Intelligent Machines 879: Alex Karp, Alex Karp, Alex Karp

It's time for Intelligent Machines. Jeff's here. Paris is here. Our guest, Rafi Krikorian, is the CTO of Mozilla.org, and he just released his State of Open Source AI. We'll talk about open AI models and why we should all get behind them. That and a whole lot more. Coming up next on Intelligent Machines. This episode is brought to you by Black Hat USA. If you listen to this show, you go deep on the technical detail. Well, so does Black Hat. For nearly three decades, it's been where the security industry's most rigorous research gets presented and pressure tested. More than 100 hands-on trainings taught by practitioners who've actually deployed in live environments, not lecturers reading from slides. And hundreds of peer-reviewed briefings that go well past the overview into the real work across the four areas defining security right now. AI and autonomous threats, cyber conflict, systemic resilience, and identity. This year, Black Hat's Briefings Pass includes all keynotes and main stage access, plus business hall entry. You also get breakfast, lunch, Arsenal live tool demos, on-demand session access, and admission to the Midnight in the War Room screening. Black Hat takes place from August 1st to the 6th in Las Vegas. If you want the depth this show gets into in person with the people doing the work, this is the room. And we'll be there, too. Prices rise on July 17th, so book before then. Use code TWIT for $200 off your briefings pass at blackhat.com slash US-26. That's B-L-A-C-K-H-A-T dot com slash US-26. Podcasts you love. From people you trust. This is Choice. This is Intelligent Machines with Paris Martineau and Jeff Jarvis. Episode 879. Recorded Wednesday, July 15th, 2026. Alex Karp. Alex Karp. Alex Karp. It's time for Intelligent Machines. This is a show where we cover the latest in AI, robotics, and all the smart doodads and doohickeys all around you. I would like to, before we introduce our very special guests for today's show, introduce Paris Martineau, who is our very special panelist, investigative journalist from Consumer Reports. Hello, Paris. Bonjour. Hello, Leo. Hello. Good to see you. Also, Professor Emeritus of Journalistic Innovation of the Craig Newmark Graduate School of Journalism. You almost got there. You're getting faster. If I jump on it, I think I can fool Benita, but never. That's Jeff Jarvis, author of Hot Type, which emerges from hot presses in two weeks. You can order it right now from jeffjarvis.com. Hello, Jeff. Hello, boss. Let me introduce our guest who actually is a fascinating fella. Uh, Rafi Krikorian is the engineer. Silicon Valley calls when something important is broken. How about that? He fixed the fail whale at Twitter. Okay. Ran Uber's first self-driving fleet in Pittsburgh. Rebuilt the Democratic Party's technology after the 2016 hack. Stop me if any of this is wrong, Rafi. Spent six years at the Emerson Collective. Lorraine Powell Jobs' wonderful kind of, I guess you'd call it almost an accelerator for great journalism. And he did a podcast there called Technically Optimistic, which I love. Since the fall, Rafi's been the first ever portfolio-wide CTO at the Mozilla Foundation. It's great to have you, Rafi. Actually, one thing that was interesting in your bio, even though you ran Uber's first self-driving fleet, you got in an accident in your Tesla under full self-driving. Your kids were in the back seat. That's right. And you were concussed. That is right. That was just a few months ago. It was in November of this year. Is everybody okay? Everyone was fine. The hardware worked great. The software, a bit of a problem. Yeah, that seems to be the lack there with full self-driving. Well, I'm glad you're up. Do you still use full self-driving? I do not own a Tesla anymore. Oh, okay. Yeah, I assume getting in an accident with children in the car is enough to... And also, in fact, the insurance on a full collision on a Tesla is more than its resale value. Total. Yeah, it was totaled. May I ask where you drive now? I'm back to my 2016 Subaru, which I have completely hacked apart, so I'm very happy with that right now. And there's no self-driving in that at all, right? Oh, there's actually a self-driving in that. Are you using Comma AI or something in that? I have hacked a Comma AI system into it. And do you feel that's more reliable than Tesla's FSD? No, absolutely not. I just can do something to tinker with, just to be clear. Well, no, Comma AI is very interesting. We've tried to get George on the show for ages and ages. is he has actually kind of an interesting take on AI, but let's not – we'll talk about that when we get him on. Let's talk about – today you delivered a talk. The first, State of Open Source AI, and people can go to the webpage, which says, by the way, version 1.0 recurring. So this is not the end of it. Your conclusion – well, I'm going to let you summarize your conclusion, and we can look at some of the data. Yeah. I mean, I think I would say three things. One is that open source models are almost a parity for everyday use cases. Wow. Like, I think we looked across all the entire model ecosystem. We looked at data from OpenRouter. We looked at all these things. And it seems like for everyday use cases, not frontier work, which I totally acknowledge, that open source models are good enough. Like, people are, and the market is shifting toward them. So I think that's number one. I think number two is that the actual untested parts of the open source AI stack have moved to what we call the agentic harness. It's no longer the models. There's lots of reasons why people like to talk about models. But it's things like open code or things like the Hermes agent or things like that, which actually wrap the model. That's where the contested parts of the open stack remain. and then the third is sort of like the sovereignty finding of just like there are a lot of countries around the world which are eager for open source AI systems because in a lot of ways they're worried about the supply chain when it comes to American technologies and what happened with the mythos shutoff and so that sent a ripple effect across the entire world on the sovereignty end. So I think those are like the three big stories that have come out as you put this together. As you mentioned, we view this as a recurring thing. Expect to see 2.0 in a year. We'll probably have 1.1 in September, and we'll just keep on updating it so we have a good definitive sense of where open source AI is. Because it's moving fast. Everything in AI is moving fast. Well, six months ago, I wouldn't have talked about agentic harnesses, and so that has been a thing in the last six months. How are you kidding me? I'm a Hermes user, and one of the beauty parts of that is from session to session, even from turn to turn, I can change the model. I have all the models in there. And I use Ornith, which is kind of a distillation of Quint, and it's running locally on my machine. 99% of it, it is all I need for agentic work. But that is, I think a lot of people would say, wait a minute, wait a minute. Open weight models are, and by the way, I prefer open weight to open source. I say open source. Okay. Because they aren't open source. They don't tell you the code they're using to train or anything like that. But you can get the parameterization. You can get the weights of the model. And maybe more importantly to us, you can get the model. Exactly. If you have enough memory, you can run GLM-5.2 locally. And that's probably really mostly what we're talking about when we're talking about open- Do you ever change the weights, Leo? Do you ever screw with the weights? Never. I wouldn't have. No, how? But it is kind of cool. You can play with it. But still, I think when you say that open weight models are approximating the frontier models, I think there's a lot of people, especially people who are spending a lot of money on Fable and Saul, who would say, what? Yeah. I think it's a jagged frontier. That's key, isn't it? It's jagged. It is not the case that the open weight models are equivalent on all tasks. I'm saying that they are equivalent for about 80% of every day task. I mean, I equate it to I don't need a Ferrari to manage my calendar. In fact, it's probably annoying to drive a Ferrari to manage my calendar. They will ship not manage my calendar. I've found that out. But I think GLM-52, which is what I'm using, works great on my calendar. And if you run a hosted version of GLM-52, I too run it locally. But if you run it on Nebias or stuff like that, It's something like 20 to 50 times cheaper than running something like a Fable. Are there any use cases for where this is true that were surprising to you? I mean, like, I do think that, like, one of the biggest questions is can these things run overnight and not be talked to while they're doing things? Like, can they actually just keep on working on a complicated task while I sleep? I don't think the answer is there yet. Like if you try to get it to write a large amount of code that would take a couple of hours, even for one of these things to go do, they'll kind of go off the rails in the middle of the night after like three hours and making that number up, something in an order of three hours. Whereas Fable will probably get me through the night. Like he'll probably keep on going and I can wake up the next morning and be like, oh, that's great, let's keep going. And so I think like these long horizon tasks are where they're still not quite there yet. Talk for a second about the methodology you used today to compare them. And we talked about this last week. The benchmarks that companies use are kind of meaningless to us as consumers. So I want to hear about the methodology you used now. But then also I'm imagining the methodology is going to change. It's not like you have a standard test. So how did you do this and how are you going to do this in the future? So two things. One is we relied on all of the different benchmarks. So we didn't pick one particular benchmark. We looked at them all to try to understand both in identity cases. There's a bunch of benchmarks we use for everyday use cases and things like that. But then the other thing we did is just like more of a personal test. So we actually have a setup where we've recorded for myself, like literally every single prompt I've given to a system in the past couple of weeks. We call the system Morph. And what it's allowed us to do is just like record every prompt to record every output so we can replay it. So we can then replay all the prompts against different models. and just look at them. So it's not the most scientific case. We're not comparing the actual outputs rigorously because it's very hard to do that. But we can eyeball them and just be like, oh, this one got it close. This one didn't get it close. But I think the first part of just recording every little thing we do was a key part of it. Did you use the same prompts sent against each model? Correct. And we can replay them on different models. So the graph on the website is chatbot arena. Correct. Correct. You feel like that's a good measure? I think Chatbot Arena is a good one for just like what everyday use cases look like. So, yes, we do think that's a good one, but we're cooperating it. So anything we've put on state of open source, we have a good gut sense that is correct because we've looked at it through this lens as well. So this is based on Chatbot Arena. Correct. Going back to January 2024, closed models led open models by 8%, which is, I think, is that significant? That's significant. I mean, that's significant. So, like, if you give it a test of 100 different things, it's about eight of them that would fail comparatively. That's pretty significant. Okay. And at the time, that was LAMA, which wasn't the best open-weight model, but it was the one. But you see a precipitous drop to August 2024 when it goes down to 0.5%, and then, as we know, this is the big watershed, DeepSea car one came out in February of last year, and that opened everybody's eyes as to what an open-weight model could do. Yep. I think, in fact, it set a shock through the American AI industry. We're back up now to a bigger gap, but you estimate that right now. I think we're getting closer, and I think we're getting even closer. I think, like, the chapter arena data doesn't count for things like J-Line Fat 2 or the latest QE models and things like that. So I think we're actually getting closer. Yeah, GLM-5-2 is amazing. Yeah, because I think these new models showed up coincidentally, and I do think it's quite a lot, I don't think there's anything nefarious about it, when the stable set-off happened. It's around the same week all these new models dropped. We got all of a sudden, and it was, that rug pull by the federal government woke up, was kind of like the deep seek factor shock, where everybody went, whoa, this could happen. And actually now we're in this kind of crisis moment because the Chinese government has started to make noises about shutting down. All the really good open-weight models right now come from China. Is that right? That is correct. And they're talking about restricting that. I've seen people on X say, everybody download all the models. There's a new, it's called Hugging Bay, kind of a pirate's bay for models. People are nervous that we're going to lose access to these open-weight models. Whereas two weeks ago, I thought that China was going to undercut the entire USAI market. I thought that was the strategy. I think they're going to be free. So was there a flip-flop, an open flip-flop in their strategy, or was it really strong? Well, I don't know if I have a good answer to that. Like, I think they called literally everyone by surprise, and I don't think they're saying exactly why. Like, I'm fundamentally confused by it because, like, I do think that, like, that China's open model strategy is a little bit of their soft power strategy. So, like, I don't fully understand what the rationale is behind it, but I guess we'll find out soon. I think there might even be multiple parties at work here. I think the companies see this as a way to get market share, and I think the Chinese Communist Party sees this as leverage, and I think there may be other parties involved in this, and that's why there's kind of a – it's not like we're a unified country in our opinions either, I want to say. So the cost of inference has fallen dramatically. Correct. And the use of open weights has risen. This graph on the right, the share of tokens routed on open router through open weight models was essentially zero to now being a majority. Yep, it's kind of wild in my opinion. I think it's a question of just that innovation or experimentation curve that people are on. When they're trying new things, they're tinkering with the big models from Anthropic or ChantGPT or OpenAI. But as they're getting closer to production, they're like, we can't afford that. We can't afford, and we can't deal with variable pricing. We can't deal with the fact that the White House might shut it off. And so they figure out at that point when they're ready to get to volume, what's the right model for the right cost? And those tend to be open-weight models, it seems. and then they pin to it so that they know it won't change, behavior won't change, it can get good pricing, good volume kind of thing. So if China cuts it off and the U.S. cuts it off, are there other places to go? I mean, right now the answer is sadly no. I mean, like, look, I don't think it's a good place for us to be in a single closed ecosystem with just U.S. closed providers. I also don't think we're in good and an open ecosystem with only one country that provides it. So, no, we actually have a problem here as well. I mean, thankfully, other countries are starting to make a lot of noise that they need to go invest in it, mostly the Europeans because they want to not have that strong. Yeah, they don't want to have dependence on the U.S. supply chain. They also don't want to have dependence on the Chinese supply chain. So the Europeans are starting to make a lot of noise on it. They start making investments with the Swiss, with the French, et cetera. So we'll see where those go. Does Mistral open weight? Mistral does have a point of policy. Okay. What about the role of Nemetron in this? Yeah. I mean, like what NVIDIA is up to could be super interesting as well. Like they're not quite there yet on quality, but they could be super interesting. And the role of – I find myself in this weird position where I'm agreeing with Alex Karp about anything in life. I know. But arguing that companies are giving away their alpha, their business intelligence to the foundation models, instead come to him and trust him by making a saddle on open source models. Is he right in terms of how, not just morally, but I mean in terms of how companies are going to think? Is that going to win the day that open source might also be like Apache in the day versus a Netscape browser or server that open source might win the commercial battle? I do think it's very likely that we can get to a place that open source wins a vast majority of the commercial battle. I don't think it wins 100%. I do think there is a world where if we think about, like, an interoperable world, if you think about your Apache world, like, if you think about a LAMP stack, like, I think we get to a place where, like, all these different components are swappable. And so, like, we can try a closed-source model while we're still in experimentation stage, but when we get to, like, real production, we want to, like, own it and pin it. Like, it's similar to what Pinterest did. And Pinterest and Q4, they switched to an open-weight model that they self-host and saved them, like, $10 million that quarter. And so it's just like they have to get to the place that they knew what they were doing and stopped experimenting that they can actually select the right model and then deploy that. Of course, there are challenges to hosting your own model. Of course, you have to have an open-weight model to begin with, but you also have to have a lot of horsepower, RAM and CPU power. And it's very, thanks to the Frontier companies, very, very expensive right now. The best, I mean, I've been using DeepSeek Flash. China's priced that at pennies on the dollar. I mean, it's a tiny cost, but I couldn't run that. I could run a Flash version, I guess, but I can't run the full DeepSeek V4 Pro at home. I definitely can't run the full GLM-5 too. Are you running, you must be running a smaller version of GLM. Yeah, I run GLM-5, I run the version that can be quantized down to what can fit on a DGX Spark. So that's what I do at my home. But I think for these enterprises, I think this is just an OPEX, CAPEX calculation. At some point, the price of the tokens gets so high that it actually makes sense to buy servers and have the SREs. If you can get it. And then you're right, Jeff, that there's a huge concern. We're going to talk about it later in the news section about privacy. It's become very clear that when you use a cloud-based model, in order to use a cloud-based model, It has to upload everything. In fact, people just found out that Grok is, in fact, uploading more than everything. And that's a huge issue for a company. It's a huge issue for everybody, privacy-wise, but it's a big issue for a company with proprietary data. So that's another concern. There's a lot of incentive for a company to do this locally. Yeah, regulated spaces, for example. They're not going to want to upload their stuff. They're going to want to have it under their control. Well, and we know the Pentagon, for instance, does not run Fable or Mythos on Anthropic servers. They run it on AWS, you know, private government-hosted servers. That's going to be another big business, by the way, is private computers. Some assurance of that. There's so much regulation news this week. Devin Sassabas, with his proposal, Politico on Anthropic is out there state-by-state doing regulatory capture on their stuff. OpenAI has been giving away equity and so on. The White House has its new thing. And they're shy to do what China does. So there's a huge crunch coming here. And there's some fear that open models could be outlawed, could be restricted in some way. How much of a fear? That's our best competition. It's not one single company to the big guys. It's the institution of open source. Coming out of this and seeing how important open source is, and I fear that your report is a bit of a red cape to a bowl to the frontier companies. It's like, oh, open source is ever more a threat to them. Yeah, he's not telling them anything they don't already know. Exactly. But they're going to push ever more for regulatory capture. So how much danger could open source be in? No, I think it's open source under a massive attack right now. In fact, to the point that I'll be very honest, to the point that I think I underestimated how much attack open source is on until we put this report together. Because as we started talking about it, the amount of, I mean, I think the reception I would say is about 80-20. 80% is incredibly positive toward it, and 20% is outright hostile. And so I think we underestimated that. But I think I want to also remind us that there's the rest of the world out there. So, like, yeah, I think we're going to have a lot of problems in the United States when it comes to open. But I think the rest of the world is actually leaning in because they sort of see the same model that built the Internet, built Linux, stuff like that. But on the other hand, I went to a World Economic Forum event about two years ago in San Francisco, and there was a contingent there that said open source, open model, open weight is dangerous because people can take down the guardrails. God knows we shouldn't have them. So there's, which sounds a little European to me, too, is the regulate mindset is we must control. And the frontier companies say, well, you can control us. We'll write the regulation. But, you know, we'll be there and you know who we are and what we're doing. I think there'll be some set of people who are worried. The moral panic level of worry about AI could hit open source first. It's not just the frontier companies. it's also others saying this shit's dangerous and some think that's the most dangerous, I think that's wrong but they all say that. I agree and I think that a lot of that's actually going to come from the big frontier labs pushing that narrative I think the question that the other countries and we talk to a lot of them all the time is going to be a race between do they think about sovereignty or do they think about those fears or which ones are going to come first because I think if they want to think about sovereignty and I've heard this directly from a bunch of them their only path to it in this world is to start to build on those open-weight models, to start fine-tuning them to work with their country needs and stuff like that. So I think, like, cutting off the open-weight models, and maybe someone needs to connect the dots with them, could be undercutting their sovereignty plans, but, like, they at least need someplace to start right now. So you, we're again talking to Rafi Krikorian. He's the CTO of Mozilla, and just today published Mozilla's first state of the open-source AI. It's all online at opensource.ai. You say open ships easy, but deploys hard. What does that mean? Correct. Yeah, I mean, it's really easy to publish these open-weight models. Like, we're seeing them all the time. Hugging Face has, what, 13 million more models that you can download. But, like, the biggest thing, we did this developer survey worldwide where we asked developers, like, are you using open-source or closed-source? If you use open-source, when did you use it? What happened? Stuff like that. and open source has a huge churn problem. I think lots of people have tried open-weight models and open-source systems, but then they gave up because it was too hard to make this stuff that we've talked about. It was too hard to stand up. They had to figure out the talent to be the SREs. They had to get access to all these GPUs that they would deploy inside. So something like 70% of people who tried it didn't follow through with it. There's nothing easier than firing up Claude Code. Exactly. I mean, like, even I have this thing. Like, when I do a weekend hack, like, I'm hitting the OpenAI API, right? Like, so that's the thing that we as an open source community need to fix. Like, if we really want this path to go, we need to make it as easy as using the OpenAI API or get as close to it as possible. Well, that's why those open harnesses are so important and the OpenAI API from OpenAI. Well, so this is – your audience for this are Leo's, are people who are going to try to install this stuff. I think the audience is enterprise more than that. Well, that too. There's a lot of people who have a more fun. But enterprise is also a broader category that's going to just require a lot more fine-tuning. The Leos, I think, are going to be the easiest adopters. They're willing to get in there, get their hands in there. And I've got nothing to lose. The enterprise is like, well, why don't I continue to renew my enterprise Anthropics subscription? It's just going to be a bit harder. Well, I think one of the things that happened there is just like, you know, We ran this experiment in Mozilla AI, one of our subsidiaries, and one of our top engineers just did the math. He's just like, well, if I actually had to do API calls for everything I did all month long, it would have been $10,000. Instead, right now I'm paying a $200 Claude subscription. Is that going to end after the IPO? Is this going to follow the same thing as rideshare pricing? So that could actually be one of the drivers, too. It is a risk, and if it happens, you better be ready. It's like you can't, if they do pull the rug, you can't just say, oh, well, now I'm going to start using GLM. Yeah. Now's the time to start planning and thinking about that kind of thing. Well, is there, like WordPress in the early days was so smart versus its competitors. By being open, it enabled its competitors to become hosts just like WordPress is. That's exactly the point I was going to make. This reminds me so much of the early days of open source software, where there was a, you know, yeah, you could always use LibreOffice, But Microsoft Office is a no-brainer. Nobody ever got fired for buying IBM, right? Well, IBM is locked down. It's the same battle, though. It's just moved forward into the AI sphere. And I think we now know, looking back, that we should have probably supported open source a lot harder. And now maybe it's time to support open-weight AI. So the reason I was asking about the audience for this, because I think you're right, it's enterprise and it's, you know, what do I call it? A hobbyist? More than that. Enthusiast. Enthusiast, right. Paris is an investigative reporter at Consumer Reports, and she doesn't speak for Consumer Reports on this show. But when do we think that AI gets to the point that consumers are going to want the same kind of judgment about the various AIs that they would get from consumer reports about refrigerators or cars or anything else? No consumer is saying, should I use an open-weight AI? Not now. Not now. If I tried to say that to my mom, I think her head would explode. And just something would spill out of her ears. You know, mom, your ears are the wrong saddle. Well, Harper Reid convinced me to buy this Chinese ESP32-based AI orb that goes right to deep sea. You connect it to your Wi-Fi. What could possibly go wrong? And then immediately connect to China. I don't think mom wants this. But mom does want. I mean, that's why Siri is so important. Apple's move with Siri is so important. Mom does want some easy way to use this stuff. Well, if you get to that WordPress hosted world where people can do things with it, when does it become a consumer industry? It's not yet. No, I mean, there is no reason it can't happen now. But I think the problem is that it's the same thing that happens with privacy. I mean, it's not great. People theoretically think they want privacy. And in their heart's heart, I believe they do. But they're going to trade convenience for it at any moment. And so the exact same thing is happening. There's a paper out of University of Maryland, I believe, just earlier this year, which did an analysis of what the chatbots recommend when you ask it to do shopping questions. So maybe the Consumer Reports, for example. And they showed that over 50% of the chatbots were actually recommending sponsored goods. And you have to ask yourself why that's happening. But most consumers might not care. So we need the Firefox of this moment. We need something that's convenient, elegant to use and things like that that also protects consumers. And I don't think they're going to do it by themselves. So does Mozilla do that? Mozilla could be one of the people who do it. But I think where I'm really focused on is that if we can build really good tools to enable people to build on the open source ecosystem, then maybe a thousand of those could show up. I worry Mozilla will only build one, but I want lots of them to happen. That's really interesting. You make the point also that open isn't a vendor choice. It's a sovereignty choice. This is not sovereignty. Data sovereignty, governmental sovereignty. It's very important. And AI is a non-trivial new technology. God, how I hate agreeing with Alice Karp. That was his message. That was his argument. Yeah. But, like, I mean, the same thing applied to cloud back in the day, right? Like, you're not going to talk to any business these days that are just like, I only build for AWS. No, they build like in a generic way so that when the AWS bill shows up, they're going to be like, well, screw you guys and we're in Azure. Like, so I think we need to get to the same mentality when it comes to like our token providers. Although somewhat this sovereignty conversation ends up being, as it is in the EU, more about which government is going to control this. And I don't think that's a solution for us. You know, one of the reasons China is succeeding is because they've had such a light hand, light touch regulatory-wise, and they are such a capitalist society, and there's so many companies who are scrambling, and even though they haven't been able to get the hardware, they haven't been able to get the chips, they've found ways around it, maybe distillation of American models, I don't know. But I think the lack of regulation has really helped the Chinese models. But a lot of times when you say sovereignty, governments just say, oh, yeah, get out of the way, here we come. Yeah, sovereignty, that's the wrong word, but the ability to have choice, I think, is really important. Because I think that, like, you know, one of the pieces of regulation that could push this around is sort of, like, data locality rules, right? Like, I don't want my data to leave my country's borders because I don't know what those guys are going to do with it. And, like, that might force a bunch of these conversations as well. Yeah. We're almost out of time. I'm thrilled that we can talk to you, Rafi Krikorian. and I'm also thrilled that you did not survive your Tesla crash, to write this, the Mozilla State of Open Source AI, which is available at stateofopensource.ai. I gave you the wrong URL out earlier. People should download it. They should read it. You can read it on the web. What would you like to see next? What do you want to have happen as a result of this report? Yeah, I mean, I think that, you know, there's a whole alliance of people who are building to the open, right? Like, Clem from Huggy Face just tweeted, I think, this morning that he's going to show up in San Francisco and they should do a rally on open source. So I think we need, like, those types of things. I'd love to see the demographics of that rally. They're all going to look like me with, like, white in their beards. It's going to be a lot of men in polos. It's going to be a really interesting mix of polos. Just take a great tiki churches and you'll be over. But I think that, I think we need more and more people. Like, we need more proof points. We need more examples, and we need more learnings, right? Like all these companies are deploying large amounts of like what they called FDEs, right? Like the forward deployed engineers that's getting their software out and embedded into Fortune 500 companies left and right. So we need the counter. We need the alternative. We need the third way to show up here. And I think the technology is ready. So now it's a deployment problem. So now we need people that are trusted enough to deploy it. Is there revenue needed for the development of more open source models? Yeah, I mean, I think the amount of money that these companies are going to spend on marketing this year alone dwarfs the entire money in Mozilla's and downright. So I think, yes, money is always helpful. But at the same time, we're seeing people even like A16Z. We're seeing investors throwing lots of money toward open source right now. It's unfocused. It's all scattered. We just need to actually, we need the lamp stack of AI. Like, we need that kind of rallying of, like, we're coming together to build software that can be deployed like an Ubuntu or deployed like a Linux or things like that. I think, honestly, we also need a lot of smart innovators to work on ways to get models smaller, to work on ways to get models smarter, to think about models that slice up the problem space so that they do specific things well. And I think that's going to take – this is what's interesting. I think it happens every time there's new technology. People talk about job loss. But I think there are going to be, in fact, many jobs created for people who are smart about this and can create something of real value. Models for mom. Yeah. I'm very grateful to Mozilla. I use Firefox. And I think if it weren't for Mozilla, we would be in a one-browser world. I think open source really is super important. And I think open-weight models are equally important. And as you say in your report, we bet on open the first time, open one. Together we can do it again. Can I ask one more question? Go ahead. So, you had a busy week. I don't know if you had a chance to read Devin Sasabas' proposal for regulatory model. So, in summary, he's proposing a private public FINRA and that there be a 30-day period of judging. and this always comes from the people who are behind, by the way. That too, yes. Slow down! Let us catch up! Well, but we'll get to, I presume, later the Anthropic commercial, which is really interesting where... With the gravestones. Yeah, the gravestones, where they're also saying, no, stop right now, we're ahead, let's just stop everything else right now, because we're there, right? And but but but I wonder. So I'm on another show. I was talking about Jason Howell earlier today and has thought about it. I'm nervous about government regulation of A.I. I'm nervous about this FINRA thing. Still, the people who now have the power establishing it. And so the fact that you came in and you judged AIs, you judged them against a, you know, how good are they? do you also come in at some point and delve into the how dangerous are they? Or because we need independent voices to judge AI on quality and risk and so on rather than I think thinking that we can create some officialdom to do it and I think to empower Mozilla and academics to do it universities to do it is, to my mind, the best way forward. So, long-winded, Joe Scarborough-like question to ask, what do you think about the various regulatory schemes that are being presented these days, pluses, minuses, and alternatives? No, I mean, I think we're all being distracted again by what's at the true edge and frontier. So I don't think for most businesses the true edge and frontier is what actually matters. So I think it's in the best interest of all these companies for us to be talking about what is the Ferrari look like. And I'm saying that most of us only need a Camry or a Toyota kind of thing. So I think we need to separate those conversations and allow us to go work on the 80% use cases and allow people to actually build real businesses and actually get real diffusion of these type of technologies. And then we can have a conversation about what the frontier governance looks like. I still maintain that open makes a lot of sense there because then we can have a conversation about what the open guardrails look like. We have questions about what does open permission systems look like. I don't want to be in a world where the decisions for the entire world are controlled by, what, seven Silicon Valley CEOs plus one person in the White House. That seems crazy to me. But if we can have a way that I actually have, what you're calling like an open edification system with open guardrails, open techniques around management of it, I think that's a way better world for us. I'll read from the state of opensource.ai webpage. Our belief is simple. The path forward is competition and interoperability. We believe in a world of many models, standard ways to plug them together, and the freedom to walk away from any vendor at any time. I think that's the world we want. This is the world we've been advocating for on this show pretty much from day one. Rafi, thank you so much for your work. It's a pleasure. We appreciate it. Thank you for joining us on Intelligent Machines. We'll be right back after this. This episode of Intelligent Machines brought to you by Gusto. We'll have more in just a bit, by the way. When you run a small business, you don't just do the job. Trust me, I know. You're also the hiring manager, the payroll department, the benefits team, and that's usually just before lunch. Gusto takes a few of those off your plate quickly and seamlessly. Gusto, G-U-S-T-O, is online payroll and benefits software built for small businesses. It's all-in-one, it's remote-friendly, and it's incredibly easy to use, so you can pay, hire, onboard, and support your team from anywhere. We love that because we are a remote business, you know, and it really does present unique challenges, But Gusto is there to help. Automatic payroll tax filing, simple direct deposits, health benefits, commuter benefits, workers' comp, 401K, you name it. Gusto makes it simple and has options for nearly every budget. Unlimited payroll runs for one monthly price, no hidden fees, no surprises. You'll save time with built-in automated tools like offer letters and onboarding docs, direct deposit and more. You get direct access to certified HR experts to help support you through any tough HR situations. It's quick. It's simple to switch to Gusto. Just transfer your existing data to get up and running fast. Plus, you don't pay a cent until you run your first payroll. Gusto is ranked number one on G2's highest satisfaction products list for 2026. That's pretty cool. And is trusted by over half a million small businesses. Try Gusto today. at gusto.com slash machines and get three months free when you run your first payroll. That's three months of free payroll at gusto.com slash machines. One more time, gusto.com slash machines. Make sure you use that URL. That's how you support the show so that they know that you saw it here. Gusto.com slash machines. We thank Gusto so much for their support of intelligent machines. Well, this was, as always, a huge week in AI with lots of news. I guess we should start with Apple suing OpenAI. This is the most public breach. They were partners. ChatGPT was the thing Siri went to when it couldn't handle your question. It still does, but I have a feeling that those days may be numbered. Apple alleges that a 24-year executive, he was at Apple 24 years vice president at the design. Oh, he wasn't 24 years old. No, he's been there. I was about to say, not 20, I thought you meant 24-year-old as well. No, no, 24-year executive. He's a 24-year tenant. No fool, supposedly. But he was a member of Apple's executive suite. He was their chief hardware officer. He had been involved in the design of the iPhone, the AirPods, the Apple Watch. I mean, he was so trusted that when he decided in 2024 to leave Apple, Apple said, good, you can take your time. We want you to train your successor. No hurry. We're not going to lock the doors behind you. Maybe they should have, at least according to this lawsuit. And, again, this is Apple's side of the story. OpenAI says, we don't steal companies. We don't need kids. We don't do that. Of course, if you don't trust OpenAI, and I think there is kind of this general thought, especially after the Ronan Ferrer article in The New Yorker, that maybe Sam Altman is a little bit slippery, this might just play right into that narrative. They accuse, OpenAI accuses, Apple accuses OpenAI of, well, so I can give you the scenario. Tang Tang left in 2024. He and another member of the Apple technical staff went to I.O., which Johnny Ive, you may remember him as Apple's head designer for many years, had started, went to work for them. And then a couple of years later, you may also remember Sam Altman and Johnny Ive walking into a bar. Looking like they were about to announce a pregnancy. Yeah. I have some bad news for you. So $6.2 billion acquisition of IO products. Along with that comes Tang Tan. By the way, we should mention Johnny Ive is not mentioned anywhere in this lawsuit. That's actually kind of interesting. Isn't it? They say that Tan, when he left, brought company secrets with him and emailed Apple suppliers and said, Hey, I might have a job for you. told job candidates interviewing people who are still working for Apple to bring actual parts from Apple to the interview. To which the job candidates repeatedly said, we're allowed to do that? Well, at least one did. One did, yeah. I think the suspicion is that that's the guy who went to Apple and said, hey, you know what they're doing here. Because I think this is what happened. I mean, the amount of details in this suggests to me more than one person turns to Apple. Yeah, but it's all Apple's point of view. But Apple is also notorious about collecting this sort of information slowly, methodically, and then waiting so that they can strike. Well, they claim that they sent a cease and desist order to OpenAI. OpenAI says, yeah, okay, your outside lawyer sent it to Wang, but he should have sent it to Chang, and because he sent it to the wrong guy, we didn't reply. With the Chinese name. You were confused. This story is amazing. It was a 41-page complaint, which reads like a great novel. I mean, you've got to read the pleading. It's just bad. It's one of those complaints that's classically written for a mass audience. Sorry to interrupt you. Well, that's an important point, because that's an interesting question, is why, Apple? What's going on here? Are you trying to stop OpenAI from making a hardware product, which everybody says they're about to do? In fact, we'll talk about what the leak says that product will be this year. Are you trying to put the kibosh on their IPO, which is perhaps imminent? What's going on? Or maybe you're just really hurt. I mean, Apple is kind of famously, if not litigious, they're famously maniacal in a way. a vindictive and maniacal in a way that borders on litigious, regardless of whether it happens in the actual court of law. Like this sort of, the amount of details and tenor of this complaint wasn't surprising to me as someone who's just tangentially followed Apple both in and out of the courtroom. Let's also remember Gizmodo and sending Apple cops after a reporter who came across a stray iPhone. Right. Oh, what a story. The interest is, of course, everything that Apple's asserting here can be proven or disproven in court with discovery. You know, discovery is always a double-edged sword, as Apple learned in the Epic lawsuit. Sometimes stuff gets revealed in your own secrets that you maybe don't want revealed. But yesterday we were talking about this on MacBreak Weekly. There was some speculation that I think Andy Yanako said they want to turn open AI upside down and shake them really hard. Yeah, and I think the thing that's just important to emphasize here is that Apple, more so than any tech company I know of, is maniacal when it comes to leakers. They will follow the ghost of a leaker to the end of the earth and then write a 40-page complaint about it just to try and stop people from following in their footsteps. They're one of the few companies in Silicon Valley that still to this day, even in an age of constant leaks and inside reporting, will really go after people for it. So I think that this being OpenAI, of course, they're going to turn what was already something up to 10. They're going to turn it up to 22. It'll be very interesting to see what happens. You know, it could, there's a number of scenarios. It could go on for five years. I mean, it could really not harm OpenAI's hardware efforts because it could go on for so long that by the time Apple got a judgment, OpenAI would be on the eighth generation of whatever they're doing. Well, the rumors we're going to get to about OpenAI, I think sounds like a really dumb product. Might this have affected their rollout that, oh, we were going to come up with a phone, but bad timing. Let's come up with something stupid instead. Remember, there's no weight of law here. This is just a lawsuit. A complaint. Anybody can say anything in a complaint, people. I will also say that. For doing from doing anything. One of the people who left fairly recently for OpenAI is the head of glasses at Apple. So OpenAI has the guy who was working for years, I'm sure, on Apple eyewear, on specs. Does, you know, OpenAI's first product, the rumor is, will be a wireless speaker. No screen, just a speaker. I don't know if Apple would be threatened in any way by that. It's unclear. I mean, I think you're right, Paris, that really Apple is doing what Apple does, which is they hate it when there's leaks like this. They think that they were wronged badly, and they may just simply be pursuing it because that's what you do. They may not have any motive, any subtext. I mean, I think it could be both. I think they could have subtext. I think they could be a bit annoyed at OpenAI kind of encroaching on the cool, slick tech company role that Apple has historically held. I mean, is there any reason why it would stop Johnny Ive from continuing to go ahead with whatever he's working on? Well, I guess I don't know, but instinctively, I'm not sure that Johnny Ive is who they're trying to get back at in this. I think that they're trying to stop general brain drain and the sort of leaks. Maybe a message for current employees. Oh, no, it's entirely for current employees, I would say. It's for current employees. It's for people who have left that are thinking about leaking information about supply chain. Basically, any of the inner working of a company like Apple, they're incredibly protective over it. It's like every company. But Apple really wants to defend that in either court or in kind of like pre-litigation demand letters. They have just taken the hardline approach. The best way to stop any of this information from leaking out is to deter people. They're doing the stick rather than the carrot. So in that case, it wouldn't – OpenAI has hired 400 Apple employees over the last – I mean, I'm sure that's – That sounds like a lot. No, I bet that they're taking a consideration. I'm sure more have gone to Meta, right? I mean, a bunch have gone everywhere. But I think part of it is them trying to be like, hey, all of you Apple employees, for everybody else. No, I mean, both to the people who are going there, but to the ones currently there, they're like, don't even think about thinking about telling them a secret code name of a project you worked on. Well, I imagine patting down employees at the spaceship door. I mean, that's like the level of stuff they do. They already do that. So you can't legally stop somebody from taking what's in here, their brain, to stop them sharing it. Yeah, you definitely can stop them from sharing it. those are trade secrets not in the state of California you can't anyway no you definitely that's what this whole thing is about okay that's an interesting question so you can't bring parts with you you can't bring documents with you you can't bring your supply chain connections that are from Apple yeah or if you know that Apple is going to next build a drone airplane you can't tell anyone that. You sign an NDA to that effect. Well, that's a different matter. So you may have an NDA with Apple. Oh, every single one of these employees has an NDA. That's what they're talking about, yeah. Yeah. But this isn't a lawsuit over an NDA. Yeah, but the point of this lawsuit is to acutely remind all of the employees that, hey, if you are trying to break your NDA, we are watching, we are ever vigilant, and we will come after you. We will define great secrets broadly. Yeah. This is something that I've seen a lot with other journalists that cover Apple. Apple is one of the companies that every once in a while you'll have a kind of big scandal or a big in tech journalism world where you'll see Apple suing somebody for leaking to a journalist, and they've tracked it through some complicated array of a work phone pinged there, a computer connected to this, and will try and come after a leaker there and then get the journalists all of their other Apple sources. So partly because of Apple, one thing that's changed a lot in California is this kind of enforceable NDA. For instance, we no longer allow non-competes in California. There are a lot in other states of the union, but you can't do it in California. And the thinking is Steve Jobs has always done this. The idea is a company could use these kinds of agreements to keep employees from looking for jobs elsewhere and improving their lot, getting a better pay package. And Steve Jobs is famous for trying to thwart that. And because of that, I think the state of California cracked down on a lot of it. And NDAs are enforceable, particularly with trade. I'm looking at a law firm now with trade secrets, recipes, algorithms, or manufacturing processes, customer and supplier information, intellectual property and proprietary business strategies. But the agreements have to be very specific about what you can and can't disclose. There needs to be compensation, exchange for signing an employment or bonus. There are a variety of laws in California. there are a lot of things that are not enforceable in California. Oh, am I aware? It's much more complicated than it used to be. A big part of my job is being, whenever I was talking to tech employees, being able to explain to them what is and what is. You know. So Apple's job is not maybe as easy as this pleading would be. But it's made easier by the fact that what they're trying to do is just scare people. Yeah. And that is so easy. They've completely accomplished that. The question is, have they thwarted OpenAI in their plans for the next few years? Maybe they've hurt them reputationally. You know, that may well be true. But I don't think if they have a hardware product they're about to release or plan to release in the next few years, this is going to stop them on that. I mean, yeah, I don't think that anything is going to stop opening eyes. You could always, I think a lot of it, you know, you just saw the back and forth between Elon Musk and Sam Altman over how untrustworthy each of them was. It was actually hysterical over the weekend. You know, Elon said, you see, this guy's a liar and a cheat. I told you so, even though his case was thrown out. to which Sam Altman said, yeah, have you looked at what Grok is doing to people? We mentioned that. We'll get to that in just a little bit. You also mentioned, and I think we should talk about this a little bit, Demis Hasibis, it's time for a global AI watchdog led by the U.S. Man, what do you think? I don't know if I like this idea at all. It's public-private. It's a FINRA. He says that you should take, that the high-end frontier models should have a three-day. FINRA is the Financial Industry Regulatory Authority, which has both independent experts and government. So it's private-public. A 30-day period of inspection, which to me is bad for open source. That's a real problem. But then he's saying, well, but only the really important models. The other one, others can be accepted, but who just should define that? Exactly. he also said this is because artificial intelligence is only a few short years away you mean AGI yeah, it's silly right, and so you've got him proposing that from the Google perspective Politico had a story today about how Anthropic is going state by state with the same law trying to get regulatory capture there. They want a kind of a federalist regulation that all the states agree on a single law that everyone agrees on And then opening eyes. Has Sebas has been lobbying the Trump administration for this. As has Sam Altman, has been lobbying Congress on this. It feels like regulatory capture. It really does. And I was talking about this with Jason earlier today on the inside. He asked, should something be done? And I said, but what? We don't know what it is. It's too soon to know what it is you're going after. What are the harms? What are the causes? And to think that there's this one body that can then tell, okay, it's taken care of now. Are we looking at privacy, at copyright, at childhood harm, at environment, at equity, at discrimination? What? What do they regulate? What is it you're regulating for? Extinction of humanity? Yes. I admit my inclination, which I know is wrong, is just let it all be and let's just see what happens. But I understand the fear of what could happen and why people feel like we need to write. Yeah, but if the discussion is on that basis of destroying mankind, then it's all stupid. Right? And the problem is a lot of discussion is there. It's not on the sarcastic parrots, you've got to hurt the environment. What if it's not destroying mankind, but what if the discussion is destroying the world we live in because of all the security flaws that will be revealed by these tip-top models? I mean, that's more concrete. That's not AGI. It's what, unfortunately, Anthropic brought upon us. But isn't that inevitable? It's just a question of when? Well, that's what our friend Alex Samos would say, is all these models can do this, and there's no way to stop them. And it may make things more secure. I also am just not super convinced about the idea that, yeah, there's going to be a day sometime soon where a switch will flip and suddenly everything will be super insecure because some random model is suddenly accessible. Well, yesterday, Microsoft issued its Patch Tuesday. Yeah, it is a patch. The largest it ever had. That's different than suddenly everyone everywhere is being hacked. Patches are good. Obviously. Patches are good. Except that these bugs were found by AI. Yeah, they weren't exploited before. Yeah. No, no, there were three zero days. And I think it's safe to say they didn't get them all. Right? Okay, well, if this technology is out there already, because it's clearly out there enough that they're patching it with Fable, and I assume that Fable isn't unique in the entire world. We don't actually know what they're using. Let's see. Whatever they're using. It's just a harness for other models. And I think it's probably... Whatever they're using is not probably unique. Or, yeah, whatever they're using is not unique to those Microsoft engineers. Why has every website in the world not been taken down then? Because I just, I feel like nothing is going to be as dire as the Doomers say it will. Yeah, I think there is a flood of zero days. Are you kidding? You're on the Doomer world now. Well, I think right now we're seeing more security flaws than we've ever seen before. Well, Theo's on the more powerful than you know world. No, no, no. But the question that is timing. I'm not making any assertion at all. I'm simply pointing out that there are more security flaws being exploited. There are more zero days now than ever before. I don't know where they're coming from. But in timing, can they get found before the bad guys find them if the tools are in these hands, right? Is there not a possibility that this is good, to Paris' point, because they can be discovered before they get exploited and fixed? Isn't that the optimistic way to look at it? Well, you know, Steve had an interesting point yesterday, and it's maybe debatable. Certainly Richard Campbell and Paul Thorat debated it today on Windows Weekly. Steve said, this is good. You're going to see a curve of Microsoft's patches. by the way, they did more than 1,000 fixes this month, that Microsoft's patches will suddenly go up, up, up, up, up, and then as things get fixed, they'll go down, down, down, down. He said within a year you're going to see almost no flaws, to which Richard said, well, here's the thing. Those 1,000 flaws, Microsoft's not looking at the entire Windows code base. They can't. It's too big. They're going section by section. and it's like painting the Golden Gate Bridge. They may never get to zero flaws. By the time they finish, they're going to have to start over. They discounted the idea that we could fix everything and make it all so good that there's no more flaws. I'm more on the Steve Gibson camp. I think, I don't know how soon it's going to happen, but I think at some point a lot of the bugs, if you look at the flaws that Microsoft fixed yesterday, most of them were memory fixes, the kinds of mistakes programmers make but AIs don't make. And I think it's – I mean, but then there's the question, well, is that it? So go back to the question. If Demis wins and you have a FINRA, what is it looking for? Is it just that? Is it just vulnerability? Here's the risk. The thumb on the scale, that's the risk. Well, if you had a legitimate agency, I'm not questioning the legitimacy. What do you think of the fantasy world where we can have legitimate agencies? The first fantasy is some measurable way to look at an AI model and say, this is good, this is bad. I don't know if that technology exists. So that's problem number one is who's going to, what metrics are you going to, we can't even do benchmarks. We can't even make a model that can't be jailbroken. but let's assume fantasy number one, there's some body that can come up with something that can find flaws in the AIs and can do it in 30 days by the way and do it with such accuracy that when the AI is released on day 31 it's safe everybody that's a lot of fantasies and everybody has faith in the thin rock and there's nobody like, I don't know, let's say the president of the United States who might say, you know, I really don't Pete Hegseth tells me Anthropic is a supply chain risk. Let's not let that one out. I think this is a non-starter, personally. Then you also get the weird thing out of the White House, and I didn't read in this enough yet. The Golden Eagle. That basically, anything that's about the same as China is okay, because China has it out. So then China sets the agenda. Well, that's not what you want either. As we talked about last week, and I didn't believe you when you said it. I couldn't believe it, but it's true. China is, the Chinese Communist Party anyway, is thinking about shutting down their open way models and preventing the United States from having access to them. So the White House has something they're calling the Gold, this should, just the name alone should raise issues. The Gold Eagle Clearinghouse for AI Cyber Threats. In celebration of our 200th anniversary, anniversary, a federal clearinghouse. I mean, first of all, I doubt there's anything really happening. This is just a fantasy. But anyway, a new federal clearinghouse for sharing AI cyber threat information between government and private sector. The Trump administration said the project is already receiving threat intelligence on cybersecurity vulnerabilities. Amazon just said something. And prioritizing patching. Gold Eagle will be managed by, oh, the Department of the Treasury. That makes sense. They're experts in all this. They're going to own pieces of open AI. With contributions from CISA, the Department of Homeland Security, and the Department of Defense, as well as open source software providers and critical infrastructure operators in industry. This sounds like what DeSibis was talking about. Under President Trump's leadership, the Treasury Department is working hand in hand. This is Scott Bessent. This is the guy literally who stopped Fable. So I don't know. I don't know. Mark Wayne Mullen, the fabulous Secretary of Homeland Security, said they will also further explore ways for the technology to be leveraged for cyber defense. My buddy Alex Karp has some ideas along those lines. I think we're reaching the maximum number of Alex Karp name drops. I think we are. In Twit history. Alex Karp, Alex Karp, Alex Karp, Alex Karp. Just trying to get there. Is it like Beetlejuice? I hope he's not going to appear. Oh God! Alright, let's take a little break. We'll come back with more. You're watching Intelligent Machines, Paris Martino, Jeff Jarvis. We are glad. You and Alex Karp are with us today. In spirit. He might be watching. I don't know. Shout out to Alex Karp out there. Get in the comments. Ladies and gentlemen, our show today brought to you by Monarch. I love Monarch. I use Monarch. But I've got to tell you, my subscription recently, I bought a year. I buy a year every year. And I subscribed and I bought a year. So my subscription came up and I thought, well, I really ought to look around and see what's out there. And I literally spent a morning installing all the other guys. And I said, what am I doing? There is nothing as good as Monarch. I love Monarch. So, yes, I re-upped. Monarch is the personal finance app that tracks everything. I put it all in there, accounts, investments, savings goals, spending, retirement plans. plans with Monarch, I'm literally relieved that I can see for the first time ever my complete financial picture. And I know where there are gaps. I know where I need to do something. It's kind of like having a financial advisor in your pocket. It's just more than kind of like it because of Monarch's AI that's built in that actually encapsulates the knowledge of real financial advisors so you can ask questions and get real advice. It's pretty cool. Most apps will tell you merely this is what you spent last month. Monarch helps you set goals. It helps you map out big purchases to see if you're actually on track before it's too late to adjust. You can ask this brilliant AI assistant. It's built in the Monarch. You can say, I don't know, how much did I spend on travel last summer? Can I afford that vacation this fall? Is it going to hit my savings or can I save up fast enough? You can spot things you wouldn't think to look for because there's also AI insights. Has your spending gone up or is that just inflation? You get a heads up on what's happening with your money. There's a weekly AI, weekly recap. It'll flag spending spikes. Things that you, you know, it sees things that you don't see because you're in the middle of it, right? Net worth shifts, upcoming expenses. There's lots of little quality of life features in there, too, like Monarch's Bill Split, which lets you scan a receipt. Everybody says, that was mine, that was mine, that was mine, and then it settles up, and you don't need a separate app to do that. But it's a little quality of life things, because these people at Monarch are really making a great product. Write your own money story with Monarch. Use offer code IM at Monarch.com. You'll get your first year of Monarch Core half off. That's just $50. 50% off your first year. Tap monarch.com with the code IM for 50% off your first year. Monarch.com. I don't know why there's a picture of Paris with a carp in the Discord. Oh, no. I summoned the carp. You summoned the carp. Oh, I get it. Not Alex Karp, the fish. Actually, this was going to be one of my picks, but I will mention it. Can we get, whoever did that do a slot picture of me holding a luxurious carp? So it's an Alux carp. I want, does anybody have, I bet you have one, Paris, a big mouth billy bass lying around? No, but I should. I'm going to order that right now. And look at this. This should be the next open AI device. It is a GitHub project by Morgan Willis. Perfect. It's called Bill AI Bass. You attach an AI to Big Mouth Billy Bass. God, can you make Big Mouth Billy Bass curse you out? Because that would be my dream. Let's listen in here. I'm going to turn up the sound. I'll take you back on this fine wooden plaque ready to drop some base and crack a few jokes. This is Billy. He's having surgery today. We're going to give him a brain. This brain will run on a Raspberry Pi and is powered by an AI agent that I created using the Strands Agents SDK. See, this is cool. This is a local AI. That calls a model hosted in Amazon Bedrock. Oh, not fully local. And for enterprise-grade fish security, I deployed an IoT certificate to the fish so that there are no hard-coded credentials. Wood and wires, pal. No wood and wires, pal. No magic here. Just good old reliable tunes. And just like the original Big Mouth, Billy Bats, you can hear the gears going, Yeah. It used to just think, don't worry, be happy. What's the thing? You have the talks that your father had on his desk full of yours? Oh, you could do it to that? Yes. The smart guy? I couldn't agree with you more completely. Thing is, he's already perfect. I mean, that's the thing. He's perfect. And I couldn't ever change it. Maybe I could get another one and mess around with that one. That would be my third one of these I've purchased because I did. How I obtained this is I was on a vacation with my parents. We were reminiscing, and I drunkenly bought two on eBay because they don't make these anymore. And I sent one to my parents and one to me. So I could buy a third. You got a backup. Send one to Malise, and the whole family will be. It's true. Like you said, Malice, the AI, want to not tell her until one day it awakens. And it's like, hello, Malice. You know, I'm watching you. The next big thing. You can get 11 hours to do the voice. You can get 11 hours to copy the voice, too. You can't. I couldn't agree with you more completely. Malice, you're right. When you're right, you're right. You're right. Actually, you could use the new chat, GPT Live. I guess we don't use the word chat anymore. This is a new generation of voice models powered by GPT voice. Oh, I've got to watch the Yentas. I haven't seen the Yentas. This is three Yentas, a.k.a. older women. Are we doing this again? Are we watching an advertisement here live? All right, we will. No, we should. We should. I'm just going to complain about it a little bit. The idea is that you can interrupt it. I don't know why, but this seems to be the holy grail for these things, is they don't just talk and then stop talking and then you talk. You've got to be able to interrupt it a little bit. I'm making a sweater for my grandson, but I don't want the needles to be too big. What do you think? If you go up, it's going to get kind of baggy. Although baggy is very popular now, don't you think? It sure is. I know, right? I'm ready. It's very natural, right? You can have a conversation. she was just being kidnapped she said don't take me Emma no don't arrest me that was a good place to freeze it anyway I haven't played with this but it is a first step into what I wanted all this time which was an AI presence on this show I'm still holding out hope that by the end of the year we will have another host who isn real who just chimes in Darren says he got this The idea is it doesn say anything unless you say something to it and then you can be like the Yenta with the needles It will talk. I put my Bloomsbury books into Notebook LM. It's a death plot we could have and query it. That's good. Yeah, that's a start. That's a start. Yeah, wait. Leo, tell us about your Chinese spyware that you got. This little thing? I like that you've gone from a little spyware device that you wear in your person to now just one that sits under deck. Go ahead, China. Take it. Yeah, actually, I disconnected it from the Wi-Fi. What is it? So the idea, and it was Harper Reid. You can blame Harper Reid for this. He's going to be on Twit on Sunday, by the way, with Alex Wilhelm. He said, oh, you should just get this. I said, what is it? He said, well, you connect with the Wi-Fi, and then it connects right up to DeepSeek in China, and you can talk to it, and it'll talk back to you. Couldn't you do that using your phone or computer? I can already do that, yes. I don't really need this. But what is cool is inside this is an ESP32. These are easily programmed. And what I'm really trying to do is put my Hermes Quicksilver, I call it, my agent, into a bunch of these devices all over the house so I can talk to it locally. How does Lisa feel about that? She has her, so she saw me interacting with Quicksilver. Where in the house were you? Paint us a word picture. Well, okay, so you have to understand. Okay, this is starting with a caveat. So you have to understand, I can talk now to Hermes anywhere. I can press the action button on my watch. I can press the action button on my phone. I can press a dictation button on any computer and talk, and it will go to the AI and the AI will respond. Now, I've been working. It's not perfect, but the theory is the AI will respond to me in an appropriate way depending on where in the house I am. So if I'm sitting at the computer, it will talk to me. It does this now. It will talk to me on the computer. And I have, by the way, I have expanded my repertoire. I had Quicksilver. I brought back Kenobi, which was the Claude code. He's running Fable. And I've introduced a new agent to the mix using the new GPT-56 saw. His name is Daedalus. So I've got Quicksilver, Kenobi, and Daedalus. They each have a different voice, so I know when one of them is talking to me, which one it is talking to me. And if I'm not on a computer, it will then use the Sonos system that is nearest to me. So demonstrate this. Well, I mean, I can do it with this Sonos system, but you won't really hear it because the speaker's in the ceiling. I'll hear it, but it doesn't, it gets cut off. But I've done this to you before. You know I've had it talked to you before. How many within 20 feet of you? She says things like there was a voice coming out of the attic, and I said, oh, yeah, I forgot. I should have shut that off. Go ahead. Within 20 feet of you right now. Yeah. How many devices are there that if you spoke out loud, an AI agent would answer? Oh, well, that's the problem, is none of them are listening. That's the problem, is none of them are listening constantly. I can say, hey, Alexa, hey, Google, and I can talk to them. But I want to talk to Quicksilver or Kenobi or Daedalus, and I've been working hard, and I haven't quite got that yet. I told you. Last month, I spent hours trying to train. I said, hey, Kenobi, literally 500 times. I said it 100 times next to the microphone, 100 times halfway away from the microphone, 100 times across the room from the microphone. And then I had to go downstairs and turn on the TV, and I had to do it 100 times with background noise going on. And then I had to do it in an empty room with no background. Anyway, I did it 500 times, and it still didn't learn. I was so annoyed. So I'm working on that one. There's this classic onion headline. I press the button. That's how I get it. There's this classic onion headline where it's Bill de Blasio to New York City. Hey, not so easy to find a mayor that doesn't suck eggs, basically. And I bet that that is what Siri feels like now. because everybody has been bad-mouthing Siri for years. The new Siri is actually great. Have you played with it? She deserves it. But, I mean, I'm excited for the new Siri. Have you played with it yet? It's the public beta. Well, we'll get there. But the one thing that Siri does have going for it is it's pretty good at recognizing, hey, Siri. At least it's got that. I was just reading about it. I'm curious as to what your experience is. I used to do the public beta. Hello, Perry. Hello, Jeff Quicksilver here from Leo. Welcome to Intelligent Machines. Hope this show is cooking. I really dislike. I was going to say, that is, I was going to say, for people not watching, Leo was talking with his mic off. Oh, now it's still talking. It's talking to the Sonos. Oh, that's so funny. Yeah, I can't control it. So I was probably, had I not been forced to update my Mac this week to whatever the new Mac OS is, Tahoe, and it was ruined my life in multiple ways, I probably would have done the public beta. But I've been so burned by the experience of upgrading to Tahoe that I'm like, God, I can't deal with Apple on my phone. I need to have at least one Apple device that works. Currently, my Spotlight has been unable to search my computer for like a week. Spotlight keeps getting corrupted just because I updated my computer. It's ridiculous. That's kind of not right. I'm just like, how are you a company that has software? I know. And yet, search isn't a functionality? What is this, Gmail? I think that maybe I'm just so used to bugs everywhere that I don't even notice. Because I am using the Apple public beta of iOS 27 on Siri on my watch and on my iPhone, and I really like the new Siri. I think it's going to make AI more accessible to people. Tell me about what it's like, because I've only kind of skimmed an article earlier today about it. I'll show you. Ask a question. What would you like me to ask it so I can just talk to it here on the phone? What would you like to know? I would say, when is the World Cup finals? What time is it here? Because I don't know. I can't make the... How do I watch it? So it does a little lozenge up at the top now. That's good. Lozenge. Schedule Sunday, July 19, 2026. Kickoff at 12 time. Well, that's not a big deal. I can't do Google for the last year. You could not do that on normal. I hope I didn't spoil this for you because it also showed who the teams were that were going to play. So if you didn't know, I apologize. Yeah, I mean, so this is pretty primitive, but I can also say, hey, have I gotten a text message from Henry in the last few days? I haven't checked my messages. It can also, did I hear that? It didn't. Oh, it crashed. It's good. Hey, that's great. I mean, I saw some sort of description today of just people describing that it is integrated throughout apps. That's what I was trying to demonstrate is that it can read my text messages. It can read my email. And apps, if they have to engineer it this way, but apps can also interact with it. So I could, if, let's say, Monarch Money, we were just talking about our sponsor, Monarch Money. If Monarch Money builds in the capability of Siri to interact with it, I could use Siri to ask questions of Monarch Money, that kind of thing. So Apple's hope is that developers of apps will turn this feature on, and then that one Siri conversation could be about, you know, anything going on on my phone. And I think that that's potentially very cool. It's going to be accessible. I mean, I'm excited to see it. I'm excited to work it. I might think about doing the public beta once I figure out whether Apple is able to correctly re-index my entire hard drive in iCloud. Yeah, I'm sorry. That's not good. You know? Yeah. And then we'll get there. Right now I'm just trying to hold out on my couple-year-old iPhone until the new iPhones are released, and I have the pleasure of paying a crazy amount of money to get a slightly better iPhone. I am currently of the opinion this is going to be the year nobody buys an iPhone because it's got to be so gosh darn expensive. I think it's going to be like two grand or something. It's going to be ridiculous. Maybe I won't even do that. I will show you what my latest AI project was. I thought it would be kind of interesting to pit Daedalus, Quicksilver, and Kenobi, each of which is currently using the top-of-the-line models from OpenAI. That's Daedalus, 5.6 Sol. Kenobi is using Fable, 5. And Daedalus is now using Grok, 4.5, the latest. We'll talk about that in a little bit. So I thought, I'm going to give them a task overnight. I'm going to give them one simple prompt, and I'm going to give them this task overnight, and I'll wake up in the morning and see what they do with it. And I just want to do something kind of silly, right? And I have been working very hard, you know, rewriting the Twit ad sales system. I want an actual update on that. That's been going really, really well. It's been really a lot of fun. In fact, Lisa is now recording videos of the old system saying, okay, this is the box I want. This is how I like it, but I don't want it to look like that because all the underlying functionality is done. And so now she's kind of got – we're trying to get the UI right. and did that in a week. It did it in quite a quick period of time. Did you, is the extra time now allotted with Fable? Not yet. They extended it. So I'm still good. All the short stories become a chapter book with Fable. It's doing a great job. But I designed it in such a way that if Fable went away or Fable got expensive that I wouldn't be spending a lot of money. So what it is is Fable does, and by the way, I would recommend this to anybody. Fable does the design. I then have Saul, Chats GPT-56, which is quite capable, review the design. I set up interagent email. I call it email, because I was getting tired of pasting the response back and forth. So I said, look, just email it. So he's emailing the agents, emailing his thoughts. They go back and forth until they agree, okay, this is the design. Then it hands it off to the less expensive Opus 4.8 to do the actual coding. and we're doing little chunks, so it doesn't have to think for a long time or anything. It's just something you can easily do. And then it gets reviewed again, not just by Fable, the high-priced model, and 5.6 Sol, but also by Grok, 4.5. So all three of them get to weigh in. And so that's been a good process. And the theory was, well, Fable, I'll only pay tokens just for the design stuff. So Nate B. Jones had a hilarious thing, a little snippet on TikTok today, that he knew the table had been extended because he saw people canceling their Tinder dates. He's joining us next week, by the way. I'm very excited. Nate B. Jones, who you introduced me to, Jeff, and I watch him religiously now. I think he's one of the smartest YouTube commentators. He really knows what he's talking about. He is going to be on next week, and we're going to ask him about all this stuff. So anyway, we're going to take a break, but I'm just going to tell you what I asked. I don't believe in one-shot prompts as you can tell from what I just described this back and forth process but it's always interesting to see what a one-shot prompt can do. People do things like and I'll show you one of my picks of the week is a one-shot Mario game that's called Super Dario but I gave it a one-shot. This is the prompt and it's kind of a weird one. Okay, so just prepare yourself. Have you ever heard of the I Ching? Yeah, you're a hippie. Gesundheit. Ching! This is back in the 60s. We were all into this. The I Ching is the ancient Chinese oracle where you toss coins or, in my case, you have yarrow stocks. Oh, for God's sake. Oh, how California you are. I feel like I don't even understand half the words that are being said. What? You know that back in the olden times. He also went to Earhart seminar training. I did that too. You know, back in the olden days, they would, you know, slaughter a cow and look at the entrails and say, well, you probably shouldn't invade the Gaul because the entrails say it's bad, auspicious. Perrots did that with alligators, but keep going. Yeah, those are oracles, right? So this is an ancient Chinese oracle that goes back thousands of years. And in the early days, they would use these zero stocks. And it's a randomization process where you'd count and divide and count and divide. And then you would, as a result of these many operations, you'd come up with what they call a hexagram. Let me see if I can find a picture of a hexagram that would have then oracular power. Because you've put all of this energy into the counting of the steps. This is not your best rabbit hole, Leo. I think it's my best because. I love when a hexagram has oracular power. So, okay, so the problem is it's a pain in the ass to do this. You know, you want the oracle, you want it fast, you don't want to have to go through a lot of trouble. So here's the prompt I wrote. I've been thinking about some sort of digital way to use the Chinese I Ching oracle. And I want you to, I gave the same exact prompt to all three of them. I want you first to investigate the I Ching, find a book of interpretations, Then take a look at the way the oracle is cast using yarrow sticks and coins. I've used the yarrow stick method. I think the idea is to influence the random throws with one's intentions, because you're supposed to form a question from the oracle, then throw the sticks while concentrating on the question. I want to make a website to simulate the process, then offer an interpretation of the result, build the site on my Cloudflare pages, call it Q-Chin for Quicksilver or D-Chin for Daedalus or K-Chin. Yeah, we got it. Yeah. Kenobi. I set them to work in the next morning, and I will show you the result. But it was a way of saying, I wonder what I'm going to get. I don't think that's how Rafi tested the source bottles. I'm just mistaken. But we will have that result. I'll show you the three sites, actually. I'm sure you will. If you don't want me to, I don't have to. No, no, no, no, no. We're very. It's okay. No, we couldn't be more. It's okay. I'm sure this is how the president makes his decisions. I'll tell you what. You think of a question that you would like to ask. For the Q-Chain. For the Q-Chain. The question should be not, what should I have for dinner tonight? He's not good at that. How, what are the vibes of the rest of this podcast going to be? Make it something about your life. That's about my life. Somewhat general. That's general. It could be good. It could be bad. I'm looking for good, bad, new. It doesn't do good or bad. It does things like the bridge is long over which you'll cross. I think that that could apply to this podcast. Well, I'm sorry I brought it up. We'll have more. There go the vibes right there. I was asking for a podcast-based prediction. Our show today brought to you by Expo. It's not Expo as you might not the Montreal Expos, but X-B-O-W. And you should know this name. This is the premier agentic pen tester. AI, we will stipulate this, right, has changed the pace of everything from how software develops to how it gets attacked, right? Engineering teams are moving faster than ever. They're creating more and more applications. The problem is security hasn't kept up. But pen testing is still one of the most trusted ways to understand real exploitable risk. But in an AI-driven world, pen testing can be the bottleneck. Penetration testing simulates how attackers would attack a system. But it takes time. Security teams end up being forced to choose between slowing down development to stay secure, to wait for the pen tests, or moving fast and then accepting, well, there are going to be gaps in coverage. We may be less secure, but Expo can eliminate that tradeoff. Expo is an autonomous, offensive security platform that runs continuous AI-driven pen testing, mirroring real-world attacks. AI-driven pen testing never tires, doesn't go to bed at night, doesn't stop for food or breaks. It pounds on your system. It finds the flaws. And Expo doesn't just scan for vulnerabilities. No, no. It discovers, exploits, and validates them so that you know you're only dealing with issues that actually matter, exploits that actually work. And that means dramatically fewer false positives and a clear view into real attack paths so your energy is not wasted chasing down leads that are meaningless. If Expo finds it, you've got to fix it. With Expo, your tests run in hours, not weeks. You're going to get complete visibility into how an attacker would move through your systems and the ability to uncover issues that traditional tools miss, including zero days and novel attack paths. Expo's results speak for themselves. Ask the application security leader at Cessnam.cz. He says, quote, even right now, after one year, I don't know any other company that's at least close to Expo in terms of agentic pen testing. The result is predictable cost, consistent quality, and stronger security without slowing down your engineers. Expo helps security teams keep pace with innovation and cover more apps more often with the resources they already have. It's an incredible lineage founded by the team that created Microsoft Copilot, already trusted by companies ranging from fast-growing startups to Fortune 500 enterprises. because XBOW Expo is quickly becoming a mission-critical layer in modern security stacks. Go to Expo.com to start a pen test today. Expo.com, the leaders in agentic pen testing. Actually, all three AI companies, I think, put out like super apps this week. Am I wrong? So, ChatGPT or OpenAI put out a ChatGPT work. This is powered by Sol. Okay. But they also made an update to the ChatGPT normal desktop that really screwed over a bunch of normal consumers. My understanding is if you had ChatGPT, just the normal desktop app on it, you, much like Claude, can normally toggle between chat and codecs, things like that. I believe, at least from what I've seen from every ChatGPT user on social media freaking out, is all of a sudden the app was changed to be ChatGPT Codex or ChatGPT, the work version, and then you have to go and download a separate one that is, I believe, ChatGPT Classic. They took out the chat. They took out the chat. ChatGPT suddenly has no more ChatGPT. It's Codex. Is it because they got rid of their browser and combined that in the app? I'm always trying to do the mom test. If you want to build a popular consumer product, you can't make deeply confusing changes like this that are going to baffle the average non-technical consumer. Darren says they mixed Codex, which is the coding tool, into ChatGPT, and now it just looks like Codex. I think we've seen this coming, which is that OpenAI sees the future not mom using chat, but the coders, enterprise, that kind of stuff. But I agree with Paris. I think the real test of this is when it becomes retail. I mean, in order for these companies to become profitable, or even break even given the amount of spending that has already occurred, they have to have extraordinarily large paying user bases. And that extends beyond just coders. Coders are not going to fill up that need for revenue. Actually, it's quite the opposite. Your mom is never going to give them the kind of money that they want and need. No, but my mom times a billion, a billion of my moms will start to fill out that hole. And a billion of my moms paying more than $20 a month for it. No, but a million of my moms paying $20 a month, those people are not even using one one-hundredth of the amount of tokens as you. It's like the sort of customer that you want at a low-cost gym. You want to get people paying your subscription fee and not using it that much. Like me at SilverSneakers, yes. It's a fitness approach, too. We don't. Line 116 says that OpenAI fell 90% short of their ad forecast. They also think they're going to be a media business with consumers and ads. And if they do that, then they've got to have a lot of bombs. So I want to point out, this is not OpenAI talking. This is some analyst. And all of the information we have about how much it costs and how much they're making or losing is at Zitrin and people like that. They're people who are guessing or trying to estimate. I mean, no, Ed had multiple years of the company's full financials, I believe, was where. Yeah, he got his data. We'll know when they go public. I'm sure we'll know better. But for right now. But the point is that if they believe they're in an ad business, then they need a scale of consumers. The way they get scale of consumers is by having Paris' mom and people like that in. Yeah. So they've got to figure out a retail business there. Or less, Anthropik could say, no, we're for coders. We're going to own that market. We're the best at it. They are. And so that's our business. It may be that OpenAI made a mistake. We don't really know what their internal models are. I mean, I believe OpenAI has said that they're going to be rolling out changes this week because of how confusing it is. Yeah, maybe. Yeah. I've been seeing this. Remember, this is Sam Altman's, you know, we're going to focus, which to me was we see Anthropic making all this money on enterprise, and we're missing the boat by having a billion users who don't give us any money. So we're going to focus more on what Anthropic's focusing on. That seems like this is more of that. And maybe it is a mistake, but I don't think we have a way of judging that. And I think that – and who knows how poorly or well-run OpenAI is. We just don't know. We just don't know. So, you know, users may be upset. They were also upset at 4.0 going away. But I don't know if 4.0 going away was a mistake on the part of OpenAI. It's like when Zuckerberg did his feed. From Znet yesterday, I loved ChatGPT Desktop until OpenAI gutted it to make room for codecs and work. OpenAI just merged the ChatGPT desktop app with Codex and removed all of my favorite productivity features. What are they thinking? Probably the same guy who six months ago says, you took my girlfriend away. I mean, I'm not sure that a staff writer, I'm not sure that a senior contributing editor at ZDNet is an AI way too guy. Do you know these people? ZDNet has been gutted. They are now run by private equity. And I'm not sure I would trust anything that they say about it. I mean, a human person is writing this, not a private equity company. That's just the people who pay them, not much. Yeah, but just take a look at the front pages. I mean, I'm just saying I have anecdotally seen 20 to 50 social media posts with hundreds of comments on it in the last three days. If OpenAI does change it back, then it'll confirm it, that they made a mistake, which they could easily have done. I mean, they may have overestimated the interest in a coding tool. I don't use any of their chatbots ever, any of them. So I'm the wrong person to ask. OpenAI may have made a big mistake, says Ars Technica, in the copyright fight with news organizations. They deleted the chat GPT logs that the New York Times was hoping to get from Discovery. They're facing calls for sanctions after fighting to keep news organizations from snooping through millions of logs. Now, I think they could reasonably say, we're just trying to protect our users' privacy. Were they under court order? They were. That's an issue. New York Times, in a sanction motion on Thursday, the New York Times and the other news organizations accused OpenAI of repeatedly lying for years to conceal evidence of infringement that could hobble OpenAI's defense. The alleged lies were exposed, again, from Ars Technica, when the court compelled an ill-prepared witness, OpenAI privacy engineer Vincent Monaco, to be rediposed during the subsequent April disposition. and he inadvertently revealed, whoops, that OpenAI misled the court for two years about the costs and burdens of searching chat GPD logs. This is, again, according to the plaintiff, the New York Times. So they want sanctions against OpenAI. They allegedly hid an 80 million log sample, two large samples, 10 million and 78 million logs. The reason the Times wants them is they hope that New York Times content will surface in those logs, that pieces of articles will show up. They also assert that OpenAI had searched those samples for content as part of its research into, quote, creating a filter that could be used to block the regurgitation of copyrighted content. So they said OpenAI was willing to search. I mean, yes, this is not what discovery is all about. Right. This is the information you get as part of discovery. But those logs are my chats with OpenAI. They are company dative. I mean, that's what happens when you interface with a private company that could emerge. I mean, the Steve Jobs email to other people was something that I just put in the chat earlier when we were talking about his policies around recruitment and NDAs in the early 2000s. And we now have copies of all of those because they emerge in discovery. And you didn't just get to say, sorry, we can't hand them over. I emailed someone else who's not part of this lawsuit. Never put it in writing, right? Yeah. Well, let's talk about the privacy issues that are raised. This is a tweet from a green being on X. Okay. Brock just uploaded my entire user directory to XAI servers. including my SSH keys, my password manager database, my documents, photos, videos, everything. And he's got the receipts here. How did he allow access to all that stuff in the first place? Well, so this is something I've had my eyes opened to. When you use, let's say I say to Anthropic, as you just did, or you said it to Google, Here's all my books. Read these. And build a nice little model for me with Notebook LM. Of course, all of those things are uploaded to Google. You know that. Generally AI, but it's also, it's Grock. What are you thinking? Well, I'm using Grock, too. Frequently, when you're using these models, they will read your files. They will read documents. This is part of, you know, what they're doing. so here's from Cereblab this is a gist on GitHub but XAI's Grok build CLI and this is their relatively new command line interface their version of Codex and Cloud Code actually sends XAI and this is a wire level analysis so they actually looked at what was being sent so for instance it transmits the contents of every file it reads including and this is scary, a .envsecrets file. So, for instance, all of my passwords, I don't give my passwords to the AI. I put them in an environment variable, right, which is not on the hard drive. It's in RAM only. You presume that it can read that, but it doesn't send it back to the home office. But wait a minute, maybe it does, verbatim and unredacted. In fact, if you think about it, it has to because it then has to use those keys to unlock stuff. It uploads entire repositories, every tracked file's content plus Git history, independent of what the agent reads. So Grok packages the workspace in your GitHub repository and uploads it. The destination, though, is not on XAI servers. It's a Google Cloud storage bucket. Wow. Anyway, I can go on and on. I think the eye-opener is, and in hindsight, I should have really realized this, when you're using cloud AI, you're sending everything up to its servers. And who knows what they're doing with it, right? Especially if it is Grok. Well, yeah. Yeah. By the way, Elon's response, we're going to throw all that stuff away. the researcher who exposed grok build uploading users entire repositories say the transfers have stopped they did it might have been a bug after a server side change and according to the register elon has separately promised that all previously uploaded user data will be deleted i'm really interested in why you chose to use grok at all seriously as opposed to it's a very good well so part you know one of the reasons i use all these different models. I don't need to use different models. I'm testing them all. I'm looking at them all and comparing them all. So I don't want to be too prejudiced against Grock. In fact, Grock is a very great model. Also, I get it for free because I am an involuntary blue check. I have a Twitter Plus account I don't pay for. So, you know, I mean, there's no reason Marie not to use it. I don't look at it. Did you think of Gemini? Did you test for Gemini? I pay for Gemini. I have an Omni subscription. But actually, Google is very cheap. For instance, they recently, I thought this was really cool, put their Stitch models up in a GitHub that I could then ingest with my AI and use Google's Stitch. Stitch. Did I say Stitch? I said Stitch Henderson. He was a famous guy on the show. Stitch Henderson, wonderful. Follow the bouncing ball. Stitch, which is what we were going to use, Paris, to design secretly British. It's their very nice design tool. So I was able to ingest it into my Hermes, but then it wanted money. How dare they? It wanted tokens. So even though I pay quite a bit for Google models, they don't give you a lot for free. Much like Anthropic, you have to use it within the Google tool and all of that. So I am a fan of open models. You couldn't look at it, but Paris told us right before the show that Miramirati's model is up. Yeah, I haven't played with it. And it must be quite good with design. Interesting. It's called inkling. Increasingly, I think the attitude we're having is that maybe the solution, instead of having a 10 trillion parameter model, that's what Nick B. Jones' fable is, you have smaller, you know, several billion parameter models that just do one thing. Well, Steve Gibson says you're going to have small local models that are really as good as coding as Fable because it won't have all that other crap. That's the on the coding's argument as well. Yeah. And then also you give it a task, it does the task, it stops. It's not trying to make paper clips until the end of time. Yeah. Would you like me to do that task again but with two extra things on it? Actually, Jaron says he pays for Gemini but doesn't use it because every time it does, it causes more damage than it fixes. That's one of the reasons I try all these models is to see how much damage they can do so I can say, don't, whatever you do, don't use this. I have to admit, GROC 4.5 is very good. It's fast. So Aikling is a mixture of experts transformer with 975 billion total parameters, 41 billion active. It supports context window of up to 1 million tokens. So you could probably run that locally. It was pre-trained on 45 trillion tokens of text, images, audio, and video. Yeah. So they're doing, I don't know why they're valued as highly as they are, mainly just because of the name, Mira Morati. Alongside it, we're sharing a preview of inkling small. I mean, this is their first iteration of anything. Yeah. The lighter weight model has 12 billion active parameters trained with a similar recipe that achieves strong performance at even lower cost and latency. Well, Apple is looking at this model from Prism ML. In fact, Apple is looking at the company called Bonsai. Very similar. The idea is a 27 billion parameter model, but because it is based on QN3627B, which is a very, very good multimodal model, but I'm pretty sure it's also a mixture of experts, which means it only loads in a little bit at a time. So it can run on a phone. It can run in an 18 model That probably too big for a phone but they even have smaller versions There a 3 one quantization version of it which probably isn very bright When you quantize that heavily, you get pretty dumb. But if it's in 3.9 gigabytes, that you could run on an iPhone. So that's what a lot of people are working on. I'm not surprised Thinking Machines is working on that as well, is the idea of how can we get a really smart model into a small space? Yeah. Because the mom is going to use it. It's just the RAM is so expensive. It doesn't even have to be mom. It's that same conversation we had with Rafi earlier. Enterprises hate the idea of their proprietary stuff being sent to the cloud. They want to run it locally, but they can't get, you know, terabyte RAM computers except at a huge cost. So if you can get it smaller, something they can run locally that's effective, especially if it's a dedicated model to assert, like, to your business rules, then that makes sense. I understand why they would want to do that. Anyway, Apple talks with Prism ML about this model because they'd love to put this on their phone. I will say, though, inkling fine-tuned on tinker produced by thinking machines is a horrifying connotation of words. It is the sort of sentence that is technically made up of words, but as it leaves my mouth, it is as if it never existed. Inkling tinker session. Yeah. No. Not really long. Fiji Simo has now stepped down from open AI. She took a leave of absence for medical reasons. She now says her medical issues are such that she's going to only be a part-time advisor. She was in charge of AGI at OpenAI. She was also in charge of product. Right. And I admire Fiji. I knew her at Facebook. She was in charge of video. She was in charge of the stream. She's a really amazing executive. And I think that we'll see her soon. She just had to take care of her health. Yeah. I mean, I think she has some very serious health issues and wish her all the best. Yep. She's brilliant. They also lost their AI safety guy. Okay. The head of safety at OpenAI is gone, but they're folding the safety into research again. Johannes Heidegger, two years as head of safety systems at OpenAI. And so OpenAI is going to reorg. Who needs safety? After all. Can we play the commercial? I know what you want. You want the... Can we play it? It's an interesting question. Can we get another commercial on the show? Mm-hmm. I'm going to say we can play it. What do you think, Benito? Do you want to play with sound off and captions on? This is the kind of stuff that they should do. Okay. That's less likely to cause a problem. Yeah, let me turn off the sound and turn on the video. And, okay, so we see a burning building. You want to turn on the captions? I think I did. Okay. No, I just turned them off again. There we go. Can AI be trusted? It's showing a lot of people's faces. Who's going to hit the stop button if we need to? How do we really ensure what we're aiming at? It's going so fast. It's showing a house on fire. It's showing really dystopian scenes. that's showing a cemetery with a bunch of two-thirds. If it ends up taking all the jobs, what does it mean to work? I'm so depressed. I'm worried. But wait a minute. Why do we have to have this stuff? Showing a data center, showing a car being built. If a machine can pretend to care better than I can actually care, how do we draw the line? Now the music goes from minor key to major key. If we all had a voice in it, then I feel like it would be better. Could AI help people? What is the point of this, Jeff? Keep going, keep going. Could AI help me build more connections? So now it's in the positive. Could it help you build more connections in the community? Oh, look, it can open a fire hydrant and the kids can play. Teacher, better teacher, better mom. Maybe it'll cure some great things. You actually want to cure some bad things. Yeah. There's a nurse with an open anthropic sticker on her laptop. We'll create a group of people that ask more questions. Will it make whales jump in the air? Will it start to be more human again? Yes. We don't want to learn the most beautiful parts of life. There's hope, it says, in hard questions. We don't have any answers. What is the point of this ad? I don't get it. Exactly. It's like, remember all the things you hate about AI? What if they actually weren't that bad? Or if they are that bad, what if we asked questions? And we've been telling you it's terrible, but in our hands, we're going to ask good questions. I really enjoyed when they showed a photo of a whale jumping out of the water. This is the duality of open AI, isn't it? This is what they did with Mythos. This is dangerous. No, this is Anthropic, not Open AI. I'm sorry. But he meant with Mythos. Yeah. So Sam Altman responded in a tweet. I thought this was satire. Kept looking for the handle to be spelled C1 on AI or something. Then Sam came back and said, hold on. That's actually very funny in time. It is very funny. That Sam Altman was like, no one will be dumb enough to point at all the stuff that people hate about our companies. So his next tweet is, hard questions are great, but only if we deem you worthy enough to not silently downgrade you or even get access at all. Beth Jaisos said, Jaisos or Jaisos? How does one pronounce that? I don't even know. Said, here it is. Anthropic is strategically trying to burn the AI vibes to the ground so people over-regulate and the game gets frozen while they're in the lead. It's ridiculous why they did this. If you go to TechMeme, it shows, you know, 100 tweets about it. People are scratching their heads. They're really schizophrenic about AI. They are. We hate it. We love it. We hate it. We love it. It's terrifying. It's going to be the best thing ever. So this ad really completely reflects. Yeah. We hate ourselves. We love ourselves. We love our product. We love our product. Yeah. Here's a vulnerability vending machine. This is not my pick. We put in tokens and vulnerabilities come out. That's an interesting idea. What else? Australia is demanding AI companies must produce more energy than they consume and stop stealing content. I think governmental regulation is going to be a really big issue in the coming years. Immediately, I think. I think Dario has to stop putting out negative stuff about AI because the job at this point is to convince government it isn't as threatening as you think. In fact, the White House, speaking to what Rafi was talking about earlier, has not ruled out action on open source AI models as well. That's what I'm scared of. Exactly. So it's not just anthropic and methos. They're also worried about open weight models. I also think, though, that this is the sort of administration that even if they had made a statement today being like, we've ruled out any action on the open source model. That could change 17 times in the next two and a half weeks, much like the next two years. Yeah, there's no point even. So Eric Schmidt wrote, there were two interesting op-eds. Eric Schmidt wrote in the New York Times, we must address the growing rage against the AI machine, which is what's happening right now. That's why he wants you to address it, because otherwise it's going to get regulated to hell. The other interesting one. Well, or you're going to get firebombed. I mean, this is the Wall Street Journal article. the hardline activists ramping up for the war with AI, the resistance to artificial intelligence is growing over fears about human extinction. But then there's these activists who are, you know... Well, they're being fed by the companies themselves. Yeah. We're going to see, I mean, I've been saying this for a while, there's going to be a schism between people who want AI and people who want to stop it for any means, by any means necessary. the history of the internet didn't help meanwhile Paul Ford wrote an op-ed in the New York Times which I think you're going to like a lot if you didn't read it I read it code is free speech, what he's arguing in terms of the open source and the regulation is, code is speech and it needs protection do you disagree with that? it's not no I don't disagree with it I don't think he made a very persuasive argument it didn't it wasn't very I wasn't persuasive in my opinion. It was an opinion. Yes, he has an opinion. There's no question about it. That's all I have to say about it. It didn't wow me. All right, let's take one last break. Picks of the week coming up in just a moment. You're watching Intelligent Machines, Paris Martineau, Jeff Jarvis, and you, dear friends, and a special thank you to all of our Club Trip members who make this show possible. About 30% of our operating costs now are paid by the club members. Thank goodness we have the club. We started during COVID because Lisa wisely realized this was probably going to be important. I always liked the idea that the people who get value out of our content should participate, should pay. That's such an ugly word. But that's what we're asking you to do. Help. Thank you. Help with $10 a month. Support the show. Go to twit.tv slash club twit. We try to give you value for money. You get ad-free versions of all the shows. You wouldn't even hear this plug. You get, by the way, in these ad-free versions, you also get chapter markers so you can jump around and skip the parts you don't want to hear or go to the part you do want to hear. You also get access to the club twit discord. It turns out a social network where people pay to be is actually great. The content, the conversations are fantastic in there. You also get all the special programming we put in there. We're going to be very busy the rest of this week. We had a great AI user group, by the way, last Friday, which I would highly recommend to club members. I think, you know, the AI user group, they said, why don't we do this more often? This is fun. So we're thinking about going to twice a month. Let us know. But Micah's Crafting Corner is coming up today, this evening, 6 p.m. Pacific. Then on Friday, photo time with Chris Marquardt. Our assignment is coastal. The Media Club, you know, we've been doing Stacey's Book Club for a while, and Micah said, why don't we do Media 2? So we're going to be talking about The Matrix. Wait, can I come? Everybody's invited, even you, Paris. You know, I bought the steel box of all the Matrix movies. Wait, can we do a special episode where we all watch The Matrix on physical media and then talk about it? Specifically, I want to do Matrix 2 in conversation with AI. Okay, because I'll be honest. I watched Matrix 1 on that nice collector's box that I bought on your recommendation. It looked beautiful, by the way, on my beautiful HD screen. And I could not bring myself to watch 2, 3, or 4. it's really they're really interesting nowadays I want to hear your take on it Matrix 2, very interesting in the age of AI Matrix 3 you can kind of take it or leave it Matrix 4, awesome always so that's in my opinion a controversial point of view but that's what the club is for so I'll tell you what we're going to do the discussion on Friday 2pm Pacific, 5pm Eastern for people who've watched the Matrix If you want to do a watch party after that, be my guest. I think that would be a lot of fun. See, this is what the club is all about, home theater geeks. Jeff Atwood's back on the 31st. He was in Berlin for the Developers Conference. We'll talk about that. His show has a geeky name, a developer joke, Off by One with Jeff Atwood. Hands on Tech. The AI User Group is coming up again the first Friday of every month. That will be August 7th. Stacey's Book Club. We've got the book, Slow Gods by Claire North. Start reading it now. In one month, we will have the book club. These are all things we do in the club. So join it. We'd love to have you. It's so much fun, and it helps us continue to do independent programming. And I think if we learn one thing in this era of AI and the Internet and everything else going on, it's that independent media is a rare and valuable commodity, but it takes your support to keep doing it. twit.tv slash club twit. It's a value add. Thank you, Mandar. I do have to warn you, a lot of animated GIFs in the club twit discord, but also a lot of Paris Martin. No. So there's that. It makes up for the animated GIFs. I'm always worth animated GIFs. That's what people always say about me. So let me show you. I mentioned Super Dario. It's pretty much, I would say, a one-shot joke, probably a one-shot AI. You will like it because if you've ever played Super Mario, you will recognize it. It says, Fable 5, trust us. This is the good one. There's Dario jumping around. Still the most powerful model this week. Watch out for that. Don't get hit by the little clawed flower. Here's the good news. You can't really die on this because there's Sam Altman. Watch out for him. He's slippery. table 5 extended by popular demand. Look at that. Through July 12th. Good news. I was able to extend it. Let's extend it some more. What do you say? Jump over. Oh, back in the hole. Jump over some Sam Altman's. Wiped out again. But wait. Table 5. Oh, your evaluation's going on. Now with 3% more reasoning extended through July 19th. If you keep playing, you keep extending it. And that's the beauty of this silly little game. Who got to show it? Super Dario. My pick of the week. I have others, actually. But I think that's the best one. There is a guy who's put up a post on how to get Claude to stop using the words load-bearing. There are certain, I don't know if you've noticed this, but there are certain tropes that keep coming up. The AIs can't stop doing it. Load-bearing is one of the most annoying. I'll leave that as an exercise for the reader. But there is a whole article on The Atlantic about it's not X, it's Y, the most famous AI writing tick. And they all do it. The funny thing is, Willow Ramis, they all do it. It's not just one model. Everybody does load-bearing. Everybody does it's not X, it's Y. Oh, one more thing, because this is important. The history of LLMs. Actually, somebody in the club sent me this. the timeline and evolution of large language models going back as far as 1950. So this is really interesting because it talks about Transformers, how they came about with the history of LLMs from ELISA to GPT, the rise of modern LLMs starting in 2018, the reasoning revolution starting in 2024. If you're interested, it's not very long. In a few pages, you can really get how we got here from there. I think it's a really well-done page, and I should give you the address, shouldn't I? T-O-L-O-K-A dot A-I. Toloka, and it's in Toloka's blog. Toloka, I guess, is a company that does AI training. And now, Paris Martin, your pick of the week. My pick of the week is an article I published today, just a quick thing. Oh, I was asking about the Cyclose Tora outbreak, which you've probably heard. You've probably seen all the headlines about how explosive diarrhea is sweeping the nation. I dug into the data. Is that the actual headline? Explosive diarrhea is sweeping the nation. Basically, they're all about how. That's the tweet now. That's the tweet. I dug into the data, and it's a bit more complicated than that. But the headline of it is, no, you shouldn't avoid fruits and vegetables due to cyclospora, just because I feel like there has been a bit of a misnomer, misconception perpetuated lately. If I cook it, is it going to be safe? Yes. Your salad, do you want your salad cooked? No, but maybe I won't eat salad. Let me give you, I guess, some general background. We have an outbreak going on of cyclospora. It's a parasite that can cause extreme diarrhea. The context, though, is every summer in the U.S., cyclospora cases surge, just because that's kind of how it works. Transmission only really happens in the summer. The U.S. has seen, you know, around 500 to like 4,700 cases of this a year. It happens in a bunch of different states. Technically, right now, the amount of states that are reporting infections is lower than it was at this point last year. However, the number of total infections is significantly higher. I was asking myself, what's going on here? Really, it seems like what's going on is you've got your normal spread of, like, a little bit of an uptick all over the U.S., plus a huge surge that's going on in Michigan and three other kind of surrounding states that seems to be related based on some genetic tests when the CDC has done, but they haven't figured out, you know, what the exact source is and what the food is. So if you're in those areas, in Michigan in particular, people recommend that, you know, you don't buy bagged lettuce, which could possibly be a source of the outbreak. Because you want to have that you can wash yourself? Is that the idea? Yeah, if you want to have salad and you're in those areas, you should get ahead of lettuce, remove the first couple of leaves on it, maybe wash the outside, chop it yourself. Generally, people recommend, you know, avoid pre-chopped, you know, vegetables or fruits, chop it, wash it yourself. However, if you are in Michigan and the three other states around there that have been identified as kind of a cluster, Michigan, Ohio, West Virginia, and Kentucky, What our food safety experts recommend is, like, maybe for the next week or two, avoid lettuce generally, if that's possible for you. I think that that's a fairly targeted recommendation. It's only because there's some preliminary data from Michigan. The state that is, like, responsible for the majority of the cases so far has kind of picked up some signals that it could be lettuce-related. But this really means for everybody outside of that cluster, you don't need to be totally panicking and avoiding eating all fruits and vegetables. There's a couple of other states that are seeing like a slight uptick in cases, like higher than average number of cases for this time of year. Like New York State has 500 cases, and in a normal season of summer, they get like 500 to 700. It's not out of normal, but it's a little higher. If you're in one of those states like New York or Illinois, wash your food whenever you're cooking it. Make sure you wash your hands, stuff like that. So Taco Bell seemed to preemptively make an announcement that it would stop selling things with lettuce and cilantro and some other things, which struck me as preemptive, like we're going to be safe. But then the stories are kind of, well, they're investigating Taco Bell. Did you come across anything about it? Well, I did. This is something – my understanding of the timeline is what happened is Michigan has – so the way the food-borne illness surveillance system in the U.S. works, it's basically kind of state-by-state basis. They don't get that much money from the federal government, but they're trying their best. Michigan's actually been really trying hard on this. They've been publishing case totals every single day, trying to give advice early on, like in the last week or two. They were like, hey, you know, I'm starting to see some of these signals around lettuce. This is something that's historically been connected here. Everybody watch out. And around this time, Taco Bell pulls lettuce off its products. They say this is just preemptive because lettuce and specifically bagged and pre-chopped lettuce, like the sort a fast food restaurant would use, is often implicated when these sort of outbreaks happen. The same thing happened at McDonald's like eight years ago. And I kind of believe them on that. I mean, the thing is, I obviously don't know what Taco Bell does and doesn't know. And none of the investigators have said it absolutely is or absolutely isn't Taco Bell. But the sort of thing is like when health investigators are looking into this, they are looking into what restaurants, what fast food chains, what grocery stores you shopped at, what you bought. And if it was something as simple as the lettuce and the Taco Bell has this parasite, I think that would provide like a very strong and identifiable signal that would be more easy to identify and perhaps have a larger national spread from this strange cluster of cases in just like one region that we're seeing that seems to be hitting a... More supplier than a... The demographics of the people who are getting sick, like the average age is 44, and it's more, it's like 60% women. Like it doesn't scream Taco Bell to me. That screams bagged salad or like herbs, which are common things. Obviously, that's just speculation. But that Washington Post article about Taco Bell you're talking about, I've read, and I mean, I never, I don't know what other journalists do and don't know, but it wasn't written to me. It was based on an honest sourcing, which could be totally legitimate. But it wasn't written to me in a sense that felt like super strong. Like, my hypothesis is that maybe, like, Taco Bell recalled, they're not recalled. As a responsible, preemptive move. Yeah, responsibly they pulled this because they're like, listen, we don't want to be involved in this. And then some investigators were like, huh, Taco Bell headlines has pulled this. You should look into it. And that seems to be the extent of what's happened. Isn't the gestation period for it also longer? It makes it harder to. Yeah, it's kind of complicated. Whenever you eat something, it could be like two weeks later that you get sick and then you're sick for quite some time. It's also way more complicated to track. Like if you're looking for something like salmonella, like it's pretty easy and fast for health officials and investigators to like both determine that salmonella is on the thing or find it in your sample and then kind of genetically test it. Everything about that process for cyclospora is way harder, plus the fact that, you know, it's a somewhat rare-ish parasite comparatively. So the average doctor before this outbreak became national news probably didn't have that top of mind if a patient came in with diarrhea. So if I wash my – you do talk about this in your article. You can wash it off, right? I mean, yes. It's important. Everybody should be washing their produce anyway, just because that's an important, helpful step to do. I put baking soda on it when I wash it. I'm not certain as the efficacy of baking soda. I mean, I looked into some of the medical and scientific research on this, and it seems to indicate that cyclospora is more difficult to remove from foods than other parasites. There was some scientists that did a study of berries, which is something that often carry a place. But less cyclospora is better. Yeah, less cyclospora is better. It is related to the quantity. They're like having your berries in a strainer under cold water for a minute, rinse them. That got rid of a sizable chunk. I think it got rid of, let's see, 11% to 69% of the cyclospora in raspberries. Raspberries are harder because there's all these little hiding spots. I would say raspberries are harder than blueberries. So water was like 11% to 69%. water with like a vinegar solution. It was like one part vinegar, three parts water. I've got a link to the part in the thing that says it was slightly more effective, but still not crazy. The most effective was like rinsing it with water and then putting it in a salad strainer and spinning it around and then rinsing it again, basically. But that still left some on it. Generally, I mean, though, removing cyclospora from a contaminated item could be beneficial because it means there's less parasite your body has to fight off. But, I mean, there's no indication that this is coming from berries so far, and some of the researchers I spoke to said, actually, the demographic data we've seen so far indicates it might not be the case because, I mean, I don't know if you guys know anything about children, but children love to consume berries by, like, the pound full, it seems, and we're not seeing a really high rate of infection in children. Again, it seems to be that. You buy raspberries every morning. I mean, my general take on this is I know it's really easy to get freaked out about stuff. Obviously, nobody wants to get explosive diarrhea. But if you're outside of these effective areas and if you're in a state that isn't seeing an unusually high number of cases, you know, you can just take basic food safety precautions. So today I went to Popeye's because I had a hankering for the Popeye's fried chicken sandwich. But I really like the new Popeye's wrap. It has some lettuce and some cheese in it, and it's really good. I decided not the time to have the wrap today because of the lettuce, so instead I had the sandwich. That was my safety tip. I mean, yeah, I think especially for people, if anyone is, like, immunocompromised or has risk factors, people particularly young, particularly old, if there's something about your situation that might cause you pause where getting a fairly severe intense bout of diarrhea could be catastrophic for you, take every precaution you want. If it's just something that would make you feel better, take every precaution you want. If you do just want to act on the data that we do and don't have, for a lot of people, just, you know, washing your fruits and veggies and washing your hands is probably a good baseline to be at right now. See, folks, how handy it is to have a food detective right here. There's an article, like all her articles, free at ConsumerReports.org. No, you shouldn't avoid fruits and vegetables. And people are so mad at me about this online. They keep being like, did the virus write this? Do you not know this either? And I'm like, guys. It's a parasite and no. I'm just like, saying that, yes, you should be allowed to eat fruits and vegetables, things you need to have a balanced diet does not mean I'm saying you can't take whatever precautions you want. I eat Wyman's frozen wild blueberries all the time. And I'm going to presume that it's like frozen. Why frozen? Why not fresh? Because blueberries last fresh about three seconds. That is true. The blueberries are faster. I just, I mean, the wild blueberries, which, A, I don't have access to, keep perfectly well frozen, and they fall out nicely. Wow, I never thought about the fact that you could have thawed them out. Amazing. It's a miracle. I've always just thought of frozen blueberries as something I put in the smoothie. Yeah, they're good frozen, as a matter of fact. Yes, again, consumerreports.org. Jeff Jarvis, pick of the week. Okay, a couple quick things. One is that Google has started a new profiles thing for creators. So I went and did it. Oh. So if you go to profile.google.com slash at Jeff Jarvis. This is kind of like About Me, right? Yeah. So you can link it to your Facebook, Twitter, Instagram blog, and so on. What's the URL, though? Is it? It's what I just said. It's profile.google.com slash at Jeff Jarvis. Uh-huh. Okay. So that's one. And then last week, or I think it was last week, before last week, we talked about the French jackets. Yes. That from Alice Carp. Oh, no. Another mention of Alice Carp today. Well, you've now broken the record, ladies and gentlemen. Introducing Alex Carp. So there's a variation on the jacket, which I think might appeal to all of us and our listeners. The Slow Learner Dusk jacket. Slow Learner. You carry a lot of books. I have seen this. It's got a big pocket for books. A huge pocket for books. Multiple pockets for books. We should all get matching jackets, guys. I might get this. It's still $200. It's not cheap. Oh, it does come in other colors. Oh, the blue one. It doesn't have to be black. Would you like orange, green, or blue? That's cute. Yeah. And it's got a big... We should all get matching intelligent machines letterman jackets. They call it anti-work wear because you should go and read. But why do you ever need to carry five books with you? Like this clown? Yeah. He doesn't even look happy about it. That man looks like he lives in Bushwick and is going to ruin your life. I have too many books. I don't want to read all these books. That man is somehow smoking three cigarettes at once in the L train. Well, does he look any happier? No. He looks more determined. Omen has four to six fine line tattoos on and on. I don't know what that means, a fine line tattoo. I'm doing Brooklyn discrimination right now. Yeah, you are. Yeah, you are. Ladies and gentlemen, we do Intelligent Machines every Wednesday, 2 p.m. Pacific, 5 p.m. Eastern. That's 2100 UTC. Although, if Congress gets its way, I don't know, it'll all be daylight savings time. We'll have nothing to talk about because there's no AI, too. Yeah, I don't know. I don't know. Next week, Nate B. Jones, AI strategist, will join us. His YouTube channel, Nate B. Jones, is incredible. The following week, finally, Henry Blodgett will show up, will join us. I'm excited about that. He's got a novel that involves AI in some way. Ah, okay. And his newsletter involves AI. Is that first convicted security fraudster? Yes. On the pod? As far as we know, right? I mean, that can't be true, given the amount of AI boys on this show. But the only known one. The only one we know about. And Philip Shoemaker, following that, he is the founder and CEO of Persona Shield. And I've forgotten why that's of interest, but I'm sure it is. There's a reason we booked him. So I'm sure that will be of great interest. Actually, at some point, I want to get Christina Warren on. She works at GitHub, where she is kind of responsible for Copilot, their AI. And she announced on the show on MacBreak Weekly on Tuesday the release of their desktop application for Copilot, which is quite nice. It's a quite nice harness. So those are all upcoming shows. You can watch this show. One more little tip this week? Yes. I went to see the invite yesterday. It's very good. Oh, I can't wait to see that. I hear good things about it. I've got to see the Odyssey. You've got five stars in The Guardian. The Odyssey did? Odyssey did. 97 on... Jeff, do you want to go to the 7 a.m. IMAX Odyssey screening with me? It's probably going to get into. Is it 70 millimeter film? It is filmed at IMAX. I know, but I mean, you have to go to a movie theater, of which there are only a handful. No, no, no. We're going to the Super IMAX one. Yeah, you need to go to 70mm film. In Brooklyn? I need to see. No, in AMC. Is it the real? So it's the 70mm. It's the Super, yeah, it's the Super 70mm IMAX one. I saw Oppenheimer on a 70mm film. Oh, did I too? It's big. It's very large. Well, no, no, no. We're talking with 70mm. Is that... So there are formats, and IMAX is usually distributed digitally. But Christopher Nolan, because he is the most influential director of his generation, is able to convince his backers to shoot it on IMAX film. Oh, yes, these are the ones that are hundreds of pounds. Yeah, it's actually the first movie to be shot entirely. I didn't realize this. Oppenheimer was not shot entirely in seven million film, but this one is. Actually, you know what I want to see? Every reel is only three minutes long, so they can only shoot for three minutes at a time. They have to make a special platter, sideways platter, to hold the film because it's so long. And it's projected horizontally. It's projected horizontally. But the movie I want to see, they just showed the trailer publicly for the first time. In fact, it was on the World Cup broadcast today. Tom Cruise, just if you want to kick, if you want to laugh, the Tom Cruise movie is bigger. Yeah. You've never seen Tom Cruise like this. No. Never, ever. It is. I was quite surprised. Yes. And I did start watching, thanks to you, Paris, Spider Noir. Oh, how is that? I need to watch that. It's good. Nicolas Cage. Never better. Really good. I just realized I need to re-watch Metropolis. I haven't seen it yet. Is it out? No, no, no, no, no, no. The original. Not Megalopolis. We all need to also watch Megalopolis. Megalopolis. Is it out yet? I don't think you can see it. Metropolis was set in 2026. Welcome to the future, ladies and gentlemen. That's why we're doing this show. It's about AI. You can watch us live, but you don't have to. You can get on-demand versions of the show on our website, twit.tv slash I am audio and video there. There's also a YouTube channel with video, interestingly enough. Great way to share it with friends and family or subscribe on your favorite podcast player. Thanks, everybody, for joining us. We'll see you next week on Intelligent Machines. Bob on. If you like what you heard and you want more of this week's top stories in tech, well, subscribe to Tech News Weekly. Every Thursday, I talk with the journalists making and breaking the tech news. I'm not a human being. Not into the animal being. I'm an intelligent machine.