00:07:45 · Jordi Hays, John Coogan
[00:07:45] S0: AI upstart is gonna exclusively use Google TPUs. That's exciting. That seems like a bombshell. [00:07:53] S1: Maybe. I mean, you gotta get big infrastructure somewhere. And if everyone's optimizing against NVIDIA, maybe there's maybe there's like a slack capacity in TPUs. I mean, you're certainly like, if you are trying to bet on a hyperscaler to hit your horse too and you see that Google's increasing their CapEx pretty significantly to 85,000,000,000, like, you you know that they are going to have, like, enough capacity to surface to to to service you to the tune of one gigawatt or something like that if you wanna do some massive training run. I do wonder how how compute constrained SSI is in the sense. Like like, much is Ilia feeling the scaling wall? Because for a long time, it was scale is all you need. Just go bigger, bigger, bigger. And that's certainly the playbook that we've seen Sam Altman run with with the Stargate project and with the bringing in Masayoshi Son, bringing in Donald Trump and Crusoe, and putting together this, like, team that can actually marshal half a quadrillion dollars. Is that right? It's 500 no. No. Half a trillion. It's it's 500,000,000,000 They they wanna [00:08:59] S0: haven't marshaled it
00:08:59 · Jordi Hays, John Coogan
[00:08:59] S0: haven't marshaled it [00:08:59] S1: yet, John. It's a lot of money. Yeah. They're working on it. It's half a trillion. But but but SSI, at least at least like the the the rumors or the memes around it have been very much have been more focused on let's advance the research. And maybe that next phase of research is not purely compute bound. It's not it's not purely, hey, we're in this scaling race. We're gonna build a bigger transformer. That's that's honestly a little bit of what Elon's been He's saying, like, we have the transformer architecture. We're gonna scale this up a ton. And and where we are great is building a massive cluster, the Colossus, 100,000 [00:09:36] S0: Yeah.
00:09:37 · John Coogan
[00:09:37] S1: GPUs. We'll build it faster, and we'll get it up to speed, and we'll train it. And in three months, when most people would think it would take a year, we haven't been seeing that from SSI. S s s we haven't seen, oh, SSI has tents with GPUs in it. That that's their that's their edge. Yeah. The the the edge with SSI, as it's been pitched at least, has always been foundational research progress, just making making progress on on the underlying research that might be an elegant solution that increases the efficiency, and you get a better model with even less GPUs. But if they need a ton, Google will have them because they're investing more in CapEx more than ever before. While talking while taking a question on agentic capabilities, it sounds like Sundar said this, when we built our series of 2.5 Pro models, it's the direction where we are investing the most. That's definitely exciting progress, including in the models we haven't fully released yet. Long answer on agentic progress. The good news here is that we are making robust progress. We think we are at the frontier here. He said that they he said they have projects running internally, but now they are slow and expensive. They see the potential and are making progress on both. Gemini has a deep research product, which I would call an agentic product. Yeah. The question is, are they going to try and match what ChatGPT launched with agent mode sooner than later? Is there a security or a risk thing? Google's been a little bit more hesitant to just throw stuff out in the wild, but they're moving very quickly. And I think they will probably be able to get something up. There's certainly nothing on the CapEx or or model capability side that would stop them. And Elon Musk just says, wow.
00:11:11 · Jordi Hays, John Coogan
[00:11:11] S0: Well, Elon is playing the enemy of my enemy is my friend. [00:11:16] S1: It's like, I I love these guys. Yeah. [00:11:18] S0: Love to see it. [00:11:19] S1: It's not a it's not a knockout drag out fight. So, let's go over to Oh, this is interesting. So Michael Nathanson of MoffettNathanson, who's been on Stratechery a few times in the Wall Street Journal was quoted, what people worry about is just the math around search. Isn't AI going to be negative to search? And they went out of the way many times on the call to say, that's not what we're seeing. And so there's this big question on can can Google avoid disruption by integrating the new disruptive technology, artificial intelligence in a way that doesn't destroy their core business which is search ads. They're putting up 80,000,000,000 [00:12:01] S0: in revenue a year on it. My personal feeling is that Chrome specifically is just so unbelievably sticky. Yeah. As long as it takes the average person to to, you know, hitting hitting two buttons to get to the point where they can do any type of search, Google will continue to dominate. Because right now I can hit command t Yep. Search something quickly. Yes. I'm sure I could update the default like search Yep. To something else, but I haven't. [00:12:33] S1: There was [00:12:33] S0: a And I love the model. [00:12:34] S1: There was a time when I bounced around. I remember having Opera installed and Firefox installed for a while. And then for the last decade, it's just been
00:12:34 · John Coogan, Jordi Hays
[00:12:34] S1: There was a time when I bounced around. I remember having Opera installed and Firefox installed for a while. And then for the last decade, it's just been [00:12:41] S0: I used to love there was a moment, like, probably five years ago that I was loving Safari. It was just so snappy and fast and simple. [00:12:48] S1: I believe it's also built on Chromium. But, yes. Course. But it's not a Google product, so they so they could in theory reroute reroute No. [00:12:57] S0: Safari is not built [00:12:58] S1: Oh, it's not? [00:12:58] S0: On Chromium. Interesting. Uses Apple's own web kit
00:12:58 · Jordi Hays, John Coogan, Alex Hawkinson
[00:12:58] S0: On Chromium. Interesting. Uses Apple's own web kit [00:13:01] S1: rendering Okay. Do I do like so it's weird. I use I use Safari on my phone and Chrome on my desktop, which feels like it totally breaks the continuity. Right? But I've just I've just kinda stuck with that and it's been fine and I don't have like tabs that float in between. They're kinda separate, but it's fine. Really quickly, this show is brought to you by Restream. One stream, 30 destinations, multistream, and reach your audience wherever they are. We love Restream. You can sign up for free. So let's go over to, Ben Thompson talking about Google earnings. He has said, I've repeatedly laid out the theoretical case for why AI is potentially disruptive to Google, and we asked him about this when he came on the show. My question was, like, is there a world where where revenue and profit from from generative AI products is actually not as counter positioned against Google in the sense that, like, if they had launched the Gemini app first and become the ChatGPT, they'd probably be seeing a drop in search volume, but then also seeing revenue spike from an even more popular Gemini that's monetizing even better. And and was it a question of them just, like, not being able to take that pill, or or was it more about, like, the risk in PR nervousness than [00:14:16] S2: The thing that Ben