPEAKER_01] Yeah. Yeah. It's pretty cool though. I'm liking the direction of where all this stuff is going. Yeah. Which is seemingly more useful, more powerful, and easier to use for normal people, which is like the right direction, right? Yes. All of this needs to be going in that direction. So I am excited about that. And yeah, there's been a lot of cool stuff. I don't know. [SPEAKER_01] Did, did, uh, GPT 5.6 come out while we were distracted? [SPEAKER_00] I don't believe so. I don't believe so. [SPEAKER_01] I kind of don't think so either. [SPEAKER_00] It did look like LM studio might've dropped a new mobile app though. [SPEAKER_01] Oh yes. I did see that. Um, you can take my personal screen down or you can hide it if you want. [SPEAKER_00] I'll share something on. GLP may, GLP ones may slow down biological aging. Those are interesting drugs. [SPEAKER_01] Yeah. [SPEAKER_00] Makes me want to go get my refill. [SPEAKER_01] Yeah. [SPEAKER_00] I mean, I transcribed 1.5. [SPEAKER_01] Ooh, Logan says we are cooking the world's best vibe coding app on Android and iOS. It's going to be so cool. [SPEAKER_00] Well, I assumed eventually Google would come to the party. [SPEAKER_01] Well, it's, it's, have you tried to make an iOS app with, um, uh, AI studio yet? I have not. Let's do, let's do a stream just on that. [SPEAKER_00] We could do that. Yeah. I've yet, I've yet to try to make an iOS app. Uh, and I know that, I know that they added it to codex as well. So maybe we, maybe we dual them. [SPEAKER_01] Yeah. We could do an iOS app on codex. Sorry. Uh, AI studio is Android only. So we could do an Android one on AI studio or you can publish it. Let's put it that way. [SPEAKER_00] That's fair. That's fair. That's fair. You will have to set up a cloud account with Google just FYI. [SPEAKER_01] Ah, yes. [SPEAKER_00] They will want all of the money. [SPEAKER_01] If only they would fix that and make it not so annoying to set up and make it really easy. That would be nice. [SPEAKER_00] Yeah. Yeah. It's weird. And, and, you know, of course there's also the element of the, you have to go through the whole app store thing. I say that with Android, you might be able to have it as a direct download from like yourself. You might be able to like bootload it somehow. [SPEAKER_01] This codex, like I told her, like I can't open this thing and it's going on like an adventure to try and fix it. It's like, no, just make it work. Just make it work. Yeah. Anyway, I am going to stop this. [SPEAKER_00] Okay. Yeah. And I think, uh, I think it's, I think we're probably good to call it a day. Uh, I, I, I really appreciate everyone who showed up and joined us. I know this was a much smaller, smaller chat than usual. And, uh, we just, we wanted to come on, on the off chance the model came and knew there were other things we could talk about. Wanted to talk a little about build anyways, cause there's a lot of interesting stuff. A lot more. I could say there too, that we didn't even touch on, honestly. Yeah. Let's, let's talk about it a bit before we wrap. [SPEAKER_01] We have 10 minutes. [SPEAKER_00] Okay. Okay. That works. Um, let's see. What was I just thinking of? Uh, Scout is really cool. Windows Scout. [SPEAKER_01] Tell me about Scout. Is that available now? [SPEAKER_00] Scout is, is right now in Frontier, uh, but is coming this summer. Uh, so you should have access to try it out. It's essentially, it's the first of their autonomous agents they're launching. And it is absolutely simple open claw. Uh, I've got a full article coming out on it, but you know, you go in and he's like, okay, I want to use Microsoft Scout. And it's like, Hey, what do you want to name your assistant? [SPEAKER_00] And you name your assistants like you're picking avatar or have it create your own. You do that. And then it's like, let's look at your email and your docs and your teams and see where maybe I could help you out. And it comes back with a bunch of recommendations. Like, Hey, do you want me making sure you're ready for meetings? Do you want me to gather this stuff for you in the mornings? And you're like, yeah, yeah. Yeah. And it'll do that. And here's the kicker. It talks to you in teams. It's just like another employee. [SPEAKER_00] So like, it'll message you in the middle of the day and be like, Hey, you got an email looks important. Uh, do you want me to reply to this for you? Or are you going to be able to make that next meeting? It looks like you're still in this meeting. Uh, do you want me to mark you as, as late and send an email reschedule? Uh, even the ability to say like, listen, I ain't dinner from five to six and I ain't changing that for nobody. And, uh, and, and if a meeting request comes, it will automatically go and send an email as itself, not as you. [SPEAKER_00] It'll be like, Hey, I'm Sebastian Grant's AI assistant and Grant can't meet tonight at five, but he does have some availability at four 30 or tomorrow at nine AM. Take your pick. And, uh, it'll do that for you. I, I jokingly asked him, I said, okay, but what if the requests from Satya? I said, I said, I said, we're thinking it comes from your team. And you're like, no, you're not scheduling a meeting in my meeting, but what if it comes from your boss? He's like, Oh, I can, I can set it to do that as well. [SPEAKER_00] Uh, Ryan's asking which platform is this Corey? I don't know if that's from right now. If we're talking about windows scout, it's going to be part of Microsoft 365. GoPilot right away or right soon. Right now it's in their frontier program, which is like their, let's roll it out and make sure it works. Get some feedback real quick. Uh, make some tweaks, but, um, it's, it's really sick. It's essentially it's, it's open claw, but it's open claw that I asked, I said, you said, do you have like educational materials on it? [SPEAKER_00] And he said, you don't really need them. He's like, it did ask you for a name. It walks you through the setup. It's just like, he's like, it's easier than, you know, opening a bank account or setting up Facebook. And, and over time it gets better. That's the other thing is it continually learns from its own data and the things you're doing. So it will constantly understand more things about you, ask you questions periodically. Uh, it'll have anything you want ready to go. [SPEAKER_00] And, um, it's absolutely open claw for dummies for lack of a better way of putting it. I shouldn't say it's, I shouldn't say for dummies. That's, that's not cool. But, but for non-technical people who, you know, right now this stuff is only available to people who either are engineers or not scared of a command line or, you know, willing to, willing to break some things. And, and the fact is that's, it's a very, very small percentage of the population and [SPEAKER_00] something that Microsoft does well, honestly better than any of them have done is, is, is taking stuff like that and making it available to the rest of people. Uh, and it's something that's always made me a little bullish on them through all of this. Cause like they have an audience that's bigger than anybody as far as like when you think of windows machines and at some point, if what they put out is good, uh, it'll be difficult to stop much like I would say about Google. [SPEAKER_00] Uh, but, but, uh, it's funny because Google's put out great models and, and project products that just kind of consistently miss the mark or don't seem to interest people. And it's been kind of the opposite with Microsoft where they focused more on the interface and getting the interface dialed in and, and its accessibility instead of the models. Uh, but the truth is the interface is pretty good. Copilot makes sense for a lot of people who maybe don't want to have multiple AI subscriptions, [SPEAKER_00] but would like to be able to use JGBT or Claude, or maybe some of these new models that they will come out with as time goes on. Cause they're very much just getting started like this, this isn't the end of this year, maybe not even this quarter, as far as releases go. I would absolutely expect, you know, much more to come over the coming months and for them to continue getting better and better. And, uh, I, I, I thanked Mustafa for, for, for, for making at least one of my predictions [SPEAKER_00] from early this year when we did our prediction. I said, I, I said, Microsoft will be competitive by mid year or something, or have a, a competitive frontier level model. And, uh, and I, I haven't spent enough time with thinking one yet to say that that is the case, but, uh, just based on how it's scored. I mean, if we're within two generations, that's, I mean, you know, and when you think within two generations, if it's scored up there with Opus 4.6, that's February. [SPEAKER_00] That's not like, that's not like, it's easy to see that and think, oh, there's those three models ago, but there are two models ago. But the truth is that means, yeah, like, yeah, that's, that's four months ago. [SPEAKER_01] That's well, four, 4.5 and 4.6 were like such a, uh, uh, difference, like such a breakthrough that I think if, yeah, if any model was at seven, that was good or open. Yeah. Seven was mid. Um, yeah, I, I got used to it eventually, but, uh, I think 4.8 is better from what I can tell, but I mean, 4.6 was all I needed. [SPEAKER_01] I think this will happen where they get really excited to release something new once they're on a roll and then they blow past something that was pretty good. And, uh, they, they've, they've both done this or something that's kind of a turd. Um, well, who, who knows? Because it depends on the use case maybe, but it just, it, it all happens so quickly that I think you got to give these things time, you know, unless it comes out and it is totally ridiculed by people and unusable. Sure. [SPEAKER_01] Then get to the next one as quickly as possible. But if not, let it, let it simmer. [SPEAKER_00] That was a big mistake with their, their, you know, constant shipping a couple months ago too, was, you know, some of those were a list feature ads. Some of them were, were pretty mid, but the desire to put them out every day. The problem is they buried the good ones with some mediocre stuff and it had a lot of things fell through the cracks instead of giving them some of those, maybe the push they really could have used to get going. [SPEAKER_01] I have, I have, they were, they were, they were cooking, but, but they were like cooking for like too many, too many. [SPEAKER_00] I have, I have a criticism I want to share of Dario right now. And this is a thing that's now happened like three times. I have plenty of criticism to Dario, which was maybe some are fair, some are not. Uh, but what I want to say is that, that when people talk about hitting rate limits and what Claude costs them, their answer is always to come out on a podcast and be like, that's because you're using it wrong. The fact of the matter is if when you create a product, you have a way you envision people [SPEAKER_00] using it and they use it a different way. People aren't wrong. Your guess was wrong. You created it with something in mind that is not the case. And it's your job to adapt the product to how the user is interacting with it. And, um, and that is a very, very common thing. And, and I, I think that blaming every developer has not been like, oh, well, you just don't know how to prompt it. And it's like, well, that's a you problem. [SPEAKER_01] To their credit. They, they released a blog post shortly after the 4.7 launch where they said, here's three things we did wrong. Um, and they went into technical detail on it. So I, I appreciate that. Very cool. I missed that. Thank you for bringing that up. I kind of read it. I like skimmed over it. I was like, okay, this is interesting, but still is kind of a turd. So I'm going to, but I'm, so I'm going to like begrudgingly use it, but, uh, I still [SPEAKER_01] ended up using 4.6 on a lot of my personal projects where I already had it queued up. [SPEAKER_00] So as a general rule, if, if, if you have a product and, and it's, and, and people are using it in a way you didn't expect, it's, it's time to reevaluate what you can about your product to make it work for the way the people are using it. If you try to get them to use it the other way, they will get frustrated. They will leave. And, uh, and, and I mean, and that goes with a cell phone that goes, you know, uh, it doesn't matter what the product is. You know, you've, you've got to assume that they love this. [SPEAKER_00] They want to use this, but it is expensive for them. How can we, how can we make that better for them? Uh, yes, yes, there absolutely is Brian. [SPEAKER_01] It's Silicon Valley episode. That's what he said. It's not an us problem about that. Yeah. Uh, the user is wrong. The user is never wrong. [SPEAKER_00] The user is never wrong. The user is right. [SPEAKER_01] Well, here's the thing. If the user is wrong, then you put out banger, like, uh, X essays like, uh, Tariq does and basically, and basically say like, this is how we use it. And then everybody adapts to the, that style. Like that, that is the way actually to solve that problem. Yeah. Um, maybe Tariq should be CEO of Anthropoc. I'm bullish on that. I would, yeah, I would be down with that. I have, yeah, I have, I have, um, I won't go into it. [SPEAKER_01] I, uh, anyway, I think we're at time, right? [SPEAKER_00] Yeah, we're at time. Well, everybody, thank you so much for watching. We really appreciate, we enjoy, these lives are the funnest thing we do. If I'm honest, uh, we really enjoy interacting with you all and trying out what's new and learning as we go as well. Uh, please make sure you like subscribe. It's a big help to us. Make sure you pop by the neuron.ai, sign up for the daily newsletter and, uh, stay tuned. We'll have some, uh, some cool stuff to announce in the near future that we can't get into yet, but you're going to want to know. And I realize I've been saying that for a minute, but I promise we're getting close. [SPEAKER_01] And, uh, if I could give you, if I could give you three takeaway points, right? It is Hermes agent is pretty sick. If you want a personal open claw like system, like a personal agent system, where you can work with whatever models you want. Um, definitely check that out. Codex is getting pretty good for non-coding. I even noticed there was a setting, which I had never seen before, um, in Codex, where if you go to settings, um, you can choose your work mode. Have you seen this, Corey? [SPEAKER_00] No. [SPEAKER_01] You can choose your work mode. So it's for coding or for everyday tap. Oh yes. Yes, I have. I have. I've not seen this before. [SPEAKER_00] So it's been maybe a couple of weeks, Max. It's, it's new. [SPEAKER_01] Yeah. So that definitely check that out. Um, but that should be like a front end feature. That should not be something that's buried in settings. [SPEAKER_00] That's kind of their, their code code answer is, is you could do it either way you want. And one's a little more. Yeah. [SPEAKER_01] Use Codex for non-coding, um, load up on all the plugins you want and definitely check that out. [SPEAKER_00] Lots of things you can't do in chat GPT that you can do there. [SPEAKER_01] Yeah, exactly. And third thing, I don't think I actually had a third thing. Um, keep an eye out for, oh, Gemma. We didn't even talk about Gemma 12B, um, which is the new local model. Um, but what I would do is I would go to, uh, your Hermes agent and say, hey, I'm going to use, um, Gemma 12B. Why don't you install that for me and run it? And it'll do it for you. So do it. Do the thing. Yeah. [SPEAKER_01] So the third thing is check out this cool stuff from NVIDIA and their new Nematron model and Gemma 12B and check that out and, uh, have a fun weekend playing with local AI. [SPEAKER_00] Grant. They said Microsoft scout can order you a pizza. Oh, that is, that is, that has been the Corey benchmark. Benchmark. Key benchmark. And it's like, you can tell it. You want me to, and, and, and one of the examples they showed was it saying, Hey, do you want me to order your dinner in? You know, and it was like, I can get you a pizza or something. And I said, how does it know what kind of pizza to get you? And it's like, you give it your door dash. And it's like, Oh, you know what? [SPEAKER_01] I have not seen a good demo of agents buying things. No. Why are they not showing demos of agents buying things? [SPEAKER_00] No. Stripes dropped a good tool for it. I mean, for, for like a wallet system, uh, lots of them are using chat. BTs using plaid now, which it must be that nobody's got it. [SPEAKER_01] Working consistently enough or like, they don't want to. Yeah. Yeah. There's some, there's something there because why haven't we seen a good demo of this? Somebody should be doing that. Somebody go, go build. Uh, I'm going to, yeah, I'm going to, I'm going to build a little agent. [SPEAKER_00] Build a wallet or we will. Yeah. [SPEAKER_01] Yeah. Like don't make me do it. Yeah. I'm the least experienced person here. [SPEAKER_00] I don't want to do stuff, especially not in banking. [SPEAKER_01] Fine. I guess I'll go make the thing that's missing that everyone needs to build. Seems like that is where you should start though. [SPEAKER_00] Okay. Once again, we're going to leave this time. Thank you once more for joining us. We had a lot of fun. Uh, we've got a cool guest going to come on here in a couple of weeks. I think I don't know the date yet, but, uh, uh, odds are if you watch AI videos on YouTube, you'll very much know him and be interested. So, uh, something cool in the works. And on that note, we'll see you back next time. Make sure to check out that interview with tools for humanity and, uh, Mustafa on next Wednesday. Okay. And on that note, that's it for today. [SPEAKER_00] Farewell for never. Bye-bye.