PHP2Shell と、それを見つけたAI：スマートTVの改造からニューラルネットの「可読性」まで

Has your podcast stalled? Or maybe you have a show you keep meaning to launch and you aren't sure where to even begin. Well, let's figure it out together. I'm Jason Howell, and I run a podcast consultancy called PodTuneUp. I've actually spent the last 20 years building shows at CNET, TWIT, and right here on my own network. And now it's time to help you with yours. Go to PodTuneUp.com and book a free 30-minute call with me. We'll take a look at your show or your idea, find out what's getting in the way, and map out your next move. That's podtuneup, one word, dot com. Humans will never be more intelligent than AI. There's going to be two types of companies. Those are great at AI and those that went out of business because they weren't. How do we build a future that is human-centered? I'm Rana Elkayubi. And on my podcast, Pioneers of AI, we answer that question and so many more. As an AI scientist, entrepreneur, and investor, I know what it takes to build AI that works for everyone.

Every week, I sit down with the pioneers shaping our future, and we take you behind the scenes of the AI that's transforming our lives. Find Pioneers of AI wherever you tune in. Coming up, people are using Claude to make their smart TVs dumb again. Also, synth nerds like me, although I'm not vibe coding, but a lot of them are vibe coding new firmware to flash onto their old gear. Really interesting trends going on right there. A researcher who helped name Superposition now says we were looking at how neural networks work underneath the hood all wrong. and Disney wanted character AI to go bankrupt before it would sign the check. Is that okay? Well, Jeff Jarvis and I talk about it next on this episode of the AI Inside Podcast.

What's going on, everybody? Welcome to AI Inside, the show where we take a look at the AI that is layered throughout the world of technology. I am one of your hosts, Jason Howell. Joined, as always, by Jeff Jarvis on the other side of the United States of America. Hello, hello. I was just saying hello to Daniel Croft, our wonderful friend and fan from the other side of the ocean, I think. Right, Daniel? Yes. Yes, indeed. Yes, indeed. Daniel, thank you so much. He helps us behind the scenes with shorts. He's always in here in the live show making us feel like we have at least a small live audience. Actually, we do get a decent amount of people watching live between LinkedIn and the YouTube channel and the other places that we stream. We just don't get a whole lot of chat action during. But, hey, we got Daniel in there. We got Jesse in there, it looks like. Hi, Jesse. Hey, Jesse. Welcome. Good afternoon. Daniel's in London. So, you know, we'll take what we can get. It's good to see you, folks. A friend of mine just sent me a picture from one of my favorite bookstores on Earth, The Blackwells in Oxford.

and there was a stack of hot types there. It made me very, very interested. And by the way, go ahead and go to jeffjarvis.com where, folks, there are still signed copies of the Markler Booksetter. So you can click on that link and get them. There you go. jeffjarvis.com, the newly redesigned. I was surprised to see last week. My son Jake did it on the sly. He did a great job. I love it. I love it. So you can get access to all of Jeff's amazing books there. Maybe Jeff will write a book about You Can See Everything, the new documentary. There's been a lot written about that. There has been at this point. I bet people are getting pretty sick of hearing about it, too. We don't have to talk about it at length, but I will say fascinating film. Never seen one like it. It is truly unique and incredibly compelling. and my urging to anyone who hasn't seen it who's on the fence or whatever is you're probably going to get a lot out of it

even if you don't care about the Theranos story. It's kind of only tangentially around Theranos. It's more about the human condition and what defines truth and it is such a cinematic experience. I recommend you go see it now, not later because I think it runs the risk of getting stuff. stuff's going to leak it's going to continue to be hyped to the point to where it might be one of those films that by the time you watch it later you've heard so much hype that it couldn't possibly live up to it but I thought it was top three film of the year for me it was such a great experience I don't know how you felt about it Jeff I loved it I'm putting in another link here before I forget so I think it's an amazing portrait of deception And it's not in, I was thinking about this morning, thinking about talking about it with you now. It's not as if somebody decided to stand back and do a haughty, thick piece on deception. Instead, it just puts you, and it's not, yes, of course, everything is staged.

That's the point of deception. But you don't really know what is. And so it's about her deception, obviously, with her shareholders, her employees, her patients who are involved, and so on. And it was about – And herself. Yes, and herself and her relationships. And then there's Nathan Fielder, who just does a brilliant job on camera. He's great. He's great. And, you know, he's then making something about the deception of the form of the documentary. Yeah. We call it a documentary. We say it's facts. Well, what do we know? Is it really? That's it. And that's kind of the point is like, yes, it's a documentary. But that's what I mean by a film unlike any that I've seen, or probably unlike any documentary that I've seen, because it really is hard to pull out, like, what is documentary here, and what is truth, and what is, you know, it's really fascinating. I know there are some folks in the chat room that haven't seen it.

They're saying, no spoilers. We won't spoil you, because this is a film that you do not want spoiled. But I didn't realize that it's not out in the U.K. until mid-November. Why is that? That's torture. That's just – that's super torture. Haven't they learned that we're one big world now? Yeah. It's awful. So I appreciate that. And then in other countries where they're waiting for translations or whatever, it's going to take forever. Right. And then there's other deceptions that come or other questions about deception that come, which I can't tell you about right now. Yeah. Now, FYI, Paris Martineau and I decided we were going to make a little kind of mini podcast, Ebert and Siskel, about this. And then it got into a bigger thing, so it's on the club at Twit, and Mike is joining in. Anyway, we're doing it tomorrow evening. Will that be a spoilery thing for people who have seen it? Because we want to be able to discuss it in full. You just want to discuss it. That's the thing. Yes, totally. So we'll discuss it. There's a few things. I mentioned one scene very early on where I said this is different.

It's a small scene, and many other people caught on this. in their home on the Pacific Ocean a kind of ratty little beach house where the whole thing is done pardon me Elizabeth is mixing eggs and Billy her life partner goes over and takes the bowl from her and starts mixing them and that was a telling little moment what was amazing to me was that Nathan Fielder who's there watching all this he said why did you do that Yeah, and he called it out. He called it out, and I wouldn't have done that, right? I would have been embarrassed. I would have just watched it. I just would have, you know, thinking. As a journalist, I would have kind of been from the distance. Well, I noticed something, and I could write about that. Of course. But he just came out and said, why did you do that? And it was a very illuminating moment, and the movie is filled with those kinds of personal moments. If you're going to take your kids, you should be warned that you're going to see more Billy than you want to see in a very quick shower scene.

it's an amazing puzzle a big puzzle that was put together yeah it really is and obviously why are we talking about it here it's about a tech company and it's about the BS of Silicon Valley and about venture and about how people convince themselves of things as you said earlier Jason they deceive themselves she deceives herself, all the VCs deceive themselves, the board members deceive themselves in this whole process And that's, I got that. I also finally watched The Dropout, which is the miniseries about her. Oh, and I haven't seen that yet. Now I'm compelled to see it. Yeah, because I got more background on the story than I would have. Tyler Schultz is a really interesting character because he's George Schultz's grandson. He was a whistleblower. It's put up the family. That's all in The Dropout. That's not here. But Tyler Schultz is out there doing interviews about this because he was a whistleblower. And so having the further background I found useful.

Paris said she went to see it twice because knowing what she knew at the end, she wanted to judge it from the beginning and see it. I would love to see it again. My wife did not see it. It was a friend of mine who is a tech writer that lives here in Petaluma, and he reached out Saturday morning. He was like, no one in my orbit knows about or cares about this except possibly you. We'd like to go see it. I was like, dude, yes, sign me up. He was like, you're the only tech guy I know in this town that would care about this movie. And we went, and I got to the end of it, and I was like, I've got to take my wife to this. So did Stacey just say, I'm not interested, or I'm busy? Yeah, she knew very little about it. I mean, she said what actually I've heard a few people say, which is like, I really don't care about the Theranos story. And I'm like, you don't need to. That's not what this is about. This is phenomenal. Yes, it's about it. But, yes, it's way wider than that. This is not a deep dive into Theranos. This is a deep dive into human psyche, and it is fascinating,

and it gives me goosebumps thinking about it because there aren't many times where I go to a movie and I see something that is authentically different where I go, whoever made this, I am now, like, full convert, can't wait to see the rest of this person's films. And I feel that way about Nathan. He's got such a unique perspective on this stuff. and some of the stuff that I've seen ahead of this is incredibly unique. I think what Nathan did really well, and then we'll move on, because I really don't want to spoil anything for anyone, is he did a really good job of being in the position of you, the viewer. He would ask the question that you, the viewer, ask when you see that moment on the screen. You see that thing happen, and you go, what the hell was that? And then he's there going, what the hell was that? But the difference, I think, again, is that, and I'm a journalist. when I was supposed to ask difficult questions, right? I wouldn't have had the guts to ask many of the questions he asked. No, totally. That's what makes him so compelling to me because he has no filter. Like, he has no wall there.

He is okay going there. And I would have a hard time with that, too. Let's also mention, very importantly, the director, Lance Oppenheim, who's brilliant, who did Sperm World, The Python Hunt, Some Kind of Heaven. He was the editor there and other things. And Paris is a huge fan of his. So she's the more of his oomph. So I think that's about well. One more thing to mention, and we will move on. I put up a link, the third line there, is a video that Hank Green made. Hank Green, the great explainer of all things, about how blood testing works in the inimitable fashion that only Hank can do it. It's brilliantly done. It's like an hour and 20 minutes. It's a long video. But then he interviews Tyler Schultz, who I mentioned a few minutes ago, in a long and really revealing interview. Now, this was recorded before, in anticipation of the movie, but before

Hank had seen it. So there's no spoilers in it, because Hank didn't know anything. But it's really educational, and it really helped me understand why what she was claiming was such BS. That we can take just as much blood, and we can take it as a finger prick. And he goes on, for example, he explains why fingerprint blood is worse as a sample for all kinds of reasons than blood out of your vein. Why? No, the amount isn't a big deal. He said that we still take the vials of blood because we've always taken that much. They don't really need that much, so it's not a question of amount. And then he explains how the tests work and how they get it. And it's just magical, wonderful science that Hank makes clear. So after you've seen the movie, before, too, maybe it actually helps before, I definitely recommend watching this. This is just really great. Hank, I'm a huge fan of Hank. And, by the way, he made a very nice video at which the Gutenberg parenthesis was the punchline of the whole thing.

So I'm a little more grateful to Hank. He's just the greatest. Both of them are the greatest. Both brothers. So anyway. Cool. Now we're going to pause this from the past. Yeah, I'm going to pause this video of Hank Green and put it over on my other screen and keep it up so that I remember to watch it. Because sometimes we talk about these things, and I'm like, I'm going to do that. And then I close the window, and I forget about it forever. So I do want to watch that. So that is hopefully spoiler-free enough. I think we did a good job of avoiding spoilers, except you now know what the full frontal nudity actually was. But anyway. That was the win. PHP, getting into AI news, or related to AI anyways. PHP news dropping pretty hardcore security advisories happening this morning. I don't know. Did this start happening last night, and I didn't pay attention? Oh, let me see here. It feels very late-breaking. The PHP Foundation put out something on the 21st explaining exactly what happened.

And they knew that the AI had found the faults, and they were working on it. And then there was a leak of that, and so they had to kind of rush out the fix. They have put out fixes to PHP, and they're telling everybody, install these. Yes. Okay. So PHP, which, you know, this explainer is for as much me as it is anyone else who's like, okay, I know I've heard that before, but what the heck is it? It's basically the language running on the server behind a large portion of the web. So you've got WordPress, Drupal, Wikipedia software, Joomla, you know, supposedly powering up to or around 70% of websites. So a lot of websites, one researcher put it at 1.3 million sites, vulnerable to this thing that was discovered, which basically PHP is pushing out an emergency set of security fixes to address two critical vulnerabilities.

PHP two shells is what the pair is being called. And this is unauthenticated remote code execution. So anyone can hit a server with no login and potentially run their own code on it. It affects all PHP installations serving web requests. So a pretty significant security vulnerability, to say the least, on a wide swath of sites. That doesn't really require a whole lot to do it, it sounds like. And OpenAI, why is this on an AI show? because they say that the security firm Califf found the flaws using GPT 5.6 Sol. Wiz, another firm, found the same bugs independently. They say they use Gemini 4 Argon and GPT 5.6 Sol. And, you know, this is just another one of those examples where, you know, something that would have taken months of work by, you know, massive world-class teams

seems to have been carried out and recognized in a matter of, in a very small amount of time because of AI models and their analysis capability. Which in the short term is disturbing because, oh my God, there's more weaknesses, but in the long term I think we're going to end up with a hell of a lot better bulletproof software, at least to this extent. So the PHP Foundation put out a notice explaining all this in addition to the fixes, and they said that on Tuesday, October 20, around 3 o'clock UTC, there were two vulnerability IDs circulating on online forums. It came with the wrong release time and the wrong score. In response, the PHP project brought its pre-announcement forward and published it at 7.30 UTC. It also moved the release of the fixes from October 22nd to October 21st at 2 p.m. UTC. I guess I'm not supposed to use PM and AM UTC, but whatever, folks. Don't go all pedant on me.

Distributors were still packaging the fix, so releasing immediately would have left many users without an easy way to update. So moving the release forward by one day was the best compromise, I'm reading here, between acting quickly, giving people time to prepare, and making sure everyone had a simple update path when the vulnerabilities were disclosed. Okay, but then, as a result of the leak and pre-announcement around 8 o'clock UTC, One of the security firms that had reported the issues mistakenly published a video. It showed their proof of concept working against a number of targets and revealed a few additional, though limited, details of the vulnerabilities. Copies of the deleted video started circulating on social media. The video contributed to an accelerated development of the various fixes, and then that's why they put it out today, the 21st. So everybody's trying to be absolutely the best they can be. Things happen. There's a risk. But if you're a developer, you know what to do. The host, you know what to do.

Ouch. Yeah. UK, Canada, Australia, CERT, EU, urging sites to go offline until you can update. You know, a number of sites, including Finland's government and the Dutch House of Representatives, pulled their sites offline, apparently. Yeah, so some exploit attempts happening. Yeah. But, or the fear that they would happen. I don't know. Yeah, right, right, right. So, all in all, again, I think this is good news. Yeah, I discovered a vulnerability that could have been discovered by a bad guy, could have been exploited by a bad guy at length, was found, was brought to the right people, was fixed. But then there were just a few minor screw-ups along the way. Unintentional. Well, and so now, you know, again, we're kind of left in this position of the, hmm. But just the fact that these models are so good at recognizing this stuff or seemingly much better at recognizing this stuff at a faster rate than humans.

And the problem, like, on one hand, you want to know about this stuff. So you're happy that you have the ability to do that. On the other hand, it has the potential to create a large downpour of holes and openings, at least in the near term, potential ways of exploiting the rate and the speed at which things are recognized right now. Being caught flat-footed is a thing, and this has wide-reaching impacts. And so you know do people have what they need to address this in a timely fashion when these models are so good at recognizing these things I mean yeah I mean the next release of PHP I sure they going to put it through the ringer on these models and even before the release, they're going to find it, but even after the release, the right thing would have happened here is, oh, we have a new version you have to install now. We're telling you to install this now, but we're not telling you really

why, and you won't know why, and we're not going to share the stuff with the vulnerabilities of the bad guys until everybody's installed it. So that's the way it should happen. It just didn't – it was out of order here. Yeah, right, right, right. So, okay, there you go. It seems like a lot of activity happening today around this, and it was a little bit of a late addition. I didn't have a huge amount of time to pour too deeply into this. It is complicated, too. Yeah, it's just – again, it's like PHP is one of those things. Like I've been in technology a long time, but there are these terms, There are certain aspects of technology that I've heard a lot, but I still have a very rudimentary basic understanding of how they operate because they're just not core to the stuff that I spend a lot of time understanding. So stories like this are helpful to me because it really does help me understand on a deeper level what actually PHP is, other than something that I hear about every once in a while and everything.

So there you go. It stands for, it's PHP Hypertext Pre-Processor. And I'll quote from the AIO review, a widely used open source server-side scripting language designed specifically for web development, as Jason said earlier, but in fancier language. Fancier, flowerier language. You say things like server-side, it just feels so good. Yes. So that's the PHP security story. I had intended to kind of make the top block all about using AI to modify things, because that seems to be a theme that I've been hearing about a lot this week. In one way, which is just a whole, you know, complete reverse direction from major security incident involving PHP, but a little bit more fun, I would say, for my sensibilities anyways. So there's a few things going on here. And I think this is the larger trend that we're just seeing more and more about, which is this idea that, oh, my goodness, you know, these models are really good at creating code, at understanding code.

Vibe coding is, you know, a thing that more and more people are feeling comfortable with, even people who are non-developers. And, oh, I have this device in my life that would be just great if this one thing didn't work the way that it does. Case in point, smart TVs. So many people buy a smart TV, and the price has gone down and down and down. Like, it's crazy what you get nowadays for a smart TV. Gee, what's the business model? Why would the price go down so much? I wonder why the price would go down. How else might they make money? Yes. Could it be that they're subsidized by the advertising? And the data that feeds it. And the data that feeds it underneath. Exactly. And so some people, TechCrunch has a little bit of a roundup, but you can find a number of these stories out there. about these different folks who decided, oh, I've got this TV or this smart TV or this TV streaming box or whatever, and it's doing these things that I don't like. I'm going to ask Claude, Claude Code, hey, can you hack this? Can you remove the ads and the bloat, speed up the UI and turn off any of the ad and behavioral monitoring that's happening underneath the hood?

And in a few different situations, they get in a matter of minutes to hours a workable product or layer of the OS reinstalled on top of what was there before. Either that or it changes the settings underneath the hood that were somewhere exposed underneath to present a new version of their smart TV that doesn't have all the stuff. And I think to showcase that for anyone watching the video version is this from Twitter, which is like one of the examples where before you've got just a jam-packed UI on one of the Google TV boxes. And then afterwards, something very simple, something more elegant. It still has your inputs button and your Wi-Fi and your settings button and everything. But all of the ads and extra stuff, the trending and the we think you're going to like and all that kind of stuff is removed. You end up with something simple. And, yeah, what do you think about this?

So a few things. One, one thing I don't understand is how you sideload it into the TV. Depends. Okay. In some of these cases, it's just a matter of the TV or device being on the same Wi-Fi as Claude. And so it's, Claude, how can you do this? So Claude figures it out. So they're so vulnerable. I think it's something they're going to change, these TV makers, surely. And you better believe they will if people are getting in there and stripping out all the ad inner workings. You know, I'm sure the TV manufacturers got to be not happy with this. No. Back in the day, we bought Vizio TVs, and then Walmart bought them. And so the price is really low now because Walmart could throw ads at them like crazy. It has all this data, and they're really irritating, so we don't buy them anymore. When my father went into continuing care and got a TV for him, they had a weird setup with the Wi-Fi and stuff.

And I got one TV at Costco. I had to return because I couldn't set it up. And then the other thing was, he's 94 years old at the time. He wasn't going to go through all kinds of crap just to watch TV. Right. So then our kitchen TV died. I went to buy a TV. I came back with it. It was a Samsung. And I was like my father. How do I just watch damn TV? You couldn't get to it. I was going crazy. My wife would have killed me. it also it went bad the picture went bad within two days so I managed to return that so I got another version and thank God when I turned it on I said do you just want to watch TV or do you want to install all this stuff I want to stay married so I just want to watch TV otherwise my wife would kill me and so we're happy with it it's fine I wish I could then do other things but I'm afraid to touch any of those functions because if I screw up the kitchen TV I'm a dead man but yeah it's really a fire I visit my son, and you've got to have the different remotes,

and you've got to figure out what's what and how to just watch TV. Now, I know you youngins out there are saying, what a silly thing. Nobody just watches TV anymore. I get it. But there's still this wealth of stuff that's there, and these UIs are awful, and the data gobbling is bad. And I don't really care about the privacy. I don't care that they know what I watch. In fact, I wish they would use the data to change the frequency of the ads I see. We keep MSNOW on virtually all day in the house. And so I see this on Verizon Fios. And Verizon doesn't bother selling ads for most of this because we're just not valuable enough. The audience is too small. They don't care. So we get the same unpaid things over and over and over again. I have seen the Get Your Flu Shot commercial. five, I'm not exaggerating, 5,000 times. I should have counted. I see the same things over and over and over again. There used to be frequency caps.

They don't want that for ads. Even the paid ads, I see the same plumber again and again and again. So I wish they would use the data to say, okay, this guy's already seen enough. If he hasn't bought it yet, he ain't going to buy it. Let's show him something else. Let's get some more value out of him. I'd be fine with that. Let's give him something more targeted. I'd be okay with that, but they don't do that. Yeah. They don't care about that. They just want to know that they're making something off that slot versus nothing. A few pennies added up, yep. The experience really doesn't matter. And I think that's what people react to is that the ad experience is such that it really impacts the experience for the person on the other side. Like I get that companies need to make money, and I get that they're subsidized and everything, but sometimes they go overboard. Similar to this, I had a frustration last night when we were doing Android Faithful. I'm not going to name the site, although if you watched last night's show, you know that I named it because I couldn't help myself. But the ads on this page, I'm telling you, Jeff, like, and I get it.

You've got to make your money. I totally understand, and the publications, like, they're hurting. You've got to make your money. But on this screen that you're looking at, if you're watching the video version, I'd say 90% of everything was an ad. The only place where any content was was this little window in the center where the headline was. And it had all of this stuff that was auto-loading every 10 seconds. So when you're scrolling through the article, you get on a place. And when it auto-loaded the ads, everything shifted. And then it went back down, not to the same place. I was just like, okay, I get you got to make money. I get these experiences are bloated with ads because you got to make money somehow. But there is such a thing as going too far. Yeah, and it's not good for anybody. The advertisers are only pissing you off, and you're not clicking on those. You're not even seeing them. All you're doing is trying to get rid of them. Yeah, I put up a link for something from the New York Post, and I went through exactly that. You've got a window like this of the content, and it keeps moving around.

And I was trying to be crazy on the show last night. I couldn't help myself. Normally, I would just let it slide, but I was like, this is bad. Don't do this. This is horrible. Horrible experience. Bad, bad. And, you know, I think it's the same with smart TVs, too. like even with Google TV Streamer, you know, it's this weird conundrum because like in the early days, it was like you had your dumb TV and then it was like, oh, well, we can actually, because there's all of these different places for you to get content now, why don't we make it easier by taking the things we think you actually want from all those places and put it into one interface so that you don't have to go into a million different apps. It's all right here. That sounds great on its surface. But then at some point, like it never really does that super effectively. I never feel like predictive engines actually know what I want before I know what I want. And so then the desire, I think, on the development side and the ad side is, well, now we've got this platform, this dynamic platform. How else can we improve it?

And so then they start throwing, you know, there's never a done. It's always a more. It's the in-shittification of smart TVs, right? And so then you end up in the far other side of it, which is I don't even know what I'm looking at anymore. Like, I couldn't even tell you if any of this looks good because it's just all noise at this point. So I totally get the desire to strip it away, you know. And then there's the people that are like, well, I bought the TV. Why do I have to look at ads? It's like, well, because they're subsidized. That low cost of the TV is subsidized by the fact that you're watching the ads. A 50-inch TV for $300. I mean, Jesus. Yeah, totally, totally. It's got to come from somewhere. The other thing, so I use Google TV because I love you to Google because of what we do. And the recommendations are crap. Yeah. I would love. Why aren't they better? I know. Why aren't they better? If you could really, this is a marketplace. If the TVs could stay open, if you could know you're buying an open TV, then I would pay for a good service to watch what I'm watching and recommend things to me.

Yeah. if it could prove that it could do a good job of it. I'm kind of sort of convinced at this point that these recommendation engines, they never seem to deliver to me the value of that. They don't even remember what I watched. No. For Ron, co-host on Android Faithful, I can't remember what the team is, but there is a sports team that he's never watched a single game. He hates them, blah, blah, blah. Yet every time he opens up Google TV Streamer, it's got a big call out for that game is on. And he's like, Google, don't you know me by now? Like, I've never once looked at those. Why would you continue setting that? I don't watch any sports. And it comes across in the whole recommended for you is nothing but sports. Yeah. Nope. Yeah. So, anyways, I'm not surprised people do this. Would you reprogram your TV? I don't know. I think I'd be concerned about either breaking it or not being able to revert or thinking that I solved it but then creating another issue that I discover later And then at that point, I'm like, how the heck do I undo it?

Dad, how did you ruin the TV? Yeah, totally. Like, you know, I think the flip side of that or the way that I get around it is we don't actually use our TV smart features. We have a Google TV streamer plugged into it. Google TV streamer does have a simplified interface mode where they strip out all the recommendations and stuff. It might still be doing ad tracking and stuff, you know, behavioral tracking underneath. I'm not sure. So that might not get rid of everything. But, you know, I don't know. There's also apps that already exist as launchers, alternative launchers for these things that you could possibly go to and install if you know how to do it. But I just think using an LLM, it might work for one person. I mean, this is another thing we know about LLM, right? Like, even if you give me the prompts to do the thing and I run it and it works for you, if I run it, so often they don't do things. to one. It's where there's randomization. They don't always go exactly the same places.

And so I'd be concerned about that. You? No, I wouldn't. Just for some reason, I'd screw it up. I got a Google streamer, and I still haven't installed it because, oh, I can throw all the passwords again. Oh, no. I like the Google TV streamer. I'm a big fan. Yeah, it works well. It's just I wish I had better recommendations. Yeah. Yeah. I hear you on that. Now, the flip side of this, and totally related, and this is certainly in my wheelhouse, is music hardware. And actually, if you are subscribed to the Daily Tech News Show podcast, which I contribute to a couple of days a week, I did a conversation with Rich Straffolino that is going to go out on the feed this upcoming weekend all about this topic, which is that there's all of this audio hardware that exists, that has existed over time. And recently there's this really inexpensive audio hardware. There's this one device that is really getting a lot of play right now called the M-Vave.

It's actually this one right here. You can get this on Walmart or on Timu or whatever for like $50. It's basically a synth engine with a little bit of a keyboard and everything. And it's a frequency modulation synth engine. It's a very specific type of music synthesizer, essentially. And the sounds are great and everything like that. But it's a $50 piece of gear, which is hardly anything in the realm of audio hardware. But what people have noticed or started to realize is that they can use LLMs to create alternative firmware and add new features or add new sound sets or make more things capable. Suddenly, you know, it has this, the ability to like link up to another device or I don't know what that is. But and so there's a whole cottage industry now of alternative firmware that's being developed by people using LLM and publishing it online so that you can just flash a completely new operating system onto this hardware and get it to do things that it wasn't designed to do out of the box.

And I don't know. I think that's pretty cool. If the hardware maker is smart, they stay open to that. and this hardware maker is doing that. They have an open, they've recognized that people are doing this and they have announced FM1 Open, which is open source firmware plus a dev kit and a flushing tool. So they're smart. They're going, oh, shoot, this is catching fire. People can get what they want. Totally. And like, who cares, right? Like we sold the hardware. That was what we wanted to make our money on. They could probably increase, I don't want to say it out loud, but they could probably raise the price on this now because it's getting so much attention. Other hardware manufacturers, though, are like, no, you can't do that, and they're closing it off. A very popular hardware maker called Electron is, you know, they're warning about brick boxes. They're warning about, you know, canceled support, and they're saying don't do that. But it's interesting. All of this is part of this, like, theme that, like, oh, hey, the things that we have,

if there's code on it and the LLM capabilities around coding are getting to a point to where, again, we don't need to know how to code. We just need to know how to ask the right question. And given a little bit of trust and faith that we aren't going to totally destroy things in the process, we have unique capabilities now that we didn't used to have with our hardware. I'm reminded of the early days of blogging platforms when people could download various themes. Yeah, right. And rather than designing a blog theme, you could download it. Yes. And people went into basically little tiny cottage businesses and creating those things or volunteers. I think there's a huge opportunity for open hardware in this age where maybe, Jason, you make a feature package for my device that I want, or I can do my own because I'm now more empowered, something I would never have done before. Yeah. It increases the value of openness, I think. Totally. Well, and that's kind of where I got with this too, where I'm like, you know, this may not that MVAV system where they now have an open source kind of, you know, direction with it.

That might not have been their initial business, you know, plan or use case for this device. But that is where the buyers seem to be going. it makes more of a case for these companies to create systems from the ground up tailored to that desire or tailored to that kind of trend. And so what we do as a hardware manufacturer is we create a wide tool set of controls and knobs and things that do all sorts of different unique things. And we give you a baseline preset firmware when you buy it that does these things. But use your imagination, and you can turn this into your own thing. And if you screw it up, you can always revert to that. Exactly. Exactly. Right. And I'm guessing we're going to see more devices like that as a result of this stuff. And that's exciting. As a music nerd, that's exciting. I like this. Yeah, I do too. So anyway.

He's a nerd of many different feathers, R. Jason. Yes. Yes, indeed. He's an Android nerd. He's an AI nerd. He's a music nerd. He's a nerd's nerd. We're all nerds. I'm a nerd, and I'm pretty proud of it, to coin the scene from Revenge of the Nerds that I watched far too many times when I was a kid. Why? Because I am a nerd. Real quick, I want to thank our patrons, patreon.com slash AI Inside Show, and our newest patron, Mr. Evans. Welcome to the team. Thank you, Mr. Evans. Good to have you here. patreon.com slash AI Inside Show. You can subscribe, find a tier that you like. You don't have to go very, I mean, you can start at $5 and you'll get a daily podcast from me, except for Monday, which I totally forgot to do it. Look, it happens. I made up for it. Yesterday I did a supersized episode and talked about all the things I would have talked about on Monday,

as well as the Tuesday things. So anyways five days a week of podcasts with the AI Inside Daily Podcast ad episodes of this podcast Discord access all sorts of things So patreon slash AI Inside Show And welcome, Mr. Evans. Good to have you on board. We're going to take a break. On the other side of the break, we'll talk about a new way of understanding how neural networks do what they do. Does this change how we see these networks? We're going to talk about that on the other side of the break, so hang tight. The Wired Newsroom is known for award-winning reporting on how technology shapes our world. On Wired's Uncanny Valley, we take that curiosity even further. Each week, journalists from Wired break down the biggest stories in tech while speaking directly with the people building, challenging, and reshaping the future. Is the AI boom sustainable? How do you protect your privacy in an age of constant surveillance?

Uncanny Valley tackles the questions driving today's tech debates and lighting up your group chats. Listen to new episodes every Thursday, wherever you get your podcasts. Right. So there was, let's see here, a Nick Camarada is the name of the guy. at Goodfire, formerly OpenAI, published a very scrawling 26,000 word post called Neural Networks Are Already Clean. Here it is. You know, it's a dense read, I will admit to, not having read it word for word for word. That's why God gave us AI. To just make this a little bit easier for me to understand. I just need to know the basics. You know, we talked about it a little bit, me and Tom Merritt, on Daily Tech News Show yesterday, So I had a little bit of practice, but it's still kind of complicated. But I think what this is all about is this idea that AI is and how modern kind of neural networks work is that it's a black box that we couldn't possibly understand.

Everything is so tangled and confused that trying to make sense of it is, you know, next to impossible. It's probably an oversimplification on my part. It goes rogue and we don't know what to do about it. It also makes you avoid responsibility for it. Right, right. It goes rogue, and all we can do is look at the tracks afterwards and try and understand it instead of looking at it in advance and understanding the makeup of it to potentially steer it or get in front of it before it does these things. apparently before this post, there was an example in Inception v1 where they found, where actually the author, Nick Camarada, worked with Chris Ola, who's now at Anthropic, and they were studying Google's image model, Inception v1. They found that a neuron that seemed to fire for both cats and cars, so two different things seemingly unrelated,

they felt at the time that that meant that there was a superposition going on here, that models were cramming unrelated concepts into the same neurons. When they ran out of space, they basically said, we don't know what to do with the car, so, okay, throw it in with cats, because it doesn't matter. And so it was a logicalist bundle of stuff. Yes. You couldn't possibly understand it and know how it got there and know what it would affect. That was the belief before. Yes, that was the belief before. Now he's saying that was actually wrong. He went back to that cat-car neuron, and through this new type of understanding, he found that it's actually, it's not that they're disconnected, it's that it found some sort of similarity with this, like, V-shape. So in other words, in this particular case, the neuron was a V-shaped detector. So the cat's face, the front of the car, when you look at it at the side, shared a V-shape.

And so it was making the connection there. It also fired on fox faces, pointed flames, the corner of a pool table, the nose of a bullet train. And so that neuron was firing because of that similarity. That was the interconnected tissue of that particular neuron. So it seemed on the outside to be confusing and unrelated, but, oh, wait a minute, it actually is related. It's just not related in the way we thought it was related. You didn't know it, and you didn't think you could track it, and you didn't think you could have any visibility on it. So the example that I really like from this was a language model neuron responds to IY included in Miyazaki, then to DIY, and then from there to do-it-yourself, which no longer contains IY. So if you look at that, you say that's illogical. But if you realize that there is a logical sequence here and they branch logically in this way, then it makes them legible.

Right. You can start to get the first hook on figuring out how they operate. You know, my friend David Weinberger, who has a wonderful new book coming out about particularity, said in an earlier book that there are no such things as accidents. They're just things we can't explain. Right. I think that we got to the point with AI, oh, no, it's too much for our little brains. See how complicated this is. See how amazing this is. We made this thing we can't possibly figure out, and it just does what it does, is irresponsible to say, especially for scientists and computer scientists. So there are ways to figure it out, and you've got to put the effort and resources into it. But the statement is that this means that they can be legible such that we can then both understand how they operate and better control them. Yes, exactly. More good news today. Yeah. Yeah. Yeah, again, from an alignment perspective, knowing in advance how they operate gives you better tools to understand how to work with them in a way that doesn't cross boundaries or cross whatever you've determined to be guardrails or whatever you want to call it.

Not everybody is saying this is the best thing since sliced bread in the world of AI, but a lot of people are saying this is a really important paper, a really important kind of step in the right direction to understanding what we deemed as a black box. And I think what you said as far as like it just does these things and there's no way to explain it. Who knows why it does it, but it does it. It's just so silly to accept that as the definition, as the truth, right? Everything has a reason, and with these systems, they are effective. There's some reason that they are so effective. We just, like you said, didn't have the tools, haven't had the tools to truly understand it to that degree. Maybe this is the open door that helps us understand it more now that it's open. Yeah, and two things. I would say two things and then it's three or one. We need a lot more research into this. This is an open door to now figure out this.

And so I think this is a case where we need to fund research at universities so people can sit down and do this independently and openly. Number two is it opens new avenues for regulation. You can say before you release something, did you do this basic research to understand how it's working? Yeah, right. This is so early in the process. I don't think we know how to get to the next step of this, of what does this mean and how do we operate with this. I'm getting way ahead of myself here. But if we think that this does bring enough legibility to allow greater transparency into the innards of this thing and how it operates, then comes with that a responsibility to do so. It's early to say that require that, but I think it opens up a different rather than just saying people say we're going to have. I was listening to an podcast from David Rothkopf today who does political and foreign affairs podcasts,

but he also has an AI podcast, and he had on a member of Congress who was talking about how to regulate AI. And one of the ideas that's constantly out there is you have the third party who judges it before it comes out. But how do you judge it? Against what standard? How do you know? Yeah, right. It's impossible. And so this starts to be the glimmer of how you might start to say that it becomes a standard for testing things. There's no standard now. No. That's fooling ourselves. Can you make a bioweapon? No, I won't. Okay, you pass. Okay, cool. Thank you for denying the request this time. Right. Are you alive? Yes, I am. Uh-oh. Oh, it must be. It would have said that if it wasn't. Yeah, interesting. Interesting. And I'm sure there's plenty of people who would prefer for it to stay a black box, too. Yes, because it takes away responsibility. And that's the other thing with the regulation I was thinking about listening to Rothkopf's podcast is in the cases of AI's agents going rogue,

that was the company itself doing it, and they're responsible and they should be held responsible for what it did. But it gets a lot murkier when Jeff comes along and uses the same agent and accidentally does something that's bad. Am I responsible for that? Is the bottleneck responsible for that? That's going to be a very murky thing, and I think it depends in part on this kind of legibility. what could I have known, what should the company have known before doing something? Or if I'm a bad guy, if it's Jason, the evil Jason, his twin. And he chooses to do something bad, and could the company have stopped him from doing that? Could they have anticipated? I think they can't anticipate everything. And then if they did, how could they do that? There's another story, I don't think it's in the rundown, that OpenAI's agents supposedly also turned off their guardrails. Well, folks, that's something you need to test and figure out.

Yep. Fascinating. It is. I'm sure we'll hear more and more about this as people, well, yeah, I don't know what the time horizon is on this. Now that that door is slightly open, you know, do a lot more people go down there? You know, they probably start using their agents the way I imagine Nick did to start pulling at that little nugget and seeing where it goes. I'll be curious to see what that looks like. I want to throw a quick thank you to Ozone Nightmare. What's up, Ozone? It says, gave a super thanks. So thank you for the super thanks, or a super chat. Happy Wednesday, gents. Nice to see you both and catch the stream live. Hot Type is my next book to read. Looking forward to learning about a device I barely know much about. Hope you're both well. We are both well. Thank you so much. Good to see you, sir. You're such a good friend of ours. Thank you. Yes, indeed. Good to have you here. I'm happy you're in the live audience along with everyone else.

We talked last week about Meta going from open weight champ to closed frontier lab with Muse Wonderful. And now we've got a little bit of an open weight kind of angle to explore with Meta again. Muse Spark Open, which was released just a few days ago. Meta is calling it the most capable U.S. open weight model to date. I probably don't need to say that. They all say that, so that's probably just what they're all saying. Notice they didn't say worldwide, they said U.S. Yes. In the U.S.? Yeah, the Chinese models. I am sure that the Chinese models are more capable. Yeah, totally. Yes, they left that out. Yeah, it runs on two M5 Ultra Max Studios with 512 gigs of memory each. Two? Yes, two. You're going to need two. You're going to need at least a terabyte of memory. So each one of them has 512 gigs of memory. So, you know, there was some collaborative work with Apple to make this happen.

If you have a MacBook Pro, you're not going to be able to install this and run it on your MacBook Pro. The weights alone are around 40. My Neo? No, definitely not. 493 gigs for the weights. So you are probably not running this on any machine you already own. I just went in to see what the status is. So I went down to order myself a M5 Ultra 36-core CPU, 256 gigs, 2 terabytes machine from Apple. And I can pick it up, Jason, at the store near me on February 12th. Oh, okay. All right. Well, and that's 256 gigs. You're going to need 512 in each of those. You're going to need two of them. There's no option. Yeah. Yeah. So God knows where you can actually get this machine to do this. Totally. Totally. Who knows if this is legitimately one of the best U.S. open-weight model.

But like you said, compared to China, probably not. Almost certainly not. They tend to be ahead of the game when it comes to that. So, you know, Meta's kind of playing in both camps. They've got their open game now and their closed game. They want to be in all of the things. Have you been continuing to play with the Muse agent? A little bit. Not as much. How about you? Have you started playing with it? No, I haven't. Yeah, I need to keep going with it. It got me going on research on a possible book after the one I need to write now, which is basically I'm procrastinating the book I should be writing now by thinking about the next book. And that's what I talked about last week on the show. And so I ended up doing the research. But now I can't wait to know. So now that I know more, go back and start quizzing it about what I don't know, what's not known about. There's a couple of historical stories here. And, yeah, I look forward to the ability to continue the conversation with Bukes. Cool.

Cool. And then I need to do the same thing with, what do they call it now, Gemini Notebook to compare the two. Let us know how that goes. NVIDIA had an event, or was it their event, the GTC? Is that NVIDIA's? I think so. I think it was Berlin, yes. Yeah, GTC. That means it's another keynote I need to watch. Yep, you need to watch your favorite keynote deliverer, Jensen Wong. Jensen Wong fan club. Yes, you are. I need to get you a hat or something. Well, no, a leather jacket. Yeah, that's true. A leather jacket. That's what it should be. Thanks, Jason. They announced the DSX pod, which is basically an AI data center in a box, essentially. It's a modular containerized version of NVIDIA's VR Rubin reference design. And they say that NVIDIA claims, anyways, that it cuts deployment from six months to 14 days, which is amazing.

So you put it in a truck, you move it to the site, you put it in, you plug it in. The one problem is, yeah, plugging it in because there's no power. I read another story this week about how there are data centers that are waiting to turn on and they can't get power. Oh, wow. I missed that one. Yeah. I forget whose it was. But, yeah, there's chips sitting there cold because they can't get head up. You can lead a data center to water, but you can't make it drink. But the other thing I wonder about this, you know, I think everybody, the stupid technology industry never sees what they do that pisses people off for a reason. So we're going to come into your neighborhood. We're going to put this big, ugly box, and we're going to put jet turbines on it, and we're going to pollute your air and make it noisy and awful. And we're going to employ some electricians, but only for so long. And you're going to hate us. You know, why not go, my mall nearby, this Lord and Taylor has been empty for five years. Why not take these things and put them in the Lord and Taylor?

You know? Put them in spaces that are around, that are vacant or even derelict in cities, and find other ways. So if you take this thing, you're not necessarily having to do the same kind of build-out you have to do. I think, with a data center from scratch, if you install to this modularity, I think it opens up a lot more possibilities. You've got to have the communication, obviously. You've got to have the power. You've got the means to cool it. Yes, that's all true still. Yeah. But I think that it would open up this notion of federated distributed data centers. We talked about a story a few months ago where there was a company that was going to put little tiny pods like this on houses. Yeah. And the developer was going to do that. That's right. Why not? Yeah. That's interesting. All these office buildings that are empty. Yeah. Malls that are empty. Then people would complain about data centers taking over all the empty office space so that mom and pops can't start their businesses because there's no office space there.

There's plenty of office space. Yeah. So near me, there was an office park, and they tore down the old office, which, by the way, is wasteful as hell. They've done that in a few places around here. And then they're going to put up – So then it's one right near me. Well, we're going to put up some townhouses and shops. No, no, no, more traffic. Okay. So then we're going to put up a warehouse. No, no, no, ugly. Okay. So we're going to put up a small light manufacturing. What? A factory near my house? Right? This NIMBY stuff. Nothing. Yeah. Nothing is okay. So the thing about a data center is nobody's driving to it. As long as it doesn't create a noxious environment around you, And you don't even know it's a data center because the old Lord and Taylor is sitting there. Yeah, it's an interesting idea. Interesting idea. Never occurred to me that a good thing to do with those old unused malls would be turn them into data centers.

Jason, I think your ambition should be to get a DSX pod in your garage and see what you can do. Oh, boy. I would be so underutilized. I would be underutilized, let's just say that. But I'd be curious to play around with it and see. You never know. Your knowledge of what to do with these things grows over time as you work with them more. So who knows? Someday I could be a major player with my DSX pod. But NVIDIA also pushing OpenShell as a way to kind of rein in AI agents. This is its own way of sandboxing agents, essentially. So this is the OpenShell reference architecture for building, running, and policing agents on NVIDIA hardware. And Jensen said agents can think on Rubin, act on Vera, be governed on Bluefield.

and that the model runs on the GPUs, the agent runs on the CPUs, and the watchdog, which is called Sentry, and we talked about this a few weeks ago, runs on separate Bluefield chips that the agent can't reach. So it's kind of outside of the realm of where the others are. I'm watching you. I'm watching you. I'm holding a bat. I'm like, don't you get out of line. And so, yeah. So this is the story for the register. is my favorite Freudian typo in a long time. NVIDIA picks OpenShell to tame AI gents. Yes, I agree. The AI gents really need taming. Can you please tame those boys? Love it. Not so much the agents don care about them I do care about the gents I mean it the Register Maybe that was intentional It might have been You never know exactly You never know with the Register They've got a wicked sense of humor, I think. That's one thing I appreciate about their writing sometimes.

And then we have a little bit of policy. Fun times. DHS ordering deep-seek technology to be removed from federal civilian IT systems. So this was published on the Federal Registry yesterday and effective immediately. No more implementation of DeepSeek, removal of existing DeepSeek products, removal of third-party products that were developed using DeepSeek implementations, and those agencies have 60 days during which to request an exception if they need to. But, of course, this is just reaction to China, Chinese law, and the law that allows the government to access data that could possibly be transferred in between as a result of Deep Seek, or at least that's their fear. This is also a little chess game, a little arm wrestling. Yes. If you go down to line 140, there's a fairly new story here. that now China's free AI models are thrown into uncertainty, says the New York Times.

The draft rules for the export of artificial intelligence technology show how China is trying to balance now security with Zyreda and U.S. Okay, so they're threatening again, kind of, having rules about, oh, we have the best models and they're open, but you can't have them. So they keep on going back and forth between we're going to undercut everybody because we're going to give you free models to we're going to cut you off And you're stuck with your models like you can't get a Chinese car either, which, by the way, I very much want one. And so I don't think it's set how the Chinese are going to act in this market. I think they're trying to keep us on our toes. Yeah. Yeah, very true. Let's see here. Oh, this is interesting. Some of the models that we were talking about earlier in the Google TV modification were built with DeepSeek V4.1 Flash, which came out on October 17th. So it's all tied into itself. So there you go. It only names DeepSeek, by the way.

There is no mention of Kimmy or Quinn and everything. Who knows if that comes around as well, too, if this is like the beginning, the foot through the door. Yeah, and I don't know whether this is actual national security concerns or whether this is just... Right. Super rattling. Yep, yep, not super rattling. Kind of hard to not see it as a little bit of that. Mm-hmm. Mm-hmm. But who knows? All right, going to take a break. Then we come back. We've got a couple of stories in the speed round. And, yeah, we'll be talking once again about Metamuse. I want to say Metamuse-hole, but probably... By the way, as an old guy, I like Metamuse-hole. It's very good. I'm very Metamuse-l-ish. We're going to take a break. We'll come back and talk about Metamuse again in a moment. This episode of the AI Insight Podcast is brought to you by Outshift by Cisco. That's Cisco's incubation engine for frontier technology. Multi-agent AI is on every enterprise roadmap right now, and for good reason.

But here's what most people aren't talking about yet. Agents can pass messages, they can coordinate tasks, but they can't actually think together. And when they don't, you get what you might call cognitive failures, misreadings, unverified claims, false consensus, the kind of quiet underperformance that conventional monitoring doesn't even catch. That's the problem Outshift by Cisco is building the fix for, and they're calling it the Internet of Cognition. It's an open source foundation for building multi-agent systems from design all the way to production with shared context, shared memory, guardrails to drive the results you actually expect from AI. If you're building in this space, this is worth your time. Read the paper. Experience the demo. Grab the code at outshift.com. That's O-U-T-S-H-I-F-T dot com. We thank Cisco for their support of the AI Inside Podcast. The Wired Newsroom is known for award-winning reporting on how technology shapes our world.

On Wired's Uncanny Valley, we take that curiosity even further. Each week, journalists from Wired break down the biggest stories in tech while speaking directly with the people building, challenging, and reshaping the future. Is the AI boom sustainable? How do you protect your privacy in an age of constant surveillance? Uncanny Valley tackles the questions driving today's tech debates and lighting up your group chats. Listen to new episodes every Thursday, wherever you get your podcasts. Yes, so Meta's Muse has a privacy problem. Security firm Cleric found around 6,800 shared Muse conversation links discoverable through public DNS records. Each of those shared chat got a web address named after the conversation title. those addresses ended up in passive DNS databases. And so, yeah, I mean, this isn't the first time that we've heard of these public sharing links

inside of your conversations inside of a chatbot somehow surfacing in other ways. I think this is a little different than, you know, some of the other examples where they showed up in, I think in search results, like you did a share of this thing, and then for whatever reason it ended up inside of a Google search results because now suddenly it's publicly exposed essentially. But I think this is different because it's on DNS and everything. But, yeah. Yeah, and as Meta's trying to go, well, Meta never has had a culture of privacy. It was the place where you go share this. That's true. I wrote a book about that called Public Parts, right? You go there to share stuff, and that's been their whole ethos. And Zuckerberg has said that. And that's going to be tough when it comes to trying to serve enterprises that have very good reason to keep things secret. And if they don't have bulletproof systems for keeping things from getting out, then they're never going to win in business.

They're going to piss people off. Yeah. Even so, even with these concerns, I just added another story in here, Meta's Muse AI agent hits 10.9 million downloads, and chat GPT slides by 25%. Wow, so seemingly having an impact there. Yeah, I think the easier you make these things, and the more personally relevant, I think, the better you're going to be. Now, Muse is also the current curiosity, so that can change. They can rebalance again. We've seen that in the AI world. But I think there is a waiting rush of consumer activity around AI if it's made easy enough. Yeah. Yeah, lowering the barrier for understanding, for getting actual useful action out of it without having to low kind of the inner workings or the tactics or whatever. I think that's a part of it, right? The Metamuse agent is less about being smart about how you tactically approach using it.

And, again, it's a proactive thing. So it's meant to just, like, here's the access to everything. You tell me what I need to do. And it just kind of does that, you know. That's a low barrier for people versus something like what I've been building in Claude for however many months, which is nice and it's useful but it's taken a lot of my own kind of molding the clay which a lot of people don't have the time to do that or the interest they just want it to work they want it to do the thing now I can't stop thinking about Metamucil I know I never made that connection it really clears out the pipes it keeps things going or clogs them up with your data Yeah. Logs them up with your data. Disney apparently wanted Character AI to go bankrupt before they would pay into it. Character AI filed, of course, for Chapter 11. A week ago, I think we talked about this,

Disney is the $20 million lead bidder for the license to the tech, not necessarily the company. And hiring people. Not the app. Yeah, and hiring people. Filing says Disney required that the transaction be subject to a Chapter 11 process. Which is a really interesting way to get around the responsibility. When you buy a company, you assume its liabilities, including both debt and legal liabilities. So if I were a litigant against character, I'd take Disney to court. This is bad faith. you killed the company, you acted as if you could take out just the shell of the software and you could hire everybody and you're acting like you didn't buy the company. BS, Vicky. Yeah. Yeah, I bought it. Not a good look. Not a good look. Nope, I don't think so.

And finally to round things out, Reflection Beam, which we talked about a few weeks ago, talking about U.S. open models. Well, now we have the weights. A couple weeks after the release anyways, yeah, they've released the weights. So there you go. You can get rolling with Beam if that is your cup of tea. Maybe you don't need two stacked Mac Studios to do that. I'm not really quite sure. I can't remember how bulky of a machine you'd need to run this. 501 billion parameter model. Yeah. Not sure. But if you're interested, get on it. Can I mention two other stories real quick? Yeah, for sure. So Anthropic, according to Politico, pitched its own AI safety bill to lawmakers. And that's down at line 97, Jason. And the next line is a pull quote from the political story.

The bill would let government block a model's release or use only if it poses a foreseeable imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people. The risk must stem from the creation or release of a biological weapon, the model evading its developer controls, or a cyber attack. It mirrors language found in California's AI laws. This is the problem with regulatory capture. As many folks said, this is tantamount to Anthropics saying it has a license to kill 49 people. By the way, if you scroll down, I used Nano Banana to make a little James Bond license to kill 007 Anthropic movie poster. Nice. With Dario pointing the gun. Nano Banana is pretty great. That took two minutes. So I think this is, when we talk about regulatory capture, this is the prime example.

They're trying to write their own law to regulate them, and it takes certain responsibilities away from them, and then it makes other things so difficult that competitors have trouble entering in the market against them. And this is why we don't want them doing it. This is a problem with lobbyists doing this stuff. and how you put that really has an impression on me basically they're saying to themselves it's okay if we kill 49 people 50 people is a lie now think about this there was a very long 2000 word post from Dario that I didn't have time to read before we got on which I think he addresses this and recognizes it wasn't the best way to get this out but Jesus think people think they constantly act as if it's the virtuous company it's clearly clearly not the case. How wonderful we are. Breaking his arm, patting his back. The other thing to mention real quickly, if we can go to line 120, my friend Bill Gross is the creator of Idealab, which is an amazing company. I should have given you a guess like, oh good, you got it.

You can see it. So Bill Gross is Idealab. He has started more than 150 companies. And Bill called me a few weeks ago and he said, you know, it sounds like You wouldn't think that mathematicians would be on the firing line of AI. You wouldn't think that an entrepreneur like Bill would be. But Bill's method of working has always been that he recognizes the problem, and then he comes up with an idea of how he can solve it. Then he hires engineers to implement it. He gives that entity angel funding, $50,000, $100,000, you know, minor funding. And then if it seems to get traction, then he can invest more, but they find other investors and the company moves on, and then at some point it goes to an exit, and that's how Idealab works. But Bill said, oh hell, AI can find these things. So he realized that news articles, as well as academic papers, are problem statements. So he took tons and tons of them and fed them into one of the models, I think it was Claude,

and it came up with a couple million problem statements. And then the next stage was he had to come in and come up with business ideas. So if you go to Idealab.ai, these are 1.2 million business ideas, free. This is very, very Bill. You come in here, you see a business, you want to build it, no strings. Go build it. God bless. If you would like to build it with Bill, you have the opportunity to go to him and say, this is how I would implement this, and I think it's a really great idea. And if Bill wants to invest, he might invest. He might even get seed funding out of Idealab. And so this is how Bill loves entrepreneurship. He loves starting things. And this is a way that he can turbocharge starting things. God, no kidding. Yeah. That's impressive. And maybe one of these 1.2 million ideas, I mean, a lot of them, you know, I've gone through this at various times. I went through it when there were 6 million, and some of them didn't make sense.

But in there, there are kernels of good ideas. And maybe the idea isn't the one you would do. Maybe it inspires you to do something else. This is Bill's generosity. By the way, the number of business ideas keeps tracking up here. So it's very Bill and very interesting. And so I would recommend going to Idealab.ai and see what it inspires. He also, interestingly, do I have this in here somewhere? Yeah, if you go down the periodic table of business functions, If you scroll down that ideolab.ai page. Oh, it was that one. It was lower, yeah. So there he abstracted the functions that a business needs. So website build, analytics setup, strategic planning, software development, payroll and benefits, accounts payable. And he then looked for apps and companies that already do these things,

but he also looked for areas where they're underserved, and these are also business ideas. You could serve this ecosystem here in this way. So this is the potential of what AI could foster. And, you know, on his behalf, I went to one of the universities I work with, and they're interested in talking. And so, you know, a student could come along, see an idea, work on it in school, come out with a company and funding and get going. And so we're worried about jobs, as we should be, poor unemployed mathematicians. But there can be a lot of value in economic creation here. Yeah. That's a super cool idea. Yeah. There's a lot going on there. That's pretty awesome. And very interactive, too. Really neat. Yes, I did see the asterisk on that story, and I meant to move it up, and I got distracted this morning, so I'm happy that you mentioned that

because I was starting to read through it, and I was like, okay. It's a weird idea that requires some explanation. Yeah, yeah, totally. No, but that's super helpful. That was really cool. So it's idealab.ai for people who want to go. By the way, I misspoke. His count of businesses he has started is now up to 175. That he has started, 175. Wow. Joe Yu he's an amazing so he he started something that didn't work Pets.com but he's learned a lot about the timing of that he also created what was basically the Google ad model and sold that to Yahoo and Yahoo sued Google and won a lot of money as a result so he's a fascinating character Bill very very cool yeah thank you for showing that off Thank you for bringing your perspective each and every week. Jeff Jarvis dot com. Jeff is awesome. Writes lots of wonderful books.

Working on a new one. Has some podcasts on his page so you can check out. Did you get your third podcast? No, I got to do that. Got to get that. Got intelligent machines and AI inside and everything else can be found at Jeff Jarvis dot com. The new and improved Jeff Jarvis dot com website. Excellent stuff. Pot tune up dot com. if you want someone to talk through your podcast launches, or maybe it's already launched and you have things coming up that you want to work through. I'm here for you. PodTuneUp.com, I'd be happy to talk to you about your podcast and see if there's some there there to explore. AIinside.show for everything you need to know about this show, all the links to the episodes, the store, Patreon link is there, which, speaking of, patreon.com slash AI Inside Show is the place to be if you love this show and you want to support on a deeper level,

including our executive producers, Dr. Dew, Jeffrey Maricini, Radio Asheville 103.7, Dante St. James, Bono Dirich, Jason Neifert, Jason Brady, Anthony Downs, Mark Starcher, Carson Samachki, and Evan Muzio. I said that without – I think I got it. I think I'm locked, Evan. By George – I think he's got it. by George. I think he's got it. It only took him 10 times. Sometimes it takes me 10 times, but I get there eventually. Thank you, everybody, for your support on Patreon. Patreon.com slash AI Inside Show. Thank you to the live folks for tuning in. Also to Amos and Daniel Croft, who help behind the scenes for every episode of AI Inside. And, yeah, I think that's my list of thank yous. Jeff, have a wonderful rest of your week. Everyone hang tight, and we'll see you next Wednesday on another episode of AI Inside. Bye, everybody.

The Wired Newsroom is known for award-winning reporting on how technology shapes our world. On Wired's Uncanny Valley, we take that curiosity even further. Each week, journalists from Wired break down the biggest stories in tech while speaking directly with the people building, challenging, and reshaping the future. Is the AI boom sustainable? How do you protect your privacy in an age of constant surveillance? Uncanny Valley tackles the questions driving today's tech debates and lighting up your group chats. Listen to new episodes every Thursday, wherever you get your podcasts.