OpenAI Astra, Local AI, and the Hardware Race

25:15Corey Noles:Let's maybe explain chain of thought real quick. So that's when you turn on thinking or reasoning when you're working with an AI model and you can see it have a bunch of words on the screen before it actually gives you its answer. That is the chain of thought. It's the thoughts, the thoughts, the words that it's producing before it goes on and gives you the final answer. Exactly. And this paper announcing it came out in February of 2025 here, talking about recursive depth and the idea of looped transformers. And, you know, it it had some concerns. Unlike approaches based on chains of thought, our approach doesn't require any specialized training data, can work with small context windows and capture types of reasoning that are not easily represented in work.

26:04So they can apparently get much, much more accurate results with this. However, that comes at a cost. Twitter, as you might expect, was in a bit of a tizzy between people who were like, heck yeah, bring on the machine gods, and people who thought it was Armageddon. But it got interesting as the whole thing kind of developed, because in a sense, there were some misunderstandings in this paper, some things like this is a really complex concept. So Jacob Pachocki, who is the chief scientist for OpenAI, came out and shared this great statement. He says, I want to prevent a race into unmonitorability kicked off by confused reporting, which is a great way to start a tweet.

26:49The depth of the computation graph for our present frontier models, including Astra, is within a factor of two of GPT-4. OpenAI has worked to preserve and utilize chain of thought monitoring since our very first reasoning models. We deeply care about this technique as it gives us a view into how model alignment generalizes from its training distribution. I do think it's fragile and unfortunately trending in a negative direction for reasons not contingent on architecture changes that I will write about soon. But there are things we can do to strengthen it, and it's a core goal of our current research program.

27:25You know, if you read the article, it goes through and it does say, like, this is used in limited ways. It is not meant to do away with chain of thought monitorability. They've apparently been very deliberate about it, but it is a bit controversial.

27:42Corey Noles:I just don't understand. So are you saying that they can still monitor the thinking? They just don't read it in tokens? What is he actually disputing here? I think what he's disputing here is that this is some move away from chain of thought monitorability altogether. together. This is like they've thrown caution to the wind and just taken this new route. And I think he's saying that it is very much not that. So will we still be able to monitor the thoughts? Yeah, but I believe that there will be elements within the system where maybe there will not be, but I doubt that that's going to be a difference to the user is my understanding, or much of what they see on the back end.

28:24It's more of a how data is processed sort of thing.

28:27Corey Noles:Think of how it reminds me of diffusion when I hear about it. When you think of like a diffusion model, instead of going token by token, diffusion spits out a full sheet of noise. And then it goes and refines that closer to what it believes the answer is. And then it does that time after time after time after time really, really excitingly to watch fast. And this is kind of, in a sense, doing a similar thing where it's going back over the same answer, double checking it and ensuring it's right and and really zeroing in um yeah then the notes that we have here says that basically it allows it to loop through difficult problems and spend more computation on them instead of a fixed amount of reasoning so it can basically do that that recurrent depth loop that you're talking about multiple multiple amounts of time uh as much as it needs to instead of like limiting the reasoning to like think for you know 350 seconds and then given answers.

29:26Yes. Yeah. Yeah. Because the truth is 350 seconds is arbitrary. That's just a limit somebody's putting in there.

29:32Corey Noles:And it's that's literally just off the top of my head. I don't know if that's the actual limit. But yeah. Well, that's yeah. Yeah. Yeah, of course. And but I think you call on something really good is that the idea is that this is something to tap into when it's needed. Like when the model is maybe struggling, knows it's on a complex thing, knows criticalness criticality is important here it has this extra tool to spin through um much like chain of thought in itself just it looks like faster and more successfully um right the other concern is that the looping architecture can actually improve exploit discovery and chaining which makes safeguards here a lot more consequential i feel like most of what's coming from the big labs is is very capable.

30:20There are definitely things you could do with them that are probably not great, but there are things you can do with a car that aren't either. So at what point do we, I don't know, take them away from the people who will get great value and great benefit out of them? Like, you know, I want it attached to my chat GPT health account, for example. I want that extra juice there. Yeah.

30:49Corey Noles:Well, it'll be interesting. We'll see how it goes. I think there's an argument to be made for they are feeling forced in a prisoner's dilemma to release this because if they don't release it first, then somebody else is going to release, you know, an equivalently better model. And then you've got to go back to training. So your model is sufficiently impressive next to what released last. Yeah. And there's a lot of like ways of like, oh, no, we're handcuffed. We have to release it, even though it's not safe. And I just some sort of. I'm a little bit skeptical of that now. I'm like, you two are doing this.

31:20Corey Noles:At any point, you can just not release this model until you know it's safe. So no one's forcing you to do this. Like, sure, the market wants a new model. We demand more AI. But what we have now is really good. And in both of their defense, I'm going to bring in Anthropic again here too. Both of them have taken pauses at moments now. Around post-training and other things. You know, OpenAI has a model. I think it's called IM4. IM4. IM4. That is now essentially chained up in a closet, never to be seen again or touched by even their own researchers. So, you know, I do think there's an element of taking that seriously because you wouldn't just do that.

32:07Like, these things are expensive. The compute to train a frontier capable model is millions and millions and millions of dollars. And I think just a willingness to show we're going to just lock this thing up says a little bit about the level of caution they're bringing to it now.

32:29Corey Noles:That's a good counterpoint. I'll take it. Grant, I also understand that Google dropped a new model. yes people people around the world let me tell you google is not dead in the race google is still google google is back as every headline says in north america every time somebody drops a model it's like is back yes it's so true okay so let me share my screen so what we're looking at here is gemini 3.8 flash which is their new gemini model and speaking of cyber capabilities 3.8 flash cyber this is google's new newest gemini model and they seem to be mostly releasing flash models i think that's because google has a lot of customers and it they can't really afford to release a big pro model really widely so they're they seem to be prior at least my perspective is they seem to be prioritizing these flash models that are very quick that are you know relatively inexpensive compared to the others out there and are quite capable so well and they work well with google's existing tool set i would say like like a smaller faster model has more place in ai mode it has more place in gemini i mean you don't want a pro model that you have to wait on over and over and over checking your email if there's a flash that can do it i would say okay so they're building on the momentum of 3.7 flash which i'll i'll tell everybody in the audience i just never even tried wasn't worth my time to even look that in that direction um from the benchmarks from everything i heard i just skipped over it but they have momentum from 3.7 flash you could say oh building on the momentum or perhaps uh overshadowing the lack of momentum um from three weeks ago three weeks ago they released this right um making it our third flash release in only six weeks today we're introducing 3.8 flash our best reasoning and coding model yet at the same speed and lower cost of 3.7 which is pretty impressive actually so okay the first 3.7 was supposedly competing on speed and price and now 3.8 is same speed and same low cost so it's the same cost but it's just a better model yeah so that's great what what is actually going on here um cory you and i have sort of gotten a little bit of benchmark blindness would you agree with that yeah like i look at this chart and i sort of say okay yada yada i don't really care like let me just try it out myself yeah and i'm like oh look it's as good as software engineer as opus 5 i don't buy it it's as good as gpt 5.6 i'm not buying it on deep suite we have like built in a bias and skepticism of Google models right now for some reason.

35:14And of anything attached to a benchmark, frankly, in my opinion. We talked about this the other day that I very much reached a point where my benchmark is my work. Give me the model. I agree. Let me do what I do every day. And if it does that better, that is awesome. And if it'll do new things, I want to know that. I want to know about its features. But as far as how it does on Val's finance agent V2 and Harvey's legal agent benchmark, I could not care less.

35:42Corey Noles:Well, here's what I'll say. I think that those actually are important to highlight because a lot of lawyers have relied on Gemini 3 Pro or 3.1 Pro, whatever the current version is, for its long context capabilities. I don't think a lot of lawyers have updated their priors in terms of realizing that Claude and OpenAI's latest models are quite good with long context as well. Because for a while, Gemini was the only one that could handle long context. Gemini was the first to the long context game. Yeah. Well, it depends. At one point, 32K was long context. That's true. I guess we're talking like 200K context and above.

36:18Corey Noles:Gemini, I want to say 2.5 Pro, was the only model that really actually could maintain consistency and fidelity of a long document at the time that it came out. And so I think a lot of people realize, hey, this is the model I'm going to use. And I think that a lot of people still use Gemini for those types of tasks. So it's good to know that they're good for legal work where hallucinations are aplenty. Yeah. Lawyers are getting, you know, fined or the case is thrown out all the time for, you know, false citations. Which, by the way, lawyers, check your citations. Make sure the case law that you're referencing is actually real.

36:55Corey Noles:Another thing worth highlighting here is that supposedly it spends more compute on hard jobs. So Google says that it works harder with extra reasoning steps and iterative tool calls on long horizon coding and agents. So this means that this could be a good agent model, which I'm excited to test. And I ribbed, but the truth is this actually looks quite good. The price is very reasonable, which matters more to me than it used to now that I'm burning a few hundred million tokens every day. Yeah, for real. So I what I hate about these benchmarks, though, is that like every time a new model comes out, it's like they've just pulled from a tossed salad of benchmark possibilities to find the numbers they feel like represent them most.

37:41You should remember when you see these cards, these cards are marketing and they are designed to position their model in your mind as better than the competition. And that is true of every company that does them. But like if we were talking about the same six benchmarks you see on every model everywhere, that would be a little different. But this is a little more like, you know, benchmark salad.

38:05Corey Noles:Yeah, I totally agree. I think if we're going to still rely on benchmarks, you need to have a report card that is an industry standard across the board. And everybody uses the same benchmarks and compares themselves on the same benchmark. And then you could really actually analyze them, which is why I like artificial analysis, which artificial analysis basically does that for everybody. They compare everybody across the same scoring systems, actually. What's missing here in my eyes, I love hearing that it's got strong agentic capabilities, but I don't see a computer use. Oh, there we go. OS World 2.

38:41How did it do? 59. That's solid. That's solid. I'll take it. Yeah, that's not bad.

38:48Corey Noles:Yeah, no. Humanity's last exam verified, 54.9. I mean, that's better than Sol, better than Opus. That's amazing. If it's accurate, I'm impressed. I do think that Gemini indexes on raw intelligence, whereas the OpenAI and Claude models, not that they're not intelligent, but they've been focused more on agent tasks and coding recently. Yeah. Yeah. Yeah, I look forward to diving into this some, too, and giving it a try, because I'm sure it's good. And I have an increasing interest in access to smaller models that can handle some of my, like, I do hours of computer use every day. And I would love a, oh, we got a new chart.

39:35Check it out. Average cost per task. It's going to win this one. Kind of, yeah. Yeah. Well, 5.6 Luna is looking pretty good over there. I don't really know how to interpret this.

39:50Corey Noles:How do I interpret this? Average cost per task on the Y. And what is the percentage? Is that like, oh, it's DeepSwee. It's DeepSwee. Percentage of accuracy percentage on DeepSwee. So like as far as accuracy goes, I mean, it's up there. I mean, we're mid-70s-ish. Yeah. Yeah. It seems to be above a lot of other models. I will also say that what I don't understand about this model is that, oh, the costs go in the opposite way. Wow, who designed this? That's what's wrong. Zero is on the farm right. Chart crime. This is a chart crime. Yeah, the zero should be at 0%, and that would make more sense to me.

40:39Now I'm following. So Luna is cheaper, a little less accurate. Luna's a good model. GLM 5.3 Flash as well. Popping up a good high score.

40:52Corey Noles:Yeah, this is kind of the area you want to be in for real work here. Yeah, I think you're right. I think you're right. I think they're in a small place. I notice GPT 5.6 Flash was cheaper. 5.6. This is the finance. I'm just skimming through this. Builds games. We'll have to test it in our next live stream. Yeah. Let's see. Flash cyber. Oh, yeah. This is what I want to talk about. So the other thing to note here is that they are having flash cyber as a separate model. And so they're sort of following OpenAI and Anthropics lead here by separating these and pulling them out. I think that's a good idea.

41:29Yeah. With like extreme cyber trained, cyber focused models. I think having those as just a separate tool, you can give to the professionals in those industries. I say give, sell to the professionals in those industries as they need them.

41:46Corey Noles:Look, if they want to give away the tokens so that everybody's critical infrastructure is shored up, we will gladly accept. This would be a bountiful bounty that you could give the world for taking all of our data. And Grant and I hold a lot of critical infrastructure. So we would like some of those tokens. Yes. I'm in desperate need. I'm in desperate need. But I wanted to highlight one thing. So Sundar Pichai, CEO of Alphabet Google, said Gemini 3.8 Flash Cyber hits frontier level vulnerability discovery and patching at flash speed and price. So this is 86.2 % on CyberGym, 47.2 % on CWE Bench, which I assume is a cyber war here.

42:33Yeah.

42:35Corey Noles:So this is pretty big. I would have to pull up the other models and where they land on this. I don't have that available right now. But it would be interesting to compare, do your own research there and see how they're doing. and artificial analysis also reported that you know this is their fourth flash in under four months it scores 59 on the artificial analysis intelligence index at high reasoning and um yeah that's pretty high up there 60s around the frontier like a bit above 60 a bit below so that's quite good like 61 to 63 or 4 is kind of where the top tier is right now yeah they also say that on the intelligence versus cost Pareto Frontier, so that's like the top 20%, it's at 58 cents per task, despite 40 % higher task costs from 30 % more output tokens.

43:30Corey Noles:So translating that, it means that it's pretty efficient, even though it produces more tokens. Yeah. Okay, that's awesome. That's awesome. Yeah. And I think that's this on this. I mean, we got to test it out. We got to try it, and see what's happening here. But good for Google. You know, you just have to release a bunch of models and eventually one of them is kind of good. Eventually one. I say that. Gemini models have been good for a long time, but they're just not the best. Yeah, exactly. They've had a good run or two, though, and, you know, did some good stuff with Nano Banana. And they've done a lot of good stuff behind the scenes with their work around AlphaFold, other science elements.

44:08Like, they've done a lot with science over at Google. And it's been really impressive. And I think you have to credit Demis Asabas with most of that, with at least leading that charge and seeing the importance of it. Now we're going to get into something fun, okay? I have a couple pieces of AI adjacent hardware that I think we need to talk about.

44:31Corey Noles:Okay. You ready for this? Let's do it. This is the micro duck from Paul and Robotics and Hugging Face. Micro duck. It's nice. Owns Pollen Robotics, and they've put out this$399 robot duck. Yeah. And it is really awesome. It is customizable. There's a lot of stuff you can do around. It's 25 centimeters tall, which would be about 10 inches, so that's a decent size. It's completely open source, so you're able to program it as you see fit. I'm certain there will be a million places you can download pre-programming stuff if you're not a developer that will help you do this. Codex and Cloud Code would probably both help you set one of these up if you wanted and probably make you a killer little quacking duck.

45:24But$3.99 for a real robot. What do you think?

45:28Corey Noles:Oh, I love it. That's great. I still want to get the previous robot that they released, which was the tabletop robot. I'm blanking on the name right now. Richie. Yes, yes, yes. I still want a mini Ricci. Yeah. Yeah. So I don't know. Maybe I should shell out and get the micro duck. I got to play with a Ricci Mini, an agentic Ricci Mini running on Nemotron and Langchain at NVIDIA GTC this year. And it's cute. It's a cool little robot, man. The cool thing about this is you train it yourself with reinforcement learning. So you can run it through various RL gyms. You can do a lot of stuff around that.

46:09But the truth is—

46:09Corey Noles:I've actually seen a lot of people making virtual models and simulators for it in Blender and training it in simulation. So, like, they train it to do things like somersaults and all sorts of fun stuff like that. Just from, you know, generating, you know, a virtual environment and teaching it in there, which is really crazy. So you train it on your computer and then you let it loose in the house. these very much give me star wars episode one the phantom menace vibes around the junkyard like you remember the robots jensen walked out with at gtc oh i was about to say the same thing yeah yeah that era star wars robot like this has those vibes it's got a little flapping bill and there's a reason nvidia just bought a hugging face

47:18Mineral Day

47:23Corey Noles:I love the sound effects.

47:28Take my money.

47:31Corey Noles:That was amazing. Yeah, that is really cool. Yeah, and if you want to order one, let's pop over here. There's a few facts on it. 15 motors, 25 centimeters tall, 800 grams to pick up. I don't know what that translates into. My centimeters are fresh, but my grams are not. Uh, looks like it has a camera plus LIDAR. Uh, seven trained moves right out of the box. 50 Hertz onboard policy loop. Uh, 25 centimeter tall biped robot, playful and educational made by pollen robotics. The Bordo based robotics team at hugging face. Their second consumer bot after Ricci mini. So this is, this is the Ricci people, of course.

48:15Um, walks with the wild. That's 800 grams.

48:19Corey Noles:800 grams is just under two pounds. I just looked it up. Joke. Yeah, so it's light. Walks with the waddle, sits, crouches, roller skates, picks objects up with his beak, and gets back on its feet from many common fall positions. Sets out to make building AI on real hardware as approachable as running a model. Complex behaviors out of the box, then an open source stack covering robot control, simulation, RL training, and sim to real deployment, So new skills can be trained in simulation and transferred to the robot. I mean, for$399, I think that's as cool as it gets. Yeah, that is awesome. And they've sold at least$2.5 million worth of them so far.

49:02Corey Noles:Last tweet I saw. Wow. And pre-orders are open now. Looks like first deliveries before Christmas. North American Europe launch. So if you're watching this from other countries, it might be a while. but four colors, cream, graphite, lavender, and sky. And yeah, sign me up, man. These are pretty sick. What did you think about the Dyson Camera Jet AI toothbrush? Well, what you should know is that I have that up here for us right now because that was my other piece of kooky AI hardware. Yeah, what is this? um well this is an ai powered toothbrush by dyson which means 500 or more because they don't know numbers smaller than that um that when brushing your teeth uses a little camera of sorts and recognizes stuff caught between your teeth and areas that aren't becoming clean and it fires mouthwash through those areas to help get your teeth clean get around your gum line i mean it's fairly cool i mean it's right is is this the ai we were promised like is this the true the true like is this what's going to bring people back from hating on data centers is this toothbrush that can actually clean your teeth that's amazing so some rich guy with a 500 toothbrush can can get extra clean between his gums and his and his tooth gaps i'm trying to imagine though like like you're just brushing your teeth and all of a sudden you just get a little squirt of mouthwash So let's read this here.

50:37Flossing is an awkward and time-consuming chore. Few of us do it even though we know we should. Dyson engineers and scientists have spent well over a decade understanding oral care regimes and how nasties build up in the mouth. Nasties. Our research has shown that people consistently miss the areas where plaque forms, the gaps between the teeth. And we wanted to solve three fundamental problems. How to see those gaps, remove plaque, and improve cleaning performance while brushing. Oh, what you should know, it combines a camera, machine learning, precision fluid dynamics, advanced brushing technologies, connectivity, and Wi-Fi.

51:11The Dyson Camera Jet introduces an entirely new way to keep your mouth open. This is... Do you feel like a Wi-Fi toothbrush is excessive? Yes.

51:27Corey Noles:Why do we need the camera? I love the idea of a toothbrush that can actually see where the real buildup is in your mouth and address it. That is a great idea. I do love the concept. So look at this. We've got two jets coming out here. Small, roundish, oblong, kind of a squished square head. What is that? Yacht radar. Yacht radar for your gums. When are they going to make a toothbrush that actually just zaps your gums with lasers? Right? I mean, like, I'm not saying it's not a cool idea. It's a cool idea. It is a cool idea. It's just RFID. Actually, I will say, speaking of sonic oscillation, I oscillate back and forth between thinking this is cool and absolutely stupid.

52:24Yes. I'm currently oscillating on the same topic. Yeah. Oh, wait, do you want Dyson toothpaste and mouth rinse? I'll bet that's cheap. Let's see.

52:34Corey Noles:Yeah, they're going to sell you so much mouth rinse. Wi-Fi connected at 2.4 gigahertz and Bluetooth connected. That way you can monitor it on your phone too, I guess. It's an integrated camera as a high quality endoscope. Oh. Allowing users to see what their dentist sees through the live viewing feature in the My Dentist app. the Wi-Fi is so that you can see what your toothbrush sees so that you can actually get in there. That's kind of cool. It's a 100 ,000-pixel micro camera, light and camera strobing at 257 hertz, just out of sequence with brush head movement, which is, I guess, important. I don't know.

53:13I'm not too up on my toothbrush technology. Brush head moves 1 ,000 times per second. Got to get a dentist on the show.

53:18Corey Noles:This is great. Hooray. This is a net positive. Over 470 ,000 dental images were collected in the development of the machine learning system. Caveats. Ooh, let's read the caveats. From assessment to publicly available market data. Clinically tested with 71 participants after four weeks of use, twice per day, with Dyson formulations versus a manual toothbrush with Dyson toothpaste. Tested in a laboratory using Dyson's proprietary proxy plaque, developed in a funded research collaboration with the National University of Singapore, Faculty of Dentistry. Deep clean mode versus using a manual toothbrush.

54:00App functionality may vary per market. Requires the internet. If it's the robot or the toothbrush, it's the duck. It's the robot.

54:10Corey Noles:Yeah, I'll let my gum health, you know. Suffer through the old analog toothbrush. Suffer another year. So, you know, as we're thinking about, you know, all of the changes that are happening in the AI industry and, you know, what's happening on device versus over the data center and everything that's happening in the memory crisis and how memory is becoming a huge deal again. Your take is that the CPU, not just the GPU, is going to matter again. For the past few years, we've been thinking about nothing but VRAM. Everything has been about VRAM. It's GPUs. It's how many graphics cards can you pile.

54:50And that still matters. I don't think there's a scenario in the nearish term future where having a killer GPU is not an important part of running your own AI or running agentic stuff on your computer even. However, what's happening is the various tools that we're seeing, like, and I'm going to use this in codex terms because that's kind of my tool of choice. I'm increasingly using computer use. I'm increasingly using tool use throughout various things like Chrome mode. It'll use my browser tools. It'll use computer use and go run Xcode and run iPhone simulations. It'll play a game on your screen.

55:31And not all of that work depends just on your GPU. And what's happening is computer use is getting faster and more capable and more capable and more capable and more used. Like today, I had it out here editing videos and doing things like that, which is a GPU intensive thing. However, when you get into general computer use, you're going to reach a point in the near future. could be tomorrow, could be a year from now, where the bottleneck becomes your computer, not your AI. Like, as this gets faster, and you can see hints of that when you're using it now, like Codex will open a browser and it'll look really still for a minute.

56:12And then it's like, blam, it's filled out a 47 line form in an instant, you know. So as that starts to happen, like, it's not all going to be about the CPU, but I do think that the CPU's relevance is about to have a little upswing because, you know, as it begins doing more and more agentic tasks for you that aren't just an LLM answering questions, we're going to see an increasing amount of need for it.

56:38Corey Noles:Yeah, not just running the model too. Because like, let's say we're trying to run a local model like QN 3.8 on your computer, you need a lot of VRAM, like a lot of GPU power to do that. But once that model is actually running, it then also needs to use your computer. So let alone running models over the cloud using your computer, if you have something local, you're also going to need the CPU. So my question to you is, is this just bullish for Apple? I think it's bullish for a few people. I think it's bullish for Apple. I think it's bullish for even NVIDIA, Intel. I think AMD, who does processors as well as GPUs.

57:17I think that you're going to see more and more of this integrated stuff like what Apple's doing, as well as, and nobody talks about this much, but it was announced in June, and this is the NVIDIA Microsoft RG Spark, RX Spark, that is essentially the Apple Silicon approach. It's all integrated on one chip. So it's like here's a chip that's more powerful than an i9 and running 128 gigs of RAM on one little chip. And I mean, I held it. It's this big. It's nothing. It's like this. It's comically small for what it is. And that combined with what Apple's doing, my gut says that what NVIDIA has done and like the Microsoft models that are going to come with it, as well as from Dell, from Acer, from Lenovo, just like with DGX Spark, where they came out with their own models of that.

58:12But what we're going to see is a flood of machines that are, if I had to guess, probably 40 % cheaper than the same amount of hardware for Apple.

58:22Corey Noles:Yeah, because they've got that Apple premium on there. Exactly. They're not going to be cheap. Like, you know, this isn't going to be, you're going to go to Best Buy, it's going to be$900 computer. You know, you're still going to be looking at several thousand. I think my guess, they still haven't announced prices on the, oh, God, is it the Surface Spark? The new one. I'm going to have to look that up. Surface Ultra? But I expect that to be in the vicinity of$4 ,000. I think that's about right. Meaning it could be$3 ,500. It could be$4 ,999, but I think they want to keep it under$5 ,000. Okay.

59:00Corey Noles:Yeah. And honestly, Apple may play a role in why you would want it under five grand. Yes, that's fair. Because they're going to have something that's like twice as much. Yeah. And do already. You know, there's an awful lot of what you buy from them that when you get into those studios, when you get into what like a graphic designer uses, like a high-end graphic designer. You know, I've got a buddy who's every three years buys like a$10 ,000 or$12 ,000 Mac. Wow. And he puts it to work? It's worth it for him. It's his day job. It's his day job, 40 hours a week, 52 weeks a year. When you amortize it over three to five years, it's really not terrible.

59:41But it sure is the day you pay for it.

59:44Corey Noles:Yeah. Yeah, especially if you want Apple's monitors. You know anyone who actually buys Apple's monitors? I was looking at them the other day. They're just so expensive. They're so expensive. It's absurd. Where I think the CPU comes into this is that the CPU hasn't, it's been a hot minute since anyone cared. Like, you know, you get your latest Apple Silicon or you get, you know, your i7, your i9, whatever the case is that you want. And it's just not a thing people think about that much anymore. Because the truth is years ago we passed a point where they became lightning and really, really fast.

1:00:21And this, using AI locally, and not even locally, not even locally, using cloud AI, but doing agentic work on your direct machine brings a new level of, you know, consumption that it's going to require to do that smoothly, to do that consistently. and I think CPU has its day coming. I mean, you know, and I mean, there's Intel, there's IBM, there's all of the obvious players in that space, TSMC, companies like that that make all the money on all the chips. And I'm not sure how soon we're going to see that, but there was an interview this past week, or there was an interview recently with Thibaut, the Codex head from OpenAI on Matt Berman's channel, And he talked about this a little bit.

1:01:10And it's something we've kind of batted around a little before. This idea that computer is about to get really fast and the bottleneck is about to be the human. It's not going to be the computer. And I think some of that is going to have to come from like, you know, the amount of work you want your computer to do in a reasonable amount of time. You know, you don't want undue strain. You know, you want to have sufficient CPU to handle the load you're giving it and have some playroom. Because, you know, in some cases you might leave and your computer just run and work for hours on end. Yesterday, I went to pick up Chick-fil-A.

1:01:44And while I was there, I had two different agents set up running, working on a pair of mobile games I've been piddling with. And I'm gone for 45 minutes. I get back. I eat dinner. I make a coffee. I take the dog out. I come in. I sit down on the couch. They're still going. Like they haven't quit more than an hour deep. And as that happens, you're less and less going to want to wait on it for things like operating a browser, you know, opening little small tools on your desktop and running basic stuff. And I think I don't think you're going to need a monster processor, but I think you're going to see like...

1:02:27And six, eight-year-old systems become significantly less usable is what I would say.

1:02:34Corey Noles:Yeah, I have a personal anecdote related to this. So I use a lot of work trees when I'm doing different features on my coding projects. And those work trees are all created locally. So not only does it take CPU, it also takes space on your computer. And so that's why a lot of people are saying like, hey, eventually, if you're working seriously with coding agents, At some point, you're going to have to embrace the cloud sandbox option, which for people who don't know, if you're using Codex or Cloud Code, you can either work with files locally on your computer or you can work with them on the cloud in a sandbox environment.

1:03:09Or on a Docker kind of tool even.

1:03:12Corey Noles:Exactly. Yeah. But, you know, when they set it up, it's like, you know, totally handled for you. You just select the cloud option. The only thing is you have to give them, you know, access to your GitHub where all your code is so that it can actually work with your files. Otherwise, you know, it doesn't have access to your files on your computer. But when you work with it on your computer, I mean, I had something like 20 to 35 plus sub-agents working on different tasks, and it was getting hot. And it's a pretty good M2 MacBook Air. It's nothing to write home about. I think it has like 16 gigabytes, 14 gigabytes.

1:03:47Corey Noles:I forget. That's what my M1 has, like 16. Unified. And it was getting hot. And this thing does not get hot easily, but it had a lot of stuff going on. And I do think that a large portion of that had to do with the CPU being pushed to the limit. And it didn't crash on me. At the same time, everything else is probably. Sure. Yeah. But it got so slow to the point where I was like, I have to spin up another instance on a different computer to keep working. I got to let those sub-agents cook over there. And you imagine this would happen if you have a local agent. that you're running. You don't think about it with cloud agents because you think, oh, you're just accessing files on my computer.

1:04:28Corey Noles:You're writing code markdown files on my computer. It's not going to heat up and burn with the blazing glory of the sun. I'll tell you what, with a Mac laptop too, heat dispersion has always been an issue for them just a little bit. Something I'd like to talk about today is about Apple's, Essentially, they're desktop data centers in the new Mac studio. In August, they dropped a studio refresh, a Mac mini refresh, which is super rare because they have this big thing every year in June. They have another big thing in September. And to just randomly drop something like this in August is a bit out of character.

1:05:09But they came out with these monster machines that have developers going wild and literally just throwing their money. I say, I'm not throwing it away. I mean, look at that. 512 gigabytes of unified memory grant.

1:05:22Corey Noles:Yeah, this is wild. So it's rare. Yeah, to your point, it's rare that they do this via press release as opposed to a big keynote example. But they have been doing that recently. They have. And I think they feel some pressure in the AI space that people want this really badly. They want it now. They're willing to pay for it. And I think that shows in the fact that OpenAI has apparently bought tens of thousands of Mac Minis and Mac Studios already, all dedicated machines for reinforcement learning and computer use agent training. How bonkers is that? Yeah, it makes sense, though. It does. Obviously, they have the scale and the billions of dollars for them to make that kind of decision and just purchase as many as they can.

1:06:07Corey Noles:Yeah. But yeah. Yeah. From what I understand, the whole Mac Mini movement caught Apple completely off guard. They did not realize how big of a deal this was going to be. You can kind of thank OpenClaw for that. Yeah. I think OpenClaw is where Mac Winnie's first went bonkers. Yeah, because you just buy a Mac Mini and, you know, it has 60, you know, you can get the 30 gigabyte, 60 gigabyte VRAM one. And you can run, you know, kind of dumb by today's standards, but fully local model directly on that computer, dedicated machine, give it access to its own accounts, and it can be your little dedicated assistant, which is pretty awesome.

1:06:45Corey Noles:And you don't have to pay per token. You know, you obviously have to pay for the electricity and all that. For sure. For anything running on your Mac Mini, and it's a safe environment, so even if you're using cloud models, it doesn't have access to all your personal files on it, so it makes it a little bit safer. It's a great sandboxed way to play with these tools. It really is. Let's scroll down. Let's take a look here. What do we got? So this is the M5 Max. Yep, all new M5 Ultra. It's two different versions. Now featuring up to 4.3x faster AI performance, 2x faster storage. That's interesting.

1:07:28Corey Noles:1.8x faster graphics. And 1.3x faster CPU speed, to your point. Interesting. I think that speed is showing up in every one of those, that all of those are speed claims is important. And I think that that speaks a lot to where agentic work is headed. Yeah, I think you're spot on. As we do things like have it go edit videos, for example. And this part's big. Up to 128 gigabytes of unified memory. Wow. Wow, wow, wow. So that's not necessarily... Yeah, and that's not the top end. Go down to see where it scales to. 512? With the powerful M5 Ultra. Wow. Up to 512 gigs. How much is this? I think you should go look because I think you're going to be shocked and probably a little appalled I know I know it'll be expensive but how much expensive I'm curious comes with black magic design nice that's cool it's a cool rig I'm not saying I wouldn't take one look at them they got four USB jacks that's rare for Mac I'm actually shocked it starts this low to be honest $2 ,400 basically$2 ,500 $2 ,299 for education that makes sense $5 ,500 for normal but that's the M5 Ultra now what you need to do is go to Apple Mac Studio and let's build one real quick and you'll see why that number is really misleading

1:09:12okay so pre-order right now yeah it's pre-order grant can you expand your screen a little you can lease it that's cool yeah all right we want the ultra of course uh 30 core 80 core let's just see how top of the line we can get it here 256 why is it 5 oh 5 5 uh 5 12 doesn't come till late october okay yeah not available we'll go 256

1:09:43Corey Noles:storage let's add 16 terabytes completely unnecessary i will i will run this again without that yeah uh we don't have anything we don't have anything comparable to trade in $18 ,000. A mere$1 ,500 a month. I mean, if you're immediately making money off of it, like with agentic workflows that can make you more than that, that's not bad. Now, let's go back up there and let's change it to, realistically, I wouldn't own it with less than like 4 terabytes. Okay, so let's do 4. That would be my number. So we're at 12 now, which is about what you said. Your friend buys, right? Yeah, I said 10 or 12, yeah.

1:10:27Yeah, they always show, like, look,$5 ,500. And then it's like, yeah, if you want the worst one ever made, which nobody wants. Yeah. I don't know that I'd do it for 90. Look, go up there now. Now change that to 96 gig. And look what the price does here. Next one. Okay. Now it's an$8 ,300 DGX Spark with less RAM.

1:10:55Corey Noles:So you're saying DGX Spark would be better in this instance? I would buy a DGX Spark before I would buy the 96 gigabyte. I mean, that's still a serious CPU. That's still a serious processor. But what if we do the Mac step? Not$3 ,500 more. Let's do the Macs, a 40-core GPU. with, let's do the max here. Okay, this is about 5 ,000. This is not bad. That's not bad. This is the one that started at 2 ,500. Yeah. Oh, yeah, you got to have the ultra if you want that. Storage. Let's do four. Okay, this is 6 ,800. First off, let's just discuss the fact that they want$1 ,000 for a 2-terabyte hard drive.

1:11:57Corey Noles:Yeah. No. That's a lot. This is where they make all their money. Yeah. Yeah. Well, and you know, the thing to remember is the people using Macs heavily, a lot of them are designers. You're graphic design people, you're video people, and they need the storage. Oh, it's true. It's true. So you're at now$6 ,000 for the regular. Yeah. And not anywhere near the Ultra chip. No. It was like$30 ,000 and$80 ,000. You're instead at$18 ,000 and$40 ,000. So it's about a third of the full price of the Ullman. I mean, I'm sure it's great. Yeah. It's better than what I have now. But yeah, yeah, same. but expensive it's so much money it's just so much money how low would this have to be for you to buy it it would have to be the price of a spark and the cost of a spark okay yeah it would have to be the cost of a spark and beat it because here's the deal for 12 grand i can buy three sparks chain them together and have 300 gigabytes of RAM.

1:13:15I mean, excuse me. Yeah, 300 gigabytes of RAM. Right. Why would I or it's actually, excuse me, 376. Let's pull it up.

1:13:26Corey Noles:Let's pull it a DGX smart. And I think they're 128 normally, the DGX smart. Yeah, 128 gig for a DGX smart. So right now they are 4600 with 128 gigabytes. They were$3 ,500 at one point. Yeah, so 128 gigabytes,$4 ,600. So you could do three of these for approximately the same price, right? Yeah. Of$1 ,800. And have almost 400 gigabytes of RAM. Yeah. You know, I'm certain there are reasons the Apple might be the better move in that case, But it's so much. Yeah. It's so much. It's hard to justify the Apple screen. And the truth is, if you're dealing with local AI, an extra 130 gigabytes of RAM over 256, which is the max you can buy right now, and it's going to be way higher when that 512 is on there.

1:14:34we were not looking with the 512 which is the creme de la creme because they didn't have the pricing on it yet

1:14:42Corey Noles:that thing is just going to be sold out I see people taking four of these people take four of these and it comes with you can buy literally link cables for them that are designed to lace four of them together into one monster what's the company XO Labs are they the one that makes the networking for that I don't know. Yeah, they're for Mac specifically, but connect your Macs and workstations into one local inference cluster. Oh, nice. Yeah, very cool. So this guy, he has like four Mac minis stacked on top of each other. This is my favorite channel for this type of content. So if you don't follow him, you should follow him.

1:15:24Corey Noles:All these guys have like four Mac mini clusters that they're all networking together. It's not always the best way to do it. Okay, so we want to cover a couple lightning round topics. before we wrap up. And one of the topics in particular, speaking of local data centers, what about the big non-local data centers that everyone is up in arms about, right? A very interesting, let's call it a very interesting update on this project was that Texas, which is known for being kind of like the come and spend all your money here, free float and regulation state, we'll build anything you want. Data center holy land.

1:15:58Corey Noles:The data center holy land. Well, no, that's Virginia, surprisingly. But Texas is up and coming data center, Holy Land. Well, it turns out that they're actually now on team pause when it comes to data center shenanigans. It turns out Governor Greg Abbott ordered data center projects to stop advancing through ERCOT's grid connection process. So ERCOT is their kind of controversial, you know, power grid system that, you know, Texas is on Texas. Texas is on Texas. Yeah, it's sort of like free rolling. And what's the right way to put it? your rates fluctuate a lot because it's kind of like free market energy over there.

1:16:37Corey Noles:And people have written up and down about the problems with it or the pluses of it, depending on your perspective. But no more data centers will be added to the grid until each request is audited. So this is more like a slowdown than a pause. What's your take on this? First off, it's been a lot of big ones going here. Musk's TerraFab, which is supposedly going to be the biggest construction project in the history of earth earth uh is going to be there and uh what i'm seeing is that they have 474 gigawatts of proposed new load right now that's five times the company's record the state's record peak electricity uh demand that's crazy uh i mean i don't know is is the idea these companies will come in and power themselves using using natural gas using solar.

1:17:28Texas would be a good place for solar.

1:17:29Corey Noles:It's flat. Yeah, solar and batteries. Lots of room. The sun is hot and very direct a lot of the year. They're a leader in that space, especially with pairing it with battery storage because of the Elon effect. Yeah, exactly. Exactly. And I think solar makes a lot of sense there. I don't know what's going on with nuclear, but I can tell you right now they're never connecting 474 gigawatts inside of the next 10 years to that system they have. Like it's had issues for a good number of years. It struggles in like the hardest summer quite often. Less so than it used to. It used to be like on the news constantly, Texas having energy problems when air conditioners were at peak use, you know, especially.

1:18:15So I don't know. You know, these are coming in from, I believe Anthropix building a new one there, OpenAI is building a new one there. Everybody's building data centers in Texas.

1:18:25Corey Noles:I guess this audit is going to review something like 250 to 300 projects. So that shows you the scale of how many people are trying to sign up, let alone - Get your lobbying checkbooks ready, AI companies. Yeah. And I think my hot take on this is that at first it was democratic politicians and democratic states who were pushing for pausing data centers. Bernie Sanders was kind of like the most famous one to come out and say, we need to halt all new data center construction in this country. and now Republican, you know, candidates and politicians are getting in board. And I just think the Republicans, they're floating with the wind on this.

1:19:04Corey Noles:They're like, yeah, sure, let's pause, let's slow down, let's audit it. And then the second the midterms are over, they're going to go right back to full speed ahead. I just don't see that pause being genuine. But, you know, you even see what, like, earlier this month, there was a comment from Trump about, you know, You know, U.S. communities that are opposing this data center development in their states are going to find themselves pretty backward and poor in a new economy. And while I don't like to delve into politics, I will say there are sides of the data center bait that don't get discussed much.

1:19:40There are definitely – I absolutely believe in a community's ability to decide whether we want these here or not. But there's a lot of things they're turning away in addition to a data center, like jobs, consistent income, upgrades to your electrical system and your plumbing as a result of having it there that these companies are promising to handle. Yeah. You know, they're making some pretty lofty promises about what comes to town when they build a data center there now. Now, Microsoft really led the way in this, in my opinion, and everyone else has followed suit pretty well that, you know, like we understand we have to earn your trust and earn the right to be in your community.

1:20:17And they seem willing to do that with cold, hard cash and some pretty strong commitments to local organizations, local infrastructure, all of the things that need to happen. So what I would say is, you know, if you're a strong community with good financial background and you don't need that, so what? But if you're a struggling community who's maybe open to some ideas like that, there's a big opportunity here for you. You know, find it an industrial park. Don't stick it in the middle of your neighborhood, of course. I know that's been a big concern. I'm in St. Charles, Missouri area, and we had one come through last year that was – I think it was already approved and then got taken off.

1:21:04But it was supposedly going to use some wells that our energy company had destroyed for human consumption a number of years ago. And it was just wells sitting there rotting with chemicals that would have been fine for cooling a data center. So, you know, there might be opportunities where you have things like that, where you have problem spaces. What if they, you know, came in and cleaned up your Superfund site, for example, in exchange for being able to put a data center? there.

1:21:33Corey Noles:Yeah, I think it's like, it makes a lot of sense. I think there's a lot of misunderstanding of data centers. And I think there's a lot of let's hurry up and push this through as fast as possible that, you know, is making people really nervous about it. I wanted to share, speaking of Trump's comments, the Vice President J.D. Vance had some interesting comments here. He said that the data centers were an important part of the AI economy. And when people build them, they have to build the power plants along with the data centers. I think probably 99 % of the backlash to data centers has come in areas where building a data center means higher utility and higher electricity for people on the ground.

1:22:13Corey Noles:I think what the companies have to do is take advantage of some of the federal deregulatory efforts that we've undertaken. If you build a data center, you should be putting power back into the grid, not taking it out. And if that is happening, I don't think the data centers are that controversial. And I think this is actually right. And I've said this before on the show, but I think that actually, if you are pro green energy, this is the most bullish reason that anyone has ever gotten behind building net new green energy, like ever. And if you think about pairing solar with batteries to power some of these data centers, that is a net positive.

1:22:54Corey Noles:And even if the AI industry goes broke tomorrow, you'll still have those power plants that can serve people. Now, exactly. And same thing with nuclear. And I just saw today Google. Google just announced a big geothermal project, possibly the biggest geothermal project ever. And that's, you know, renewable energy. So there's a lot of reasons for why it's a positive thing that they're building power plants to power these data centers, even if the data centers end up being, you know, being useless in five years. From the green energy sustainability perspective too, I'd like to mention that earlier this year, I went to a startup open mic night in Silicon Valley one night and like 30 companies come up.

1:23:39And I would say three out of four were young startups focused on how to create more efficient data center cooling, data center energy, data center thermal reduction. All of it was around piping renewable energy into data centers and making them a much more energy efficient system that exists. And I think that was really exciting. It's a whole lot of companies coming up in sustainable energy and sustainability. And I think that's a great thing for the environment, the country, the industry. I think it's good that the industry seems to be recognizing there's a desire to do these in a way that is less taxing, both environmentally in terms of land and in terms of sound and all of these different areas that are things that have upset people along the way.

1:24:32I think what we're doing is we're just watching an industry be built from nowhere. And the fact is it's going to make mistakes and it'll get better as we go. I think we'll continue to see data centers become more of a thing you might want in your area. I mean, like I said, you don't want it in your yard, But if you have an industrial park on the edge of town and they're going to either build their own power or upgrade the whole grid running into your city already, maybe that's not a bad idea. You know, a lot of small towns especially have real struggles with income, you know, as far as like small towns don't have the amount of business they used to.

1:25:09So the dollars coming in and the dollars going out aren't as nicely balanced as you'd like. So I think it's just about it being the right communities. For the right communities, these are probably a smart deal. For the wrong community, there are others.

1:25:22Corey Noles:I think also trying to set up these deals in such a way so that they're taxed appropriately and those taxes go back to the local community would be a net positive because then – A profit share of sorts even. Yeah. I've heard stories of cases where the data center coming into town has revitalized the city. However, I am skeptical of this being a long-term trend. So I think we have to restructure some things to make it work. I'll say one more thing on this, which is the data center watch research group found that grassroots groups have actually blocked or delayed at least 75 data center projects that are worth about$130 billion in the first three months of 2026.

1:26:07Corey Noles:So I think for me, let's not just block it for no reason. Like, let's say exactly what we want in order to allow it. If it brings tax revenue to our community, we want it. If it brings green energy or just energy of any kind that lowers everyone's electricity costs, we want it. if it is sustainably done so it's not draining our resources for no reason, but actually creates net benefits by helping cure cancer and these sort of things, we want it. We just don't want it if it's going to come in, pollute the area, it's going to make all this sound pollution, and it's going to not end up benefiting us.

1:26:45Corey Noles:And the last point I'll say on this is that there needs to be, somebody needs to make the case for why this is a net positive for white-collar workers. Because I think there's a really good case for this being a net positive for blue-collar workers, But if you are a desk job, a white collar worker, and you think about all the data centers being built, that just to me reminds me that like, oh, they want to automate all knowledge work. So somebody needs to make the case. What are all the white collar jobs going to be? Like how are the data centers coming to your community going to benefit white collar workers?

1:27:17Corey Noles:That's what I'd like to know. And I think telling the story to the blue collar workers is important as well because, you know, the sheer volume of construction you're looking at associated with, I mean, even smaller data centers are still massive undertakings. They bring work to carpenters. They bring work to HVAC teams. They bring work to electrical engineers, for God's sakes, electrical engineers, all of the electricians, plumbing, all of these areas. is they bring in a ton of work for, which is good for the companies building the supplies that create the building. It's good for the local workers.

1:27:51And I think looking into that is a really important thing, especially in a time when, like, home construction is a little slow. You know, I mean, there's probably a lot of people who would take the extra work right now. My last note would be pay attention to who's building it. When you see a proposal come up in or near your community, pay attention. Is this some small investor group? Is this one of the big names you hear about on the news pretty regularly? Like, what are they offering? Like, if it's one of the big companies, I would listen because they're making some pretty attractive offers right now.

1:28:29If it's somebody you've never heard of before, doesn't have a track record, I would be more standoffish about it maybe and really want to see them prove what they're bringing to the table.

1:28:40Corey Noles:Yeah, because I think there's a lot of like dumb money, for lack of a better word, trying to get in on the gold rush here. Get in on the data center rush. Yeah. Yeah. And I think like that's where all of this surplus, you know, concern, all the surplus concern is coming from. And I mean, man, we could talk about this for hours because there's a whole other can of worms around this, which is how much debt is being used to finance a lot of these. And I'm still skeptical that we're going to need, you know, not that the demand's not going to be there, but that like the current crop of data centers are going to, you know, we're not going to have better chips in five years when they're actually built that, you know, potentially could be tremendously times more efficient than this.

1:29:21Corey Noles:And perhaps the demand situation flips. I don't know. I'm skeptical of that, but I'm still open to it. Something else I'd like to talk about today is, you remember during the Super Bowl, there were all these ads about ads in ChatGPT. Anthropic put out the commercial that was the guy at the gym, and there were the others. And they were funny, a little tongue-in-cheek, but about how you won't see ads in Claude. And I think that's fair. But we just heard the ChatGPT ads, which launched then, has already hit a billion-dollar revenue runway. That is insane. And a major inroad and brand new still. I would also like to add that as a paying customer, we've had them for, what, nine months now, eight, nine months?

1:30:11I've yet to see a single ad. Like, these are literally fed at free level. um they're not i'm not seeing things pop up in my feed i'm not seeing things pop up in my chat i have seen no ads uh to this date and i think it's just wild to think that in such short time that business was built propped up sold i mean because can you imagine how many people would

1:30:36Corey Noles:be like i would love to advertise in chat gpt yes yes sign me up uh and it ain't cheap no it's expensive my only point with this is that i think is kind of interesting and i'm not sure i want to put this idea out there they could probably raise prices on the$20 pro plan or plus plan whatever it's called they probably could at this point and then like do what the streamers are doing which is every year they raise it by about a dollar or like i don't know it's every half or seven dollars like netflix yeah and then it's like yeah you can you can downgrade to free but you're just gonna get ads yeah yeah and you know they gotta make money uh and they are making money this we're not at a point anymore where these companies are bringing in zero dollars like it was two years ago i wouldn't say like there's a significant amount of billions coming in for both of them now uh are they spending more absolutely they are spending more but but it's not like oh look they made 300 million dollars last year and spent 84 billion it's it's a little less crazy than what it was uh that's true this says the ad business crossed that threshold less than 200 days after launch suggesting unusually rapid advertiser adoption.

1:31:49And I mean, that's just getting started. There's a ton of people that use the free platform. There's tons of impressions they can serve, I'm sure. While my whole deal with ads from the very beginning, as I've always said, has been that it needs to not impact the answer the AI gives you. That's where Google went wrong with search, is ads started becoming the answer. If ads are the answer, it's not going to work. But if ads are adjacent to in some way that is more clear, like the fact is you might scroll through eight sponsored pages in a Google search before you get to something that is actually the best answer to your problem.

1:32:34Like it can be buried most of the first page in some instances.

1:32:40Corey Noles:I've never even been to this page ads.openai.com that's where you can buy them I get pitched by them occasionally I get a little hey you want to buy some ads? that's funny we should do an experiment we should actually try to buy like you know I've got a mobile game yeah I wonder if I wonder if something like that would yeah I'd be willing to throw like $100 at it as a as a video test just to see what happens. They might tell me there's a$5 ,000 minimum buy-in or something too. I don't know. Well, this says you get 500 ad credit when you spend$500, so about$1 ,000 in value. So if you spend$500, you get$1 ,000.

1:33:23That's pretty standard. Most, like when you get with Google AdSense, when you get with X ads, meta ads, when you're new, they'll almost always kind of match you up to a certain amount and regularly come through with like, hey, here's a little credit.

1:33:38Corey Noles:You know, if you're if you're continuously spending, you just kind of think like who is what are people asking JGPT about? That would make sense to advertise. Yeah. Like what's like the most common prompts that people probably ask? And do you have a product or service that's related to that? Well, everyone, thank you so much for joining us today. It's been a lot of fun. Grant, I appreciate it, man. Yeah. And you too, Corey. This was a good one. If you haven't yet, please take just a minute to like and subscribe to the channel. Make sure you check out the neuronacademy.com. go visit theneuron.ai and sign up for our daily newsletter we'd love to have you as somebody there and on that note always like and subscribe to this channel yes like subscribe do the thing it helps us I promise it's important and it means a lot on that note hope you all have a great week farewell for now humans