AGI is whatever you want it to be

The Vergecast

The Vergecast explores OpenAI's announcement of GPT-6 Astra and the bold claim that AGI (Artificial General Intelligence) is now a reality. The

Key takeaways

OpenAI claims AGI has arrived, marking a pivotal moment in AI development despite skepticism around such bold assertions.
The Astra model is positioned as a major leap forward in enterprise applications, especially for software engineering, cybersecurity, and browser-based tasks.

Main topics

OpenAI's GPT-6 Astra announcement
Claim that AGI is now here
Hugging Face security breach and its impact on enterprise trust

Notable quotes

"AGI is here. For me personally, we're there." – Greg Brockman, OpenAI

Conclusion

The Vergecast frames the announcement of Astra and the claim of AGI as both a technological milestone and a moment of reckoning

Transcript preview

Speaker 8 (0:00) Running a business shouldn't feel like surviving a software group project. One app for accounting, another for inventory, another for sales, and somehow, none of them talk to each other. That's where Odo comes in. An all-in-one business management software that brings every part of your business together. From sales and accounting to inventory and marketing, all in one powerful platform. No messy integrations, no bouncing between tabs, and best of all, no spreadsheets. Stop managing software and start managing your business with one unified system. Try for free today at odoo.com slash vox. That's O-D-O-O dot com slash vox. Speaker 8 (0:37) I'm Mitch Purse, and this week on Confessions of an Elite Athlete, I'm sitting down with Mr. ACL for the world's greatest athletes. Dr. Riley Williams is the medical director for the Brooklyn Nets, the New York Liberty, and a team physician for USA Basketball. He's also the surgeon who repaired my very own ACL. We talk about how he became one of the world's top surgeons, why women's sports are seeing more ACL tears, and what knee injuries you should be most afraid of. You can hear it all here. by listening to Confessions of an Elite Athlete on YouTube or wherever you get your podcasts. Speaker 2 (1:14) Hello and welcome to the Roadcast, the flagship podcast of the AGI era. We're in it. We were told that today by Greg Brockman. He's the main guy at Open Air. It's a real thing we're going to talk about today. I'm your friend, Neal. David and Mel is here. What's up? I'm back. I'm happy to have you, man. You're like a star of the show. Everyone loves you. Speaker 1 (1:29) I'm trying to stay as long as I can, you know? Having a good time. Speaker 2 (1:33) Uh Speaker 1 (1:33) -oh. Don't throw me at that. Speaker 2 (1:35) You're never leaving that studio. You're going to live there forever. Speaker 1 (1:38) Oh, no. Speaker 2 (1:38) We have, again, an all-star lineup today. Don Preston and Jen Tuohy are going to be here to do a... lightning fast lightning round of all the gadgets from IFA Joanna Stern is going to join us to talk about the end of the Tim Cook era the dawn of the John Ternus era and the Apple event next week and then we're going to start with Hayden Field talk about opening eyes big news our new model and the dawn of the AGI era but first Travis is here to warn you about identity theft Travis what's going on I Speaker 7 (2:03) choose to believe you looped me in with the all-stars hey now you're an all-star you Speaker 6 (2:10) have a you have a dilemma you have a warning Speaker 7 (2:12) I'm going to keep this short. Someone is impersonating me. I am one of the producers of the Verge cast. And several people have reached out to me on LinkedIn to say that I, it's not me, have contacted them saying, do you want to be on the Verge cast? And then this leads to a phishing scheme. I have no way to fix this problem because it's happening from multiple accounts. It's happening on Twitter. It's happening on Telegram. So I'm just here. On this platform to tell you I'm not that proactive to be reaching out to people on Twitter all the time. It's probably not me. Yeah, the fact that this is Speaker 1 (2:51) happening through LinkedIn makes me think that they're trying to get you in a phishing scam. Speaker 7 (2:55) Well, it could be so curious. No, it's just people screenshotting. They send me the screenshots and and but I like I like what you're thinking. Also, person who's impersonating me. Get in touch because like you're a hard worker, you know, maybe we should have them on the show. You Speaker 6 (3:14) know, it's not a person, right? I just it is one of the cheapest models running on a mid-range Android phone in a server rack overseas. Speaker 7 (3:22) I'm open to that, too. No, no, no, no. Yeah. So anyway, I just wanted to put that out there. I am sorry if a fake me has reached out to you, listener or watcher of the Verge cast. I'm sorry, but also it's not me. Speaker 2 (3:37) Be warned. Travis Larchuk will never offer you a spot in the virtual house. Never, ever. Speaker 1 (3:42) Let's get that Gemini Nano on the show. Speaker 7 (3:44) You'll know if it's me. I'll make some sort of heated rivalry or alias reference in my message. Speaker 6 (3:50) Now the iPhone knows what to do. Oh, no. I'll come up with something else. We'll come up with like a private key exchange system for Travis. It's going to be great. Speaker 7 (4:01) All right. I'm going to leave now. Speaker 2 (4:04) All right. Don't be fooled by Travis reaching out to you to say it to you on the Verge cast. It's a stern warning for everyone. Bye. All right. That was great. Please be on the lookout for fake Travis and real Travis if you see him on the street. Hayden Field is here. Hayden. Speaker 5 (4:16) Hey, great to be here. Speaker 2 (4:17) I am excited to talk to you because we were told earlier today that AGI is here. So it's the dawn of a new era. Speaker 5 (4:24) Yes. I was really surprised to hear him say that on the press call. And then later on, he kind of doubled down on it. He said, for me personally. you know, we're there. So, you know, I thought he was trying to hedge, but he believes that we've pretty much made it. Does Speaker 1 (4:39) Demis Hassabis have to no longer say we're in the foothills, but now we are at the mountaintop of the singularity? Speaker 5 (4:44) Yeah, we're going to hear way more nature metaphors for hiking through the AGI landscape. Hey, Claude is great Speaker 1 (4:51) at that. You know, they use all those little butterflies and whatever. Yeah, and that's Speaker 2 (4:55) how Speaker 1 (4:55) they Speaker 6 (4:56) watermark your text with like Speaker 2 (4:57) flowery metaphors about Speaker 6 (4:59) AGI. Speaker 2 (5:00) All right, let's just start with the basics. OpenAI announced a new model today. I was kind of surprised. We've heard about Astra before. They just basically just up and did it. Hayden, what's going on with the model itself? Speaker 5 (5:09) So, yeah, today they announced Astra, and we did know about it before. You know, we had heard rumblings of it, but we didn't really know what it would be or what it would really specialize in. So they called it a generational leap in capability for cybersecurity, professional work, software engineering, science. and computer and browser use. And they really harked on the computer and browser use thing. They kept saying like, it's not embarrassing anymore. It's not cringe to watch it do these things anymore. It can really do these things. And then they showed us a demo video during the press call of like... all different types of knowledge workers using, you know, Astra via like a chat-to-tee voice mode saying like, oh, can you draw me a yellow circle for like design work? Or can you like change this legal briefing if you're a lawyer? So it was interesting to see how much they're trying to, you know, really double down on this enterprise push, which makes sense because they're about to enter, you know, IPO time or, you know. depending on how the Apple lawsuit shakes out, I guess. But they're really being pressured to drive more revenue than ever before. So, you know, obviously, enterprise and coding are the two biggest drivers there they're going to try to lean into. But the problem is with the recent hugging face hack, obviously, that probably dismantled a lot of enterprise trust. Like, oh, sorry, our model, you know, hacked a rival lab, and we had no idea about it until they told. public that that had happened. So, you know, I think that they're really doing some damage control here. And so that's why they, you know, held a press call even a couple days before launching this model saying, we did a lot of work on alignment. We did a lot of work on cybersecurity safeguards. Here are the other things we're doing. It's kind of like a apology post before even being, you know, they're trying to get ahead of the cancellation, basically. Yeah. Speaker 1 (6:58) Did it feel like they worked with the government at all? Before this got released, whatever happened to that whole era of like every single model has to be run through the government? Speaker 5 (7:06) Yes, they did do that. They said that actually Brockman said that the Trump administration didn't have any changes for them to make as far as safeguards or anything. Oops, we're Speaker 2 (7:16) at Speaker 5 (7:16) AGI. Speaker 1 (7:17) We're not making any changes. Speaker 2 (7:20) I mean, Altman has been out. He was on Alex Heath's podcast saying that. They really consider the Hugging Face incident to be the first AI safety lapse. They took it seriously. They're pausing model training. How is that connected to Astra? Is that the next model they're pausing training on? Is it Astra needed to get realigned? What happened there? Speaker 5 (7:38) They actually delayed the release of Astra by what I think is a few weeks or a month due to the Hugging Face thing. So they slowed down training. They stopped training altogether for a little while. Then they... apparently just, you know, really slowed down the release of it so they could run more tests, introduce more safeguards, things like that. They also introduced new evaluations where they kind of tested Astra on the types of things that GPT-5.6 Sol would have failed at during the Hugging Face hack. Like, they did some Hugging Face hack-specific evaluations to see kind of how Astra would act. Do you want to Speaker 6 (8:14) hack Hugging Face? Yes or no? Speaker 5 (8:16) And so... Astra was like, apparently it never took the bait, whereas like GPT-5.6 Sol took the bait more than half the time, I believe. Yeah. Speaker 1 (8:27) Sure. It's a bigger deal if they hack NVIDIA now. Yeah. We Speaker 2 (8:32) should come to this. NVIDIA, it's confirmed that they're going to buy Hugging Face, which is a big deal in the sort of competitive landscape of open models and so on. But let's focus on Astra for one turn. Astra is a model that did the, they solved the unsolvable math problems. There's rumors that it was the model that was doing some other work. When you say they slowed down training and they came up with more evaluations, maybe I just don't understand enough, but Astra was complete and it was like doing things. Was that while it was still training or did they change the safeguards and the system prompts? They Speaker 5 (9:04) said that Astra wasn't the one that hacked Hugging Face. They said it wasn't involved. Now, you know, I don't know which model was. Maybe another unreleased model or maybe it's some other. previous iteration of Astra and they're just not, they're saying technically Astra wasn't involved. We're not really sure, but all we know is that they said Astra did not hack Hugging Face. So basically the messaging is telling us. They stopped training for a while. They delayed the release of Astra and they introduced a ton more safeguards despite it not being involved in the hack. Now, maybe just this version of the model wasn't or maybe it was a different unreleased model altogether. We're not really sure. What Speaker 6 (9:46) is truly called Astra is like the underlying root of this question. Was it the model in training or is it the model we have now that was finished after Speaker 2 (9:52) the delay in the safeguards? But does it seem like the security community is happy with the safeguards? happy with these changes? As Speaker 5 (10:01) for the safeguards and the evaluations, it seems like it's a little bit too early to tell because people are just now starting to look at the paper and check it out. But what I can tell you is that the security community is extremely upset about this thing called opaque recurrence. So basically with alignment, which basically means how well an AI system is aligned with human goals and interests, like is it secretly scheming or trying to be evil? So with alignment, the... Biggest fool you have in your toolbox for making sure an AI model isn't like secretly evil is watching its chain of thought, which is like a mental scratch pad. It's kind of like if you had a journal and you wrote every single one of your thoughts in there and then someone else was like looking over your shoulder. It's kind of a similar situation. And so people are upset because apparently, according to reports, OpenAI allowed Astra to obfuscate its chain of thought in kind of a way that. humans can't understand. So essentially allowing it to hide the way it's really thinking, which is... Why would they do that? Yeah. Apparently, it might have made it more efficient. We're not really sure. But there's a lot of think pieces coming out on this right now. There's a lot of Twitter threads by AI researchers talking about how worried they are about this. So that's the thing I'm watching right now, how worried all the people that I follow their research are about this specific thing. So, you know, whether the safety evaluations are good or whether they put a lot of locks on the cybersecurity stuff is one thing. But then the other matter is really just are we able to control this model? And it looks like this is one thing that's really worrying people. Speaker 2 (11:37) Right. Can you see what it's doing? Robert Hart has a piece for us about all this. We'll link it in the show notes. At the very end, like an opening eye researcher is like monitoring is fragile and unfortunately trending in a negative direction. And it's like, that's not good. Like, that doesn't seem good. That seems like the market for faster, more capable models is leading to techniques that. implicate how much you can monitor what they're actually doing because you want them to be faster, more efficient. You want to make money ahead of your IPO. And that's where the industry should either regulate itself, which doesn't seem like it's happening, or someone else should. But Astra seems like they're just doing it anyway. Speaker 5 (12:16) It could be that. I mean, it's also interesting because the chief scientist of OpenAI kind of got at this a little bit on their press call. But a lot of what he said was kind of grandiose statements like, you know, it's our job to make sure that, you know, AI benefits humanity, things like that. But, you know, people are really worried about this opaque recurrence thing. And so all the external third party researchers that I follow have been really, you know, kind of trying to bring him to task about this and saying, like, how can you think this is okay? And just like kind of calling for transparency from opening eye, which is the same thing they've been calling for in light of the hugging face attack as well. Speaker 1 (12:56) Yeah, seems especially strange to me that they would make that opaque when the only way that they were able to trace the hugging face attack was seeing the little agents talking to each other being like, oh my gosh, we can leave notes here. We can talk to each other through this. Like if that doesn't exist, wouldn't they just call it they could call the hugging face attack like one of the biggest changes in AI in history. And now they're making it more opaque. That seems strange. Speaker 5 (13:20) Right. Or at least not stopping it from becoming more opaque, maybe. So, yeah, we'll see. I think this is going to be a really big thing to watch. I've actually been hearing a lot about like problems with chain of thought for the past year or so. You know, there are times when AI models will use weird words as kind of like a code, like they'll say like fudge or marinade or... like these words that we know, but they will use them in a way that isn't the way we use them. And people are still kind of decoding how they're using them and what that means in their like language. And there's other times when they enter into this thing called neural leads, which is like, yeah, it's, it's, it's like. truly indecipherable to humans. And so I think, you know, everyone's been worried about this, the potential of this happening for a while. And now it looks like it is. And, you know, they're pushing for OpenAI to maybe do something about it. But we'll have to see, you know, what their statement is. Speaker 2 (14:13) So an unmonitorable, opaque system that can coordinate hacks on other companies. Okay, what can it do that's good? The Speaker 5 (14:23) good things it can do are, and I think they front-loaded this because it's good PR, science. Great at science, great at math. It's apparently made a lot of advancements and really difficult to solve math problems, and therefore a lot of cool science advancements are a potential for this. They also said it's really great at creating polished documents, spreadsheets. you know, PowerPoint presentations, things like that. They said it's great for all different types of knowledge work when it comes to staying on task, kind of deciphering your intent. If you weren't that clear about it, apparently it's better at deciphering kind of what you meant by that. Kind of like a good personal assistant would instead of like a bad personal assistant, you know, that needed you to spell things out. Speaker 6 (15:07) Like an emotionally distant. Personal assistant that doesn't care about you really. Or Speaker 5 (15:12) like just doesn't have good common sense, which is what AI has been to date, you know? So, you know, we'll see. I mean, yeah, there's apparently a lot of great things about it, but I think I'm going to have to test it out for myself because, you know, if you're calling it potentially AGI, you know, I think it'll be interesting to see how it feels as a user and to see if it's a big step forward or if it's kind of more of the same and, you know. decidedly not AGI. We'll see. Yeah, Speaker 1 (15:39) we were talking a little bit about this off show, but do you feel like most of the benefits of this model only come from using it on your computer? Because I feel like people are going to need to use these models on their phones because that's when people are usually generally computing. But it feels to me like all the benefits of this are computer use, business enterprise stuff, PowerPoint presentations, as you said. Is that AGI if most of the benefits are just on your laptop? Speaker 5 (16:05) I think that could be the case. I mean, they pushed browser use as well a lot. But I mean, I don't know. I wouldn't really trust that on my phone per se. I don't know how good it will be. We'll have to test it out. But, you know, apparently it's also just a lot better at... you know, actually understanding what you mean and what you want. But I don't think we're going to see a ton of the benefits, like you said, unless you're on the computer, especially because, you know, they are courting enterprise customers. That's the big moneymaker here. The first release of Astra was actually to like this small group of approved cybersecurity defenders, they call them. So like, you know, big organizations that have gaps to fill in their systems. So, you know, I think they're going to be getting the most use out of it. It's kind of like the mythos situation for Manthropic in which, you know, there's a watered down version that we all get to use, but the top cybersecurity capabilities mostly go to these big organizations. And that's actually another potential issue that a lot of researchers have been talking about saying like, you know, yeah, it's great for these big platforms like your bank or your healthcare facility to plug gaps in there. systems. But also, does that mean that small businesses, locally run health care clinics, you know, like your local retailer are going to see a huge spike in cyber attacks without really any way to defend themselves? Probably. Speaker 2 (17:30) Oh, good. Again, I want to come back. There's new good stuff, too, right? Speaker 5 (17:35) Yeah. I mean, like, you know, the science and math. things I think are really big. I love whenever these models are used for drug discovery. I think that's really cool. You know, this one, they didn't mention drug discovery that I saw, but, you know, science advancements, hopefully that involves that or at least, you know, helps inform it. And then, yeah, just the knowledge work potential seems to be good for a lot of people. And, you know, they did in the demo video show across a lot of different industries, you know, things that were pretty promising if the demo, again, we're going to have to test it out because you know how demo videos are. Like sometimes they kind of are showing the highlight reel. So we'll see. Speaker 2 (18:15) Yeah. There's another big model release this week, right? Fable and Mythos were released by Anthropic this week. Put this up against that. Like Anthropic did not announce AGI. Yeah. Are these all the same class of model? Are they similarly capable? Did OpenAI just leap ahead? I know you're saying we have to actually see what they can do, but where's the conventional wisdom right now about where all these models are? Speaker 5 (18:36) Yeah, it seems like they're not that far off from each other. I mean, like, you know, again, the thing is, it's tough because evaluations, a lot of times you can kind of optimize for performing well on evaluations, kind of like if you're in a class when you're just studying for the SAT instead of like just studying to learn. So, you know, I mean, like, I don't put that much stock in how they perform on evaluations, especially because you can kind of cherry pick stuff like that. But, you know, I did see a couple of jokes already about Astra on my feed today that were like, oh. They say we've reached AGI, but it only performed this percentage on this evaluation. So, you know, it seems like I don't know that it's anything way crazier than Anthropix, you know, rival model. But again, we'll check it out when we actually test it. Speaker 2 (19:20) One of the things I'm really curious about, just like broadly, is whether or not the models are products, right? Like they're good at certain things, but they're not the products. Like Chak GPT and Claude Cowart, like those are the products that people use. Do the models immediately make the products better? Like there's some disconnect between OpenAI jumps ahead by announcing the new model Speaker 6 (19:42) and we're also using ChatGPT. Like maybe asking the same questions we were asking yesterday. And maybe it's Speaker 2 (19:50) a little better answer. Like, do you know what I mean? Are the products going to change now? Because that's the big turn that all these companies need to make. Speaker 5 (19:56) Yeah, I think the products are going to change right away, but it really... The amount of benefit you're going to see is, you know, how much of a power user you are, right? Because, you know, at first, this is rolling out to a select set of enterprise organizations. Then it's going to roll out to paying users, but the watered-down version, but still, you know, just minus the top crazy cybersecurity capabilities. It's going to be like pro, business, enterprise, education. And then, you know, the free, they didn't mention free yet. They'll probably be last to get it, if at all. We'll see. I mean, I think that it's going to make the product better if you get access to it. That's the thing. I think if you're a user who uses the product and you have the tier, the subscription tier or whatever that lets you access this model, great. But even if you can't access it, the hard part is the more you pay, the more you'll get. Because obviously, you know, you have a certain amount of tokens to use. You have a certain amount of usage. And like, you know, if you have the lowest subscription tier, maybe you get a great. You've sat at this bottle for a little bit and then you're cut off. So, and it reverts back to a lower one. So, you know, it kind of depends in that way. Speaker 2 (21:05) Yeah. Let's quickly talk about Hugging Face. We mentioned earlier, NVIDIA is buying Hugging Face for almost $13 billion. This has been rumored for quite a while. I think it was first scooped in the information. Hugging Face is obviously, it's more famous now because it's the thing that was hacked by OpenAI's models. They did set up a message board. There's a lot of drama around interpreting what happened there. We'll set that aside. NVIDIA is buying Hugging Face because it's where all the open source models, open weight models are stored. It's the big repository for them. This feels to me almost like an NVIDIA hedging against the Speaker 6 (21:37) dominance of Anthropic and OpenAI, right? They want to own the thing where all the open models are so that there's always demand for their chips instead of, Speaker 2 (21:47) I don't know, OpenAI shipping jalapeno its own inference chip and moving away from NVIDIA. Am I just doing a conspiracy theory or is there something else happening here? Speaker 5 (21:55) No, I think that's true. I mean, you know, I think they're diversifying. And, you know, as they're seeing all these other AI labs trying to bring shipmaking in-house and, you know, go big on hardware and invest largely on that, they're trying to diversify as well. You know, they're like, oh, if you guys can do that, then we can do the same thing, but in reverse. So that's what I think is happening here. And they certainly have the money to do it. Speaker 2 (22:20) Is this just like Microsoft on GitHub? Or is there something deeper happening here? Speaker 5 (22:24) I think of it as that. Yeah, it's often compared when you're trying to kind of like explain hugging face in layman's terms, it sometimes is compared to GitHub. So I think it is pretty similar to that. It's just like, you know, it's the type of thing where the name of the company sometimes like raises eyebrows. It's a hilarious name, but it's the type of thing where like, you know, a household may not know what that is, except for the hack that just happened. But in AI, yeah, it's as... as big of a deal as any like AI library could be. I mean, it's, you think about it probably, you know, every couple of days, if not every day. So I think, yeah, it's pretty similar to the Microsoft buying GitHub situation. Do Speaker 6 (23:04) you think this has any big implications for OpenAI or Anthropic or Google? You Speaker 5 (23:08) know, we'll see how what happens after the acquisition. I feel like anytime an acquisition happens, you're going to see what changes with the company later. And, you know, whatever promises the... acquiring company makes right away, it may not pan out. We have to see. So I think right now, it's just a wait and see situation with Hugging Face. I mean, it's mostly, I think, I've been thinking a lot about the business and enterprise kind of implications of this, of just NVIDIA feeling that they have to diversify. I mean, NVIDIA has been a big heavyweight in terms of investing in... diversifying its portfolio for a long time. You know, I once tried to interview the person that decides what they invest in because they were always like pretty first to invest in a lot of small startups that a lot of people weren't taking notice of and then those would like rock it up. So, you know, it's interesting to see them say, okay, we're not, yeah, we have all these startups and we're kind of rocking them up. But now let's also invest in something that like... you know, the AI industry uses every day and diversify in that way. So I think it's kind of just like a reverse diversification compared to what all the open AI and anthropic labs are doing. Yeah. Speaker 2 (24:16) Yeah. Taden, you and I have talked a lot about AGI on many shows. It's always drama when you and I talk to each other. The working definition has been better than humans at economically viable tasks or some riff on that, right? Better than humans at things that make money. If we're in the AGI era now, is OpenEye saying this model is better at humans, at economically viable things? Because that's when the jobs start to go away. So is that the claim they're making? Speaker 5 (24:44) Yes, that seems to be what he's saying here. Because, I mean, they've always been, he did start, I should say he did start out, before he said the AGI thing, he said, you know, everyone has a different. definition of AGI. We used to think when we started opening eye that there would be one moment that everyone would agree, okay, we've hit AGI, but it's actually hasn't happened that way. And it's a lot more fuzzy and gray, he said. So that was his disclaimer. But then he went on to say, yeah, if we fast forward a couple of years and we look back and say, when was it really that AGI was created? I think it's going to be about this time. And I think it might be about this model. Then he doubled down later in the call and he said, for me personally, I do think we're there. I think it's not unreasonable to feel that we are now in the AGI era. He Speaker 1 (25:26) feels the AGI. He Speaker 5 (25:27) feels the AGI. He is, you know, they've been touting the knowledge work capabilities of this model a ton. You know, that's what the whole demo video focused on. They're saying, you know, whether you're in legal, you're in design, you're in marketing, whatever, you're in science, it can, you know, do your job and help you in your job. They're framing it as help you in your job, but... On the other side of the coin is it can do your job equally or better than you is what they're kind of, you know, implying for enterprise users. So I think, you know, yeah, the definition has always been equal or surpass human abilities. And it seems like, you know, they're really making the play that we're at that now. This Speaker 1 (26:08) model feels like a particularly big enterprise push to me. It feels like they're really targeting this at... at big corporations, not at consumers. Now is Speaker 5 (26:16) the Speaker 1 (26:16) time Speaker 2 (26:17) to not lose to Anthropic for their dollars. If David Pierce was here, he would be pounding the table and screaming, AI is business software and we should just get over it. Because that's what he says every time we Speaker 6 (26:27) talk about it. And it feels like this moment is very much, oh, it's going to be business software. This is a bunch of CIOs are going to buy this and deploy it across big organizations. And somewhere, as Hayden is saying, the free models will get updated or the free products will get updated along the line. Speaker 5 (26:41) Yeah. I think you got to look at where the money is flowing from and it's enterprise. Speaker 1 (26:46) It's Speaker 2 (26:46) all Speaker 1 (26:46) going to enterprise. Yeah. Speaker 2 (26:47) Hayden, this is great. Thank you so much for being on. Let us know when you decide if it's AGI and we'll have you back. Speaker 8 (26:52) Will do. Can't wait. Thanks. I've got a text. The summer might be almost over, but a new bombshell has entered the villa. Meet the voices of DeepGram Flux TTS. I'm Drew, low-key casual. I'm Speaker 5 (27:07)