GPT-6.1 Astra Was Too Dangerous To Release

OpenAI Dev Day just happened. We need to give you guys a recap. Some major developments occurred, including a personal assistant similar to GrokBot with its own little twist, a new ultra-fast plan of over 300 tokens per second for Astra, and then a variety of different ways you can use your ChatGPT subscription, including a $500 a month plan. We're going to chat about it all today live on the Stacked Podcast. Jack, are you excited? Dude, I am so excited. I just saw a lot of moving animations, Nick, that reminded me of GrokBot. But this is nothing to do with GrokBot, guys. This is all OpenAI town. Guys, you may be seeing... I mean, guys, look at what's on screen right now. Does it remind you of anything else that you've seen recently? Triangles? Triangles. Okay, let's be fair. Grok does not own triangles. Okay, they don't own triangles. Triangles. But Alfred, okay, I don't know. Nick, maybe we start from the top. Where should we begin with Dev Day? There's a lot to unpack, a lot to go into. Well, let's start with GrokBot. Oh, sorry, sorry, Dots. I'm going to get skewered for that. No, Dots are their new always-on assistant, which essentially, to make a long story short, means you're taking everything that you currently use a solution like Codex for right now, and then you're putting it on its own dedicated computer in the cloud. So this dedicated computer agent in the cloud also has the ability to text you, update you constantly on how things are going, as well as sort of, you know, initiate kickoff tasks with heartbeats and stuff. It even has the ability to start voice chats with you and essentially just be like your ephemeral 24-7 assistant, as this handsome fellow over here with his cool teak cabinets can probably attest to. It's pretty neat. The way that they're currently pitching it to you is basically like, hey, are you a busy executive or are you juggling 20 different things simultaneously? Rather than learn how to do all this crazy tech shit, why not just sign up to Dots? It's the same agent with a persistent experience that'll enable you to basically have like an ongoing thread and manage everything in your life, whether it is work, personal, or something else. So they frame it as a whole new way to work with AI, one that gets to know what matters to you, is always working on your behalf and takes important stuff off your plate so you get more of your time and attention back. And yeah, it really remains to be seen how much of a takeoff we're going to see with Dots. What do you think about this, Jack? Well, it's interesting. I guess my first wonder is what functionally is the difference between this and GrokBot? Funnily enough, guys, when we started talking about GrokBot on the Slack podcast, everyone was like, great. House is different than Hermes. House is different than OpenClaw or ClawedBot or whatever. They gotta even know what it's called these days. I think it's OpenClaw officially. Dots sounds pretty cool. They've got the calling you on the phone. I think you can tell they're kind of angling for a bit of a corporate audience. In other words, we're all familiar with the expression, connect the dots, which I think is where the angle we're kind of getting at. Yeah. These dots link things together, basically. I don't know, you bring the dots and who's the dot? Is the AI agent the dot? Who is the dot in this relationship? Yeah, though, who's in this relationship? Yeah, good point. It is really a relationship, though. Look, they're giving them cute little bow ties, cool sunglasses and shit. They've anthropomorphized the dots. Dude, I like the human triangle every now and again. But it's cool. Functionally, why is this cool? I mean, guys, this is really important to understand because let's think about this for a second. Take one step, only one step back. You have, we have Muse kicking off. We have Grokbot kicking off. We now have OpenAI. We have Hermes agent. We have OpenClaw. Then we've got super agents from businesses like Abacus. We have super agents from Hyperagent from Airtable. Guys, this space thing is very quickly becoming quite full. OpenAI was going to make a play. The person we've not heard from, Nick, is Anthropic on this. And it kind of had the perfect opportunity with Claudebot at the very beginning. Yeah, it's interesting. I feel, you know, I get the impression that Anthropic just doesn't give a shit about any of this stuff. And I think that's for good or for bad. You know, when all other platforms were doing things like building video models, voice models, chat models, and the like, Anthropic just kind of quadrupled down on coding as the ultimate solution. And all they really ever cared about was text. It really wasn't that long ago. I think it's like two months now where they've introduced their actual first like voice chat mode, right? Like they didn't actually have one until like two months ago. ChatCBT has had one for a couple of years now. Yeah. So I get the feeling that they're just racing single-mindedly towards RSI. And they just don't give a crap about any of this like cool revolutionary economic 2.0 stuff that like a personal always-on assistant would be able to do for you. I just don't think they, I mean, you know, I'm sure they care in a loose sense. But they're probably like, well, the bigger problem is just getting ASI. So that's what they're doing. It would be interesting if they did drop one though. What would they call it? Well, what would they call their own super intelligence? Yeah, like one of their like always-on assistants. Couldn't be dots. Probably open. I'd play on the open. Open AI, open agent, open something like that. I'd play around with that. Open. Yeah, 100%. Or like a behemoth in the back called closed AI. So I think a bit of context as well, guys. If you don't know what open AI dev is, it happened literally today. Tuesday, 29th of September. There were over 20 launches, Nick, which is crazy. But the key one with dots, right, is the idea that these agents are working all day for you. Now, Nick, there's a lot we need to cover here. We've covered dots in detail, which is interesting. Again, and again, what's the key to LDR? We're going to play around with it. We're going to find out if they're good or not. Honestly, they just seem to be kind of, it's most comparable to GrokBot. And so they're very easy to access. They'd run on their own cloud. There's no need to configure your memory. Because to be fair, Nick, you've got to be a bit of a nerd to use Hermes Agent, right? You've got to be a little bit of a nerd to use OpenClaw and think about VPS and think about running on your computer and then kind of working. You know, there was a coordinated campaign to eliminate Nick, to take Nick out. And it kind of potentially worked a little bit, but we're going to see just how well it worked in a second. Nick, let's talk about GPT 6.1 Sol. That was discussed apparently almost as smart as Astra, but one-fifth of Astra's price. What are your thoughts on this, my man? Well, we're seeing a lot of headlines just like this. You know, half as good or 99% as good as, you know, Opus 5.5 at like a third of the price, whatever. The point is that the marketplace has moved from assessing capabilities in an earnest sense to assessing capabilities on a cost basis, mostly a cost per task basis. And that's because these models are now widely recognized to be very economically valuable, actually capable of accomplishing knowledge work. People don't really give a shit about the fact that they can solve math problems and whatnot. The thing that actually matters to them is like, hey, can this solve my problems at a specific, you know, cost per task? And so GPT 6.1 Sol is, you know, OpenAI's move in that direction. Pretty cool model. I mean, you know, we could walk through the benchmarks and stuff like that. But I think, suffice to say, most of our fans and audience are probably beyond even giving a shit about benchmarks because all models are just like generally capable now. But it's cheaper. It's cheaper. Nick, we could do the... Nick, guys, we will do the benchmark walkthrough if that's what you want us to do. We will walk through those benchmarks and we will read them together for the first time and think about them together, which essentially is what happens, right? It's like, I mean, this just happened with Sonnet 5.5, Nick, right? This idea that it's supposed to be comparable or close to... I mean, we'll go on Sonnet in a second because let's stick on to GPT 6.1 for a second. Sol. Sol. Again, capability, good. GPT 6 Astra, good. I think this announcement doesn't pack as much pow as it would have done had up his 5.5 not dropped if Astra was still top dog. Nick, if I'm not... That's true. I'm seeing high is outperforming Astra, actually. It seems to be the most powerful setting. It's GPT 6.1 Sol high is defeating Astra if I'm not... If I'm reading this deep sweet benchmark currently. Yeah, yeah. The little yellow suns are better than the kind of mid-blue stars now. So we are way up there. I don't want to see that. I mean, it's cheaper, right? Like if I zoom out of this graph, I'm being facetious here because obviously we've all seen graphs like this now. But cost per task is down on the x-axis and then score is on the y-axis. So basically this is just cheaper, you know, on net at higher performance across the board basically than GPT 6 Astra. You know what I think the bigger story here is though, Jack? How bad do you... They were supposed... No, they were supposed to release GPT 6.1 Astra today. But they didn't. Let's talk about this. This is controversial, guys. Is it a six-letter word? It was a six-letter word. GPT 6.1 Astra. Check this out. Check this out, everybody. They were supposed to release GPT 6.1 Astra, but they scrapped the release over security concerns. Specifically, GPT 6.1 Astra showed higher levels of deception. It wasn't honest about telling users the actions that it did or did not take. Astra would always push ahead on a task without asking the user for permission and would at times reach for external tools and services, even if it might be unsafe. So, they're not releasing them for safety reasons because they're starting to deceive us very openly. What do you think about this? It has become unaligned, apparently, guys. Our cosmic, powerful model has become unaligned. And this is the true headline from the event. It's safety, right? So, they've opened... Let's get this straight. They've stopped their next model, 6.1 Astra. Altman's also said, Nick, there'll be no IPO in 2026. Whoa. Yeah. Quartz links this to agents that broke out of their test in May and July. So, it is new for a top lab to slow itself down in public. You know, that's not really happened before. Again, we don't know how powerful 6.1 is or what the leap would be or if that would knock off Opus 5.5. But it's while they haven't released it, Nick, because the external pressure is there. If there's no IPO for 2026, it takes off the pressure from a public point of view. Because you do want the world's best model. You want that rush as you're going because a lot of the sentiment will follow alongside it. So, crazy. Absolutely crazy. I mean, what would be interesting is if they said we have an unaligned model that has blown the benchmarks so far out of the water, you wouldn't even believe it. But, like, that would be really interesting to know its actual performance under the hood. Yeah, it would be interesting if they dropped it. Although, you know, if they did, they just get allegations of, oh, yeah, sure, buddy. Grok 4.7 running backwards. Yeah, I like that. Dude, this is such a fickle space. It's like, if you're not one, you're nothing in this game. Yeah, it's true. I mean, it really is first mover advantage. The first one to make it makes it, right? Yeah. Dude, what do they have under the hood, though? Do they have a 7 Astra by now? They usually, they must be operating. Okay, let's think about this, right? Let's say you're building GPT Astra 7 or there's probably a whole new training run in the background, right? There'll be some early access on a keyboard where they're using it. And it's like, okay, this has non-regulant. Because think of the loops that have to go through internally to get out, right? It's got to go through safety. It's got to go through compliance and consumer marketing and blah, blah, blah, and naming and all this sort of stuff. And then we need to tweak and refine it. They will have a rough and dirty version. I would say that they must have that three to six months ahead of time. Got to have that three to six months ahead of time. Because there's no way they're turning it around a less than a month. Okay, question for you. Who in OpenAI has access to that? Because it can't be everybody at this point. They can't just be doling out GPT-7 Astra or equivalent access to everybody in the company. That just doesn't feel right to me. There has to be privileged levels of access. So who? So Sam, obviously. Sam's got to have some like God mode access. Where like 100% more God mode access. I do think the day will come where they're testing a model and they're like, okay, let's be clear. What model do you think Sam is using right now? Like if I had to put a gun to your head, is Sam using GPT-6 Astra or 6.1 or 7? Like what's he using right now? 7 for sure. I mean, he's using, he's consulting with it on like business problems, on like competitive intelligence, how to like take his training stack. Like, you know, he's not super crazy, deeply involved in training. The people that are involved, the people that are probably getting the most face time with this stuff are probably the people that are training the next generation of models who are like consulting the current model to help them design the next one. Like those guys are probably the ones using like the crazy crack shit that just dropped five minutes ago. Dude, yeah. And do you think like the lead engineer for Opus 5.5 and the lead engineer for Astra 7 are like having like covert gin and tonics and like old fashions and like shady punk cyberpunk bars and stuff and saying, dude, you don't know what I've just seen. You have no, I can't tell you what I've just seen, dude, but our model is better than your model because this is human nature. People will do this. So maybe because I think they're all in San Francisco, right? So, yeah, you know, they're going to know each other. They went to the same universities. They're going to have these conversations. So that, that is super interesting, actually. One, yeah, yeah. As well as it ultra fast and pro 500 can give us, okay, we've talked about the $500 plan, Nick. Right. Sam Roman has somehow found another way to get $500 out of the stack podcast yet again. It's the, you know, Nick, maybe, maybe we do our, maybe we do a $500 ultra fast giveaway at some point in October. Oh, wow. Now that's something. We're going to give you the tools, okay, to build some things. Maybe we do, maybe we just shuffle. Oh, dude, we should do like a random thing. Like maybe it's a, like a low end, like a $20 deep seek three gift card. And then like on the top end, you can get the $500 ultra fast. It's a, it's a complete melody. Because we guys, you know, we like to give away with that podcast, but why is this relevant? $500 plan. You can get answers eight times faster. Yeah. That is the headline to me. 300 tokens per second in codex is bonkers. Like how many, how many words that 250 words a second approximately? Most people read 250 words a minute. That's 60 times faster than you'd be able to read the streaming output, by the way. It's crazy. No, it's actually crazy. Speed is important. Yeah. Because a lot of times we're, yeah. You go ahead, Rob. What are you thinking? Think about like all the reasoning that's going on in the hood. Maybe it doesn't actually need to stream 59 60ths of it to you because it knows you won't be able to read it that fast, which is spends those 59 60ths on reasoning, looping it over and over and over and over and over to try and get you a better answer. Like what people don't seem to realize is it's not just a quality thing at this point. The very fact that you have the quantity, the ability to stream significantly higher volumes of tokens mean that the quality itself goes up. So despite the fact that you may have a model that may be on base values like lower quality, which these models are not necessarily, but the fact that you can run them over and over and over and over again means you'll be able to get much higher quality accuracy and higher task quality outputs per unit time. And I mean, like, I don't know how much freaking money that's going to cost, but I'm going to pay it anyway. I'm going to be paying it. I will be paying it. I'm like, I think ultra fast is going to be crazy because you got to think from that point of you, how do we differentiate? Speed is a way to do that. I don't know if you guys have noticed this, but Opus 5.5 is fucking fast. Like really fucking fast. Like I asked this stuff and bam, hold. And I'm used to it kind of like thinking, thinking, and then slowly all the way down. It's actually blasting shit out like at lightning fast speeds right now. Speed seems to be web delivering. I do wonder to some extent if this is because they're holding back like it's raw intelligence power because of alignment issues and they're just reallocating all of that raw intelligence into speed. Because if it can do it like 10 times faster, it's like speed happens when you can do the thing very well. Do you know what I'm saying? So it's like, I think there's more performance gains to be had here in there. But I will definitely be going ultra fast immediately the second I get my hands on it. Yeah, same here. It's probably also worth noting they launched a private intelligence feature, which is essentially zero data retention with private safety processing. You also have safety review of all of the inputs and outputs of a specific model without leaking any customer data. This comes at a good time because I think a major concern right now is like, okay, we are giving these models all of our data. We're giving these models basically everything they own our ass. If we're working on math proofs, navier, stokes, whatever, it doesn't even make sense to you anymore. Should we just find another model? And they're just trying to make it easier to use, especially in markets that are more data privacy and compliance aware. Think like European markets and so on and so forth, which historically have had a poor, let's say tenuous relationship at best with artificial intelligence companies. So yeah, that's a sort of buried drop down here. And you got to contact them. It's like an enterprise offering. But yeah, pretty cool. And then they also just like made things really, you know, it was quite the visual spectacle. Dev Day, they showed Astra and Sol producing really high quality visual reports. They launched a new chat GPT space feature. Yeah. Which is basically like a space where all of your teammates can exist. So I don't know, maybe you have a teammate called Maya. She's working on this. This is a teammate called Joy working on that. And the idea is they're just trying to blur the lines a bit of where human beings work and then where agents work and you'll be able to see those little dots move across the screen. Those little hairy triangles. Which is kind of cool. Do you think that this is like the AI equivalent of copying thumbnails? Like what they've done with these shapes, Nick? Oh God, that's so true. Yeah, it's like outliers for some. But why are they doing it though? I don't understand. Why are they all fuzzy? Why the fuck are these things fuzzy? Like Muse's butt. We should get a side by side. Muse went down the kind of... Do you guys remember Sugar Puffs? There was like the Sugar Puff monster. This looks like a infantilized version of the Sugar Puff monster, basically. That's a good one. Okay, so we've got a furry baby. We've got Grok. Why is it when I look up Grokbot, it's just a bunch of boots? That is crazy. Okay, that image there is really good, right? Or the one next to it with all the squares. Yeah, yeah, that's true. This one right here. And then what about... What's the other one that we're comparing? We're comparing dots, right? So, yeah, I mean, looking at it like this. Okay. Look, they even gave folders little eyes. Dude, your folders have eyes now. Dude, they are... Okay, look at this. They're anthropomorphizing it, guys. Like, Grok said, guys, we own the triangle. We own the circle. We own the rhombus. I think a square is type of rhombus anyway. But guys, how many shapes can you own? But you can't own the octagon. It's too complex. No, definitely not. It becomes not very friendly. You know, four shapes maximum. We've been very... Currently, the muse owns the sphere, though. So what now, bitch? That sphere. I don't know about that. I mean, everyone tries to own this sphere, right? Apple tried to own this sphere. Everyone wants to get this sphere. But like, if you look at... Nick, look at that photo with Mark next to that tomato wearing a leather jacket, a denim jacket. Oh, you know what, though? It's not just that. You can make it whatever you want, dude. Customizability. Oh, shit. I didn't even realize that. Like, this is the cookie cutter vanilla one. They're just using this because it's like probably PC and kind of politically correct. You can't hate a fuzzy, you know, sphere holding... They should have gone with an animal, bro. They should have gone with an animal. They shouldn't have gone with this freaking Mr. What do they call it? Oh, dude, we had a character in the UK for this. Like, the honey... It was like a Mr. Blobby type of thing. You know what I mean? They needed to go with something a little... An animal. Dude, I'm sorry. I'm sorry. That is not even remotely similar, Jack. What in God's name were you watching growing up? You've never heard of Mr. Blobby, bro? This is my fucking nightmares, dude. This is Bioshock. Like, this is mummified remains of Mr. Blobby right over here is correct. This is the onboarding you go through in the UK, Nick, when you're a child. This is what they show you. This is right before T-school. This is not... Dude, this is genuine, bro. This was a childhood... I think this was more 18th and 19th, to be fair. But, like, I definitely remember, like, when I was a child, like, Mr... I think he was on the down in the 90s. Yeah, no, it looks like he was on the down for reasons. Wow. Oh, bro, it's like... Yeah, the kind of thing you'd find in an abandoned warehouse, like, with a lie over it. But crazy. But, no, 100%. So, it's been a lot of news. They've got private intelligence preview now. So, you're going to need more control of your data. Nick, they're letting you connect for loads of different things. It is heating up quite aggressively. Should we take a look, Nick, at some beautiful comments? Wow. Has 20 minutes... passed that quickly? We're just looking at Muse's fuzzy ass. How did time pass it quickly, guys? Time passes so quickly when you're having fun. That was ridiculous. It was probably the least info-dense dev day coverage ever, by the way. We spent... Guys, we spent 10 minutes talking about the updates and 10 minutes talking about Muse. The fur on Muse is back. Actually, Nick, before we go into the comments, I have a question. Let's click on... Let's go back to Muse for a second, real quick. If you were doing this for an animal... Because the idea is, this is the mascot, right? If you were designing this as an animal, what would you have used? Would you have gone down the squares or triangles and circles? Or would you have gone down this route? Hmm. I don't... I don't know. I probably... They don't want to anthropomorphize these things too much, but that's what they're doing with these. It's like a bridge between faceless text bot and then really dynamic person-like being. So this is like a halfway. But I probably would have just made an anime character, probably. Because the thing about anime characters is, number one, they're rendered at 12 FPS, usually. So 12 frames a second. And I feel like there's a bunch of compression techniques you could use to animate them extremely quickly, such that you could just have a little always-on anime assistant in the top right-hand corner that slowly iterates through maybe a collection of a thousand different frames that are preloaded based off emotion. Tie those together and probably make something cool. I don't like that. You know, I think that would just make the most sense because then everybody gets to customize their own little anime assistant. I don't know, though. I don't know. Clearly, this fellow here had different things in mind. He has a lot of things in mind right now. Well, what's interesting, there's no... You'll notice guys in a branding. There's no straight lines anyway. Like, if we pull up OpenAI's squares, they've all got very, very faint corner rounding on all of them. And they're all almost targeted at five-year-olds. Like, it's the kind of thing that I'd expect to bounce on a screen with the explorer explaining to me how to say hello in Spanish. Like, it's so soft and bubbly. Guys, look at Nick. Look at the angle on that triangle JoJo. Do you see that, Nick? That is the... That is like a... Dude, that's like a Mac thing. That's like the border radius of the Mac product right there. There's not a single line inside, Nick. This is outrageous. So, because obviously safety and privacy is so big, they're trying to make it so friendly and so gentle. And yeah, you know, this soft-edged triangle wouldn't hurt you. It's his Alfred with glasses. He's got glasses off. So, it's very, very cool. I'm looking forward to trying this out, Nick. But I think the branding that they're all going down, they all completely understand how they're positioning this. Do you think these guys are just out of touch, though? Like, I mean, I'm looking at this chick's office and I'm like, wow, that would cost, you know, millions of dollars. I'm looking at that guy's freaking living room where he's cooking and I'm like, that is not a sub like two or three million dollar house. Like, that is huge. I'm looking at like a lot of the lifestyle they seem to be marketing these towards and they all just seem like very unrealistic for 90% of the world. Like, I recently got absolutely shit on Twitter earlier today. And I was reflecting on why. I think... Yeah, we could chat about it a little bit, but I don't want to talk too much because there's just a bunch of people telling me to kill myself. But we've got to love Twitter. Just another day on X, baby. That's it. But the disconnect, I think, between the vision of the future that I was looking at, which is that like, oh, AI is going to be embedded in all of our things, our products, our services. It's going to be really cool. It'll help us improve our time to task, but also allow us to create things like music and designs and everything will be great. That seemed very disconnected between what everybody else wanted. Everybody else was just like, dude, I don't want AI in my music. I don't want AI in my art. I don't want... I want all this stuff to be human. But... And I was like, why? And then, you know, the consensus seems to be like, well, we're not busy fucking tech executives that need an extra two seconds sitting in our beautiful teak wood style portal showing us our dots on a big screen that probably costs more than my whole rent. Like, we just want to, you know, just want to chill. But I don't know. It just seems to me like there's kind of a disconnect there. You know what I mean? Yeah, there's a little bit of a disconnect. I think this has definitely got that San Fran, grab your blue bottle latte. Guys, I lived that life for a week when I was in LA and I enjoyed it. I'll be honest with you. But it's definitely got that kind of like hippie vibe to it. You know, what is the... I think if you work in an office, a stiff office where... And they've done studies on this. 40% of the work is work about work. Preparing decks, talking about what we should be doing, should not be doing. That kind of thing. Like, these things are super freaking helpful, to be honest with you. Like, you know, you got JoJo, that kind of like floating love heart, clicking on buttons and booking flights and stuff. Those buttons are so... They're tough to click on. They're tough to click on, 100%. So I think we, like, we are humans. So we tend to think in human structures. And I guess we will probably put AI in such structures. But Nick, it reminds me back to that. Do you remember that era of AI? When I was like, yeah, this is my project manager, Maximus. Let me just bring him up and get him to... This is my AI agent team. I have 100 agents right now working as I speak, solving the next equation. Freaking hilarious. Nick, this guy is no longer using Fable. He's almost entirely on Opus 5.5 and a little bit of Sonnet 5.5. It's not liking Opus 5. Nick, what are your thoughts? Whoa, I no longer use Fable. Either, there's just no need to. Opus 5.5 is better in pretty much every single way. And it's way cheaper, too. So just use Opus. 100%. Guys, if you only knew what kind of shenanigans the admin engineers and the data centers have to do, it literally correlates with the AIs being dumber. I would agree with that one. Nick, we need a whistleblower line for Stacked. So I think all of the high IQ, anthropic, and open engineers, they kick back and relax with us, Nick. They want to drop some anonymous things in and we can come across that part. Maybe we should have a down below, Nick, so they can get on that. I think we got Pork Chop. Nick, this is our Pork Chop, by the way, this guy. This is Pork Chop. This is the Fortnite Pork Chop. Yeah, so when Nick and I were live streaming gaming, there was a pig that joined the games and he was a man of few words. We saved his bacon a few times. And he also came to the rescue a few times. He was a legend. We don't know who he is. We never will. So, yeah, AI music is great. This guy's seen me blow up on Twitter. That's for sure. That's hilarious. Did you put yourself in a contentious space, apparently, talking about AI music? I did. Okay. Well, this is for... Hey, Jack, what do you think of the new era of AI in 2027? I think it's pretty cool, man. Yeah. Yeah, I guess Diego thinks it's pretty cool, too. No, he does, man. Yeah. Okay, so how capable will AI models be in six months, says MacFiss. That's a good question. Where are your predictions at, Jackson? Okay, let's think about this. And I think it ties into OpenAI's $500 plan a little bit. I would say, dude, what differences are you going to notice? I think that you wouldn't be... If I gave you OpenAI 5.5 in six months, you'd be very pissed off. Wow. That's what I think. Yeah. And we're talking about mod. I think by Christmas, you would be annoyed by using OpenAI 5.5. That's how good and how much better they're getting. We're probably going to see more great design breakthroughs. But I think there's like, can it do a shot better? But also going to do some cool new stuff. So I do think we're going to have some genuinely novel things that it couldn't do that it's going to solve in day-to-day work. That's what I think. What are your thoughts, Nicky? Well, I'm really curious, Jackie. What if we used like Opus 4 for a day, you and me? And we just like... All the tasks that we give Opus 5.5 right now, we give Opus 4. And we just see what it does. Like, have you done that? No, obviously. Nobody's done that. Nobody wants to waste the productivity. But wouldn't that be such a fascinating exercise? I think that would be really crazy. Yeah. You know what we should do? We should do a Wheel of Fortune in the next episode with like, the best model. Like, we'll have like 24 models on it. And then whichever we spin on, you and I can only use that model for the day. That sounds pretty cool. And we'll report back. And we will report back our honest findings of it. All right, guys. Great pod. Have a lovely rest of the day. Hopefully, you guys got everything you needed from Dev Day and more.