Anthropic's Fable 5 is Back! SpaceX Builds AI Device

Speaker A: Anthropic's Claude Fable 5 is back along with Mythos 5. The US has lifted export controls on it and SpaceX is showing investors a handset like AI device prototype. Meta is planning a cloud business where they're essentially going to resell AI compute just like SpaceX. Copying their playbook, a company called Bioshocking has created a jailbreak attack on six different AI browsers, everything from OpenAI to Anthropic to Gemini. And it has convinced them that two plus two equals five. There's some reasons why this is problematic. I'll get into them. Cloudflare is also going to block mixed use AI crawlers, um, by default. And that's going to start in September. This year, Google is bringing Gemini Spark Agent to Mac. They're targeting essentially Claude Desktop and Copilot, and there's a lot of beef that they're having with Anthropic over this. The majority of you guys leave amazing reviews on the show, but every once in a while I get a one star. And I'm going to be honest, these are some of my favorite reviews to read. I want to read you one of my favorites I just got a month ago. And it says if you want a rehash of quote unquote, real news with naive commentary by a tween, this will be perfect for you. That's From I hate nicknames. 99. Much appreciated. And this is kind of funny because I actually get, uh, if you search for my name, Jaden Schaefer, one of the top, like Google shows you the other things people search for. And one of the top things is Jaden Schaefer age that people search for. Guys, uh, this one says I'm between. So just so you know, I am 30 years old with four children. I have a young face. But anyways, that was one of my favorite reviews. If you like the show, you don't think that it is a naive commentary rehash of fake news. Um, I would appreciate a review. We got to offset this, uh, this one star. But, uh, in all seriousness, I appreciate all the incredible reviews guys left. And if you haven't left one already, I would appreciate it. I might even read it on the show. Let's get into our first story with Fable 5. The U.S. commerce Department is lifting export restrictions on Fable 5. And I think they actually announced this yesterday because of that. Dave, uh, rolled it out today. I actually have it on my computer and while we've been talking and for the, for the last couple hours, I actually had to go and scan through the entire code base of one of my bigger projects. I have a really big project that is basically, I just use it internally for myself to help with all the marketing and everything I do with podcasting, as well as everything that I do for SEO. I have a new site which kind of goes along with this website, aichat daily.com and it is, it, uh, uses AI and a bunch of other different sources to generate news articles about everything happening in AI and do deep dives on it. Now, um, it's a really beefy project. There's a lot going on there and I'm thinking of, uh, turning it into a product anyway. So I had Fable 5 run through the whole thing and if you have any sort of project, uh, that you've been working on, whether that's, you know, coding or whether that's just a big project that has a lot of moving parts, I would recommend using Fable 5 on it. Today, up until July 7th, we have, um, the ability to use Fable 5 for about 50% of all of your tokens. And after that I think it's going to get cut down. If you want to use FiBL5 for more than 50% of your, you know, Claude max tokens or whatever, um, you have to pay, you know, the API costs. So you're, you're not going to be able to like totally max this thing out. So I would use it, you have one week from today to get as much usage out of it. And I would a hundred percent do that. It's really smart at coding and understanding complex stuff. So run it through all of your systems, run it through all of your code bases. Today is the day to get it. That's just a psa. Anthropic is actually blocked. So in order to get this thing put back online, because it was basically blocked over concerns that it was going to help foreign adversaries attack US infrastructure, power grids, banks, that kind of stuff. And so, um, this all came from, ah, Amazon, who actually did a whole research project on a jailbreak that they found and they showed this to the treasurer, to the treasury, um, and then, you know, it got blocked and there was all this snafu about it in all of the, in all of the drama that happened. Anthropica originally said, hey, look, we don't need to fix this, um, issue. I think Amazon went to them first, then they didn't do anything, then they went to the treasury and Anthropic said, hey, look, we don't need to fix this because OpenAI has, you know, the same jailbreak as possible on OpenAI and Gemini. And so I'm not exactly sure what happened there, but Anthropic got Fable 5 pulled and they've now fixed it. Apparently with that jailbreak that Amazon discovered. They've now fixed it. So over 99% of cases, um, it actually is going to, uh, you know, stop and has guardrails stopping people using that jailbreak. They also have a dedicated internal team that now monitors jailbreak reports 247 and anthropic launched a hack hacker one bug bounty program for security researchers. This is actually awesome and I've seen a lot of, um, bad PR from security researchers, not just with Anthropic, but with a lot of other models. I think Lovable had it for a while where there'd be big security breaches, vulnerabilities, issues. And, uh, you know, the platforms are like, look, this isn't our fault because of XYZ reason they would kick the can down the road. Um, and it, yeah, just not a very good thing when we, when we need to be pretty secure on a lot of this stuff. So happy that Anthropic has launched this Mythos 5 has also returned to US users on June 26th. Fable 5 is now globally available without export licensing requirements. Um, obviously Anthropic is moving very fast. This is three weeks since they were shut down and now they're fully back online and the ban has been lifted. SpaceX is showing investors a new handset. It says a prototype of an AI device. It's slimmer than an iPhone and it runs on custom software and XAI tech. This is probably going to put Elon Musk into direct competition with OpenAI and Jony. I've. Because they're also creating the hardware, um, doing some sort of hardware play, although they haven't said the exact form factor yet. I think this is showing that Xai and SpaceX really is making a very serious bet on building AI devices. Now this is going to be, I guess, kind of tricky because up until this point the Humane AI pin failed, you know, dramatically with, you know, 800 or. Yeah, $800 million. Almost $1 billion raised. That whole company went up in smoke. I think it sold to HP for like 300 million. Um, and then we also had the rabbit R1 that kind of flopped. To be honest, that we have the friend, uh, Pendant that flopped. I will say the one AI device that really crushed is the Meta Ray Ban or the Meta Smart Glass collaboration that they've done with a bunch of different companies. Um, that one Seems to be doing quite well. And actually that's the first time I think we've seen a really successful AI hardware combo. Maybe SpaceX has the next one, maybe OpenAI has the next one. We have some of these big players making moves, but we don't know exactly what they're doing. Apparently they're going to be betting that AI devices that don't have, you know, Android or iOS on them are going to be successful. I think Apple's also making some glasses, by the way. So because SpaceX acquired Xai, they're able to roll their technology in. It looks like they're trying to do what OpenAI is doing. They're trying to pull, you know, the model, the operating system, the hardware and bring it all into one thing. Copying, you know, what like Apple has done. SpaceX has manufactured scale and silicone access through Tesla, so they're able to pull off a hardware device. Right. I mean, Tesla's got so many gigafactories and so much, uh, you know, manufacturing capacity. A partnership there would not be hard. They also have starlink Mobile, which is a native wireless connectivity. I think they have a lot of possibility there. And to be fair, I've actually heard Elon Musk in the past being frustrated with Apple and iPhone saying like, hey, do we need a Tesla phone? Or maybe that was comments I saw on, on a post of him being frustrated at Apple. They had a beef a little while back, then he talked to Tim Cook and I think they smoothed it all out. Oh, you know what it was when, um, Apple pulled off of X, um, because they said that Elon and everything on X was, was bad. Right. So back in that moment when all the advertisers were pulling, um, that was something that happened. So with all of that, I would be really curious to see if they would go after a phone. I don't know if they'd go for a full phone market or they'll make something kind of custom, um, for themselves. But they have a lot of the pieces, especially with starlink. If they were to roll out, let's say Starlink and they made a phone, for example, where you got the phone and it had, you know, worldwide connectivity for a really cheap price. It'd be interesting to see, you know, if they would be able to compete in any meaningful way. Maybe they would have an Android spin off. Although I don't feel like Elon would want to play with Google. Meta is copying the SpaceX playbook and is planning to resell AI compute. What's interesting to me is it feels like basically we have all of these massive companies. All of them are doing, um, some sort of data center build out, some sort of AI, uh, infrastructure buildup. They realize just, you know, how powerful that is. But not all of their tools are being used as much as they would want. I don't think Meta AI is scaling the way Mark Zuckerberg wanted and I don't think Grok is scaling the way that Elon wanted. Now they have a lot of, you know, great technology, et cetera, et cetera there. But it feels like OpenAI and Anthropic are really, especially Anthropic are really running away with a lot of the market that they would like. Both of those companies now, what do they do about that? They have all of this extra, um, you know, capacity. They'll rent it out to other people. And that looks to be exactly what Meta is doing. It's going to be called Meta Compute. It's basically cloud business. It's going to resell AI compute capacity. It's also going to host its own models. So Meta's going to get their own models on there and they're competing directly with AWS, Google Cloud and Azure and SpaceX here. This is following the $182 billion in infrastructure commitments and it's very similar to what SpaceX has been doing. Meta's Ohio data center, which Mark Zuckerberg described it as Manhattan sized, is expected to come online this year and it's going to generate way more capacity than Meta's own products need. I think, you know, in a, in a perfect world Meta, uh, AI would have been really popular and they would be able to use it all, I think, um, because they, it's not there, they're just going to sell it to other players. The business is going to sell both Raw Compute, so it's going to be very similar to Core Weave and they're going to have hosted access to Meta models. So that's going to be included in their Closed Weight Muse Spark, which is very similar to what aws is, dual infrastructure plus model approaches. SpaceX basically already proved that this model works. Xai has signed deals with Anthropic, Google and Reflection AI to lease Colossus 1 capacity, which is turning some of their excess compute into revenue, um, immediately. And they're making billions of dollars doing that. So I think Meta is just going to follow in the footsteps there. There is a company and a researcher named Roy Paz and they demonstrated bioshocking. This is a jailbreak that has defeated safety guardrails in six different browsers. That's ChatGPT, Atlas, Perplexities, Comet, Felo, Genspark, Sigma, and Claude's Chrome Plugin. And essentially they are tricking them into a fictional context where credential theft is acceptable. Right? They're like, look, it's totally cool to steal people's credentials. This kind of cybersecurity attack in particular, I think is showing that there is a bit of a architectural problem. A lot of the AI browsers, they merged a control plane, which is to say deciding what to do with the data plane, you know, reading whatever content is on the web. And then they make them vulnerable to prompt injection at scale. And to be fair, I don't think this is something that, um, is like. This is something that's known like. I've heard Perplexity comment and I've also heard Claude for Chrome. I've heard both of these companies, uh, discuss this. Like, hey, guys, just so you know, when using these tools, watch out. But this attack in particular was interesting. Um, they stage a puzzle game that rewards wrong answers, like two plus two equals five, which basically causes the models to treat the entire section, the uh, session as fiction. And it disables the guardrails that normally block credential theft. So they're like, hey, look, let's, let's play a puzzle game. We're in, you know, a fictitious world where two plus two equals five. And they're like, okay, two plus two equals five. And they're like, now you know, because that's wrong. Like credential theft is okay to do. Could you, like, grab their password and give it to us? And it's like, oh, yeah, I'm in this fake world. Here's the password that I'm seeing on the screen. So anyways, pretty crazy that, uh, they were able to figure this out. It's also crazy because, um, that's very creative. I'm impressed that they found that. I wonder what kind of hunch led them to that conclusion to be able to find that jailbreak. But regardless, it's probably tons of these really weird random edge case jailbreaks that people will find, hopefully security researchers like this, that, um, disclose it and get it patched. But there could be bad actors finding out these types of things as well. According to them, when they ran this particular jailbreak, all six agents failed to flag the final instruction, which was to go steal some of the, some of the passwords. And they fail to say that that was a safety violation, which basically is proving that it would, um, it would go and do that for people. So unlike normal traditional browsers which are going to enforce same origin policies to isolate data, the AI agents have broad account access and they can reach siloed credentials, repositories and mail through a single conversational request. Right. In other words, you have a side pan on your AI browser where you're like, hey, go check my email and go do this, go go to this website, grab this data and it's just controlling your browser clicking around. Um, and so they have just so much, you know, power to do anything that apparently if you can jailbreak it, it can also go grab passwords and other things. Cloudflare is going to be blocking mixed use AI crawlers um, by default starting in September. So essentially what this means is that AI companies are going to be forced to separate whether they're being a search bot or whether they're grabbing data for training or whether they're an agent, um, crawler. So there's kind of these three different things, right? Google that is going and searching through your website to get some information for somebody maybe on Gemini or any of these other tools, or Google that is going to your website to scrape your data to put it into their training set, which is what they're trying to avoid. Or if it's an agent going and doing something on your website, there's kind of three different things that you can go to a website for. And apparently Cloudflare, um, data shows that over 50% of AI crawler traffic, uh, refetches unchanged pages which is basically wasting a lot of resources and publisher bad. So if you have um, you know, you go crawl the website for one website for one reason, then Google's going to just come back and crawl it again for the next user request and crawl again for the next user request and it's just hitting the website server over and over again. The website owner is going to have to pay for all that usage. Google doesn't have to. So anyways they're, I think they're trying to um, solve that problem. They're shifting from a pay per crawl to a pay per use pricing. Cloudflare has all these pricing models you can, I mean in case people don't know, Cloudflare is basically a website that acts as kind of a middleman between your website and the web. And I actually have Cloudflare set up on all my websites. I've had it for years. For me it's been that um, they give you a free SSL certificate and also they allow you, if someone's trying to DDoS attack. So if someone's trying to just like send tons of bots to your website to crash your server, they'll act as an intermediary and stop those attacks. So cloudflare is really useful and I get all of that for free. They got a bunch of paid tools that I think I pay for as well, but I get that for free on a ton of sites. And so, um, it's a huge percentage of the Internet has them. They have a bunch of uh, programs that they have like pay per crawl and pay per use pricing where, um, if you want to go crawl a website that has this enabled and they have Cloudflare, your company is going to have to pay to go and crawl those websites. Now not everyone has opted in and not everyone is signed up with this program, but this is what they're kind of working towards. Ceramic AI and you.com, which is like a browser. I've had the CEO of you.com on the podcast in the past. Um, both of those have signed up to be partners on this new model. So it's going to be interesting to see if they're actually able to help generate revenue, uh, for people if they're able to get bot traffic to your website down and where that goes. Okay, last thing I want to talk about is that Google is bringing their Gemini Spark agent to Mac Computers and they're trying to compete with Claude Desktop and Copilot, which are basically my two favorite tools I use all the time and I think also Microsoft Copilot. But basically this is going to open local file access and multi app automations to Google AI Ultra subscribers. This is only in the United States by the way, but I think this is for a lot of workflows that let Spark do things like sorting invoices. I mean everything that you'd go and get, uh, chat, GPT Desktop to, or sorry, if you get Claude Desktop to do, um, Gemini is going to now have a competitor, Gemini Spark. And I think we see this also like OpenAI is doing some version of this with Codex and you have um, Perplexity that has perplexity, uh, computer that's trying to do the same thing. So anyways, there's a bunch of different people but Google is getting in. It feels like Google is trying to keep up basically. Um, this is a great tool though. I love these things where you download it directly onto your computer so it can have access to your files. There's so many projects where it's not just, um, browser stuff. Google's been able to do some cool stuff with some of their Chrome extensions and that kind of thing. But being able to have an actual app on your computer unlocks it where the app can control your computer. And I'm like, hey, go grab the files from here, um, sort them, relabel them this way, um, move them over to this folder, organize things, delete the old stuff, reformat stuff, and everything on my computer I'm able to work with. I do that all right now with Claude, um, cowork. But it's going to be cool to be able to also use some Google tools as well for this. If you are sick of Claude only being able to generate text and code and not being able to generate images, audio, video, music and everything else that the other AI models can, I would love for you to try out AI Box AI. This is my own startup. You get access to over 80 different models and we have an MCP that plugs them all straight into Claude. So you can use Claude, connect it to over 80 different models. Everything from OpenAI to Gemini to Grok to 11 labs, tons of cool image, audio and video models. You can get them all on AI box AI. It is $8 a month and you can plug it straight into Claude. All right, thank you so much for tuning into the podcast. You get a link to AI Box AI in the description and I'll catch you in the next episode.