Loading summary
Transcript1 lines
- [00:00]
A
Today on the show, Anthropic has taught Claude how to dream so overnight it can improve itself. Red Access, a security research firm, just published a report that thousands of apps built on lovable REPL and base 44 are leaking medical and corporate data right now. And Google has quietly killed Project Mariner just two weeks before Google IO. Anthropic is on an absolute tear with cloud compute deals, doing one with Google Cloud and also with C SpaceX for 300 megawatts, with a possible future deal where they'll be paying SpaceX to put Orbital data centers into space. Let's get into some of the Google stories first. With Google IO just around the corner, they are making a bunch of changes and it feels like they ship things and aren't really making huge announcements, but they have some big, some big impact. So Google Mariner, which is the basically agentic browser, it was kind of like a Claude coworker, right? It was supposed to take control of your browser and click and fill out forms and book travel for you, blah, blah, blah, uh, kind of everything that Claude Cowork is doing, Google has shut it down. There wasn't like a big announcement about it, but all of the tech that was in Project Mariner is getting folded into the Gemini Agent and kind of the whole agent assistant ecosystem. What I think is actually happening here is that Mariner basically didn't make u unit economics sense. The browser agent that can take, you know, it's taking screenshots and it's visually parsing the DOM the they're very slow, they're very expensive and command line agents like Codex and Claude Code have basically, they basically do the exact same thing, but it's a way lower cost. And so I think Google saw that and they kind of made this call. Now what's interesting is that Claude Cowork and presumably what some of the things that Codex does, they do a lot of the same things. So when I use Claude Cowork, you know, sometimes I'll say, hey, go over to, you know, go over to like Audacity, my audio editor, take control of my screen and do this, you know, this chain of fixes on the audio and it will go and literally click through. And in order to do that, it's got to take screenshots, it's got to make adjustments, and it can do some pretty incredible stuff. What I will say is there was a recent report that came out and it said that doing that, which is called computer use, is about 85 times as compute intensive as doing something through just directly through cloud code or you know, asking it questions and it kind of giving you a text response. So this obviously is way more use useful but 85 times more compute intensive. So I'm not shocked that Google kind of pulled this out and stuck it over into Gemini. I also think that they're seeing how a lot of these other agents are doing it and kind of realizing this might be the direction. Now Claude Cowork for example, is actually doing something very clever and even Claude code where most of the time when I ask it for something, it doesn't need to do the strictly computer use screenshot tasks, although they do I guess kind of take some screenshots in the background as they go. A lot of what it does, it tries to figure out the most code way to do it. So if I tell it to go edit a video in capcut, instead of opening up the Cap Cut app, I've had better success with it going to the Cap Cut file. It opens up the file and just edits the code on the file to make the edits that I want. It's kind of a crazy robot way to do things, but it actually works really well and it's way more efficient. So I think Google might move in that direction a little bit. But it's interesting that that got shut down. The other thing that's interesting is that Google just launched a hundred dollar Fitbit. This is the first Fitbit hardware that we've had in four years. It's screenless and it basically pairs with your Gemini powered health app. And they even have a $10 a month AI coach that is kind of built into this. Next up we have Red Access is a security research firm and they have found thousands of vibe coded apps that are leaking customer data. And specifically they're doing things like corporate data and health data. All sorts of stuff you wouldn't want to have leaked. Red Access is an Israeli security forum. They basically scanned 3, 380,000 publicly accessible apps built on Level Replit, Base44 and then Netlify. And I think about 5,000 of those had no authentication. So about 40% of those or about 2,000 apps were exposing sensitive data to anyone who happened to find the URL. So what kind of data? Like I mentioned, there was conversations between doctors and patients at a long term care facility for children. Ironically, they also found a security company's incident response data. They found a school app with lesson recordings and student records. They found a vacation planning guide that detailed some of the customers details. They had a customer chatbot logs that they were able to actually just see all of the conversations the customers were having. They had internal corporate strategy decks, and basically anyone with their URL could go and read all of that. The platforms right now, Replit specifically is kind of pushing back on this report. And he's saying, look, you know, Red Access only gave them 24 hours before going to the press. And like, he's. I guess it's kind of true. So. But he. But he's like, why didn't you guys give us more time to, like, fix it before you went live with it? So that's kind of what he's upset about. I have a base 44 account and I think I actually saw an email from them saying something like all of the. All projects on Base 44 made before a certain date were automatically made private. If you want to go make it public again, you can go do it, but you got to do, like, some security stuff. So I think they are kind of working on it, but this is definitely a issue that they all encountered. And I think there's a huge difference between AI building you a working app and AI building you a secure app that isn't going to leak your data. And so that's kind of the next big phase that a lot of these companies are having to work on. And I think some of. I know specifically with, like, lovable, there is another kind of security incident just like this not too long ago. And it was like every lovable project that was made before, you know, November 2025 had issues, but everything after didn't. And so sometimes it's kind of the older legacy projects. It seems like the best course of action if the users haven't updated the security, because right now the security is getting a little bit better when you're building stuff is they just make them private again. And if you want to make it public, you got to go and fix those security things. So really wild times with the world of vibe coding. Next up, Anthropic is teaching Claude to dream. Now, it's kind of what they're. They're naming it Dreaming. But it's a really cool feat, really cool functionality where Claude Code is going to be able to basically, in its idle downtime overnight or between sessions, dreaming. This new feature called Dreaming is going to kick on and it is a process that the agent basically reviews up to 100 of its previous sessions. It's going to look for patterns that it missed in real time, and then it's going to identify reoccurring mistakes and it will restructure its Long term memory store based off of that. Right. So like if you. I actually have a whole bunch of things where I'll like tell it to go and edit certain files or certain websites or certain code bases a certain way. I kind of read the logs along as it's going sometimes and it's like, oh, I've tried this. Oh, that didn't work. Let me try this. Oh, this worked. And sometimes it's yeah, the same problem over and over again but I'm like, whatever, it'll figure it out. I don't want to sit there and imperfect my prompt. Maybe that's bad, but I'll just shoot off what I want it to get down and it will get it down eventually and I'll move on with my life. This is really awesome. It's going to be number one more efficient. Number two, it's going to use less tokens and number three, faster. And speed is I think one of the things that everyone can get behind as you know, making some massive improvements and it's not training, it's just memory consolidation. So what Anthropic is kind of saying is that this is kind of the AI version of what your brain actually does during slight sleep in their own internal testing. The related outcomes feature which uses a separate grader agent to score task completion, improved task success rate by up to 10 points over their standard prompting loop. So outcomes that they this new feature they have called Outcomes plus, this new Dreaming plus multi agent orchestration are all now in public beta on the cloud platform. I think this is some majorly underrated agent features that are really excited that they've been shipping and just every other company Google and opening AI really have all their work cut out to try to keep up with what Anthropic is putting out there. In addition to this, another story for Anthropic is that the information just broke this and basically anthropic just committed $200 billion to Google Cloud over the next five years. Google is dedicating 5 gigawatts of server capacity and this is exclusively for Anthropic. And the new TPU capacity from the Broadcom partnership that Anthropic just has is going to come online in 2027 and I think between those two, that accounts for more than 40% of Google's disclosed cloud revenue backlog, which is massive. Basically I think the market noticed this. Alphabet's after hours price briefly pushed past Nvidia's market cap on the news. So this is the first time in years that that has actually happened. Alphabet was at about $4.6 trillion. Nvidia is at 4.8. And they, you know, they were able to kind of bump above Anthropic for a little bit. When it comes to Anthropic's relationship with Google, I will say I like to be fair, Alphabet is putting up to $40 billion into anthropic directly. So Google's kind of paying Anthropic to build on Google and then then Anthropic is going to go and pay Google for compute. And it's kind of interesting the way this works because 40 billion goes directly into Anthropic and then anthropic makes a $200 billion commitment back. So it looks like positive ROI. That's also a commitment. It's not actual cash. And it's over five years, so you got to divide 200 by five. So there is some interesting things at play where it basically feels like positive roi. And assuming Anthropic continues to grow, like it all makes sense, unless some sort of bubble pops or something bad happens, then a lot of this stuff gets put into limbo. And of course, this is just more of the kind of circular cash flow that Wall street loves in the, in the whole AI market, which is interesting. All right, if you are already paying for ChatGPT and Claude and Gemini and Grok and maybe like 11 labs for audio or on any of the IM models, I would love for you to check out my startup, which is called AI Box AI. You don't have to pay a a dozen different subscriptions. It is $8.99 a month and you get access to 80 different AI models all in one platform. You get everything from OpenAI, from Claude, from Gemini, from Grok, 11 labs for audio. You get Google VO3 for IM video generation, tons of other incredible models. It's only $9 a month and it's consolidates everything into one place. You don't have to have tons of subscriptions and tons of different platforms to log into. I hope it saves you a ton of time and a ton of money. It is how I am making it so everything is all consolidated into one place. I love Anthropic. It's what I use for basically all my thinking and writing stuff. But it doesn't have image generators, it doesn't have audio, doesn't have video. So I use video with Google, I use image with OpenAI, I use audio with 11 labs and in order to get all of those into one place, I use OpenAI to do all of it. So you can See the link in the description to AI box. AI okay. Anthropic has announced a massive deal with SpaceX. They're going to be locking up the full compute capacity of SpaceX's Colossus 1 data center. It's in Memphis. It this kind of was like famous and controversial for a while because when they first built it they just grabbed like a random slab of concrete and filled it with servers and got a bunch of power generators and portable cooling units and they just stood this thing up super fast. And people complained that the generators had like a lot of fumes and stuff. So I think they got a lot of that crap figured out by now, or at least I hope. And with all of that though, Anthropics is going to get about 300 megawatts and they're going to get about 220,000 GPUs, Nvidia GPUs. So there's H1 hundreds, H2 hundreds, and then they have a really dense deployment of GB2 hundreds. Anthropic's own blog post about all of this said basically because they were able to, and I think Dario was like on stage in San Francisco and kind of announced it as well and kind of said the same thing, whatever it was on Twitter. But basically what they're saying is they are doubling CLAUDE code's five hour rate limits for pro and Max users immediately because of this. And personally I'm thrilled because I keep getting rate limited and I'm tempted to make multiple Claude max subscription accounts. It's a $201, so that doesn't happen to me anymore. But in any case, they're doubling all of that because of this big deal. They're also removing peak hour throttling entirely, which was something that was so brutal. But I mean they had no other option, right? They just had a certain amount of compute. There's nothing they do. So with this they'll be able to remove that. And so this is kind of a massive upgrade for them. I also read a tweet that Elon put out about this. He basically said that he spent a bunch of time with the leadership at Anthropic and some of the people there and talked to them and he feels like he doesn't have any evil vibes from them, so he thinks that Claude will be generally good for humanity. Anyways, it's really funny the, the way some of this stuff works, but with all of that, another really interesting aspect that Bloomberg was writing about is that SpaceX and Anthropic have apparently had some interest from Anthropic side in developing multiple gigawatts of compute capacity in space. And I'm not sure if that's why they made this deal with, you know, SpaceX or if that was just them trying to get more compute and this was the best way for them to do it. I'll also say from a lot of people were saying, oh look, this is kind of like exit the the downfall of xai. They don't have more compute capacity than Anthropic anymore if they're literally selling it to Anthropic. Elon has had a bunch of, you know, recent press and stuff talking about how they are going to be building this massive giga Colossus, you know, data center thing. And I think they already have one called Colossus 2. So in his post about it he said, look, we already have Colossus 2, which is massive. That's what we're training on now. So we don't really need the Colossus one and we're happy to just go give it to somebody else. Which is, it's just fascinating that the companies that are investing in the actual compute are renting it to their competitors. We're seeing this with Microsoft and Google and evidently this is why OpenAI wants to spend so much money building out data centers and compute. Also, I can't help but feel like all of this is Elon Musk trying to get back at Sam Altman. So he's just giving his biggest, you know, Sam's biggest competitor more compute to compete with him as he's going through court, you know, against Sam Altman. One bank of America strategist put out a note that CNBC picked up and basically it was warning that the upcoming anthropic and SpaceX IPOs are going to pull money out of the same basically AI trade that's been holding up the entire S&P 500 from all of the top, you know, Google, Nvidia, all of these top tech firms and all of the AI money that's being made that's lifting up the S&P 500. Anthropic's private market valuation hit $1.2 trillion in early May. If that's, you know, according to CoinDesk, their IPO is now reported going to be coming in June. So this isn't just a compute deal, it's also kind of a pre IPO thing that's happening right now. So if you're a builder, one of the takeaways I would get from all of this is that Claude Code's rate limit is now double. That's amazing. And also, Opus Inference is going to feel a lot faster this week. If you're an investor, watch the anthropic IPO filings. The bank of America kind of, you know, drain on the market warning is the first time that a major bank has framed the AI IPO wave as a liquidity risk to the broader S and P. So that's very interesting. But that's it for the show today. Thank you so much for tuning in. If you enjoyed this episode, make sure to go leave a rating or review wherever you get your podcasts. It helps the show out in the algorithm a ton to be found by more amazing people like yourself. So over on Spotify it's the about tab. You can leave a comment over on Apple I read them all. I appreciate them all. And as always, make sure to go check out AI box AI if you want to get access to over 80 different AI models in one place for 8.99amonth, there's a link in the description and I'll catch you guys all in the next episode.