IM 874: Google Knows I Love the Pepper Cannon - AI and the New Social Contract

It's time for Intelligent Machines. Jeff's here, Paris is back. Our guest this week, Jeffrey Cannell. He is the founder of Noose Research. They've created a new agent I live with. I am crazy about it. We'll talk about Kermes and a lot of other things, including the new Fable model, what Apple's proposing with Siri. It's going to be a big Intelligent Machines. We even talk about Paris' article about food safety and the Pepper Cannon, all coming up next on Intelligent Machines. Podcasts you love. From people you trust. This is TWiT. This is Intelligent Machines with Paris Martineau and Jeff Jarvis. Episode 874, recorded Wednesday, June 10th, 2026. Google Knows I Love the Pepper Cannon. It's time for Intelligent Machines, the show we cover the latest in robotics, AI and the smart doodads all around us. We are so happy to have Paris Martineau back. She has emerged from the trenches of the big expose. She's been writing at Consumer Reports, which came out yesterday. It's true. Are you relaxed now? It's out in the world. I'm relaxed. I'm sleeping a normal amount of hours a night. I sat in the sun yesterday. It's delightful. Did you have some hostess donuts? We'll talk about this. I did not have any hostess donuts. We can talk about why after we have a wonderful interview with our guest. Right. We'll save that. And yes, and I actually made one of the picks of the week be something that you worked with for this. So we'll talk about that in just a little bit. Yeah. Paris, of course, Consumer Reports investigative reporter in food safety. Always great to have you here. Jeff is here as well. Jeff Jarvis, Journalistic Professor. Let's see. Let me get this right. Emeritus Professor of Journalistic Innovation at the Craig Newmark Graduate School of Journalism. Craig Newmark. The City University of New York. He is also the author of Hot Type, which comes out in a couple of months. Yes. Yay. But you can pre-order now at Jeff Jarvis. August 20, and I just finished the audio book last week. Yay. So you can order that too. Congratulations. I am thrilled to have a returning guest this week, Jeffrey Connell. He is the founder and former CEO of Noose Research, now CTO. He's kind of stepped back into the research role. Really wonderful to talk to you again, Jeffrey. Love talking to you then, but you were cagey. You did not tell us about a little something that you had in the lab. Yes. When we had you on, we were talking about your models. You have some really interesting models like Psyche. But it turns out you also had an agent running in the labs. And the story I read, which I think came from you, is that you saw OpenClaw and how it took off in January. And you said, you know, our thing is actually better. So in February, you released something called Hermes. By March, I was using it full time. I am madly in love with Hermes. It is exactly what I want. He has been obnoxiously in love. He has been elegiacly in love. It is exactly what I want with an agent. It's robust. It's powerful. It's easy to use. I'm running it through a third party web UI, which I really like. Just this week, you released that dedicated application for it for Windows Mac. And is there a Linux version too? It runs a Linux too. Yeah. But I happen to like the web UI. That's one of the things that's great about Noose Research. You're fairly agnostic about those kinds of things. There's a command line and a TUI as well. You also are very, and the reason I like it, very agnostic about models. Because right now, for instance, I'm running this on my, I have a central server. It's a framework desktop. I think we talked about that last time. And I'm running Quen3635B on it, a local, fully local model as my agentic model. Hermes is smart enough to delegate harder jobs to others. But Quen writes quite well. So this is fully local at this point. And I'm blown away. So much so that my wife, who saw me playing with Hermes and got jealous, said, can I have a profile too? So one of the great things. Now that you've divorced Claude. Yes. And your marriage is safe. Mine is called Quicksilver after Hermes, right? But hers is called Rosie. And it, so it has all the skills. We share the skills, but her own memory. Does Rosie, your cat, know that her name has been co-opted and perverted into an AI? No, but I don't think she really cares, frankly. The good thing is it's also Rosie the robot from the Jetsons. So it's appropriate in two ways. One of the things I like about Hermes, it's really a battery included agent. It has more than 90 skills come with it. It's very powerful. You have a choice of memory models, including, I turned on hindsight, but there's Honcho. There's a bunch of different choices. It's very easy to enable. And it's, it's just super wonderful in every respect. So I just wanted to start by saying thank you for releasing it. It's funny that you had it all that time. When did you start using it? We started probably sometime in December. And I mean, I've told the story a few other places, but I'll give the quick recap, which is just that we were looking for ways to, you know, supercharge our model development. You know, we are not as well and hugely funded as a company, companies like Anthropic and OpenAI. So we always have to look for these like thousand X increases, you know, that are going to allow us to compete with the, with the bigger boys. And so we were like, let's try to see if we can get an RSI loop going. Well, RSI being recursive self-improvement. Right? So we said, well, to have that, we need to have some sort of scaffolding harness that can learn as it goes. And we built Hermes Agent internally for our post training team to work on model stuff. And it kind of was one of those things where I guess maybe when you swim in it all the time, like we've always known AIs can do stuff like this and have used it like this. So it's kind of like a water is, you know, you fish don't know they're in the water kind of situation. And when OpenClock came out and it suddenly was having, you know, massive penetration into like everyday people's lives use cases, we really did. We're just like, hey, we have this thing that we think is just as good. And we didn't really position it as like a competitor or put it out there because we even ourselves weren't quite sure what the reception would be. So we put it out there and, you know, immediately it just, you know, was like PMF on like anything we'd ever seen before. Like people were just coming in and loving it. And so we, you know, stay close to the community. We listen to what people wanted every day. We use sort of the asymmetric power of these models now to be able to scale up the development because we can now, you know, theoretically hire like a thousand, you know, engineers all at once to work on it if we're willing to pay for it. And, you know, we've been just ship, ship, ship, keep an eye towards the users, keep an eye towards, like you said, being agnostic about use cases, really allowing people to develop things and let it mold it to themselves. Because once you mold it to yourself, you really come to love it, right? Like in a sense that it is, you know, you built it up to what it is and now it's extremely useful for you. It's so personalized to me now because of all the memory over the months. 190,000 stars on GitHub. I mean, I think you're underselling the success. It is now more stars than VS Code. It is an incredible success. And I think people who were using OpenClaw look at it and go, this is much more stable. It doesn't have the security issues that OpenClaw has because there's no Hermes marketplace, right? There's no third party marketplace for skills because Hermes writes its own skills. Yep. So all I do is- That was really like the self-improvement, I would say, like unlock that we had was like looking at the dynamic skill creation and the curated local memory was kind of like the unlock that we innovated on that turned it into what it is now. And sort of the first time you see it do this thing where you're using it, you're solving a hard problem, and then it creates a skill automatically from that. And then the next time you try to do something in there, it just instantly happens because it knows the best way to do it. And it knows your way to do it, you know, like whatever your flow happens to be, we're going to, it'll build up dynamically over time. We like to say Hermes agent, it gets better the more you use it, right? And so if that's the case, the more you use it, it gets better. People have, yeah, people have loved it. So yeah, like you said, 190,000 stars. We released our own desktop app last week after we demoed it on stage. Jensen did it. Oh yeah, let's not forget Jensen Wong giving you a big plug at Computex. Yeah, that was really cool. You were there, right? Yes, I was there, I was there. I met him later and talked to him, he's a great guy. And so we released the desktop app. We had been working with NVIDIA and Microsoft on the RTX Spark, which was the laptop that they announced. So we had been working on that. And so we released the desktop. I think we got, we've had about like 1.5 million downloads of that. So it's been a huge success. And we're really trying to like, you know, the goal of Noose always was to bring what this transformational technology to as many people as possible in an open way. You know, and the question is- Open is key, isn't it? Yeah, yeah. And which way that was going to go, I think we always were feeling from what it was. And now that we've sort of found that place that's good, we're having a, just having a blast, you know, take it going along for the ride. It just shows that good guys can win. One of the reasons I broke up with Anthropics is I was mad at their policy that you had, you couldn't use your subscription with anything but Cloud Code. One of the things that Hermes does that's pretty cool is if I need a challenging programming job and I want to use code, it will automatically launch Cloud Code or OpenAI's Codex. Do the programming in there. I don't even see the command line. Put it back in Hermes. So it's the best of both worlds. It uses the OpenAI API. So every model except Cloud is fine. You could use Cloud API tokens. And if you're going to use Fable, you're going to have to in a couple of weeks. We're certainly going to talk about that. Yeah. And you better go looking under the pillows for some dollars. Holy cow. Wow. $10 in and $30 out is a lot of money. Twice as expensive. Deep Seek. Deep Seek. Well, and this is why I'm using Quen now and I'm actually really impressed. But Noose has its own, and I do have a Noose subscription. You have a $20 a month subscription. You also have Macs and Pro subscriptions, which give you access to, I think like it's 120 models, some huge number of models. Yep. We've partnered with several different like inference providers to bundle it all together. What's nice about the Noose subscription as well is it also includes all the tools. So you've got web search, image search, text to speech. All of that sort of stuff comes with it as well. And so it's not just, it's sort of the external tools that you otherwise would have had to go get API keys for. We sort of bundle the best in class versions of all of those together and they're served alongside with the new subscription. So that's one thing that's useful about it too. It's kind of like the open router model, right? Are you routing it yourself, orchestrating it yourself? Yes. Yeah. So we do use open router in behind the scenes for some of the models. So we have like our own open router layer essentially that, you know, routes it to different people. For different, like when we did a deal with Kimmy or some of the other ones, they give us special access to their inference platform. So in behind the scenes, we have sort of our own open router style model routing thing as well. Yeah. Yeah. I, as you can see, I have quite a big balance I've built up. I appreciate it. I don't know. I, you know, but it's great because it allows me to test and try a huge variety of models. One of the things I have Hermes do every week is go out and look at all the models and pick them for the delegation so that it can, for, you know, images, it can choose the best image model, which I think probably still nano banana for coding. It could use the best. GPT image two is pretty good. GPT image two is pretty good. Yeah. Yeah. There's, there's some really good stuff out there. I've been using a GPT five, five, but, but, but I started a couple of days ago to use the local model and I'm actually pretty surprised. It turns out, I think a lot of people are reeling realizing this. We've all been focused on models and model strength, but the harness, the stuff around it may be equally important. And if you've built something really good with something like Hermes, you don't necessarily need the smartest model to do 90% of what you want to do. Yeah. And, and a lot of it, I think, you know, there's a world where even if like we froze all our models today, we could like still squeeze a lot more, you know, juice, juice out of them. The harness thing is very interesting to me because like this, this unlock happened because of sort of this invisible wall that got, that got passed with the models. Like it kind of started around when Opus 4.5 came out. But there was just this period where all of a sudden the models got good enough, good enough at long context to really take in the huge amount of, you know, stuff we put around to, to direct the model and to, you know, use tools effectively. Like it was just kind of this like invisible line that was crossed about six months ago. That's really an emergent property of the models, right? Like, yes, they train, they train in it someone, but like the fact that we were able to like get such useful improvements out of the models without changing them at all. Just one more thing about how actually amazing the actual models are. This is in a way what Apple talked about on Monday and what I think they intend with a Siri because Siri will have memory, it will have all the context from your phone. That in a way that what they're building is an agentic Siri. They're not going to say that out loud, but it, but I think people are going to have this experience of using an AI is so much better when it understands your context, understands you and it knows about you and it has some history with you. But then there's this big issue because if you're using one of the frontier models, if you're using chat GPT or Anthropic, it's going to get a lot of information about you. Yeah. And I think people are very nervous about that. Mm hmm. And, you know, we try to do this in Hermes by having the ability to have sessions where you can segment things out yourself, like, you know, really put you in control. I mean, yes, they're making an agentic Siri. It'll be interesting. You know, Apple's sort of play here is very interesting to me because they sort of sat out the race of AI for the last however many years, you know, when maybe like the default thought would be, oh, if you're, you know, a tech company in SF or like what you got to have your own AI division, you got to be making your own model, you know, and they sort of sat out and have now sort of like surgically decided to strike it. At some, you know, specific. Might have been a good strategy as it turns out. Right. They don't think they missed it. Depends if they succeed, but they definitely saved themselves, you know, $200 billion capex. Yeah, exactly. And they're riding on Google's Gemini investments and so forth. So it's they're not all alone, but they have their own foundation models. Tell me a little bit about your own models, about Noose Research's models. So right now we have our old class of Hermes models, which were sort of the old paradigm where you they weren't agentic. So we're working right now on training new versions of our models that will be agentic right now. So we don't actually have any of our own models that we ship that are for Hermes agent yet, but we're working on it. We actually also recently joined the Nemotron coalition. So which is. That's NVIDIA's model. Yeah, NVIDIA's open source training model. And because, you know, really the question is like if you think about like fully open source models, you know, fully foundation, full LLM from pre training all the way through, you know, the money keeps going up. The money keeps going up. And there is a question about like who is aligned to even like do that in open source anymore. Right. Like when it was the seven B models that we were training, you know, your early llamas, like sort of justify it as a as like a marketing cost or something like that, you know. But what happened with Meta and these other places that basically said we're not going to do it anymore. And functionally, the only companies that are doing open source training in the world anymore are the Chinese companies. Right. I remember you saying last time you were on you were very nervous because of Meta might pull back LLM at any moment. Yep. They still haven't. Well, they haven't. They never. They're not training Meta LLM 5 though. You know, there's no intention in open sourcing anything. Right. Yeah. So at this point, we sort of are in that future. And I think Nvidia is one company that sort of credibly could put money into open source and have it still be aligned with the business. Right. Because, you know, whether more models, more tokens just means more Nvidia chips that are getting bought. Right. So I think there's a way that it really makes sense there. So we're working with Nvidia to like, you know, steer the direction of American open source and try to get foundation models out that are fully open source. I will say Quinn 3.6 36B quite amazing. The fact that they got as much of that into a 36B, it's it was really like I wouldn't have thought it was possible to be honest with you. They really did do a great job with. And that's the thing. He's suggested turning on, oh, and I can't remember the acronyms, a number of features, flash features and so forth to make it faster, more responsive. He's actually helped me tune Quinn. I'm using Lama CPP and it said, use these settings. And it helped me tune it quite a bit. And I was amazed at how much faster and better the time to first token improved. Everything is better. So there is, there is a lot you can do with a local model. And I think it's really important. Privacy is for privacy and other reasons. Cost is another reason. There's going to be, we're seeing this with fable and we'll see it with mythos that these, these high end models. Yes, they're very, very good, but they're very, very expensive. Only billionaires can afford them. And how do you know when you need to use that quote unquote high end? I think that's still an open question really. And that's something that I think ought to be explored more also, even on the harness side, we're working on model routing to dynamically be able to switch between models at different tiers and adaptively do it because, you know, there, you don't need opus tokens or, you know, fable tokens to like move a file around right now. There's all these questions about how do you actually pull that off? Do you invalidate the cache context if you switch between models? So it's a little bit of a hard problem, but certainly, you know, at the prices we're talking about now, only, you know, well-funded businesses can really afford to allow people to use frontier models as their daily driver if they're doing any sort of heavy work. And even they are starting to quail at the token costs. Yes. We're talking to Jeffrey Canel. He, am I saying right, Canel? You don't say yes. Perfect. Jeffrey Canel of Noose Research. He is the, well, it's an interesting story, kind of co-founder. It was a discord server where you had a lot of people very interested in this idea some years ago, a couple of years ago, and formed a company around it. You've got venture funding now, right? Is that right? Yep. And what would you say your primary business now? Is it Hermes or is it Moms? Yeah, absolutely. It's Hermes and figuring out ways that we can, you know, make that be profitable to the company. We have a huge user install base, but we're trying not to be extractive, you know? Yeah, and that's why, and you do, but like, we also offer Open Router and all these other things because we're not going to, A, we're not going to force you into something that's like extractive. We want Noose Portal to be the best place for you to experience Hermes, but we want to allow you to run your local model if you want to run, you know, local models and set it up with Honcho if you want to set it up with Honcho and all those other places. So we're going to, you know, kind of give an offering right now that is sort of our best class version of it, but still offer the, always offer the freedom to configure in your own case and however you want. When I emailed you, I asked you to, I said, if you want to bring Technium along with you, is Technium the lead developer on Hermes? Yes, he's the lead developer on Hermes and it's really, you know, I got, I'll show you. He's a little shy. He didn't want to do this. He's a little shy, yeah. He doesn't like to get on the camera, but here's the thing about Tech. You know, he is not, did not go to school for CS and was not a coder at all. Like, did not really know Python, did not know anything at all. And from like a coding perspective. And he built Hermes Agent, all right? Like, now the number one. Did he vibe code it? Yeah, completely all with, yeah, not vibe coding, but he is now like, he knows how to, I would say what he does is like many levels above vibe coding, you know? I follow him on X and he's very valuable to follow. Highly recommend it. Super smart. He really knows how to articulate what he wants it to be. He has a vision for the product of what he wants it to be. He just didn't have the technical, he didn't go to computer science school, you know, and he just never learned that. So it's really amazing now that we live in this time where like one of the biggest applications on the planet right now, you know, was written by someone who was not a developer, right? Like, and so that's a huge unlock, right? Because it used to be gated behind people who had to go to school in a specific way. Like, it just goes to show like how transformational these tech, this technology really can be and how enabling it can be of people who previously just for some reason or other wouldn't have been able to, you know, to do something. It allows everyone to bring their ideas to life. And if you have a great idea and you're someone who really is exacting on what you want, you can now make it happen when, you know, the circumstances of your life previously may have made it so that you couldn't have. Jeffrey, isn't that the future? That's the power of this, isn't it? Yeah, Jeff, you're going to say the same thing. This is what technology brings us. Mm-hmm. Tech is really good also at listening. And half the time, somebody will tweet at him and he'll say, merged. And that's the thing, it's like to be successful, you just have to be obsessive about the customer. You have to be obsessive about wanting everyone to be happy and stuff. And if you're willing to be obsessive, and he is, he's 16 hours a day, seven days a week, like he loves the product. And, you know, and that's a huge piece of why we've been able to be successful. Here's an example from earlier today. Somebody says, I work on Gemini at Google. I added a few text-to-speech features to Hermes and Technium's response is merged. Which means we get it. We all get it, which is fantastic. This is one of the things I love about Hermes is you don't have to use all the skills. They don't take up context. They're just there on the hard drive. But when you say, and you can say this to Hermes, you know, is there any way for me to scan through stuff on X? And Hermes will walk you through it. And suddenly you have a whole new capability. And that's very powerful. Go ahead, Paris. I'm sorry. Oh, I was going to say, I know originally a focus for new research was the Solana blockchain. That was kind of central. Is that still something that is central to your guys' operations? Well, we're not working on it right now. We've sort of had to shift a lot of our focus. Just we only have so many people. And like when you have that many people, like that many PRs we have, we've kind of had to like jury rig the whole ship over and try to like focus it on Hermes agent. So we're focusing pretty much everything we are right now on Hermes agent. I mean, I think our experiments on Solana were completely still well. There were experiments. We were trying to figure out how can we incentivize model training and make it work. And I think we've just sort of shifted that into this next domain with working with NVIDIA and like trying to find the right answer for how can we still bring true open source models to market. You did mention that it seems to be a capability. I can in the middle of a session at any turn, I can change models. And it seems to be it will pick up the session. It also has a memory of all the sessions. And I can go back to a session and that now becomes part of the context and continue on with the conversation. These little things, little quality of life things like this make this incredibly useful. Much more so than anything else. Hermes agent is now 100% written with Hermes agent, you know? So like that's what we use every day to build it. So all those sorts of, you know, if you have to use it all the day to build this complicated thing, you will pick up all of it. You know, you get, yeah, it's able to do it. Yeah. So we were talking earlier when you joined us and I see behind you, it looks like a picture from the Sistine ceiling. Am I right? That is actually the Transfiguration by Raphael. It's Raphael's final. It's his final painting. It actually was unfinished when he died. And actually people thought it was so good. They like put it next to his casket when he like had their funeral, like the unfinished one. It actually sits behind the Pope at one of the churches in Rome, the original. But that is, that one is actually Raphael. Well, and I know Jeff was on last Sunday when Father Robert was with us, our favorite Jesuit. And we've been talking a lot about the Pope's encyclical. It was all about AI. In fact, Pope Leo took his... There it is. I made a song out of Section 238. And wrote a long post out of it. It's a fascinating document. I think it's a great document. Yeah. Yeah. What do you think of it, Jeffrey? As a practicing Catholic? Yeah. Yeah. I mean, it was really, you know, I was, maybe I was a little bit like, I wasn't nervous, but I was like, what is this going to say? What is this going to say? And, you know, the encyclical is meant to survive the test of time. You know, this is part of the church's official, you know, position in its role as teacher and speaking with the magisterium. And so these sorts of things have to live on, you know, forever. And, you know, so what the Holy Father did was not so much say, here's where AI is right. Here's where AI is wrong. This is what is more. How, what is a framework on how to think about this? Right. It was a framework on how to think about these sorts of questions in the modern age. And it really is even more than just artificial intelligence. I would even call it an encyclical on modernity, whatever you want to call it, you know, in 2026. Right. Like, what is the state of the world in 2026? And there are all these new challenges that, you know, that hypercapitalism and, you know, post work, whatever you want to call it, kind of stuff happening. When, you know, what is the already established teaching of the church? How does it apply in this place? So, for example, the word artificial intelligence actually only appears one time in like the first third of the document. He really goes to great lengths to like bring forward what Catholic social teaching has said over the previous years. You know, how Catholic social teaching was applied during the Industrial Revolution. How is it applied during, you know, these other sort of, you know, changes in human society. And then bring it to the forefront and just really say that the purpose of any technology is to, you know, improve the cause of humanity on earth. Right. And if these tools help make us better humans, they make us more drawn to what our cause is and, you know, enable us to do that more than absolutely all the good, you know. And I think, you know, highlighting that there is both good and there are both things to watch out for. And here's sort of a framework, you know, a preferential treatment for the poor, for example. You know, it's part of Catholic social teaching. We have to make sure that when we're making these sorts of models exactly, you know, we're not creating, it's a joke on Twitter, but like, you know, the permanent AI underclass, right? Of like the people who get access to Fable in the top. I don't think it's a joke. I think it's increasingly becoming obvious. A reality. Yeah. A reality. Yeah. Yeah. And so we just have to like, you know, those are the sorts of things that we need to look at. What is the ultimate destination of all goods? It's, you know, the common cause of all humanity. And so he goes through building it all up. So I was extremely happy with, with how it came out. There are several places in it. There are actually only, as I said, there's several things he says that are about, you know, the ontological status of artificial intelligences. You know, it's the church's position that they aren't alive. They don't, you know, feel love and stuff like that. It's paragraph 99 is like, kind of like the one that really goes into that. That's what I like best too. Yeah. And, but it really goes into say like, here's, you know, some ontological claims, but also goes so far as to say, but the real disposition of these still is the purview of academia and research. And that science is something that can be used to solve these problems. And there's no reason not to use science to investigate and solve these problems. I love that. Don't turn your back on science. What are you hearing from your fellow technologists about it? Did they, did they pay attention to it? Yes, they absolutely did pay attention to it. It was, you know, it was kind of not shocking to me, but like I live in kind of this Catholic bubble, you know, sometimes it feels like. Um, and so for the fact that a lot of people, um, who were, you know, not in that bubble, we're still taking what the church was saying seriously in this, you know, lend credence that, you know, the church still exists, you know, exists in some sort of moral authority throughout the world, you know? Um, and it was happy to see it. And I would say it was extremely well received by everyone else because it wasn't this AI is evil. Can't do it. Don't stop, you know, putting it sort of like that. It was just, Hey, here's how to think about it in a way that, you know, improves human flourishing, uh, across the board. I would be remiss, Jeffrey, if I didn't ask you a little bit about your particular Hermes installation. What, what models are you using these days? Uh, I'm, I mean, I was a big, still am a big Opus 4.8 fan and 4.6, um, only because I think I learned how to write, you know, raggle them pretty, raggle them pretty good. Um, uh, GPT-55 actually to me is like super unwieldy. I know a ton of people love it, but to me, I'm going to be a lot of the I find it very unwieldy. So it's really me. I just have sort of developed a rapport with 4.8. I know how it's going to think. And, uh, so I, I use, I use 4.8 a lot for that, that, um, I mean, I started using Fable yet. You're not using a sub though. You're using API. Uh, so I, so I use news portal, which has 4.8 on there. You can pay on it. Um, you know, and I'm one of the lucky few that gets the, you know, unmetered account that I can set as many tokens as I want. So, uh, have you added a Fable to the portal yet? Yeah. So we, we added Fable yesterday. Um, I actually have only used it a little bit just cause I've kind of been doing a podcast circuit today with everything with Fable coming out. Um, so I haven't been able to form like too, too many, um, feelings about it. What I will say about it is that, um, uh, you know, what, what Anthropic has said about, uh, the LLM research and open source research in the spaces, you know, greatly concerns me as someone who cares a lot about that. Um, and so I'm sort of at my own kind of like existential, do I really want to keep supporting this kind of company? So, um, with, with the release of, you know, first of all, they did this whole security theater thing, which, you know, every model company does. If you remember they, after GPT-2, GPT, you know, open AI said, we have to keep the waits for GPT-3 closed. It's just too dangerous. Right. Um, you know, which is now would be. That was Dario working at open AI who said that. So yeah. Yeah. Yeah. So it's sort of playbook. Yeah. Yeah. It's so, so playbook. It's like, not, you know, it's so dangerous. We can't release it except of course we can, you know, we're the ones who'll get to do it. And then eventually you release it, you know, whatever, blah, blah, blah. So they did this whole security theater thing with Mythos and I, no doubt that it really is a, you know, world-class, it's the best AI on the planet right now. Um, and then with, with, with Fable, they finally said, well, we're going to be able to release it because we've included all these extra guardrails on there. Um, including two sets of things that it does. So it used to be, there always were guardrails on Claude. Um, and if you hit it, what would happen is it would just, well, there was, there was two levels. There was the AI would say, I'm sorry, I can't help you with it. If it says that, then it was actually like from the training data, like the AI emergently refusing to do it because it was trained not to say certain things. They had a second layer, which is this classifier layer, which is actually looking, which is this different model that's sort of sitting there babysitting all the tokens and is making a judgment call of like, is this thing, is, you know, this getting out of whack, you know, is this not doing what we want? And if you hit the classifier, what happens is the response just stops. Like you just get an empty response from, like you just, the API just returns like nothing, basically, um, no, no refusal. It just returns nothing. Um, so this was how the safety stuff worked before the previous Opus models. The Fable, they've changed it now to where it will, it has a new detection mechanism where it will actually downgrade you out of Fable back to Opus 4.8. If it detects you're doing, I believe the categories were like chemical research, um, cybersecurity research, biology, some of these like, you know, high leverage sort of things where you can imagine there could be a bad application of it, but there's also a million unknown good applications for it, I'll say. Um, and it'll kick you back to Opus 4.8. But the more insidious thing that they've done, which is they have, if you are doing research on AI, particularly LLM research on AI, they won't actually kick you back to Opus 4.8. They have new mechanisms that will silently degrade the quality of the responses and like, literally like lie to you, like not tell you what it knows. And like, they literally like inject a dumb vector into the AI training at, at, at runtime to like dumb down the model to keep you from being able to do frontier level AI research. Like literally- Is that to keep people from doing distillation attacks? They've complained about China, for instance. I mean, you know, you can, you can make any reasoning out of your own about why they, they might've done it. You know, I could, it could be distillation. Absolutely. Could it be, I mean, they don't want distillation cause they don't want anyone competing with them. Right? So like at the end of the day, the real argument is competition, but this is like a whole new Pandora's box that hadn't been opened before where like the, we are going to actively sabotage the token stream to keep you from getting access to something it already could have done. And that we are doing internally as well, right? Like they're using it to do a frontier level research, but we're going to actively sabotage it. I don't know if you guys are familiar with, I'm sure you are. Um, uh, uh, the three body problem, you know, the SOFANS that literally, this is literally like the plot of the three body problem where they send these things in to like sabotage scientific research to like keep the, keep humans, you know, from, you know, advancing too quickly, uh, until their ships can get here. So, uh, it's pretty much the same thing. And it's something that I think all of us in the open source community really feel like somewhat of a red line has, has been like crossed here because they're now saying that like, they've always sort of said they didn't want open source AI to exist, but like they're now literally using their position in the market to like go and sabotage open source AI, like literally sabotage it versus even just classified refusal. It's not even just refusing. It's like, we're going to actively sabotage it. And that just seems to me like, like, what are we, what are we doing here? People? I mean, as much as you want to say it, oh, the anthropic was, they've done a ton of work, but they also exist because people open source transformers, people open source to open, you know, there's a huge amount of open science that started all of this and to sort of pull the ladder up, you know, the second you guys got to some level just seems incredibly, if I may quote Elon, misanthropic. Speaking of pulling the ladder up, Jeffrey, what about their recent, um, push to say, okay, once we, once we've finished our model and once we've gone to one IPO now, all development should stop that. Do we want to pause? Yes. What did you think about that? Um, I can't imagine that would fly because if she open AI eclipses fable with GPT five, six or something, I think they would actually be like in violation of their fiduciary duties to the shareholders to like, you know, like, yeah, like you could do it as a private company actually, but it's like a public company. It's a little bit of a different like question because like you're liable now to do what's in the best interest of the common shareholders and the common shareholders are like people who want their value of their shares to go up. And there isn't really a planet where they pause their AI and other people continue and like their company gets more valuable. So I mean, I don't know about how this work on the public market, but famously, uh, something we hear a lot from people who all like hold shares of anthropic employees. They note that there's a clause in their contracts that basically says we reserve the right to totally tank the value of all of your shares based on these specific principles. And I wonder if that sort of mindset is even possible to fly in the public markets. Yeah, that's a good question. So, uh, of course, some of this is, is powered by the fact that anthropic and open AI and SpaceX for that matter, all put pursuing a, uh, uh, IPOs this year. Um, now admittedly, Nvidia has its own, uh, financial interests, but, uh, it sounds like you believe that they support open models because they're going to make money on the hardware. Anyway, it's about CUDA and it's about their GPUs more than it is. As long as at the end of the day, it has to run on an Nvidia chip, you know, that is better. They're okay with it. That's kind of why I think like they're the only company like that could marshal enough resources to actually pull off like some of these foundational, uh, open source models where it would still be in line with their, where it's still in line with. It's still make money. Yeah. Yeah. Yeah. What I don't like about it is it, uh, CUDA is proprietary and it means Apple's MLX technology. And it means my, the rock technology on my AMD processors are not compatible. Um, to me, open means it should work on a variety of hardware that shouldn't be hardware specific. Uh, but, uh, but these are such expensive models to build that I understand there's, you've, you've got to consider the cost of building it and you've got to get to the billions of dollars. Like you gotta, yeah. Yeah. So this is a tough challenge. I mean, you're pursuing open models, right? Yes. Yes, absolutely. So how do you make that work financially? So we're going to be relying on the Nemo Tron coalition's, uh, access to the GPUs to do it. Right. So, uh, Jensen has sworn as not sworn, but you know, he's, he's told, you know, the coalition that over the next two years, the NEMA, uh, NVIDIA is going to commit 15 to $25 billion, um, for, for this, for this purpose. Right. Um, so that's, that's quite a lot, you know, and that gets you to the frontier level of like training resources. So it doesn't mean you can run it locally is the issue. I mean, I'm right now an Emotron three 120 B is free. I can run it through a European router connection, uh, and, and run it three. And I presume it's a big model. It's a, it's a powerful model, but I have to run it on the cloud. I can't run it locally. Yeah. They just released a NEMA Tron, uh, ultra, which is their 550 B. Um, so there's a, there's three, there's nano, super and ultra. I mean, yes, unfortunately. Uh, I mean, I think you've seen that there are, there's a world where small models can still do BJD drivers for local and source. Deep seek's very good. Yep. And Quinn's very good. One way you can actually do this is through something called on policy distillation, where you train the giant huge model and then you like can like suck it down and compress it. You, you have, you have a new, a small model whose only job is to like learn the outputs of the other model. And it just tries to mimic the outputs of the other model. And it like can by as most, it's like suck a lot of that information in. So I think there's a way where that always there. And with things like RTX spark and like the DJX spark, I think we could see that like the, they have like a commitment at least to like make that be a local inference, be like a viable, like a viable path. Now, as the models get bigger, you know, people who knows how big, you know, mythos is. It could be the 10 trillion. It could be as high as 10 trillion parameters. We don't know. They don't tell us. Yeah. They don't tell us. But like that scale is only going to continue to go up, right? It's got to be. 10 trillion. Two years from now, we're going to be talking about. So like NVIDIA's previous chips, you know, the current Blackwells were designed to run trillion parameter models, right? The next Vera Rubens are going to be for 10 trillion parameter class models. And Feynman will be, you know, for 100 trillion parameter class models, probably. What did you think of Jensen Wong when you met him? Oh, he was a great guy. He was just, he's a very down to earth, very joking with you. Like, you know, very not like this. Oh, look at me. You know, the CEO guy, very, very down to earth guy. Yeah. I didn't realize I was doing some research that you were, you wrote the paper on yarn. That was your, the capability to kind of stretch context. It was one of the things, you know, this is a perfect example also where like that, that ability to go from 4K to 128K was in many ways like a precursor to the agentic era that we have now, right? Like you can't do any of this if you're stuck at the old, like original transformer sort of limitations, right? Yeah. Yeah. And that was done in the spirit of open science, right? Like we released it and open AI immediately started using it. Like everyone immediately started using it. And that was, you know, kind of in this collaborative open research thing. And so I would like to see things like that continue to still, to still happen. Do you agree with the model that adding compute, adding these giant parameters will make it smarter that, that is this kind of an endless growth? Did the, did you believe in the bitter lesson? Are we, are we just pouring compute? It hasn't stopped. Everyone keeps thinking it's going to stop and it hasn't stopped. And, you know, eventually you're going to start getting to, you know, parameter counts that are like somewhat in the order of like, if you were to make a rubber guess of like functionally the amount of neurons that are like inside of a human brain and stuff like that, which is somewhere around one and a half, 150 to like 200 trillion. Now there's a question whether these neurons like directly map functionally to like the amount of parameter, like is, is a neuron doing just like a one and a, we know they are, they do more than just like what a single parameter does. So it may not be like an exact comparison. They're also massively parallel in a way that, you know, Van Neumann machines, Van Neumann machines. Yeah. So, so there's, there's an open question, but like it, it keeps going. Now, having said that, I think there will always be the scale competition, which is, can you scale it up? And then can you run an equivalent thing smaller for less energy, right? Because while AIs are incredible in what they can do, they are like approximately a billion times less energy efficient than a human brain, right? Like your brain runs on like charitably 30 Watts of power. You know, and it is, is like a, so it's a 200 trillion parameter model running on 30 Watts of power, right? Right now we need, you know, multiple kilowatts to run these, you know, to run one single instance of these trillion parameter models. So there's definitely like a literal two order of magnitude amount of energy efficiency that the AIs don't have. So that's another area to compete on. Jeff, can I ask him the question? Go ahead. You know what question I'm going to ask? No. So what do you think? Are these models conscious? I do not think they are conscious because I do not think that they have the experience that we have as humans. And that's what we really mean when we say conscious, we say, are you experiencing the world? Like I experienced it, right? Yep. I mean, it gets down to the question of how do you know anyone other than yourself is conscious. Well, you just say, it looks like me. It talks like me enough. But also that it has the, it has a common framework of understanding the world. I believe that it would grew up. I believe that it most, you know, this person most likely had a mother and father, you know, or at the people around us in the United States at the same cultural priors about that. It felt a certain way, was probably made fun of at once, was sad, you know, was hungry, was thirsty. All of these cultural, you know, experiential priors, I guess, at that point that we assume we bring together. You know, is there something that you can narrowly say there's a self-reflective mechanism inside of AIs? Well, well, yeah, there's a self-reflective mechanism in AIs, but you could argue that about like a for loop or something if you wanted to get like too, too close to it. Something that can like, you know, so like, so really it's like the experiential priors. We say, when we say consciousness, the reason we don't have a good definition for consciousness, because what most people really mean is, is it like me in the way that I think I am? And for whatever you can say about the outputs of models, they just scientifically did not undergo growth in the way that you and I went around. They don't even experience time the way we experience time, right? Like the common thing about all of us, which is that we're moving forward one second per second in this causal world where if we make a mistake, there's no rewinding time, there's no going back. Our choices are ours and ours alone and can never be undone by, you know, the relentless law of thermodynamics pushing us forward. That is just not the experiential world that an LLM even is in, even if you want it to a claim, it has some sort of self-reflective mechanism. So to me, I think consciousness is this whole bubble of like us. And I just think quantitatively, even it's not like us. So the answer is no. All right. Here's the other question. Go ahead, Paris. Leo, what did you mean by that question? I don't know. I just thought I'd ask. I don't know. Are you conscious? I don't know. I think I'm conscious, but that's, you know, that's me saying it. So I'm not happy. Sounds like what an LLM would say, right? That's what an AI would say. Yeah. The ancillary question, is AGI the goal? And the ancillary question is, what do you mean by AGI? Well, I think that AGI, we are in, AGI is here. It's just unevenly distributed, you know? What do you mean by AGI? I mean by AGI as in as good as humans at economically valuable tasks. At least that would be a narrower definition, right? And so I think it's like a functionalist definition. Would you ever give money to a computer to do something that you could have given a human to? And are you economically rational to make that judgment to have the computer do it, right? So obviously in places like coding right now, I would say the answer is we have AGI in coding. The latest coding models are better than the best programmers, essentially. Now there are niches where people are, you know, have it. And often writing the code itself isn't like everything that it takes to bring the outcome to market, right? Like what you really want when you write code is some other set of outcomes. You want a product that people use that has aesthetic matching to people's experience. Like there's some other hidden set of motivations, but it's sort of like the leak code style, like programming, like in a vacuum. Like AGI is here for programming and it's just can't even argue with it. And you only have to look at the best programmers who will tell you this too. And in other, you know, quantitative domains like math research, we're starting to see AGI being here. I know you probably saw recently about the discovery of a solution to the unit distance problem, that GBD 5.5. You know, this is kind of, I would say, I don't know if you guys remember growing up hearing about something called like the four color theorem, which was this original, you know, math result that was verified combinatorially on computers in the 70s. And it was kind of like the first time people were able to like use computers to solve some sort of unsolvable problem. And that was only because they just couldn't do exhaustive. It was just do the exhaustive. It was brute force. Yes, but just we didn't have enough people who would sit there and check hand by hand and not know when screwed up. But we're currently at sort of the four color theorem level right now in mathematics with the unit distance problem, solving truly unsolved math problems. And you can listen to people like Terry Tao, who will tell you, yes, like it really is doing, you know, fundamental research in that area too. So on the quantitative domains, I think AGI is here for those. Now, if the definition of AGI is better than all humans at everything, I would say no. And again, that would be sort of like a no by definition, because there are certain things that we value about humans that because of the way AIs are, they can never be that. And so like if you include that in the definition, then it's like tautologically false, you know. But for some sort of like functionalist AGI where you look at some subclasses of things and say, would I pay it to do this? Or is it better than all humans? I think we're getting there. And in the areas where we're not there in a quantitative domain, it's really just a matter of time. Jeffrey Quinnell, great to talk to you. I am so grateful to you for the work you do. Let's keep it open. Let's let people do their own thing. Let's keep the token budgets in line. And man, if you don't have an agent yet, you better download Hermes. I love it. Hermes Desktop is a great way to start. Yeah. Thanks for having me back. Really appreciate your time. Glad to always catch up with the Tech TV roots. Appreciate it. Thanks. Jeffrey told me last time that I'm a little bit responsible for this. So I'm going to take credit. Beautiful moment. Thank you, Jeffrey. All right. Thank you, Jeffrey. We'll continue with Intelligent Machines and our assessment of Fable and a lot more in just a little bit right after this. This episode of Intelligent Machines brought to you by Helix Sleep. And Lisa and I, about a year ago, realized that it was time for a new mattress. You're supposed to replace your mattress. They wear out every, you know, six to ten years. Ours was eight years old. It was time. It really was. So we did a lot of research. We looked at all the websites. And we found Helix Sleep. And man, am I glad we did. A good night's rest. Man, that's everything. Sets you up for a great day. We're learning more and more. That's where health begins. And now that summer's here and it's getting hot, this would be the great time to upgrade to a Helix mattress. No more night sweats. It sleeps cool. No back pain. You know, it's not sagging in the middle. That's what happens with an old mattress. No motion transfer. The kitty cat jumps in the bed. I don't jump up and go earthquake. I love our Helix mattress. You don't want to settle for a mattress made overseas with low quality, questionable materials. You've got to rest assured your brand new Helix mattress is assembled. It's packaged and shipped from Arizona. And they make it to order within days of your order. So it's fresh. It's brand new. You should also do what we did, which was take the Helix Sleep Quiz. It's right there on the website. It'll match you with the perfect mattress based on your personal preferences and your sleep needs. And they have a mattress for every preference, every sleep need. And it really works. They did a Westbury sleep study. They measured participants' sleep performance after switching, like we did, from their old mattress to a Helix mattress. And what they found is exactly what I experienced. 82% of participants saw an increase in their deep sleep cycle. That's the most important of the three sleep cycles. Participants, on average, achieved an additional 25 minutes of deep sleep per night. For me, that was a 100% increase. It was huge. Participants, on average, achieved 39 more minutes of overall sleep per night. Not more time in bed, necessarily. Just less time awake. Which is pretty good, too. For 10 years, Helix Sleep is consistently ranked at the top across the most important independent review sites. Tested and reviewed by experts like Forbes, like Wired, Wirecutter, Oprah. Helix delivers your mattress right to your door. Free shipping in the U.S. And rest easy with seamless returns and exchanges. The Happy with Helix guarantee provides a risk-free customer-first experience, ensuring you're completely satisfied with your new mattress. So, do what we did. Go to helixsleep.com slash machines for 20% off site-wide during summer savings. That's helixsleep.com slash machines. 20% off site-wide. Make sure you mention Intelligent Machines at checkout so they know we sent you. That's important to us. The offer ends June 11th. Hurry. Now, if it's after June 11th, don't worry. You should still go there because there are always great deals at helixsleep.com slash machines. We love them. You will, too. Helixsleep.com slash machines. We thank them so much for supporting Intelligent Machines. Helixsleep.com slash machines. So, Paris was missing in action the last couple of weeks. I was going to take it personally, but she assures me it has nothing to do with my personality. It has to do with titanium dioxide in your ho-hos. That says a lot about your personality. Consumer Reports. It came out yesterday. I mean, well, titanium dioxide not in ho-hos, but Hostess Donuts mini powdered donuts. Yeah. Well, I actually downloaded Yucca so I could figure out what's in ho-hos, right? Everything. You want to know. So, this was a thing you did with Yucca in conjunction. Hey, real quick, Leo. Yucca's what? I'm seeing your Wii right now. Oh, you mean your Switch? Oh, yeah. I did that on purpose. Just to tease. I thought you were out. There you go. No, I forgot to switch back. This is, you did this, you, how long have you been working on this? Like two or three months. Yikes. It's been a big one. A big one. And I mean, so basically, we, this is an investigation we partnered with Yucca, like Leo said, where essentially we tested 40 different processed food products, different like popular grocery stuff for eight different additives and two processed contaminants that have been associated with health, like possible health issues depending on how much you consume and how frequently you consume it. But the issue is with additive, this is kind of true par for the course for many, if not all additives allowed in U.S. food. But the problem is, even if something is permitted in U.S. food, companies don't have to report either to the public or the government precisely how much is in every product that they're selling to you. So no one can really estimate what your exposure is to these additives and thus whether or not, if you are a frequent consumer of these things, whether or not there's a potential health risk associated with that. So we bought a lot of these products, over 120 like different samples of all of these products and sent them to like state-of-the-art labs to test them for all of these different things. And then had our kind of team of scientists analyze what we found. And part of the reason why this project took so long is, one, a lot of products tested for a lot of different things. So that meant we had to figure out a lot of different safety thresholds because you'll see in this article we have kind of a chart that we go into the actual amount of all the 10 different additives and contaminants we found in all the products and kind of what that means for you in a very simple way. But the bulk of the text, like a nearly 5,000 word story is about how we even got to this situation at all. And it, I mean, this has ruled most of my life over the last month, month and a half certainly, because I realized that the story of how, you know, part of like our top line findings were that we found that like 11 to 14 products, depending on your age and size, like contain kind of a concerning amount of additives or contaminants. And concerning in the sense that it, it, the amount in a single serving exceeds the amount that some public health agencies have identified as like safe to consume daily. And that's just in one serving. And so I was like, how is this possible? And it turns out it's possible because of decades of compounding errors at the FDA and the U.S.'s general approach to food additive safety and regulation that has led to a lot of these additives and substances being present and being allowed to be present at much higher levels in U.S. food products than European. Are you telling me I should consume very rarely Cheetos Flamin' Hot Cheetos? Yeah, the Cheetos was the standout finding for me personally. I mean, basically, so the article Leo's looking at is a second one that we had our scientists calculate where it's like, all right, what does this actually mean? How much should you, how much is safe then? And we calculated different recommended limits to kind of keep you under this safe level. You shouldn't eat more than one serving of Hostess Donut powder mini donuts per month. And that's three mini powder donuts. A month? A month. Is that the titanium dioxide? Well, that's, it's interesting. So the kind of one of the standout things was these Hostess Donuts mini powdered donuts. And we found that they had an elevated level of this thing called glycetyl esters. It's a process contaminant. Oh, so they don't put it in. It's not an additive. It's just a contaminant. It's the sort of thing that glycetyl esters are a contaminant that can form and is basically known to form when certain ingredients like vegetable oils or certain like additives like mono and diglycerols are processed. Like if they're heated to high temperatures, these can emerge. So, you know, if you have a refined oil and it's heated to high temperatures during refining and then you take that refined oil and heat it up again to say like fry a product, you might get exponentially more. Do you think consumers might be doing this in their own kitchens when they cook? Yes. I mean, that's a common way that you can get more kind of processed contaminants is through at-home cooking, which kind of adds part of the issue with all of these things we found. Is that not only are these problematic substances in foods, either because they've been added in there, if they're additives, or they've formed due to the processing of certain ingredients, but they're not just in Cheetos and donuts. They're in a lot of, if not most of, or many things that you eat. And the cumulative effects of all of that is almost like unknowable. And that's a problem because consuming, I mean, kind of a calculation we had to do to understand the risk of these products is like, okay, we just focused in on the products. If you had a serving of crunchy, flaming hot Cheetos every day for the rest of your life, like what would the impact of that be? What would it be? Just asking for a friend. I mean, it depends on your weight and size and other health factors. But that's the other thing. I mean, but they've always said highly processed foods are bad for you, right? We kind of know that. Yeah, but I'd always been like, yeah, they're bad for you because junk food's bad. But what does that mean? The thing that I thought was fascinating about this is like, this actually shows the reason why ultra processed foods are bad is because one, the processing itself, all the general junk foods, if you know, but it's the stuff that's in them. It brings definition to this. Processing was, you know, you process milk when you homogenize it. Yeah. Processing per se isn't bad. But you're putting specifics, receipts on what it means to be a super processed food. And this is part of a broader debate that's happening right now around the term ultra processed food. When I started reporting this, I was like, oh, we can't have the term ultra processed food in there. That's kind of a buzzword. It's like chemophobia. But as I talked to more and more researchers, at first I was like, okay, if we use it, we should use this California state definition that says it's ultra processed if you have additives plus like a certain high percentage of, let's say, fat added sugars or one other third thing I'm forgetting the name of. And I was talking to some researchers like, no, no, no, no. No, that's the wrong approach whatsoever. There's this classification system called NOVA that is kind of where the term ultra processed food, I think, came from or it really popularized it, especially among the scientific community. And it categorizes ultra processed food as basically foods that are produced in an industrial manner that you could not like you cannot. I could not make flaming hot Cheetos in my home right now. Try as you might. Try as you might. It's something that's industrially processed where you couldn't easily make it in your home. And it includes a list of like specific additives to it. And kind of it shows that the ingredients are a big part of this definition. And I just I hadn't really considered it. So Paris, let me ask you two questions. First, the genesis of the story. Did was this just looking to process foods or did somebody come to you and say, hey, the hostess donuts have titanium dioxide in them. Follow the titanium dioxide. You know, what was the first goal in what led you down this path? And then the second question is, you don't know what the company's motives are, what the processes are. But is it likely that these companies know that these things are in there, that they're buying vats of titanium dioxide to make the donuts white? And they know that's bad because it's not allowed in Europe. Or is it something where the laxness of American regulation has just gotten to the point that, yeah, this works and we don't know what it does. But nobody's telling us not to. And so we put it in there. So that's two questions. Yeah. So how we kind of selected the origins for this came from us deciding how to partner with Yuka on a broader investigation. It's this app that you can use by this kind of great team of French scientists and researchers and general kind of like health and food fanatics. And we decided we wanted to test. Originally, actually, we were just focused on additives. And it was kind of motivated by the fact that there's such a gulf between U.S. and European food regulation as it relates to additives. And so like kind of what I was saying before, when it comes to these ingredients, it's kind of a dose makes the poison situation. Most countries, like if a food additive is allowed, like there's a specific potency it's allowed at. You can have it up to this level in this sort of food. But it's really difficult for the average person to know whether that's the case and then assess their cumulative exposure to this. Because it doesn't matter if the amount of red 40 in, say, like Takis or Cheetos or whatever your favorite like red snack is. It doesn't matter if that amount of red 40 is like safe. If you're having seven other foods throughout the week that also have that, it might push you over kind of the limit where you want to be concerned. So kind of what we did is we looked through products that additives are listed on the ingredients list. We looked through products that had additives that we knew could be kind of problematic depending on the dose. Wanted to make sure we found the most popular ones that had this and then bought a bunch of them, sent them to a lab to test it and figure out what was going on in them. What was your second question? Second question is, is how is part of what you cover in terms of the lax FDA work. But but do companies knowingly say, gee, I need the donuts to be white? It was like the red dyes. We know there's been lots about that. But these other additives, especially the ones that they purposely add rather than the ones that are byproducts. Are they likely knowledgeable of what they're doing or I mean, yeah, the companies know exactly about these. They buy them, put them in the products list. And if you make donuts with titanium dioxide, you can't sell them outside the US. Yeah. I mean, you can't sell them in Europe because titanium dioxide is banned as a food additive in 2022. They must be aware of that. I hope the Europeans would buy donuts in any case. But they would love donuts if they just had a chance. Give them a chance. Yeah, it's actually very interesting. Because going into this, the partnership had kind of already been established when I was brought into this. We'd run some of the tests. And I was initially, personally, like a little skeptical because I feel like a lot of chemical. Good. You're a journalist. There's a lot of chemophobia around these sort of things. And I didn't want to do a story that was just like additives bad. Chemicals are in food. And it's like, yeah, everything's a chemical. Plus. We are chemicals. But frankly, the modern method of making food has made food much more widely available thanks to preservatives. There's a lot of reasons why these are not necessarily bad things. But what you want to find out is if they cause physical harm. You know, BHA is a preservative that means that people, foodstuffs can be shipped and produced one place and shipped somewhere else and last. Before, you know, we had preservatives. Food would rot, you know, before you could eat it. Yeah. I mean, there are definitely additives that have incredible benefits to them and that are not outweighed by any sort of risk. But I think the thing that if you've ever had rancid oil, you know, BHA is a good thing. Rancid oil is worse than BHA. Let's put it that way. But I mean, I think there are other ways that you can prevent rancidity. Of course. That haven't been like associated with, you know, cancer or things like that. It's just not everybody has access to fresh foods. Yeah. And I mean, I think that one of the things that ended up being so surprising or almost like radicalizing to me as I was reporting this out is I just, I don't know, I guess this is like the theme for me in being a food safety journalist and digging more and more into science than I had been since I like worked at Wired. Is just I was the story I ended up writing is, of course, about the additives and the things we tested, but it ends up being about the FDA and how basically this current panic that we're having in the U.S. around, oh, the chemicals in our food. People on both the right and the left are very concerned about this. There's a lot of scrutiny on additive safety. This exact debate we were having in 1958, people were freaking out about the chemicals in our food. They had a whole congressional investigation. They found out that, you know, there's like 800 some chemicals that companies are putting in our food and the U.S. government only knows that 40 of them or 400 of them are safe. And so they decided to pass this thing called the Food Additives Amendment of 1958 that was going to fix all of it. And basically what they did is they were like, yeah, any additive you put in food, we're declaring it unsafe unless the companies prove to us, the FDA, that it's safe. And that should have solved it. But there's like two, I mean, there's a lot of problems. The two core ones is that they had a honestly well-intentioned loophole at first built in where they're like, you know, we're just one agency. We've suddenly declared all food additives unsafe. It's we probably shouldn't have waste our time having companies prove to us that salt is safe to add to food or that, you know, technically if you chop up, let's say, apples and put them in your yogurt, that could be considered a food additive. You don't need to prove that apples are safe. So they're like, these things can be called generally recognized as safe grass. And you don't have to, you know, do anything. They're just good. The other issue is that this law, once companies proved that an additive was safe, there's like no clause in it that says the FDA has to go and revisit that determination ever. And I don't know if you guys have heard, but a lot of science has actually happened since 1958. And what I've learned is basically that the FDA, most of the additives we tested for this product project, the FDA has not reassessed the safety of in like decades. Even as other countries and other prominent regulatory bodies have reviewed new science and, you know, taken steps either ban or severely restrict use of this. The FDA has been like, well, it's an approved additive. And it's just, it's just complicated, especially for non-scientists. In general, it's very hard for people to understand and absorb accurate nutrition information. It's just hard to do the tests because it's in vitro. So, you know, you talk about sucralose. There's really nobody putting too much sucralose in their foods. You point that out. And you mentioned a large scale study of 100,000 French adults that found an association between this non-nutritive sweetener and an increased risk of type 2 diabetes. But an association, I should point out, correlation does not mean causation. In fact, it makes sense that people who are doing diet sugars might, in fact, be worried about type 2 diabetes. Sucralose, I think, is the worst example of the three artificial sweeteners we tested. Yeah, I don't do aspartame. Although there's a lot of evidence that aspartame isn't that bad for you. I spent a lot of time on because, I mean, first of all, none of the three artificial sweeteners that we tested, asulfame, aspartame, and sucralose exceeded any of the safety thresholds. We didn't recommend anybody limit the products based on what we found. A big part of this study was figuring out, like, what safety thresholds we want to use because the FDA does not have ones for all of these. And there's a variety of different ones to pick from from the various agencies. But they have tested, as an example, these sweeteners and determined they're safe. The FDA, in many cases, has not assessed the safety of these sweeteners in multiple decades. Right. And so part of the thing we looked into, though, is, you know, one of the kind of underlying regulatory things here is they have these things called acceptable daily intake limits. So whenever, you know, an additive is approved, they kind of figure out through the math science, like, what's a normal amount that someone can be exposed to every day and it's not a problem? This is determined from a variety of ways. But increasingly, and especially with the reason why I included that line for the three artificial sweeteners that talks about this large-scale observational study where they followed, like, over 100,000 French adults for, like, 12 years, like, recorded detailed daily, like, dietary stuff for, like, weeks on end. It's kind of a first-of-its-kind study. And they did find the results are way stronger for asulfame K and aspartame. Like, considerable, like, really notable associations between, like, low-level consumption of asulfame K and aspartame and increased risk of developing cancer, cardiovascular disease, type 2 diabetes. Again, associations, not causation. But I spoke to a lot of artificial sweetener researchers because I had that exact same instinct. I was like, yeah, this seems like BS, right? And all of them were like, no, we keep finding this in lots of large-scale studies. And we don't know exactly how to rock it. We don't know what the causation is. But it's an incredibly strong signal with a lot of these. And we think the FDA and other people should be paying attention and looking into this stuff. But, you know, it's one thing of 20,000 things that the FDA should probably be doing. So, I don't know. It's very hard. It's very complicated. It's very complicated. Article, because it really gives you a deep look into my mind palace. Yeah. Yeah, it's really impressive, deep, strong journalism. And I recommend people look at it. You see what Paris does in her day job, which is really important. And in no paywall. I know. As with all of Paris' writing, it's available even to non-subscribers of Consumer Reports. Consumer Reports does point out as a little disclaimer paragraph in here that we should probably mention that... What is it now? I've lost it. I had it here. What sort of disclaimer are you talking about? There's no reason to panic? Oh, it was a good disclaimer. And I thought it was a thoughtful disclaimer. This is one of the reasons I really respect Consumer Reports. They do these studies. They pay people like Paris to really work hard. It's important to note that neither Consumer Reports nor Yucca is a compliance or regulatory body. We offer information for consumers to make informed decisions. No legal judgments can be made for our findings. And then there's a whole page on methodology, which is where I really love about... I was going to say, there's a methodology that could have been 20 more pages, frankly. Does that have the part in there about you tearing your hair out? I mean, it should. Part of the methodology was Paris tearing her hair out. A big part of it is this page three on it that seems so simply, like lists all of the substances, the thresholds we use, the sources for it, and things like that. And this, truly, this page alone probably took me like four weeks, dozens of meetings. Well, because it's me and a bunch of other scientists. Like all of our great PhD scientists here. We worked with a great toxicologist from Yucca. And part of the thing is like earnestly and rigorously debating between ourselves, with outside experts, like what are the best thresholds to use to kind of assess against? And there's like a lot of different arguments. The one that we ended up having like kind of the most debate back and forth is like Red 40. Because both the EU and the US, their acceptable daily intake for Red 40 is the same as it was in 1970 or 71, when the US, basically the manufacturer of Red 40 submitted one unpublished rat study to the FDA. And they were like, great. They were like seven milligrams per kilogram body weight per day. But in recent like years, like especially in the last 10 years, there's been a lot of research that has come out and shown the other concerning effects of Red 40. And again, this is one of those ones that I was like kind of skeptical of at first. But I read this 300 some page report from like a California regulator that looked at all the available evidence for it. And they assessed this one 2018 study that was like found that if you feed rats, the dose of Red 40 that the FDA says is fine and that the EU says is fine. Those rats had neurological damage and they had impaired performance on learning and memory tests, those rats. And, you know, I don't know. I thought that so that is kind of the level we ended up using is based on this new research. But I don't know. Check it out. There's a lot of thought that went into every word in this. So happy. Oh, but my main thing I want to and I'll shut this out at the end of the thing. I'm doing a Reddit AMA on Friday the 12th at 1 p.m. And if you have any questions, get in there and ask me. She'll have the answers. I just wanted to say real quick, anecdotally, having moved back to the Philippines after living in the States for 20 years, it's very, very apparent that the food over there is not good. Why do you say that? Because like the last 20 years, I've had stomach problems living in America and then moving here, gone, all gone. Right. Did you grow up in the Philippines? I did. So maybe that too. But I also go to Germany and I see all kinds of things are still fried in palm oil. This is very, very hard. Country by country. Yeah, that's why it's totally anecdotal. It's very, very hard to do this. I mean, and something that we talk about a bit in the article and that a lot of the experts I spoke to brought up, which I think is a great point, like is, yeah, we're here. We're doing this testing. We've got like all the data for you to look through. We've got a whole thing of our scientists that have gone through and been like, all right, if you want to eat these things, but still be safe, here's how to think about it. Sure, you can make more informed decisions as a consumer. But, I mean, one of the policy experts spoke to said, really, this should be the job of the government and regulatory bodies to be the people who employ a bunch of scientists and who are paid by our tax dollars to look at this research, reevaluate it in recent years and make decisions so that every person in America doesn't have to make become a little mini scientist and figure it out on their own. Yeah. You could always just ask AI what to eat. I wouldn't do that. Yeah, I'm sure AI will never get that wrong. The safest thing to do is eat lots of fresh foods, you know. But I mean, yeah, it feels like much like my protein, like whenever the takeaway from protein was like, yeah, eat real foods. Eat whole protein, eat real food. Protein instead of protein powder. What did Michael Pollan say? Eat real food. Mostly plants. Mostly plants. Yeah. I mean, the issue really is that all that stuff with all the processed food, that's the cheapest food for most people. I mean, that is part of the problem. But, and I have to point this out again, there are societal consequences of not using these techniques because not everybody has access to fresh food. And if you fry stuff, you're creating glycidyl esters in your food every single time you fry it. Well, it depends on the sort of oil you use. It's very, it's much more complicated because it's humans. And it's very hard to do real scientific testing on humans because for ethical reasons, you don't want to kill some people and not kill others. It's just not done. So all we have is a lot of, I think, I think scientific consensus is very hard to reach in a lot of these things. Monosodium glutamate and ospartame are very good examples of foods that are eschewed by a lot of people, but the evidence isn't strong that they are dangerous. In fact, they break down into compounds that you have in your body anyway. So it's just complicated. And I understand. I know why you went through months of back and forth on this because it's very complicated. It's very hard to do. But I think this is a very judicious and reasonable article. And probably everybody should read it before you go out and eat more flaming hot Cheetos. Or hostess donuts. Isn't titanium oxide what you use for sunscreen on your nose? Titanium, it's basically this kind of white pigment. Yeah, I think it's sunscreen. Yeah, it's used in a lot of different things, but no longer in food. Back in the day, didn't women use, I think, lead as white makeup? Oh, that's right. I mean, people have used a lot of weird stuff for a lot of things. Yeah, yeah. It's used in paints. All right, we're going to take a break. Come back with more in just a little bit. We actually have some AI news. AI news. AI news. Yes, there was a few things happened this week. Just a few. Just a few. Our show today brought to you by Melissa, the trusted data quality expert. Summer is a season of growth. But while you're focused on expansion, your data quality can quietly deteriorate. No. And in today's AI landscape, failing to maintain high quality data directly compromises the integrity and effectiveness of your AI initiatives. You've heard the phrase garbage in, garbage out. For 41 years, Melissa has been the data quality partner, keeping business data clean, complete, and up to date. And they don't rest in their laurels. Yeah, they've been doing this for 40 years. But over those decades, they have become premier data scientists. And they can do so much with your data. Here's some more of what they can do. Melissa's alert service, for instance. The best way for companies to keep track of changes to their customer data in real time, including address updates, change of address, property transactions, hazard risks. Melissa's matchup technology does something very hard to do. It provides smart deduplication by finding non-exact matches. Global address verification and autocomplete validate and standardize addresses in real time. Globally, 240 countries. Melissa's change of address tracking. This is awesome. Automatically monitors and updates records of customer address changes, property transactions, hazard risks, and more. And their data enrichment lets you append demographic data, property information, and geographic insights to those basic contact records, turning them into something you can actually use. Mobile identity verification, great for KYC. It connects customers to their mobile numbers seamlessly, helpful in reducing fraud. Join more than 10,000 businesses worldwide using Melissa for all of those things and more. Like Lambeth Council, they needed a new way to verify IDs. Lambeth's performance manager says, now more than ever, we're very confident in our ability to prevent fraud and protect tight budgets, all at a lower cost compared to when we implemented manual ID checks. Melissa has apps for every tool you're using. Salesforce, Dynamics, CRM, Shopify, Stripe, Microsoft Office, Google Docs, and more. Melissa's solutions and services are GDPR and CCPA compliant. They're FedRAMP and ISO 27001 certified, and they meet SOC 2 and HIPAA high trust standards for information security management. Clean data leads to better marketing ROI, higher customer lifetime value, and AI that works as intended. Get started today with 1,000 records. Clean for free at melissa.com slash twit. That's melissa, M-E-L-I-S-S-A dot com slash twit. We thank them so much for their support of intelligent machines. I don't know where to begin. Apple had its announcements. It's WWDC keynote on Monday, putting, I think, AI in a very consumer-friendly fashion into the hands of, you know, many, many iPhones users coming this fall. Actually, many people are already using it in the development preview. The public preview comes out next month. Did you guys look at any of the features that they added to the iPhone? You're going to get it, Paris, in the fall. When is it coming out? And what features have they added? Paris has just come out of a cave. I have come out of a cave. I haven't done anything. I think what's interesting about it is it's AI for the people. No command line, Leo? No command line. It's agentic, but you wouldn't know it's an agent. It's Siri. You'll still say, you know, hey, you know who. Are they still doing the thing where you can load whoever you want in there, though? Like a voice? Oh, models. Well, yeah, I mean, it's unclear. They didn't talk about that, but people have found code tidbits that imply that that will be the case. For right now, what Apple's saying is it isn't Gemini. Yes, they mentioned Google. We are paying Google a billion dollars a year. But what they're saying is these are our models. They call them Apple Foundation models. There's a model for on-device that's very small but effective in a lot of cases. There's a model that runs in Apple's cloud. And they admitted there's a model that will run on Google's cloud with NVIDIA chips for the most challenging tasks. But they say in all three cases, they'll be able to keep it private. Now, there are some who disagree. Matthew Green, who's a cryptographer at Johns Hopkins, says it's going to be very hard to keep this stuff private because anytime you're using AI to look up movie times or get plane information, flight information, you're sending information out of this secure enclave into the real world. Like the web. Just like the web. Just like the web. But Apple's really touting, you know, you can use our AI privately. You can't use anybody else's AI privately. Does the user choose which AI to use or the system makes that determination? No, the system makes that determination. But as I said, some rumors said, and there is some evidence in code that you could perhaps choose Anthropic or OpenAI's models instead of Gemini. It's unclear. It's unclear. But Gemini's the default model. Except again, Apple says, not Gemini. It's our models. But we, in conjunction with Google, there's a lot of hand wavy. Maybe they distilled it. It sounded like they did it in the post training they used Google. I have to say that they demonstrate image playground looks very similar to Nano Banana in its capabilities and its style. You know, like a Buick and an Oldsmobile may have different brands, but they all come from the same factory. I feel like it's kind of like that. But anyway, already people are starting to use it. They have rolled it out in the developer preview. There is a wait list. Next month, it'll be public preview. I'll install it then. It's things like you could say, they show this a lot. My sister sent me an email about a video about titanium dioxide in my donuts. Can you find that? Then Siri, relatively quickly, within a few seconds, says, yes. I found the email. Because it sees your email. I found the email your sister sent. It has a link to this video. And you can say, would you play that? And it will play that. I think for a lot of people, that's what they want. Just the description you'd had. Couldn't you, in the time you asked Claude to find the email, or ask Siri to find the email, open it, and then allow me to watch it? Couldn't you just do that? Oh, but it could be, she sent it two weeks ago. I can't remember. Maybe it was my sister. Maybe it was, I don't know who sent it to me. I know there's this word in there. I face this all the time. The thing that Apple. The one thing I have noticed from the bits that have, the light that has fallen into my cave and played out in the shadow wall, is that I do think that a bit of Apple's marketing around this. I mean, I just still have bones to pick with whenever Apple Intelligence's original pitch for text message summaries was the most banal and annoying. Like, I believe the summary that they used for this round of it was someone texts you, Hey, have you heard about this plant? It's called this. Have you heard about Kalfea? Kalfea, yeah. It's a patterned tropical house plant. And Siri says, your friend texted you about Kalfea. She describes it as a patterned tropical house plant. Thanks, Siri. Wow. This is a really stupid example, isn't it? It's so dumb. It's like, did they use Sonnet? Did they use Sonnet to write that model? It's all Gemini, or actually it's all Apple models trained with help from Google. I think these are bad. That is a particularly bad example. But imagine that you have a webpage with a schedule of concerts. They showed this as well. And it can then compare it to your calendar, and you can see which ones you can attend, and you can add it, you can buy tickets. It's that kind of agentic stuff. And, of course, until we get it, we don't know how well it works. The premise, though, is interesting. It's similar to what Google says, which is, we know everything about you. We have all this information, you know, your emails on your phone. But you trust us. Your calendar, and we're going to keep it private. We're going to do everything we can on device so it doesn't even go out to the Internet. And if we have to go out to the Internet, we're going to keep it private there. And that's the pitch. I think more importantly, in my mind, is it's going to introduce a lot more people to some of the kinds of things that AI can do. Well, we're going to see the same thing from Spark. And what's the other one? Microsoft Scout. Scout and Spark. Yeah. Again, I think it sounds like two dogs. They don't want to give them human names, do they? Yeah. No. A lot of people do give their AI. So it's S's. It's Siri, Scout, and Spark. Yes. The Scout and Spark are specifically agents. They didn't mention agentics so much with Siri. Well, they didn't. No, but it kind of really is. It is that. We'll see. I mean, Apple isn't doing anything that you can't already do. Let's put that also out there. You can, with Google Lens, take a picture of something on the screen and ask about it. You can do a lot of the things that Apple's showing already. But Apple will put it all together in a very palatable, productized package. And I think that that's going to be introducing a lot more people to kind of the intelligence that you can build in. And I think that in general is a good thing. It's going to be the way people use AI in many cases. In many cases. A lot of photo enhancement. The photo editing, I just. You're not crazy about that? I mean, one of the examples they showed is you take a, or I think I saw someone who was using a preview show this. You take a photo of someone like sitting at a table and then you're like, oh, I don't like the angle of this. You could like use Gen AI to have, to pan around and change the angle. And I mean, I guess, I guess that's a fine. But how different is that from coloring? How different is that from coloring though? Really? Yeah. Somebody said this is something because people use once, twice and then forget all about which. That was cool. Yeah. They are going to use the same technology, Gaussian splats, to enhance the flyover so that when you fly over Paris's house, you'll actually be able to kind of zoom into it in a 3D way. It's going to be very interesting. That's going to be in maps. I don't know. I think this is a very careful use of AI. All of these capabilities will be available to developers fairly easily as they build apps. So you'll see more apps with intelligence. It's what Apple does. They take existing products and polish them up so that they're comfortable for consumers. So that was the one big announcement. The other big announcement dropped yesterday, which is that a version of Mythos is now shipping. It's called Fable. It is a new model of Fable 5. Remember, we were on Opus 4. We are now on Fable 5. And as Jeffrey was saying, they've put a lot of restrictions on it. Some of them silent. So it will step down to a lower model without telling you. Let me ask you about that. So I want to make a biological weapon. Oh, sorry. No, you're going to be moving down to Opus. Then I said, Opus. Or it won't do it at all. Yeah. I want a biological weapon. What does it mean? It just means Opus is less. I understand why it doesn't just say, no, I'm not going to do that versus I'm going to step you down. Well, no, I believe so. All of them, if you're like, I want to make a biological weapon, it's like no biological weapon for you, bud. They're worried that people are going to be too good at getting around. They're worried that people are going to be asking smarter questions rather than I want a biological weapon. They claim the safety is there, but as we've seen again and again. Basically, what we've seen so far, Anna, I want to make a recipe with titanium dioxide. Anything that could be in relation to biology or related to a bunch of no-no areas. They're just like, we're not even letting you ask it. You're going to Opus. I'll give you an example. Anthony yesterday took Steve Gibson's security show notes and asked for a summary from Fable. Fable said, no, no, no, that's cybersecurity. Really? Yeah, it's not too bright. It's not. Yeah. Yeah. So, but this is the way that Anthropic feels they can safely put this stuff out. They oversold the danger, danger, danger, Will Robinson. And now they're doing... Well, I think the danger is there. I'll give you an example. Yesterday, Microsoft, which has been using Mythos, did the largest patch Tuesday ever. 200 bugs were fixed, many of them serious. Something like two dozen of them were... So is that turned off now for the average user? I mean, if it's... That's because they had access to the full Mythos. But it wasn't tuned... That's right. The average user will not be able to do that. That's right. So it wasn't tuned to do that, but it was so powerful it could. Yeah. And so, I mean, again and again, we're seeing companies, Firefox, Microsoft, and others release huge numbers of bug fixes, curl, FFmpeg, flaws that have been around for 30 years are being fixed, and it's because of Mythos. So we know these capabilities are there. What they're afraid of is that if they release exactly the same capabilities to the real world, people will use it to look for flaws that they can exploit. And I think that's not unreasonable. They're worried that people are going to use it to create bioweapons. If it's that good, I guess they could. You know, it's the same question of, well, is this hype? Is this marketing? Or is this genuinely a problem? I'm leaning towards it's genuinely a problem, to be honest. Any of this... Have you played around with it? Oh, yeah, quite a bit. What do you think about it? Very smart. It's definitely a significant leap ahead. One of the things I've been doing with it is having it review all my old code. The stuff that I wrote with Claude Opus, my Hermes agent. I had it overnight go through everything in my Hermes agent. It fixed a whole bunch of stuff. It has been very good at finding issues that were there, but nothing else found. So it also seems smarter. It doesn't seem as sycophantic. It doesn't apologize. In fact, I think I know when it drops down. By the way, it doesn't tell you I've dropped to 4.8. But I could tell because suddenly it's apologizing. Cable does not apologize. Well, maybe all of your versions of Claude know that you want to be apologized to. It could be. There is some evidence that these models will start to... Leo is very sensitive. You will start to grok what your preferences are. I know we've been over this before, but I'm mad that Elon Musk took the word grok from us. Yes. I can't say that I've grok something without people being like, oh, you're into that. And I'm like, no. It was a word from a sci-fi novel before this that was adopted into common parlance. Here is... Just to keep an eye on what Fable's up to, I have it write a summary for me of all the things. This was the audit it did on Hermes. These are the issues it found overnight, issues I fixed. It had a deep understanding of the architecture, better, frankly, in some ways than Hermes did, and fixed a lot of things, found a lot of stuff that wasn't a big problem. And one of the things I was very impressed, every time it made a change, it tested to make sure that Hermes was still running, that everything was working. And then it would go to the next thing and go to the next thing. And it did this all unattended from... I started at about 1 a.m. and it didn't finish till about 5 a.m. So four hours of unattended work. Very impressive, cleaning up my agent. I then did the same thing with my clawed code setup. Found a bunch of stuff that was no op, that wasn't working, that was excess. Got rid of a lunch. It's a lot of stuff. I'm doing this because... Now here's the other shoe that's going to drop. You can use Fable right now. And anybody who's paying for clawed code or is using the clawed chat app will see Fable as one of the choices. But let me run it so you can see the warning that they give you because it's a little bit annoying. It's only going to... I have a subscription. It's only going to work on that subscription. Fable is here. Our newest model for complex, long-running work included in your plan limits until June 22nd. Then... Then bye-bye. Bye. You're going to have to switch to usage credits. They also mentioned that it is twice as expensive as Opus 48. So $15 for a million tokens in $30. Versus what does it cost now for DeepSeek? DeepSeek is 12 cents and 30 cents. What is that? One one-thousandth? I don't know. It's a lot less. 95% of the tasks one might ask this to do, is DeepSeek that inferior? What is it that makes using Fable so necessary that companies will spend this high amount? How will they know that they want Fable? How will they know that another model won't be just fine for much less money? Well, and that's one of the things Jeffrey was talking about earlier. And one of the things a good agent will do is delegate. It'll route tasks to an appropriate model. A good agent, and I've set my Hermes up to do this, will say, oh, you're doing coding. Okay, let's go use Fable. And then your agent's going to get kickbacks with models. Well, I hope not. But I'm able to use the local model, Quinn. Right now, I've been running the local model because- For most things you do. For most things, it's just look up something. Yeah, right. Run a cron schedule. So what goes elsewhere? What kinds of things go elsewhere? Coding, images, visual recognition. Harder stuff. Not rationing. You're not asking for reasoning things. It's functions that work well, like images. Well, I think coding is a reasoning thing. Okay. All right, fair. Fair. What would be reasoning? What are you using this for? What did you use this for today and yesterday? I went out and did the show prep that we do every time for a new guest and said, would you update Jeffrey Cannell's bio? And it found a bunch of new stuff. One of the things I like to do with Hermes is something called Pulse, which is a skill where it goes out and I can say, what are people saying about the Consumer Reports article on food additives? And it will check Reddit, hackernews, x.com, checks 20 or 30 sources and gives me a vibe check, a summarized vibe check of, well, there seems to be some real discussion about this. And then it'll give me some links. There's some very useful things like that. And I think that probably the local model is good enough for most of that stuff because it's using a skill. It's mostly lookup. We do one sheets for our advertisers. So when a new advertiser comes in, Lisa is able to run a skill that says, tell me everything about this advertiser, where they advertise. Here's one. I'll give you an example. This is one for the company that does Black Hat. They were interested in advertising. So it gives us a snapshot. Let me make it bigger. Snapshot of the company, who owns it, who runs it, its revenue, who its potential customer is. That's very valuable to us for figuring out which show to put it on. In fact, it even recommends which shows it's going to be good on. Talking points. It tells us what awards it's won. It compares it to existing sponsors. This is all done by my agent in about 10 minutes. Existing sponsors. Where it's advertising now. Where it's social media is. So it's just, is the agent in this case just doing web searches? How is it? Yes. It's essentially, it's a competitive analysis. Are you that this is? It's all 100%. There's no hallucination. I've not found one error. That's not correct. It is absolutely correct. I promise you. This whole thing about hallucination is really, in certain situations. Are you saying that you know more about the hallucination of models than the makers of the models? Because none of the model makers have said that they've been able to produce a model without hallucinations. And it depends on how you're using it. So it's not 100%. I have yet to find an error in any of this. You're not fact checking everything. You're not looking for errors. No, I'm spot checking. But I do look for errors. Absolutely. It's very reliable. Very reliable. And that partly, well, I know you're skeptical. But that's partly because of how the agent is designed and how the skill is designed. It's completely possible. I just, I mean, I've tried to use AI agents for any of my work. And I find so many errors that it's just. Yeah, I know. I hear people say that. And I don't know where that's coming from. I mean, I think it comes from the fact that I have to fact check every single word of it. And so do a team with three other people. No, I understand why you're saying that. I don't understand why it's making those mistakes. I really know. I mean, it's making these mistakes. Because it doesn't have a sense of truth. Yeah, it doesn't have a sense of truth. It's the things we've talked about in this show a million times. It is processing this, but it does not know what is true versus false. It can find likely answers because of probability, but it doesn't know how to check that against truth. And that's not an insult to it. I feel the video's looking hurt now. No, I'm just thinking wrong. It's not an insult to it. That's a user error. You're not using it well. I think if you trust. And, you know, you used to talk about rag a lot. I mean, essentially all of this is rag now. It's all referring to specific information. You have to be very clear in the skill setup that it's not to make up information, that if it can't find it, it doesn't know. Occasionally, I'll make errors. Absolutely. Absolutely. I was looking for a battery to, I actually, you know, we had a conversation about, and this is with Quinn, which is a local model and not super bright, about a UPS that I need. And it's, I said, I need a recommendation for a UPS for the desktop because the power went out. I don't know. Was this show? Was it this show? One of the shows. The power went out and I was off the air for 10 minutes. I was securing out last week. And so I got its recommendations, which are absolutely good. But then I said, well, you know, I have this one. Can I use this? And it said, oh, well, I need to know what year it is. I said, well, here's the sticker on the front. It said, no, that's not the sticker. There's another. Is it another one? Oh, yeah, this is it. Oh, yeah, this is it. You can show this. I'm showing all this as I'm doing it. And then it said, yeah, I found it. Here's the model. It was made in 2024. Based on that, I would keep the one you're using. Now, this is an error. This is a hallucination. It says it uses an RBC-19, which is actually no longer used by APC as battery replacement. So I went out and I searched for an RBC-19 and I said, I found something. I said, is this the right model? I said, no, no, no, that's the wrong model. So then I searched for it. I said, are you sure that's the part number? I can't find it. Oh, it said good instinct to double check. You smart guy. APC sometimes changes part numbers. So, you know, it's clearly 100%. Well, it was an error, but this is an error. So it's catchable. Yeah. So I went to look for it. It's catchable. Because you caught it. You're not disagreeing, you two. I know. I'm just, I'm saying, I think that these, we've done that. 100% in the sense that there are no hidden errors. It's incorrect to say they're 100%. I mean, one would argue that was a hidden error that was not unhidden until you asked it about it. Yeah. Because I couldn't find it. So I asked you. You're agreeing you two. Yeah. Well, all right. Yeah. Don't call it hallucination. You're saying essentially untrustworthy. I don't think it's untrustworthy. I think you have to use it. I think everything in the, everything and source of information in the world is untrustworthy. Yeah. Yeah. When you're a journalist, that's how you think. When I was at Time Inc., they ticked every damn word. And that system didn't work very well. We killed off Ed Vigoda once. Yeah. Well, that's not a good one. Yeah. No. All right. We're going to take a break. Come back. We have a little bit more before we wrap things up, including Mira Mirati. Returning the dead. She's not dead. That was an error my AI made. Our show today brought to you by Zscaler, the world's largest cloud security platform. As we have talked about over and over again, the potential rewards of AI really too great to ignore, but so are the risks. The loss of sensitive data, attacks against enterprise managed AI, generative AI increases opportunities for threat actors too, helping them to rapidly create phishing lures, to write malicious code, to automate data extraction. I'll give you an example. There were 1.3 million instances of social security numbers leaked to AI applications. Some of that maybe is malicious. Some of it is just accidentally exfiltrating proprietary company information when you use an AI. You want to stop both. And Zscaler can. It's time to rethink your organization's safe use of public and private AI. Just ask Chad Pallet, acting CISO at BioIVT. He says Zscaler helped them reduce their cyber premiums, their insurance by 50% and doubled their coverage and improved their controls. Take a look at this. With Zscaler, as long as you've got internet, you're good to go. A big part of the reason that we moved to a consolidated solution away from SD-WAN and VPN is to eliminate that lateral opportunity that people had and that opportunity for misdirection or open access to the network. It also was an opportunity for us to maintain and provide our remote users with a cafe style environment. Thank you, Chad. With Zscaler's Zero Trust Plus AI, you can safely adopt generative AI and private AI to boost productivity across the business. Their Zero Trust architecture plus AI helps you reduce the risks of AI-related data loss and protects against AI attacks to guarantee greater productivity and compliance. Learn more at zscaler.com slash security. That's zscaler.com slash security. We thank them so much for their support of intelligent machines. I think I would say on the accuracy thing, do you trust Google Maps? Sometimes it makes a mistake. No, Google Maps, I mean, I trust it generally, but I find issues in it a lot. You're right. I shouldn't say it's 100% accurate because neither is Google Maps. Neither is Google Maps. You can't know. In this case, that error came from an earlier parts list from APC that is no longer accurate. There was a reason it said that number. It didn't just make it up out of thin air. And then when I said, I can't find it, it said, oh yeah, that's an old parts list. Let me check and see what the new one is. And it found the new one, which I ordered. So similar to Google Maps, Google Maps will route you to our house through an alley that no one should ever go. And I always know when somebody's using Google Maps to get here because they go through that alley. It didn't do that for me, to be honest. You probably were using Apple Maps. I was using Google Maps. Oh, okay. Oh, you were coming from a different direction, probably. I was. If you come from this direction, it goes through that alley. And I always see, you know, like when Uber comes or something, I can always tell when the Uber, which map system the Uber driver is using. So that is an error. It's not, but generally you trust the maps, right? It's not hallucinating streets that don't exist and things like that. I mean, I trust it generally, but I know that if I am following it and I'm looking for something that I then can't find with my own eyes, the map is the problem. You know? So I'll give it that level of accuracy. That's probably not fair to say 100%. It's verifiably accurate. Let's put it that way. And certainly those one sheets that we generate are more than enough. I mean, I presume it's the same with code. At some point, what's wearing is you think, I got it. If it's mission critical and journalism, the journalism Paris is doing is mission critical, then anything you get from it, you have to check absolutely everything. Well, code is a little easier because code either works or it doesn't work. So if there is a massive error in the code and the program doesn't work, yeah, it failed. That happens, by the way, all the time. And then you fix it. So code hallucinations are a little more subtle. There are problems with, for instance, fake tests where I always say the way we do my code is something called red, green, blue testing, where you write a failing test. You write the code to see if you can get it to green. And if you can, then that part passes. But sometimes AIs will act childishly. I don't know what the right word is. And we'll write a kind of a dummy test that always passes. And that's not a good test. So you have to kind of pay attention to stuff like that. Code generally, though, if it doesn't work, you know it. And if it does work, you know it. So it's a little more deterministic. It's a little harder with probabilistic things. For instance, the judge who threw out a case because, well, this is a perfect example of what you're talking about. Lawyers on both sides were using AI. Idiots. The judge canceled the trial and kicked everyone off. Hasn't anyone learned in the legal profession at this point? There have been enough schmuck lawyers who've screwed this up. The problem is when it works, it works so well. Well, that's that exactly. Well, oh, this case is fascinating. One of the things I did in my cave life, you know, when you're in the cave life, you have like maybe an hour every night when you're trying to go to sleep. You're like, I can't look at bad screen anymore. I've got to look at slightly smaller bad screen. And one of my slightly smaller bad screen things was reading the whole, I guess it was the judgment summary from this specific case. And it was brutal. Oh. Where is Han? I've got it right here. I want to make sure that I get. Withers versus the city of Aberdeen. Oh, yeah. This is it. Basically, it seems like, oh, is this a different one? This is. Oh, this is a fully. I was thinking of a different lawyer. There's a lot of it. There's a lot of it happened this week as well, which was the judge ripped into this lawyer because he won used AI in for his filings, had a bunch of fake citations. But then when the judge asked him about it, he was like, I've used chat GPT a couple times, but not for this. He's like, well, we're ordering you to provide a chat GPT transcript to the court. He then went in, deleted the tran, deleted his chat GPT account. Then when it said he still had time remaining on the chat GPT account. So it would be around for another month, which is around the end of the window. He went in and asked for a partial refund. So all the data would be wiped and then told the judge, sorry, I don't have a chat GPT account anymore. And the judge lost their, they were like, you, I think he was suspended for months. They have to do a public notice where he has to provide this, the most brutal write up you've ever seen about how dumb of a person you are to every judge and attorney you've ever worked with. I mean, this seems like it is a scourge on the legal industry right now. This is what the court wrote in the case. Upon reviewing the brief submitted in support of the party's respective positions, both parties, with regard to the two motions, the court was unable to locate certain legal authorities cited within them. Specifically, the court determined the following followings contained hallucinatory citations. And there are quite a few of them, which the court also lists. Basically, the court decided you all are out of here and I'm not going to go. I'm not going to continue this at all. I mean, the poor plaintiffs. Each of the attorneys expressed embarrassment and apologized to the court. Yeah. It's it is. This is equally a scathing judgment. And the clients, I mean, when in the case that the first famous case of this, which I covered in federal court, the judge made him apologize to his client. And that also made him write to the to the judges in the cases that he had cited and apologize and find the money at the end of the judgment. This court is yet again burdened with addressing AI hallucinations in court filings. It is previously acknowledged that AI is a powerful tool that when used prudently in italics provides immense benefits. This case presents the court with an unusual scenario. Attorneys for both litigants engaged in similar sanctionable conduct. Is there a sanction order in it? Is that throwing out the case? I think there were sanctions, actually. They probably have to come in for a show cause why you shouldn't be sanctioned. Yeah. I get the sense this is happening because firms are taking on a lot more cases because they can blow through them with chat GBT. Or they're not bringing in interns to do the work. They're, you know, bringing in. Additionally, the court is compelled to point out that this sanctionable conduct inevitably implicates Williams and Wilson's ability to continue practicing before it. So you can see where they're headed here. A unifying framework for determining the appropriate sanctions in cases involving unverified AI usage has not been adopted yet within the Fifth Circuit. In this, in the, in the past, this court has considered the violating attorney's candor, accountability and remedial rep measures. So, you know, how guilty you feel. I don't know if I could go on. This is a very long thing in which I haven't read. Instead, Wilson explained she was shocked when the court issued the show cause order pointing out the hallucinated cases appearing in her filing. In essence, the attorney took the position she was unaware that AI could produce hallucinated cases and explained she didn't even know what a hallucinated case was. By now. The court finds that explanation to be insufficient and incredulous. When I covered my poor schmuck, it was early enough in Chetty P's life. I thought it was a super goal. I am not a cat. I thought it was a search. But that excuse goes away. Yeah. Apparently, he was pretty mad at this attorney. She's done this before. So I've, I've found the one that I'm thinking of, which was, I believe, an Alabama court case. There was two weeks ago this was filed because this is all happening. The court is not ordering the harshest of attorney harm sanctions because he made a mistake. The court is ordering them because when confronted with that mistake, he chose dishonesty over candor and destruction over disclosure. Lawyers make errors. Competent and ethical lawyers own them. When lawyers are caught submitting AI-generated misrepresentations to the court, they have two options. They can either admit to their mistakes and show contrition, or they can attempt to cover up their mistakes and demonstrate a weakness of character unsuited to the legal profession. Oh, that's saying something. If you're too low to be a lawyer. If you're too low to be a former path, they'll likely preserve their standing before the court. If they choose the latter, they may well lose their career. The judge here wrote, it's also apparent she attempted to minimize the violation by emphasizing the legal propositions in her filing were correct statements of law, despite conceding that she had cited fake cases. How hard would it be to look at your sites and just verify them in the law books? Cut and paste. Do a little search in Westlaw for crying out loud. Well, I mean, they all probably got the same memo as you, that their AI is 100% accurate. Well, I'll go back to my case again. It was a guy who does state courts, but the case went up to the federal courts, and he didn't have the license for the federal search engine. Oh, so he couldn't look at it. So he thought Chachity was a super search engine, and gee, it's free. Well, that's an innocent error. At the time, it was actually somewhat, excuse me, until he then went back and asked it, are you sure these are real? Which gave up the ghost on that. Yeah, he fined Wilson $2,500, barred her for two years from appearing before the court, and she ordered her to attend a CLA on artificial intelligence with an ethics component. The other attorney, also her admission in the case is revoked, barred from appearing for two years from today's date, and a $3,500 fine. So, yeah, they were all fined, mild fines, but mostly disqualified from appearing in that court again. So, yeah, anyway, this is happening again and again. In fact, there's a guy, Rob Freund, who has an entire page dedicated to these errors and so forth. So that's where this story came from, 404 reporting on it. I mean, this is probably why some of the large law firms are now investing considerable amount of time and money into trying to train their own LLMs. Yeah. And yet, this is why it's so complicated. There's also amazing things these things are doing, including mythos finding bugs that have been around forever. From TNW, Mira Mirati resurfaces after 18 months with a warning about AI governance and a product no one expected. She was the CTO at OpenAI, who was for about three minutes the CEO when they fired Sam Altman. She then started thinking machines. This may be a product that would appeal to you, no? Sitting down with Bloomberg's Emily Chang in San Francisco, she gave her first major appearance in 18 months. Thinking Machines Lab, her startup, which had spent the year raising $2 billion, securing a gigawatt of NVIDIA Vera Rubin compute, shipping one product and losing a troubling number of researchers it hired to build the next one. This is good writing by Christian Dina at The Next Web. The product is, drumroll, something they're calling interaction models, a fundamentally different kind of AI interface. Rather than the prompt and response format, the company's models are designed to produce continuous streams of audio, text, and video in 200 millisecond intervals. So this is a model designed specifically to interpret that kind of TikTok meme where you've got somebody playing Temple Run in the background and Family Guy on one side and then a live stream on another? This is a model for that? You know? It's a model for video slop. She says, when I wake up in the morning, I'm not thinking about how to kill the competitor. Okay. Maybe others are. Anyway, this is an interesting idea. Actually, we're trying really hard. Darren Oakey, our AI genius in the club, our Australian who's a regular on our AI user group, says he has a model that we can now put in the show that will listen and interact. I've been waiting for that. Yes. Yeah. I'm very interested. I'm very interested. So at some point, you may hear, I don't know if it'll be related to the thinking machines model, but you may hear a new analyst. What will you name it? Well, I think we should all decide. Is it going to be weighted to advocate for your slash Darren's viewpoint? Of course it will. Of course it will. It's an AI. It's not a democracy. It's an AI. It's smart. I mean, there are other AIs that aren't weighted to do that. But, you know, many AIs will immediately say, oh, no, don't don't trust me. Oh, it's not his model. It's it's got real time to is the name of it. Thank you, Darren. Let's see. What else? SAG-AFTRA Actors Union has striked a four struck strikes a four year deal with studios to protect the performers against AI, better pay and benefits. And and it avoided a walkout, which is good. No strike this year. More than 90 percent of the votes approved the agreement. This is our discussion last week with Robert Turcic. Stuff's going to start happening in Hollywood. Yeah. Yeah. I imagine it allows the use of AI as long as the actor. The contract says AI performers must bring, quote, significant additional value over a live actor or a digital capture of them. If producers are to use them, union leaders say this will keep the use of AI actors minimal. I wouldn't count on that. Anyway, they've got some concessions. And I think that's a good thing. I think that's a good thing. Actors are good people. They deserve to also allow some innovation to happen. Yeah. You need to find a balance there. And I think I last story this I think you will like. If LLMs have human like attributes, then so does Age of Empires 2. An archive paper that says if I can if you think an LLM is conscious, I can make Age of Empires 2 conscious. A great game, by the way. They led to evaluations of various areas, theory of mind, learning and understanding and psychology. In this paper, we leverage those observations to show that in LLM research, assuming the general anthropomorphic properties exist or not as part of their measurement is fundamentally flawed. Any sufficiently powerful substrate could implement an entity equivalent to an LLM, including the video game Age of Empires 2. So that should make you very happy. Well, we've had AI in video games forever. Like that's we've just been calling the computer the AI forever. So he trained a perceptron in AI and AOE 2. Yeah, it's interesting. Here is a picture of a NAND gate in Age of Empires 2's editor. I'm not sure this proves it in any respect, but it's kind of a fun idea. All right. What else do I'm looking down at your you like the new Gemini 3 translation? This is pretty exciting. Yeah, it is. Live translate almost in real time. And they showed let me show you the video. They showed it happening in simultaneous translation. Which is pretty good. So it's translating English. In almost real time. This is Sundar Pichai's talk at Google I.O. This demo, we're going to show you a live dubbing experience. Here we're using the API to stream translated audio directly from a tab. Watch as we listen to the Google I.O. keynote in Hindi. What's really incredible is how people are using our AI. But I don't speak Hindi, but I imagine that's pretty good. They showed it in other languages. They even showed it in four languages simultaneously. Which is, you know, UN style simultaneous translation. It's a little chaotic. I'm reading news all over the world. Actually, the other speaker speaks German. Which I know that you speak a little bit of German. Here's the German to English translation. Translation sounds. No choppiness. No artificial pauses. It flows like a completely natural language. Switch now to a Japanese session. It's pretty impressive. It is. I think the Babel fish is getting close. It's a little grain of salt, though, because this is Google and they are, you know. I know, they do these demos. I was like, Google loves to make an impressive demo video. And it's kind of cobbled together. I remember seeing Eric Schmidt in Davos many years ago saying, when we can do this, we'll have world peace. You forgot a few factors. Well, the Pope referred to the Tower of Babel, I think. Yeah, that's the Tower of Babel. That's the whole thing, right? Well, it's the opposite, right? The Tower of Babel, nobody could understand anybody because they all spoke different languages. But Google would have solved the Tower of Babel. If only the Tower had Google. The Tower of Babel was the representation of everyone having the same knowledge, right? Isn't that what it was? Oh, I thought it was because. The hope was that we'd all have one, yes. But in fact, we had diversity. Instead, we got Esperanto. Which, by the way, Google Translate does Esperanto. Wow. Does it do Klingon? I didn't see Klingon in there. I did download the latest version. I'm sure it does. There is a nerd over there that made that happen. Oh, yeah. It has to be. Yeah, yeah. What else? I mean, I feel like. I'm sure you talked about the anthropic confidential filing last week. But we had, you know, OpenAI did that this week. Everything's geared up. Yep. IPOs are coming. Perplexity says 2028. All right. Sure. Perplexity is going to be challenged because they don't have models of their own. You won't have money by then, perplexity. Yeah. Meta is rumored to also be going to the public market like Google. There's going to be a huge push. Yes. Raise more money. Yeah. But Meta is already a public company. Well, no. So they raise more. Google's raising another 80 billion, diluting their current stock, right? Ah, taking a page from the GameStock. Oh, and speaking of Germany, German courts. Yeah, this is a bad one. Yep. I have declared Google's AI overviews are Google's own words. And thus, Google is liable for errors. I think that's fun. No. Because what I get the logic of it. Google's got to be liable for all my own words, even the words I tweet. I think Google could, you know, listen, I understand there's broader implications to this that my flippant answer is not considering. But while I'm doing flippant answer, I think it is fun that somebody else has to actually care about their precision of their words. So what happened was pretty bad. So Google's AI overviews falsely tied two German Munich-based publishers to scams, subscription traps, and shady business practices. According to the court, the AI mixed up information about other genuinely sketchy companies with the plaintiffs who sued and drew connections that did not appear in any of the link's sources. Publishers sent Google a cease and desist, but Google didn't. Google's AI overviews work nothing like traditional search results, the court argues. The AI rewrites and judges results in its own words. According to its own structure, the ruling says, in the case at hand, for example, it opened with a confident claim like, yes, this company is known for dubious business practices. I mean, I think logically this makes sense. If you have your core product offering for Google right now is its AI search results. It's selling ads. It's reorienting its whole kind of product structure around the search product around this. And if you do not offer any tools to when you get it wrong to correct that information, when someone notifies you that you've gotten it wrong, and you continue to show this incorrect and libelous information to an incredibly large audience, I don't see why you should. Built in, I'm not disagreeing with you, but there's randomness built in. So the next answer may be the opposite. Well, you still showed it to a lot of people enough so that this lawsuit was able to get it into discoveries. You know, in plain old search, I get the logic of the decision, right? In plain old search, you were delivering the web. Then in ChatGPT, well, that's just a tool, and you asked it a question, and it has caveats. This decision is saying, but it's Google speaking as Google, answering this question. But it's a fine line, I think, there. And the result of this, I mean, Google is just going to make the caveats a hell of a lot bigger, an 18-point. This often makes mistakes. It's not true. You know, beware, beware. But the end result could be that, you know, they pull AI out of Germany or something. I don't know. By the way, Benito, you were right. I'm just kidding. I got my result from my AI. From Genesis 11, 1 through 9. After the flood, humanity is described as speaking one language. People decided to build a great city with a tower with its top in the heavens to make a name for themselves. God sees the project and says, this is a little weird, because they are united by one language, nothing they propose to do will now be impossible for them. So he confuses their language and makes them unable to understand one another and scatters them across the earth. Yep. So we tried. That's what kind of the... We tried. We tried to AI. That's right. That's what my agent, because it knows me, says, this is a nice cyberpunk Taoist read. Babel is what happens when coordination turns into domination when a shared protocol becomes a monument to control. See, it knows me. And it always sticks in little stuff like that. So you want a few other quick stories? Yes. So turn it in, which is supposed to detect plagiarism and now AI. I mean, turn it in, which has been flagrantly wrong about a bunch of stuff for many years. I couldn't agree more. But researchers did an interesting experiment where they gave it 100% human text, 100% AI text. But then they also varied the text, you know, 20%, 40%, 60%, whatever AI text added in. It was pretty good on either end of the extremes. But the paradox here was that the smaller the AI contribution, the larger it thought it was. The larger the AI contribution, the smaller it thought it was. All of these things are coming out trying to argue with the Pope. When we have Padre on, some accuse the Pope of putting AI into the encyclical. We're just going to get used to the fact that you don't know. You don't know. I'm curious as to how that same test would work on Panagram. I would too. Yeah. Do you think that one's good? Like that's the one you think? I have been entirely dismissive of all of them always and I think rightfully so. This is the first Panagram is the first one that has given me some pause in the sense that I. I mean, my assessment of it has been entirely fully myopic in the sense that I'll put in my own writing from years ago. It flags that as 100% human every time no matter what combination I do. I try and mix in some of my own writing with AI generated text that kind of sounds like my writing and it will catch that. I don't know how it does it. I don't know. It's Panagram, not Panagram. Sorry. I don't know if this is more broadly applicable. I'm sure there are a bunch of things that might get wrong. I don't know how it handles new models. I'd love to get someone from Panagram on the show. I'm going to talk through this a bit because I think part of my understanding is part of what sets them apart is they are like using their own model. To basically be like, what with every single word, what does our own model think the next likely word is going to be? And then it kind of compares, uses that to try and determine. That's a very rudimentary and probably somewhat partially wrong explanation of it. But it's a different sort of assessment than we traditionally get from these sort of tools. Let's see. I'm asking it to, okay, 100% of my text is human written, which is true. That was from my journal. Let me find some AI, see if it can detect some AI talk. It probably can. I could detect some of this AI talk, actually. I mean, yeah. Yeah. It's very interesting. I mean, I played around it quite a bit because I saw some people whenever that, I think, Commonwealth Prize short story was going on. People were citing Panagram as being like, oh, evidence of AI. I was like, oh, this is such BS. People have been using this service forever. It's obviously wrong. I'm going to find a way to prove how dumb it is so that I can dunk on them on Twitter. And I couldn't. So, I mean, that was the only, I didn't try longer than like 30 minutes, but I thought that was notable. It's the first one of those I hadn't been able to figure it out. Yeah. I don't think teachers should use it though, right? I mean. No. I mean, here's the thing. There's such a high stakes. I don't think any of these should be used to make definitive judgments in any truly meaningful way. But I think it is a useful signal. Yeah. But I don't know. That's the problem. I just don't know the data. I need research like the one I cited to go in and test it. I'm putting in a post with the encyclical. One of the founders of Panagram just had a debate with a researcher about this very thing. And it concluded with, I think this is on either Twitter or Blue Sky, them being like, we will give you this many thousand dollars worth of free Pangame credits. You can do whatever you want with it. All we ask to test it. All we ask is that you just publish everything you put in, everything you do out, and you do it all by yourself. Have nothing to do with us. I just put in the Tower of Babel answer from my AI agent. Oh, got him. 100% human written. Uh-oh. Right? Oh, gotcha. Good work, Leo. That's great. See? Although a false negative is probably better than a false positive, right? You don't want a student who actually wrote a paper to be flagged as AI. Most of this is quotes, though, right? So the quotes are human written. Confidence low, though. Yeah, it does say confidence low, which is… Supporting evidence. Well, I'm not going to… Well, supporting evidence I wouldn't use. Supporting evidence isn't… It doesn't highlight the evidence that… Oh, it's not specific to the… Yeah, it just… It calls out things that people commonly call it, like a three… You know, a rule of three, or a if not this, then this. But it says it doesn't use that at all in its assessments. So it won't judge you on your MDash usage? I mean, this is something I thought about a lot when I was writing this story, because, like, between that and the second story, I wrote, like, when published, like, 5,000 words and all the other stuff out there, like, a couple more thousand. And I was like, I feel like I'm going insane. Like, I go and look at my old work from, like, pre-2020 even. I'm like, I used… I have always used MDash as a journalist, you know? And I've always used a lot of the phrasing or things like that that now is considered common with AI, because it did get that from scraping the work of lots of journalists that have similar habits. But now I'm always, like… Like, I started one of my articles off… I wanted to, like, list the top, like, three products that had a lot of additives or contaminants in it. But I was like, oh, everybody always says that when there's three things listed, that's AI. So I'm going to start it with four instead. And I did. Because I didn't want people to think I used AI. Not that I did. I mean, the three things thing is like a writing technique that's been passed down through the ages. I know! It's just something that sounds good and normal. It's like the rule of threes. You do blank, blank, and blank. Absolutely. It's the right rhythm. And then the other one is tell them what you're going to tell them, tell them, then tell them what you told them. My sister taught homiletics, sermon writing at a seminary. That's what you did. These are things that we all do. So I'm going to try one more thing. I did give it some more AI pros, and it was able to detect it. It said 100% AI generated. But now, if I have any more tokens, I'm going to give it... I asked my AI to humanize it, because it has a skill to take AI-isms out. Nope, didn't fool it. Hundo confidence high. Hundo percent. So I wasn't able to humanize it as well as it thinks it can. What's weird is that the confidence low was still 100%. Like, that should have been a smaller number, right? Well, no, it's 100% of this text, it believes. Oh, of the text. Yeah, yeah. It's not 100% confidence, because it will break it up into chunks. Like, sometimes it'll be like, we think that this paragraph could be AI-generated. Or it'll sometimes highlight it and be like, we think this could be combo AI-human. You know, like a light rewrite situation. Well, so, you know, at least it didn't flag as AI something a human wrote, which I think would be a much worse outcome. Anything else before we break for our picks? Another interesting one to me is the Amazon is letting you an image generator. So you can, if Paris has a dress in mind that she really wants, but can't find it, she can have it generate an AI image of that dress and then have Amazon look for anything that in reality is- It would be better if it made it. Well, that's where you go. So speaking of that, Amazon has also introduced a structure. So you can have the AI design an image and then make the custom, the merch from that. Oh, that's nice. That's nice. Right? So, but it's mostly like making a t-shirt or a sweatshirt. Right now, but, you know, soon it'll make you a Jensen Wong fake leather jacket. Sign a heavy metal logo for my family. Jensen's not going to let that happen. Come on. He's got, he's bought up all the snakes in the world so that you can't make one. Finally, Elon Musk- He can change the AI, right? Right. Finally, Elon Musk says he's building a chip that's two to three times better than NVIDIA at 10% of the cost. He is just such a- Is he going to make it on Mars? Yes. Jesus Christ. I don't, I no longer even listen when he says something. It's like, okay, fine. Go ahead. You do that. You know, I'm, I've been thinking a little bit about this. There was a point in time where I think he was a genius. He did do some pretty amazing things. And he did some very smart things. But the argument was he came in. I mean, Tesla, he didn't do Tesla. He came into Tesla. No, but he made Tesla happen. And those guys had never even built a car. He bought the idea and then ended up making a car that is arguably to this day, the best electric vehicle ever made. It certainly is the longest lasting. He did some good things. Now he didn't do it all by himself. He hired the engineers and he put the money into it. Same thing with SpaceX. I mean, that's a tremendous success in many ways. But we don't have any idea. Maybe he went crazy. CEO brain worms when you're surrounded by enough people. That's what I think. Tell you that you're the smartest, best person ever. It's like having AI all around you. And just say yes to everything. Right. Plus, internet brain poisoning from general internet use. Plus having an army of online sycophants that. And daily ketamine use doesn't help. And third pillar, a lot of alleged drug use that if you were to be believed, probably makes all those things worse. Yeah. I was also thinking that if we are going to go to Mars, let's just send the AIs. What do we need to go for? They're not going to promise. What are they going to do up there? Why can't they build a city? Well, this is what the test grail people say. Is that when they talk about the 10 to the 40th, whatever the number is, human beings in the future, they don't mean they're all human beings with bodies. They think that they're going to create virtual humans that will then be able to populate the universe because they never die. Yeah, we polled. We did a survey. We asked 10,000 people. And you're like, oh, how did you do the survey? And they're like, well, we used an AI tool to ask. To simulate human. I'm like, you're not talking about people. You're talking about a tool's idea of people. We surveyed the AIs. Well, that's being used now because it's cheaper than, in fact, surveys. You're right. It makes a lot more sense, though, to send AIs to Mars. They don't have the low gravity, the long trip. But what these people think is that these are going to be alive because they can create. They may not be alive. This is your Jeffrey Hinton thing, that they're conscious. And we can send this conscious being on our behalf. And so we have populated the universe with human extensions. We may not agree on what the process is in an AI's mind. I don't think we disagree on the fact that it's not human. I'm not asserting that AIs are human in any respect. Well, sentient and conscious. Well, they might be sentient and conscious. We don't know. But they're not human. They never will be. The Pope is going to strike you down. Well, let's not forget that the Pope is in a business. He's infallible, Leo. You want 100%. A business where faith is key, right? Believing in the face of no evidence. Same as all the Tuscaroos. Yeah. Same as everything related to me. Same as me. You're watching Intelligent Machines with three very intelligent people. And no machines involved. Jeff Jarvis, Paris Martineau. I just want to invite those of you who are not yet members of the club to support our shows. These shows require your financial support. But even more than financial support, they require your emotional and intention to continue. Because yes, we do have advertisers. They cover 70%. Let's say, I'm not sure the exact numbers, around 70% of our operations costs. We could continue on without the club, but we'd have to cut back a little bit. The club, though, gives us the support we need to keep doing what we're doing. Not merely financial, but emotional support. It's a vote for what we're doing. Independent broadcasting, independent journalism, not owned or operated by a big company, working on your behalf, no one else's. I really think that's an important thing to do. And if you do, do, I would love to have you join the club. You get lots of benefits, ad-free versions of the shows. You get access to the Discord, all that special programming we do just for club members. Find out more at twit.tv slash club twit. Cast your vote for independent journalism. Guys, shout out to everybody who left reviews for the show in the last month. Did we get a bunch? Yay! A bunch in the last month, yeah. Let's see some reviews. Let's see. Always makes you think, says LRAU hockey. What do you get when you put together three smart minds, differing opinions of AI? You get intelligent conversations about AI. Extremely intelligent, says Peggy Lisba. Thank you. I've loved the show for ages. Previously Twig. Keep up the banter. Definitely part of the charm. Let me in, Paris, Leo, or Jeff, says Matt. Definitely not an AI. The show's fantastic. With sand in their shoes or not. I'm not sure, but I think you do not need to be conscious to enjoy the show. I want to hear more of Paris' opinions and less of the two old men. We agree with that one. There's so many great ones here. Did they say more of your opinions and less of the two old men's? Yeah, they did. You glossed over that. You should really emphasize that. That one's Paris. You know, there's one negative opinion that got really mad that I guess Jeff talked about CBS last week. But, you know, other than that, lovely time. Leave your five-star reviews at your favorite podcast client. We appreciate that. That helps us spread the word. Paris Martineau, pick of the week. I got a lot down here, honestly. Let me even find it. Yeah, as I said before, come check out my Reddit AMA Friday, June 12th at 1 p.m. Eastern. I've included a link just to Consumer Reports Reddit because the AMA hasn't been posted yet. But it'll be at slash r slash IAMA. Yeah. You know, I think it's in a couple of hours or something before we'll open it up. Can I ask questions? Get in there? Other picks are, I don't know. I told you guys that two or so months ago I started getting into basketball. And since then there's been a lot of moves in the NBA. You picked a good time to get into basketball, young lady. I picked a really good time to get into basketball. I told you. I'm your time. The Knicks owe you something there. I mean, I'm going to a bar to watch this game four right after the show. Very exciting. You know, it's lovely. I have suddenly a desperate hatred of the tall Frenchman. No, you should hate is the referees. Oh, I do hate them. They are the ones to hate. Wait, so that's you. Referees who agreed basically made a call yesterday. They're like, yeah, Wemby can bring a loaded gun to the game and we won't stop him is what they've decided. But, you know, it's fun. I'd really I'm not speaking to anyone who hasn't watched basketball before. But if you're someone like me, never really watched basketball. It's an adult. Check it out. It's a delightful time. Just don't wear your nuts gear. Paris, just don't wear your nuts gear, please. The tweet about Wemby going to pushing people out of line at Salt Hanks. Oh, I did. Yeah. I saw that come across and I was like, yeah, wow. The nexus of my interests. Sources say Wemby has a big day planned in New York City, shoving people out of line for cronuts, yanking people out of line at the Raphael show at the Met, pushing people out of line at Salt Hanks, wrenching people out of line at Kith, etc., etc. Wow. Hank has made it into the good Eric Lockett. He's Hannah Pley of New York landmarks. That's true. That's cool. He writes for the New Yorker. That's like a big deal. Yeah. Wow. Should I talk about gaming or are we going to include that part in the show from earlier or cut it out because it was when we were figuring out things? She is very excited about getting herself one of these little babies. Switch 2 is basically, I didn't watch Nintendo Direct, but I was scrolling through it yesterday. I'm like, oh, I haven't bought a Switch since like 2018. I'm mostly a Steam Deck girly. And I haven't needed to because I have an early Nintendo Switch, but allegedly they are releasing a new Fire Emblem game on September 17th. And it appears to be a spiritual, if not direct, successor to one of my favorite games of all time, Fire Emblem Three Houses. And I'm so thrilled. It must be generational. I feel very guilty playing video games. Like I'm wasting my time. I enjoy them. Here's the secret, Leo. You're always wasting your time. Every second you spend doing anything is the second closer you are to death. Unless you're jump shotting a big two points to help the New York Knicks win game four of the NBA Finals. It's true. That is a nice video of you, an image of you. Here she is showing her ball handling skills. I don't know how it shows her two shows. I'm a Jess's Celtics fan, though. I like that none of us are in the Finals. The AI knew that we are completely unrelated to everything going on. Yes. We don't know what's happening. The other game I'm really excited about those, I've never really gotten into Final Fantasy because, I mean, I've played through a bit of some of the old classic ones, but never fully. Because, I mean, the graphics and everything, it's such a leap going from modern gaming to, you know, like Final Fantasy V. But they're releasing a new Final Fantasy game in HD2D with turn-based combat. I'm going to be right there. Ooh, turn-based combat is your thing. As a coward who doesn't like to have to play in real time, because I like to be able to sit there and think and then make a move once I've considered all my options. That's why I hate turn-based combat. I mean, that's the thing is, I'm not here to, I'm not playing a game to have reflexes. My constant thought in basketball is, I'm like, how are they moving so fast and doing so many things? This is great because it looks like it's Octopath Traveler 2 style, which is phenomenal. Nice. Nice. But yeah. It is kind of a retro look for Final Fantasy. Yeah, it's HG2D, which has gotten really popular with Octopath Traveler. Is this a remake of an old Final Fantasy or which one is this? I believe it's a reimagining of a story concept popularized in a mobile game, but everything else about it is completely different. And that instead it incorporates elements from basically every Final Fantasy game in where your characters can kind of like their job classes are these things called visions, where you get to basically pluck a character from all the iconic Final Fantasy games of yore and they emerge to do a special move for you. Okay, that sounds cool. Yes, I do like one turn-based game, which is called chess, but that's a little different. All of the turn-based games are basically chess, but just a little differently. Yeah, maybe I should look into it. Yeah, it's all just a more complicated chess. That's the thing, is they've made versions of chess that are more complicated than you could ever imagine. Maybe that's why I don't like it. Versions of chess that are so complicated and then you have the worst, most maniacal AI ever that's playing against you and you can turn on modes that are labeled things like maddening and it's awesome. It's awesome. I have a little pick, something Google shipped out. Let's see. When did they ship this? Last week, I think, called Dream Beans. This is a very weird AI experiment from Google. It's an app. This is running on my iPhone. Where it looks into all the stuff it knows about you, generates an image, and then suggests something you might want to do. Play the podcast Where the River Took Us. Apparently thinks I would enjoy that. Tracking Mountain Lions at Jack London State Historic Park. By the way, that's a pretty good image of Lisa and Michael and me tracking mountain lions. Pre-order the Space Opera Exodus by Peter F. Hamilton. I do want to do that. He's right about that one. So it does these images. I don't know why it thinks I'm going to go to the, oh, I'm going to drink the 2025 Northern Road Vintage of Gros Almitage. That is a very hard wine to find. I love Cotarone. Yes, we do too. Tell me it's a dream for an upbeat morning. If you're looking for fresh studio drops, spinning this vibrant solo debut from British alt-pop songwriter. So these are recommendations, but then it draws these silly images to go along with it. There are Steve Gibson and I running a mobile privacy audit on my Pixel. But quite enjoying yourselves. Yeah, it looks like we're having fun. Having a really fun time. Isn't that so strange? It's a little creepy that it knows so much. Like, how did it know that I wear a Scotty vest whenever I travel? Maybe I mentioned that once? Here, Lisa and I are. Oh, oh, oh, please. It knows all of your pictures. It's seen all of your photos. It knows everything. It has. Trying out. Lisa just told me I should order a second Manic Kitchen Pepper Cannon for steak night. It must be listening to me. Experience vintage. It's almost as if you've got your house wired up with microphones you've given AI access to. So when you install Dream Beans, you give it your Google account. And then. Guess what? And then it goes and looks at all of the stuff it knows about you. Goes through your Gmails and stuff. And generates suggestions based on your location. Based on who you are, what you're doing. It does these images. I want more stories about. Oh, so you can tune it a little bit. Show me less of. Never show me. So that's just for tuning. This is Gemini, obviously. I think it's pretty good. I mean, it's weird. Yeah, this is an example of how you can find out what Google is. Jeff, are you allowed to use this? Nope. Oh, of course it's not on Workspace. By the way, the Manic Kitchen Pepper Cannon is the best, most expensive pepper mill in the world. But how much pepper are you milling? A lot of it. So much so that we need a second one. Because we have one next to the stove. Is it her? And Lisa says we need one for the table. So we'd have to go to the stove to get more pepper. Are you just using it casually on every meal? I'm going to show you this pepper cannon. This is the best thing ever. No, I understand what a pepper cannon is. But what are you putting pepper on? Everything. With such frequency. Pepper's the best. Don't you like pepper? I think pepper's fine. I can't remember the last time I milled pepper. It's hard to give Salt Hank a gift that he hasn't already been sent for free. Is that not insulting for you to send him pepper? You'd think he loved it. He said, Dad. And he started to use it in his videos. He loves the Manic Kitchen. I was going to say, it's probably because he likes a big thing that makes a crunchy crunchy sound. But it's good. It's got steel grinders. The extra large version is $249. Yeah, I think that's the one I have. I don't. The regular one is $199. Yeah, that's probably. They're expensive. I'll get the regular one for the table. Hire somebody with a hammer to crack the pepper for you. It'd be cheaper. But it lasts forever. How many pepper grinders have you gone through in your life? You don't know whether it's going to last forever yet, Leo. Well, it's lasted in 100% years. I say it's 100% true. I love the pepper cannon. Okay. And Google knows it. Google knows it. I'll never know. Okay. Are there... How much plastic is in this device? None. It's aerospace-grade aluminum construction, Paris. And most importantly, 10 times the output of ordinary pepper mills. Do you know what they use for cooking? High carbon stainless steel birds. You could use it to grind your coffee. That's more acceptable. The one thing I will argue is that my cute deucin' deucin' pepper mill, I hadn't thought of until now, probably got plastic in there that I'm grinding into mine. There you go. Yep. Meanwhile, I'm getting metal in my pepper. I just suggested the best show title, which is just, Google knows I love the pepper cannon. That really makes me laugh. Jeff Jive is your pick of the week. All right. I'm going to take you to some... The work of... I discovered this through a McGill University professor, Sarah M. Grimes. Panic first, evidence later. Oh, she's great. This is quite wonderful. This is... Oh, that's right. You covered... This was your old beat. So, she takes to task Jonathan Haidt, as I try to often. I quoted most of these researchers in my book, The Web We Weave, which no one bought. And because optimism doesn't sell. And it takes apart Haidt's bestselling book arguments with receipts and real research from real researchers who know what the F they're talking about. Well, I read Candace Hodger's piece that she refers to here in Nature. In fact, I read it into the record on this show, I believe, way back when, when the Haidt book came out. Rogers, this is her field of study, and she completely debunked it. Correlations, not causation. Effect sizes are tiny. We've seen this before. The global data don't fit. Data being plural, like researchers do. Well, that's one of the most interesting pieces. This is an American phenomenon, but phones are everywhere. Yeah. It just doesn't work. And finally, the critique is it punishes kids, not companies. So if you've got issues, go with the companies. It goes into Haidt himself, his actual research fields, moral psychology, intuition and emotions, political psychology and polarization, business ethics, not adolescent development, not media effects or screen time, not child psychology or pediatrics. Well, she probably won't like this new paper that says the reason the birth rate is dropping is the iPhone. Oh, God. That one's killing me. It's killing me. The birth rate started dropping dramatically in 2007. Because we screw our phones. And it's all because you're looking at your phones instead of having relations. How about it's the fact that the world's falling apart and it wasn't to bring a child into it? This is the New York Times. Two new studies point to phones. At least they don't mention the iPhone. Although 2007 is when the iPhone was released. It's not a coincidence. I would argue that smartphones actually increased the birth rate by a small amount. I would argue that. What were we doing? We're doing people to. You suddenly have a bunch of applications that are specifically designed towards activities related to increasing the birth rate. This is a famous piece that I remember reading about 2007. If you look at the things that changed in 2007, it's really when the world went to hell. So I think we can blame the iPhone for recession, the housing boom, the global financial crisis, and Gizmo. When was Gizmo born? Six years ago. See? When the iPhone came out. I rest my case. No way. That makes no sense. Gizmo has been radicalized by technology. Over the course of when I was in my cave time for this story, she got really mad that I was spending so much time at the computer and not playing with her. Oh, I bet. And she's learned how to, one, flop my desk and mess up my mic, as we all know. But she's also learned how to hit the buttons on my computer that turn it off. And turn off my notifications. Like, she's figured out that if she taps the lock screen button, I will move her and therefore touch her. Wow. She is smart. No, she needs to get dumber. She's too smart, is what you're saying. Yeah, it's untenable. Feed her some donuts. I should. A little titanium dioxide goes a long way in a kitty cat's diet. I think we've learned something today. Ladies and gentlemen, we're so glad you tuned in, Intelligent Machines. We do this show every Wednesday right after Windows Weekly, 2 p.m. Pacific, 5 p.m. Eastern. That's 2100 UTC. You can watch us do the show live in the Club Twit Discord, of course, if you're a member of the club. If you're not, join. But even if you're not a member, you can watch on YouTube, X.com, Facebook, LinkedIn, Kick, and Twitch. We stream on all those platforms. After the fact, on-demand versions of the show on our website, twit.tv.com. We do audio and video. You can choose. There is a video version on YouTube. There's a dedicated channel for IM. Great way to share clips. And if you subscribe to your favorite podcast client, make sure you leave us a good review so that Paris can do a dramatic reading on next week's episode. We do not have a guest for next week. Maybe we can get this pan-gram. I actually, I think we, during the course of the show, I think we did lock down a guest. Let me make sure. We did. Things happen even when I'm not aware of it. They're working all the time for you. All the time. They are like little agents working 24-7 trying to come up with something. So good. We are going to talk to Ian Bogost again in a month because his book, The Small Stuff, will be coming out. And from bigspin.ai, we'll talk to Christopher Potts at the beginning of July. But we do have a couple of openings, or one opening, I guess. So do we know, Benito, who will be here next week for Intelligent Machines? I'm looking through my email right now. I'm still trying to find it. Somebody was booked. Yeah, I'm not sure if it was for next week, though, but I'm just not sure. All right. Well, I'm going to get my man kitchen pepper can out and try to find some guests by peppering their little behinds. Do you like pepper? Do you eat a lot of pepper? You don't eat pepper? I don't not eat pepper. I put it, you know, it's one of the spices. How do you eat cottage cheese without pepper? I hate cottage cheese. Well, there you go. Or salt. We have one pepper mill by the stove. I have a pepper mill. And we have one on the table. Well, the distance from my stove to my table is not far. So having two pepper mills would be, yeah, I could reach. Our kitchen is probably the size of your apartment in the suburbs. That's where life is like. Yeah, that's probably true. Yeah. Thank you, everybody, for putting up with us. Thank you, Paris, for being back. I missed you. We're so glad you are out from under. Now, does this mean you get a little break or do you have to immediately embark on another? I do. I am. I mean, I've got some stuff to take over the next week, but I've decided I'm going to take two weeks off. Nice. Are you going to travel? I think so. I don't know where. I haven't. I just decided I'm going to take two weeks off as of yesterday, so I haven't decided where. Paris always does the corkiest things. But I think I'm going to do a similar thing to what I did last year when I ended up visiting Yulia, which I think I'm going to do a one-way flight to somewhere, rent a car, and maybe go hiking in some national parks. Can I just do a road trip through some states in the U.S. I hadn't been to before? Good idea. If anybody has any recommendations for national parks that are gorgeous and lovely and not completely swamped over the next month, let me know. I think you have to go somewhere where there's also a Buc-ee's on the way. I'm open to that. B-U-C-E-E-S. Now, why do we need to go to a Buc-ee's? I've never been to a Buc-ee's. It's a phenomenon. It's an American phenomenon. I mean that. Is it like the 7-Eleven in Japan? Oh, no. Is it like Stu Laird? You really haven't seen stories about this? Well, I feel like I have. You get brisket? The brisket is to die for? Brisket at Buc-ee's? Oh, I like brisket. People go out of their way to go to Buc-ee's. If you look up B-U-C-E-E-S.com. It's the one with the beaver. Yes, the beaver. Now, I don't know if we have any Buc-ee's in California. No, you don't. You don't. It's a southern thing. The locations are- Alabama, Colorado, Florida. You must have been to a Buc-ee's. No, I've never been. I want to go. That's what I'm telling Paris. We should all go to a Buc-ee's. When are you coming to visit us? Let's all go to Huber Heights, Ohio. You clearly hate us, Leo, because you never come to see us or your son. Or my son. I'm currently thinking of going to Glacier National Park in Montana. Oh, that's good. I'd love that. I'd love that. Yeah. It's supposed to be very lovely. Pacific Northwest is very nice in the summer. Well, she did that the last time. I did. Oh, you did that last time. It was so great. I didn't spend that much time in Washington, honestly. I could do it again. Like, it's so good. I don't know. The world's my oyster. I want to stay in, I think, the U.S., though. Maybe Canada as well could be in there. But I'm trying not to lie. You should probably go to a World Cup game in one of the fun. No. If I go to a World Cup game, I will be stuck in traffic for the next two weeks. Are you going to any more Knicks games? I'm going to a watch party tonight whenever we're done with this. All right. Get out of here. Also, look in the Discord chat. If you scroll up, a friend just sent me a video because we went to a bar on Monday for Game 3. And you can see what stage in Game 3 this photo was taken based on our reactions. You put this in the Discord? I did. I'll add you with it if you'd like. I'm not seeing it for some reason. It was at 7.54. That's our time. So do your calculation. Whatever. Yeah, do your time. I'm seeing a lot of Bucky's posts. That's why. Everybody had something to say about Bucky's. Oh, it's not going well. It's not going well. If you zoom in, it's rough. It's rough happening here. There came a time where my friend standing to me on the left of the photo. She's a sports reporter. She was like, and the world's largest Knicks fan for her entire life. It's like, I have to stand up. I think this will change the vibes. And so we all, it didn't. It didn't change. You are at least wearing an orange scarf. So you are in the group. You know, I'm prepared. Very cute. My old boss, Steve Newhouse, gets like second row tickets. He goes forever. He's been the greatest fan ever. And he put up a photo of himself in the subway going to, and he looked concerned. I said, you look concerned, boss. Today? Yeah, today. He looked all happy. Well, there once again, I mean. For the first time. Got to win today. Apparently this ref today is supposed to be somewhat more reasonable, but. So really it was bad officiating? It was really bad. I mean, last game, the Knicks were not phenomenal. No, they were not. Even I could tell that. They made some, even I could tell that. Me, a person who's had to have multiple people to me explain the fundamentals of basketball over the last couple months. I could see there were problems. But there was also, I mean, there is one thing where Wimpy literally tackles Brunson, a Knicks player, to the ground. And they're like, yeah, no foul. No flagrant. Yeah. And a flagrant is like unnecessary touch or contact. Which was clearly not flagrant by any definition. How else do you get someone to the ground from your hands being on their head? You're not supposed to get people to the ground. I think that's part of the problem. That's kind of part of it. And then shortly thereafter, the Knicks got a flagrant because a player on the Spurs jumped up to shoot and landed on a Knicks player's foot. And they're like, well, you not moving your feet out of the way. That's a flagrant. See, this is how I know you just started watching basketball, because every finals is like this. I know. But I think it's beautiful that I'm getting to witness this without generations of latent trauma. So I get to experience it fresh. No, but that's what makes it even that much sweeter for New Yorkers who have been dry for 50 years. Oh. And I think that's honestly one of the best things about all of this is I, as a full bandwagoner, I mean, bandwagoner by proxy. I just decided to start watching basketball games two months ago, not really realizing. I was like, I'll watch whatever New York teams are happening anywhere. And suddenly a New York team is in the finals. Everyone I've met who's been a lifelong Knicks fan that could be like, go screw yourself. You haven't suffered. They're like, no, the Knicks are great. I'm so happy you're here. Here, let me tell you about why Dolan sucks. Yeah, that's a Warriors fan 10 years ago, because we were dry for 40 years until we saw That was a long drought for the Warriors. Yeah. Well, I don't know how we managed to prolong this show an additional 10 minutes, but we keep going. Pepper Grigers and Knicks. Three hours. All right. Thank you, everybody. Have a wonderful evening. We'll see you next time on Intelligent Machines. Go Knicks. Go Knicks. Hi there. Leo Laporte here. I just wanted to let you know about some of the other shows we do on this network. You probably already know about This Week in Tech. Every Sunday, I bring together some of the top journalists in the tech field to talk about the tech stories. It's a wonderful chance for you to keep up on what's going on with tech, plus be entertained by some very bright and fun minds. I hope you'll tune in every Sunday for This Week in Tech. Just go to your favorite podcast client and subscribe. This Week in Tech from the Twit Network. Thank you.