Hermes Agent, NotebookLM & LiveKit Founders on the AI Agent Race | TWiAI 17

If Dario keeps saying you're going to lose all your jobs and Elon says work will be optional and we'll have universal high basic income, people seem to be spiraling here. Yeah. Their predictions. Why were these kids who are the AI generation booing it? We told everyone that like the knowledge work of the white collar was the pinnacle of achievement in society. You know, scaled up this thing that said, oops, actually, sorry, maybe that wasn't the contract that we're going to live under anymore. That knowledge work is this, you know, the pinnacle of what you need to be working on. If you leverage these tools, like what are we really preparing people for? I really think that like these AI agents are just going to automate away a lot of the entry level work. So I think that that's creating this almost like disconnect between like what college is trying to preparing you for, like the job market is also like tightening on the other side. So everyone's getting squeezed. Yes, the job, we don't know what's going to happen in the job market, but those skills will be valuable in whatever market ends up in March. Thanks to our friends at PayPal, the exclusive sponsor for This Week in AI. Try the payment and growth platform that's trusted by millions of customers worldwide. PayPal Open. Start growing today at PayPalOpen.com. All right, everybody. Welcome back to This Week in AI, episode 17. We're cooking with oil. This Week in AI.ai. If you want to sign up for the email or get quick links to our YouTube, Spotify, Apple podcast, all that good stuff, follow us on X.com, formerly known as Twitter, at This Week, the letter N, AI. We dropped the I because somebody else had squatted on the name. We've got a lot to talk about today, lots of the news and three great roundtable participants to walk us through everything. Jeffrey Cannell is from News Research, N-O-U-S, not N-E-W-S. They're an AI research lab creating open source AI, something I keep asking for, and they are famous for creating the ERMES agent, which is getting very popular. How is ERMES agent doing? I understand you may have just passed OpenClaw, and OpenClaw is getting a little rusty. Yeah, I mean, you know, we prefer to focus on our own thing, so I won't, you know, I won't hit on the lobster too hard. No, yeah, Hermes agent probably. I think it's like number one on OpenRouter now, so seeing a lot of growth. We just put out a desktop app a couple days ago, too. So it's been sort of like a whirlwind over the last three months. You know, we sort of had this, you know, it started with CloudCode and OpenClaw, where suddenly, like, all of a sudden, both the models were good enough, and when you paired them with an agent, you know, you got this emergent behavior that sort of is like taking over the world, these agentic harnesses. So, yeah, it's been a lot of fun to just head along with the ride and try to keep open source AI out there at the forefront. Yeah, and I mean, listen, you may not want to talk about the competitor, but you guys are contemporaries. Let's just call it that, you're contemporaries, and it feels like the fit, the finish, the polish on Hermes agent has won you a lot of fans. Is that part of the explicit strategy is, hey, we're OpenClaw's a little rough around the edges, we're going to make this a little bit more polished on the UI UX, and that's the early win? Yeah, I mean, we really started out as building this for ourselves. We actually had a version of it before OpenClaw came out. It was our internal tool that we used to help our model researchers, like, automate some of their work. So, it really came from, like, this obsession with using it all the time. Like, if you're using it all the time and you love the product, we just have a class of people at Noose who would be doing this if they weren't getting paid for it. Like, it's what they love. And so, sort of that obsession with the product and loving it so much, you know, made it so that, made it into this thing that was really easy to use. And we do, you know, make an attempt to keep an eye towards the aesthetic as well. You know, it's something that I think can get lost a little bit in the tech spaces. But, like, you know, we are people, too. We're not just lines of text on a computer. Like, we have all of our senses. So, we try to, like, engage people across, like, all that whole spectrum, too. So, we do spend a lot of time making sure that, like, the touch and the feel is good and just, you know, making sure that it doesn't break no matter what. But, you know, you update it always works to the point that we have finally, I felt, like, with these agents, I've sort of, like, broken out into, like, I guess you could call it, like, normie bill. You know, like, there's a lot of, like, non-AI tech people who are finally seeing the agentic harnesses. Let them see AI as something more than what it was before, which is kind of just, like, Google with, you know, like, Google search on crackers. You know what I mean? It was just, like, asking questions. Like, the agentic harnesses, when presented in the right way, people kind of like, oh, now I see why you guys are so excited about AI. You can, like, do all this, you know, actual interesting, useful stuff. Yeah. And there's something with that persistent memory and the skills building that really resonates with the hacker crew. With us again, my old friend. And I mean that in both senses of the word. We both got old and we've known each other since the 90s. Ancient. Ancient Stephen Burlian Johnson is, of course, the editorial director of Google Labs and the co-founder of Notebook LM, which was the OG. I don't know how to describe it, but just a great application for projects, you know, is the way I think about it. When you're working on a project and you want a big context window, Notebook LM is just, like, the nuts. How do you describe it today? Because, man, you've had this thing out for, I think, close to two years and you were so early. Now there's a lot of people now, Stephen, who have your contemporaries with a lot of people, including Ermey's agent and OpenClaw and Claude Cowork and Perplexity Computer. There's a lot of options now. Where does the product stand in your mind in terms of the playing field? Yeah, I mean, what we tried to do with Notebook is, you know, from the beginning it was predicated on this idea, which was pretty novel at the time of like a source grounded experience with AI. So you're giving it the knowledge you need to do whatever project you're working on, whether you're a student trying to study for a class or whether you're doing a startup and you have all these documents you're trying to deal with. So the idea was that the AI would always be grounded in that experience in those documents so that you had a sense of trust. And we had, you know, kind of state-of-the-art citation systems. You could always go back and read the original passage that was important to you. And then we rolled out audio overviews about two years ago where you could take the knowledge in those documents and turn it into an AI podcast, which was kind of our crazy viral moment, which is really fun. So it's really, you know, a tool for anything that involves a complex knowledge base that you are trying to understand and do things with. And increasingly, we have gotten more and more agentic in the kinds of things that you can do with that knowledge base, including a big announcement yesterday that is kind of the most significant change to the product. But we see a lot of people using it in research mode. I mean, it's heavily used by students. Yeah. It's our kind of number one user base. But, you know, anybody like a lawyer that has to manage information and make sense of information and transform information that's scattered across potentially hundreds of documents, notebook we think is the best tool, the best surface, the best application for that kind of project. It's really fascinating. I forgot. I wanted to get your take on this. We might as well do it now and go around the horn. I want to get your take. I'll introduce our other panelists. But think about this one. And I want to get all your takes on why AI got booed at commencement speeches when young people were the ones who are applying this technology more than anyone. So just keep that in mind while I introduce Russ Dessau. Russ Dessau, of course, is the CEO of LiveKit. That's the open source real-time voice and video infrastructure company. Been around for a while. And they are the transport layer behind ChetCpt's voice mode and many other people's. I don't know if you can – and you were, of course, on This Week in AI back in April. I'm not sure what episode number. But tell us how the business is going. And you have some other high-profile customers. Do they want people to know that you're powering their voice? Or do you have to keep that quiet, Russ? For a lot of them, we have to keep it quiet. But there's definitely some that we talk about externally. So Spotify uses us. They build the kitchen on top of us. Also Tesla, all of Tesla support centers, service centers, roadside assistance and robo-taxis, Grok Voice, all of that stuff built on top of LiveKit as well. What are some other ones I can talk about? Be careful here. Do they pay two different prices if they want it private, it costs more? If they want to be public, it costs less? Or is it just their choice? We give some discounts if they allow us to kind of tell the story. But most of them – I wouldn't say that there's reticence to have it be known that LiveKit is powering it. But I mean, in other examples, we power all of Agent Force Voice for Salesforce. Oh, nice. Joule SAP just launched Joule, their platform and application a few weeks ago. We were in Orlando for that because we're powering voice for them as well. So lots of really large enterprises and thousands and thousands of smaller startups as well, all building these voice AI-based applications. So you can interact with the AI like you interact with the person. I think that that's the real big novel kind of unlock that has happened here. It's certainly become less annoying. I literally – on Sunday, my old Tesla Roadster, the original one, I opened the door and the battery was down to 20 miles. And I had been on vacation or traveling for a week or so. And, you know, we're supposed to check it every week, but somebody forgot to check it. And I was like, oh, my God, I went into a full-scale panic. My battery is going to break. I called the service center. And I had a delightful conversation with an AI. It was so obviously an AI. But it was so much better than talking to a human because a human feels the obligation to do like a little small talk and warm you up and how are you doing and da-da-da-da. And just it was so efficient and it actually worked. Like if you interrupted it, it worked. Which has always like been the problem with these voice gel systems, you know, press one, press two, da-da-da-da. Just get to the point. But are you also the CRM level for them or you just provide the voice layer and then it goes into whatever CRM they're using? Because it does seem like some people are bundling this voice level with a CRM, with a, you know, gosh, what's the name of the famous – oh, Zendesk or something like that. You know, people use those kind of tools. So are you in competition with those people, cooperation with them, and the line's getting blurred? How do you think about that? No, yeah, I'd say I don't think we're in competition. I think we're just at different layers of the stack. So folks like Zendesk, there's other ones out there too. Even Agent Force from Salesforce, you know, has a product here. And what they kind of build within a vertical, right? So customer support, it's not technically a vertical, but it's kind of like an area where they have integrations with CRM and other types of tools. Like what we really build is a platform that allows you – it's effectively the agentic framework or the harness, I think, is a word that a lot of people are using. But I think people are mostly using – harness in the way that Jeffrey describes harness are effectively like the software layer that wraps the LLM that allows it to do work autonomously on its own. Or the software layer that wraps the LLM that allows you to have an interaction using your voice and computer vision with that LLM, right? So handling things like turn detection, understanding when you're done speaking and when the AI can start speaking, handling interruptions automatically, letting you use any model for STT, speech-to-text, LLM or TTS, being able to expose tools. So like what an agent force would do is they would build the agent on our platform using our framework and software, and then they would wire it up to tools that they have on their side. So CRM or another product that agent force wants to integrate. And then – yeah, so you build it, and then we're the cloud platform where you can run it as well. Just amazing that there are like so many companies, Stephen, going after each vertical and solving each of these very granular problems. It was like the early days of Web 1.0 where there were just so many cruft and little nooks and crannies that needed to be polished, and then we just take them all for granted today. I do think of those times a lot, Jason, but it's different in the sense that there's – obviously there's so much more attention. Like we were able to kind of develop the web in somewhat secrecy for a while. Like it was an insider kind of experience, and you would have many conversations with people where they would be like, hey, I don't even know what this internet thing is, or I don't know what the web is, and you have to explain it to them. Whereas now, to your point about the boos at the college graduations, everyone at least has an opinion about AI right now. Yeah. Yeah, and there's 3 billion people using it. Well, 3 billion people using it, and I think when you last used it and how deeply you integrated and how deep you went forms your opinion for better or worse. Because if you used this 6, 12 months ago for three days, your opinion is completely based on an outdated experience. But unpack for me, Stephen, what you think happened at those graduations. Why were these kids who are, let's face it, the AI generation – this is the first generation who used AI in school, right? It was three, four years of ChatGPT. ChatGPT, they were using it, their teachers were using it, but they're booing it. What was your take? And then I'll go to you, Jeffrey. I think there are a lot of things at work there, Jason. I think on the one hand, there is the concern that generation has about the impact of this on the job market, right? That they're worried that this is – there are kind of leaders of some of these companies who are saying this is going to eliminate X number of white-collar jobs in the next five years, right? So we have that concern. Like, is there future being put at risk at this? The other side of that, which is something I've been – I just wrote a piece about this actually called Cognitive Uploading, which is this idea that we've focused so much on the ways in which AI helps you kind of bypass getting a good education and helps you potentially like have the AI write the paper for you one click instead of actually doing the hard work of thinking and processing. And I think we haven't spent nearly enough time, you know, and it's on us as well as, you know, other people to really walk through the ways in which actually if you use it properly, AI can actually help you become a deeper thinker and expand your understanding of the material you're trying to work with. And so for me, the way that I use it all the time is, you know, I'm constantly going in and saying like, give me something new to think about here. Here's a problem that I'm trying to understand and here's a piece that I'm trying to write. Help me go find the information that I need to understand it and then help me explore that information and make new connections. And so I feel like it's an extraordinary kind of amplifier of my cognitive processes, right? Like I, of course, I'm a richer kind of deeper thinker, but I think the public discussion of it and also to your point about people basing their kind of model of what it's capable of doing on chat GPT circa 2023. Yes. That they're just like, oh, this is a tool for plagiarism and cheating. And it's just, you know, it could be used in that way. But of course, there are also all these other ways in which it can be used to really enhance education. I think we just need to do a better job of walking through what would be the ideal engagement for a student, say, to really get more out of it. Yeah. And that's what framing is critical. What we try to do at Notebook is build a scaffolding that actually encourages learning rather than allows the student to kind of bypass learning. Yeah. And you mentioned Berkeley has like their rule set for AI. And it was essentially when I read the first half of your piece right before I got on air, one of the producers sent it to me and I'll finish it after I get off. But essentially, they are saying, like, don't use it. Yeah. Basically, maybe you could talk about what Berkeley said and what they got right or wrong. Yeah. So there was an interesting contrast between the Berkeley law. This is Berkeley law. And then our old friend, Larry Lessig at Harvard Law, who had let Larry integrate teaching a constitutional law class and everyone had to use a notebook. Like he put all the cases that they were studying, like 115 cases that they're studying in this class. And they made a notebook for each of those cases. And then the students would add additional knowledge. And it was just a way of like diving deeper into the thing they were trying to understand. And on the other side of the country, Berkeley, literally their default policy is that the only way you're allowed to use AI in any form is to help you find sources. You are not allowed to upload documents or sources as in some kind of research partner mode. You're not allowed to write with AI. You're not allowed to get, you know, outlining or brainstorming with AI. Individual professors can choose to opt out of that. But it's a very kind of blanket policy. So one person is framing it as the card catalog. The other one is framing it as your researcher who is at your beck and call at any moment to quickly go run and use the Dewey Decimal System, bring you back stuff and brainstorm with you. Hey, maybe these are some angles. Jeffrey, when you look at the framing of AI and the young people booing, obviously. Yeah, if Dario keeps saying you're going to lose all your jobs and Elon says work will be optional and we'll have universal high basic income. I think maybe and Sam's talking about UBI and maybe giving half of open AI to Trump and to the government. I mean, people seem to be spiraling here. Yeah. Their predictions. And then the reality seems slightly different. But I don't know if these kids are being precocious and think it's funny to troll AI or if they have a legitimate fear or disgust with it. Because I kind of got the vibe that they were like, ChatGPT kind of ruined my college experience. It wasn't organic. It wasn't like, I'm the first generation to lose college. I don't know. That was just my projection into their thinking. But what do you think, Jeffrey? Well, I mean, certainly there is always the argument between your stated and revealed preferences, right? I think people will, you know, college is a great place to be loud and have, you know, loud opinions and speak truth to power and so on and so forth. But what are they actually doing? You know, what are they doing in their personal lives? So there's probably a bit of that going on. But I mean, obviously, you know, this is a question that I think, you know, if you only view people through economic terms, right, which is where we stand today in a lot of society. And then you introduce them to this new thing that says, you know, for the last 40 years, what was the plan? It was go to college, knowledge work, you know, learn how to code. Remember 10 years ago, 15 years ago, coding boot camps. You know, we told everyone that, like, the knowledge work of the white collar was the pinnacle of achievement in society. Right. Yeah. And, you know, within, you know, two to three, you know, really within the last six to nine months, if you want, like, the truth. But the timelines are a bit longer to people who aren't paying attention. You know, we basically, you know, scaled up this thing that said, oops, actually, sorry. Maybe that wasn't the contract that we're going to live under anymore. That knowledge work is this, you know, the pinnacle of what you need to be working on. And it, you know, it doesn't surprise me that if you tell people that their worth is, you know, where they go to school, what job they can get, how much money to make. And then you create this new thing that says, I can do all that better than you. That, you know, there will be a natural to imagine there'll be some sort of social pushback. So I think that is just a natural, a natural outcome. But we need to really, you know, there are people, for example, like Jensen and NVIDIA, who, like, pushes very strongly back on this, like, unemployment is coming kind of, like, argument that, you know, work is going to be optional. And his argument is basically just every time humans have had sort of these step unlocks of technology, whether you start from the printing press all the way through the Industrial Revolution, even to the Internet. Like, we've always just found more ways to do more work, right? Yes. Monkeys like tools. We are monkeys. Given a new set of tools, monkeys will use tools to entertain themselves and solve problems and murder each other on the margins. It's just basic monkey behavior. Russ, what was your take as we go around the horn here before we get into our series 17.0 discussion? Yeah, yeah. My take, I think there's a few different dots, a little bit scattered. But I think the Jensen take, I think, is maybe right for this current era of AI. But, you know, the line has to end somewhere. It can't just be that every single phase shift, there will always be new jobs. There's some end of the line at some point. Is that for this current era? I don't think so, personally. But, you know, this is one thought. The second is that, you know, college is kind of like this almost like connective tissue between like, you know, when you finish college, it's supposed to kind of prepare you for the world out there to get a job. I think that was like, at least when we were probably all in college, that was like the promise of it. And so I think to like, with this new technology, we have to think about like, how do we assess what college is preparing you for? Like, what is the success criteria when you finish college? I think today that hasn't really changed much. And we suddenly have like a TI-83, but for thinking. It's like, I mean, I remember when I was in high school, you take the SAT2 math and like the entire test can be done on a graphing calculator. But that didn't stop people from continuing to do math or anything like that. Right? It just, the calculator can handle the mechanics of it. But you still have to think. And I think like with LLMs, it's something similar that it can help you refine ideas. But ultimately, I think like it's not the final polish. Right? I think that humans are still doing that. And at least when I use LLMs, I'm not letting it, you know, kind of just generate the thing and spit it out. And that's the final product. I always go through a review. Even programmers go through a review on all the code that is generated for the most part. Well, at least we do. And so I think you have to assess like what are students doing in these colleges? And like what are, like if you leverage these tools, like what are we really preparing people for? That's, I think, yeah. On the other side of this, you know, like my partner, she has two younger brothers who are like 23 years old and 25 years old. And, you know, they're just graduating and they're out in the job market and they're having a hard time. Like it's insanely hard for them to find jobs. I think they've been looking for over a year, you know, different industries. But it's for the entry level work, I think like a lot of companies are kind of looking down the road and thinking, okay, well, you know what, I'm just going to stop hiring or slow down hiring because I really think that like these AI agents are just going to automate away a lot of the entry level work. So I think that that's creating this almost like disconnect between like what college is trying to preparing you for, like, you know, maybe it's preparing you for the same stuff from 15 years ago that it was preparing us for. And then, like the job market is also like tightening on the other side. So everyone's getting squeezed at that age. Steven, just summarizing everybody's views here and working it out in my own mind, there seems to be a tension between what is the point of college? Is it to actually prepare you for a job or is it to teach you, hey, to find what you're passionate about, to learn how to learn, to get those last couple of years of maturation before you leave the nest? If you viewed it in that latter as opposed to the former, you're like, okay, that's a good use of time, right? You're just learning how to learn. You're getting well-rounded, bit of a luxury. It's kind of like a luxury vacation. If it's costing you $50,000 or $100,000 a year, it does seem like kind of crazy to spend that amount on your last couple of years of maturation unless you can afford it. But if you look at it like it's there to prepare you for a job and the first two runs of the latter have been removed, now it's an even bigger gap, right? Yeah, I think that's right. To me, one of the things I came back to in writing that piece that kind of shows up in the second half, which maybe you'll get to read, Jason, after this is over. I will, absolutely. It's kind of building on what Russ was saying, which is like there's a kind of easy rule of thumb I think that students could take to like figure out to feel good about the AI experience that they're getting in school, which is fundamentally to just imagine that you have a world-class like tutor, editor, and researcher like at your side. And just treat that AI editor, AI researcher like you would a human, right? Like if you had a great tutor, you wouldn't sit down with your tutor and say, please write my paper for me, right? Like you just wouldn't do that. But you would say like, hey, here's some ideas I have, like let's brainstorm how I can make this a better thing. If you had a great editor, you'd be like, take a look at this paragraph, like how can I do this a little bit? I do that all the time with editors, right? I get so much out of the dialogue I have because I'm lucky enough to have access to actual human world-class editors. Am I cheating when the editor gives me feedback and improves my writing? Like, no, that's an incredible like blessing that I have because I get to be a professional writer. So I think if we have that as like the kind of common sense framework that like ask the AI to do the things that you would ask a valuable, knowledgeable assistant to do for you, but still do the thinking yourself. You will at least get that kind of learning how to learn, learning how to think, learning how to engage with the world deeply. And then, yes, the job, we don't know what's going to happen in the job market, but those skills will be valuable in whatever market ends up emerging. Yeah, I think this speaks to Jeffrey, self-reliance, radical self-reliance, as opposed to I'm part of a machine and I get funneled, you know, from high school based on my performance into whatever college is best for my degree slash ability. Then I get funneled to whatever training program at Goldman Sachs or go to Google and then they take me along the road. You're on your own. But the good news is, even though you're on your own, you have an unlimited number of career counselors, coaches, mentors, and researchers at your fingertips. If it's stunning to me how people, certain folks ask me questions like people who work for me, how do I do this before they're asking Claude and Gemini and Grok and perplexity? Like I am asking those tools. What questions am I not ask, you know, asking here? Give me some blind spots. I literally type that in. Tell me what blind spots I have about this health issue, about this business issue. So, yeah, kind of an interesting moment in time for rugged individualism and self-reliance. Yeah, Jeff? Yeah. I mean, I think you need to, AI should make you better today than yesterday and better tomorrow than today. And if it can do that, then it's serving its purpose well in the world. Yeah. It should be inspiring. All right. Let's get to the news and the docket. Apple has been MIA when it comes to AI. But at Tim Cook's final keynote, finally, Tim Cook is done. It's not personal. I just want to see some innovative products come out of Cupertino as opposed to the end of Steve Jobs' roadmap and squeezing out every last nickel out of it. Apple unveiled a rebuilt Siri AI. It's a dedicated standalone app. I shouldn't have done that a long time ago. It's conversational. It's powered by Google's models. And it's shipping later this year. It has the ability to perform cross-app actions, i.e. it has the ability to use the things on your phone, which it doesn't even know how to spell my last name after a 20-year relationship and tens of thousands of dollars of iPhones bought. Apple bolted Google's models onto the iPhone instead of building their own AI. I don't say bolting, but integrated. They're paying Google roughly a billion dollars annually to customize for a custom $1.2 trillion parameter Gemini model. Apple, Craig Federighi, accuses rivals of, quote, pursuing AI for the sake of AI without clear regard for the people it's ultimately meant to serve. So I guess playing into our previous discussion there about being a victim of AI versus being empowered by AI. They introduced Apple intelligence back in 2024. It was completely a dud. So here we are, and obviously there's the CapEx discussion to add to this. Russ, what we should take when you saw Siri AI, the original agent, along with Alexa from Amazon, both of those are disgraceful shells of what we see in the modern era and completely worthless and disgraceful products coming from those two great companies. That's my opinion. What do you think, Russ? Yeah, I don't disagree with you. I think that even now, like I watched the keynote, and I think it suffers from the same fundamental user experience problem. So look, like the voice has gotten better. They had these two sliders. They had like pace at which this voice can be customized to talk, and then they had like expressiveness or emotion or something like that. But they didn't move the emotion slider at all in the demo. And I'm like, huh, I wonder. Maybe that part's not ready. Yeah, not shipped yet. Let's say, given the benefit of the doubt, they have like a voice that feels more emotional, feels more human-like. So that's cool. And capability-wise, I think, you know, building on top of like the stuff that the Gemini folks are working on is awesome. It's like a level up in the capabilities now by partnering with Google on that. I still think, though, that there's a UX disconnect. And it's been like this ever since like the 2010 or 2011 launch of like Siri slash Alexa. And it's that you don't exactly know what you can use the thing for, right? Like so all of their demos were demos. They're kind of in a vacuum. And, you know, like one of the examples that Rockwell had was like he said, oh, okay, like there's like a lottery for some tickets to this event. And then that was it. And it wasn't like, well, how do you know? Can you ask the agent, can you put me in the lottery? Right? Can you go and do that for me so I don't have to go and open the website and fill out a form and all of that stuff? Like can the agent just go and do this via computer use or some other, you know, facility like that? You know, that wasn't shown. And it's also like if it was available, you wouldn't know it. And so what's punishing about that experience is as soon as you ask, and it was like this in 2010 too, like or 2011, when you ask Siri to do something and then it's like, oh, I have a bunch of web pages I can show you to do this thing. And it's like, oh man, forget this thing. It's literally slower and more cumbersome to use Siri than to use a browser. And I think the other problem is that like it doesn't feel human-like to interact with, at least in the demos. It still feels like this like transactional thing where I like ask a thing to do something. I wait, it goes and does something and then comes back at the response. I'm like, oh, okay, the next thing. It doesn't feel like a conversation with another person like a lot of the other kind of like voice-based AI systems do nowadays. Yeah, they're definitely behind. Steven, obviously you work at Google, but you don't speak for Google in all ways. But this does seem to be a partnership because when I use Gemini, I specifically route any travel, local, flights. You know, there are certain maps. There are certain categories of things I just associate with Google doing a great job. Obviously, Google Maps. Obviously, Google Local. Obviously, Google Flights. Google Travel. All that stuff is dynamite. And it's integrated extremely well into Gemini now. It wasn't always like a year or two ago. It wasn't. The hooks weren't there. But it seems like the hooks and what web services and what data services you can get to from your voice agent is the new paradigm. Not having to route through a phone and an app. So it's almost like you're forcing – I'm trying to think of an analogy here like putting airbags and three-point seatbelts onto a horse and a buggy. You're better at these analogies. But the previous paradigm of apps doesn't work in an AI era anyway. So, you know, that seems to be a blocker. How would you frame what needs to happen here to make a Siri or an Alexa or whatever – or OK Google work really well for customers? It's so funny, Jason. I actually had almost the exact opposite take on it in a way. OK. Here we go. Which is that – and I didn't watch the keynote very closely. But the thing that caught my eye that I was excited about – and this is maybe more about my long history as a fan of Apple and as a customer of Apple. Sure. That I was happy they made a Siri app. And the reason I say that is I feel like Apple has a great history of making new, you know, path-breaking, like, applications in different spaces. Like, and, you know, going back to HyperCard, which changed my life in 1987. But, like, think about what GarageBand was like. Like, it was such a revelation of, like, what you could do with music, music production. iTunes was kind of like that. These were both apps that were kind of modified from things that they'd acquired. But, you know, they have a great history of making new paradigms for what applications can be. And I continue to think that there are, like, new application types that are going to be possible because of AI. That's what we started with with Notebook LM. It's like there's going to be a new – you know, there are going to be a whole new kind of visual metaphors we're going to have and a whole new framework for how we think about information. If you know that there's an AI at the center of it, we're not just adding AI to a word processor. Like, we're building something from scratch. It's genuinely new. And so the fact that they're, like, introducing this thing that maybe could become more than just a standard issue, like, chatbot – it didn't kind of look like a chatbot in the screens that I saw. I don't know. It's just I was excited because I want to see them playing in that space. Like, they do make – they have a history of making – As a fanboy, you want to see them in the game not – like, they're not making the playoffs. Like, Apple not making the playoffs. It's like the Knicks not making the playoffs. Yeah, they sat out two seasons, Jason. They didn't even play. Yes. I mean, on some level. Like, I hate to say it. But, like, there wasn't a product that was there. And so – and I know there are such creative people there. Like, and, you know, like, we – I think at Google, at DeepMind, at the Notebook team, the Gemini team, like, we're inspired by things that Anthropic is doing. We're inspired by things that OpenAI is doing. We like to see ideas circulating. And, like, we just weren't getting anything out of Cupertito in terms of the consumer products. It's kind of a bummer, right? You want them in the game. I like that. I like to see them, you know, creating. I'll tell you what the coolest thing was, in my opinion, that they showed in terms of, like, a new kind of interaction model was the Siri integration on the Vision Pro. You don't actually have to, like, say anything to engage it. It's just always floating there. And then you just, like, look at the thing and you just start talking to it. And it understands when you're looking at it because of the eye tracking. And it just engages. And so that felt really cool, like, as an interaction to kind of see happening right there. Here's what I want to do, Jeffrey. I want to get, and this would be against all privacy and all of Apple's, you know, ethos, I think, at the moment. But hear me out. This could work. They get the iPhone to the level with Apple Silicon and the amount of memory in it that it can run. Let's assume they, this Google relationship, which is now, you know, a very old relationship with Surge or whatever. They get Gemma on the phone. And I can tell it, record everything on my phone. And they already have, like, a little box that says Apple Intelligence. And you can flip it for different apps. And I guess theoretically it's learning. And everything that occurs in its learning is encrypted on my device. And then, you know, obviously you can put it in my iCloud, but it's encrypted. If it's, they get asked to crack it or give it to the government, it's just like the San Bernardino shooting where they're like, yeah, we don't have the keys. It's, you're going to have to go find somebody to crack the keys for you. We're not capable of doing that. But this would be amazing if it recorded everything I did on my phone and understood, like, okay, yeah, I play chess during these time periods. I have this many open games. These are my weaknesses. I have these flights coming up. I use this app. I used Hotel Tonight when I was in Los Angeles, but I opened Condé Nast Traveler on my browser and I bookmarked three or four things. Like, the insights it would have combining my desktop with my phone, with my watch, you know, with my habits in Apple Music and iCloud, my photos. Man, it could be a truly powerful assistant, but I would have to give it access to everything. And you know all about that, Jeffrey. Running Hermes where people do have that concern, et cetera. So what's your take on what would make Siri extraordinary as Stephen wants it to be as an Apple fanboy? Well, it certainly has something to do with, like, the end user product experience, right? Like, that's the one thing that we know that they've been able to consistently deliver. You know, I will comment as a side note before I answer. You know, it has been interesting when you talked about how they sat out two rounds of, like, the AI race. You know, people kind of, you have to wonder, like, was this either, like, the world's greatest bag fumble or were they, like, playing 10-D chess and, like, really understood that, like, there wasn't, like, because what would have been the standard playbook, right? Everyone was like, we got to go, you know, we have to train our own models, too. We got to go spend, you know, raise the trillion dollars, you know, to go build all these data centers. And, like, did there need to be a fourth or fifth player? You know, some business people would say, well, we need that intelligence in ourselves. Like, we have to be able to do it ourselves. We can't rely on an external partner. But maybe, you know, as, you know, sort of the model costs for training these frontier models just continues to get higher and higher. You know, it was possible that Apple was, like, you know, has made the right choice by sitting out. We're not going to play that game. There is some historical context here. You know, you remember, I don't know if you're old enough, Jeffrey, but Bill Gates was running around when he was CEO of Microsoft with a tablet, with a stylus, with a pen, showing off a touchscreen Windows device that crashed. And then they had their little Windows HP personal assistant. What was that called, Stephen? The Windows on that personal assistant? Windows CE, maybe? Yeah. Windows CE was the mobile platform. That was it. Okay, great. So, Jeffrey, you are old enough. Either that or you're a student of history. I'm 38. I'm old. I'm old enough. So, you were, like, eight years old when this all went down. But anyway, putting it aside, maybe you were playing with Windows CE at eight years old. Steve Jobs was like, you know what? I'm chill. Oh, there it is. I mean, put it back up on the screen for a second. Let's all just ponder what we're looking at here. Why was that not a gigantic hit? I mean, it just... Look at it. I mean, I feel like... I feel like I'm in the army. And I'm, like, setting up a cannon. Calling it a nuclear striker. Yeah. Yeah. The Newton as well. Look at those chiclet keyboard. Oh, my God. Those were brutal to touch those keys. It took, like, three or four times to get a space in that space bar. What a disgraceful product. That's it. But Steve Jobs saw that, and he was like, yeah, not ready. And then all of a sudden, he's like, yeah, you know what? Congratulations on your, you know, all your special books and tablets. I'm going to come out with a tiny one called the iPhone. Shout out Blackberry. Shout out Nokia, which were N95. All that stuff's great. But, yeah, he sat it out for four or five years. And then, boom, iPhone. Perfect. That could be the play here is their partnership with Google gives them access to lots of compute. And they did the same thing with search, right? They're like, we don't need to be in the search engine business. We can just have a great partnership. And they become a key player in driving distribution. Yeah. They do. So, distribution, great. They actually are also, you know, focused somewhat on the open source side a bit with their MLX platform with the Max, which basically is their own platform for running local models on Apple Silicon. So, they have invested in the open source side somewhat. You know, there's always this question about, you know, can you run stuff like this totally in the cloud? Or, like, is the latency requirements going to essentially dictate that a lot of things move onto the device? And so, it'll be interesting to see which way it goes. But certainly, the open source models right now aren't in a space where they could record everything and have all your context. But as we've seen in the space, it's legitimately only a matter of time, right? Where does your project and Siri conflict, overlap? And how do you think about it? Because I would look at what you're building with ERMES or OpenClaw as the most open, extensible platform for what Siri and Alexa should have been. Sure, a little more complicated. But when I introduced you on the program, hey, fit and finish and polish is something you're working on. So, how do you think about getting ERMES agent onto iPhones and beating Siri? How do you beat Siri from taking your business, Jeffrey? That's a good question. Now, Apple has made it clear that if they want to win, they will win because they own the device and they'll make the final call, right? You're saying they'll uncle your product? Yeah. So, really, I think a lot of it is just about where we're framing. We look more on the desktop application power user side right now as sort of like self-modification. So, for example, one thing that we put forward is the agent able to change itself dynamically, where it's actually able to modify its own code, change features like that. That's something that probably doesn't fit within a pure customer user, like a full iPhone experience. So, I think they can be complimentary for now, but certainly the thing that they showed off about it, like controlling all your apps, that's something that we're desktop and computer usage is something that we're working extremely hard on. Because if you think about it, like we spent the last 15 or 20 years forcing all of the white collar economy into these that can fit onto a screen, right? Like everyone does their whole job just through a laptop screen, essentially, right? So, theoretically, you know, all the tools are there on your computer to like do everything that people consider to be valuable, right? So, I think that, you know, we're going to be entering into this new era where we have the scaffolding of the apps and all the old programs, the websites that were meant for humans to use, right? And the scaffolding has been built up. And now the agents will be using this scaffolding for some period of time. But it may be that this is not like the final form factor. And I do think there's going to be an incredible amount of innovation on the UX side. And I didn't even see the keynote. Like, for example, what you guys mentioned about, you know, just looking at the agent, you know, looking at a specific piece, part of the screen and that like engaging it. That's like the kind of out of the box thinking that I think can drive new, can not only like make AI more, it makes AI more useful to people because it gets them into the funnel without having this sort of janky experience that we're used to. Ross, who's going to win this agent race? Is it going to be a third party who crosses across all platforms, doesn't own the platform, like OpenClaw, Claude Code, you know, pick your favorite agent, Hermes, obviously. Or do you think Apple winds up winning or maybe people have both? You know, they use Google Office, Microsoft Office, and Apple has an offering. I don't know what it's called. I'm sure. Well, you know, I think like I have a- Some people use it. Apple, right? What is the Apple word processor called? Pages. Pages. Pages. Pages. Yeah, yeah, that's right. I think there's one person on this call who still uses pages. Do you write your books in pages, Stephen? I know. I'm writing them in notebook. I don't know, actually. But that's a whole other story. Fascinating. I'm going to double click on that later. Apple pages. Go ahead, Russ. Who's going to win this? I have a- I don't know if it's a controversial take, but I have a take. All right. So here's my take. It's a bit higher level though. So I think that the one that transcends devices for agents is going to win, not the device specific in Apple's case. They're incentivized to be device specific because they got to sell devices. That's like their bread and butter and that's their cash cow. But so here's my take. Let's say that the goal of like this current era of AI continues to get better and better is to automate jobs that we don't want to do, right? Let's just imagine for talking purposes, that's what the ultimate goal is. Agentic AI, so digital workers, AI agents that are confined to the digital world, right? They run inside computers. They can use computers, all of that stuff. The jobs to be done there, to be automated are all related to work, right? Like it's all the things that humans don't want to do is all this work stuff, right? And then in your personal life on the consumer side, the work that you don't want to do are like the dishes and like folding laundry. Chores. You're referring to chores, right? Like the rest of your time in your personal life is mostly just entertainment. Apple builds amazing devices for consumption of content for entertainment, right? But like it really boils down to agentic AI is really for the workplace and the enterprise and all of that. And then physical AI or embodied AI in the form of humanoid robots is what's going to automate all of the work in the physical world, in your personal life. And that's where I think the two big, like call it like economic value drivers, they're going to kind of accrete to those two specific buckets. And so in that world, if that is true, then I actually don't think that Apple really has a play here because they're not really focused on the enterprise from agentic AI perspective. And they're not even really focused on robots from a kind of like personal life chore automation perspective. And so that would be the ultimate mind blowing win, huh? Stephen Berlin Johnson. If the new CEO, I forgot his name of Apple, who is a real hardware engineer. Hardware guy. Yeah. Yeah. He just said, yo, by the way, we're making a robot. We're in the game. You're going to go to the Apple store and there's going to be robots walking around selling you watches and you can walk out with one. Maybe that's why they're slow playing AI so much, right? Like the AI agent stuff. They don't really care. They're in the lab working on the robot already. Who understands consumers better, Stephen, like for chores? If we take Russ's framing as like, these things are great for chores you don't want to do. So will Apple get into the human robotics space and who would you rather buy a human robot from? Is there a brand you would rather like have a Samsung or a Toyota or a Tesla? I thought you were really excited about Elon's robot. I'm excited about all the robots, frankly, because I have a ranch and there's a lot of work on there. There's a lot of brush to clear. Literally, I was out there cutting the grass, which I love doing. It's cool. You're like Reagan. Mowing therapy. No, no. You know, you're going to find this hysterical as the Brooklyn grinder now living on a horse ranch out in Texas. I love getting on my ride on mower and just riding around and cutting the grass and making a path. Living the Hank Hill life, huh? Oh, I love it. I love it. It's therapy. You could ruminate and go to a- I would love to time travel back to 25-year-old Jason. In black. At coffee shop. At prom. This is your future, man. It's just literally like I'd be like, what? That's some corny shit. I think you're confused. I'm so confused. Take us home on this one with agents. Are agents just WYSIWYG boxes and web pages and, you know, just a commodity? And we're so enamored with this concept that we're thinking somebody has to win it when, in fact, it could just be like, like I said, a WYSIWYG editor. Every product has one and it's a commodity and they all work relatively well. Yeah. I guess I was kind of leaning in that direction, Jason, when you were asking like, who's going to win the agent thing? It's like, it's a bit like asking like, who's going to win the website thing in 1995? Yeah. Like, they're going to be everywhere. Like we, so I'll tell you how we're thinking about it in terms of notebooks. So we just rolled out this thing I mentioned earlier yesterday, which is like the biggest change to how the product works. And effectively before we had kind of three agents, right? We had an agent that was really good at like looking at your sources, helping you understand your sources, finding the information you need, summarizing, explaining. We had an agent that was great at making things like audio overviews and slide decks and things like that. And then we had a research agent. We had a deep research agent that would go off and research. But the problem was they were three separate things and they didn't really talk to each other. And so we've integrated them all into kind of a single agent that you can experience in the chat. This is as of yesterday, just for ultra users, but we're coming to more users. But basically, so now the agent knows your whole chat history and what you've researched and all your sources and the artifacts you've made. And so it lets you do precisely the thing you were talking about before, Jason, about like asking about blind spots. So I can go in there when I'm in research mode and I can be like, hey, okay, I'm trying to figure out this problem. Like, what are we missing here? Like, what's the thing I haven't thought of yet? And because it knows, you know, weeks and weeks of our conversation on the chapter I'm trying to write, or because it knows the sources that I've already assembled and what I've done, it's able to actually like help me see around those blind spots and find the precise thing that I need, which is the fact that you can do that kind of negative search. Like, I'm searching not for a keyword, but I'm searching for a topic. I'm searching for a thing that I can't tell you what it is because I don't know what it is. And it actually is able to fill in those gaps for you. Like, just as a researcher alone, that is such an incredible god sign. It's like some weird... The genetic work is so exciting to me. It's super exciting because it's like having a personal coach or a professional coach who's like, yeah, you're getting better at chess, but here's some heuristic. It's not even in your mind. Let me blow your mind by telling you how to think about the openings, for example, in chunks, right? And the great chess players are not thinking three moves ahead. They're thinking three chunks ahead. Okay, we're going to exchange these pawns. Then that's going to unleash the rooks. Then that's going to unleash, you know, this pin that I've got blocked. And they're thinking in those chunks of heuristics. That's like, what's going to happen with these agents? They're just like, hey, you're thinking about this in a very ABC way of playing poker. Let me explain to you. Poker is even the better analogy because you learn it by losing money, et cetera. Just what am I not thinking about? And in some ways, it's almost like what super intelligence would be, Jeffrey, if we define AGI as like being like, I don't know, the smartest human at Google, the smartest human, you know, at news. Okay, great. You're the smartest human here. Congratulations. Everybody hates you or people think you're weird. Fine. But being so super intelligent that human intelligence is not comparable to what you're doing. In other words, we can't comprehend how you're looking at the field. That's kind of upon us now, I think. We're starting to emerge this year. Yes. I would say certainly we are like what I tell people we're at like a functional AGI level. It's unevenly distributed among the tasks, but on certain tasks, like it's as good as the best, you know, people that there are. And you only have to look at, you know, the people who actually are at the top, who are, you know, your tarry towels and stuff like that, who turn around and say, like, look, I'm the best there is. And I'm telling you, this thing is as good as I am. You know, like you sort of have to have that level of like cachet for people listening. But I mean, personally, myself, you know, I'm like mentioned earlier, I'm 38 years old. I started writing basic when I was like eight is what I did all day, every day. You know, for the last 25, 30 years of my life, it was, you know, what I centered around my own professional identity on everything that I did, you know. And only within like the last three to six months, like I no longer really write code in the same way that I used to, you know. And that's really like and that's because it finally got as good as I am. And but the thing is, rather than viewing that as like this defeatist moment of like, oh, my gosh, what am I going to do now? Now I get to be like an even better version of myself because now I can do 10 projects when I used to only be able to focus on one. And so like rather than viewing it as like a it's like I get to rewind the clock and be eight or nine again, you know, and I get to start over. So I just think people need to come at it from that way. But speaking, Stephen, about what you said, you know, I do think it's interesting because like all of that is just an emergent property of the model. Like we're the agent harnesses we do are kind of like, you know, we're putting some stuff around it. But all of it is completely pointless if the model itself can actually do it. And so many of these like new use cases just sort of silently get unlocked as the models get better. Like there's no like specific spot. But all of a sudden it's good enough to notice what it is that you haven't said. And it has, you know, I'm sure functionally a million window context length. So it actually can take everything. And whereas like two years ago, you know, you had your 32K context length if you were lucky. So you had to do all these other like, you know, hacks and stuff and just things sort of like frog boiled to the point that we are functionally at AGI at several categories. And yet it's interesting because like the world is different, but it hasn't completely broken either. Right. And, you know, maybe, you know, we actually are on like a trajectory where it's all going to be OK, even with the hyper ASI sort of things coming. Feels like it's going to be OK, Russ. This week, and I went down the rabbit hole of building apps and I was just thinking all the things I had asked people to do two or three years ago and was like, yeah, we don't have the time or budget or team to allocate towards that web interface, agent, researcher, SDR, et cetera. And I just with my pedal, which I have a pedal under both of my desks here and whisper, I just started talking stream of consciousness into perplexity computer and Claude Cowork would cut one into the other and just said, bill me these. And then I looked at the two outputs and I put like three or four jobs into each. I'm paying for the two or three hundred dollar version of both. I put two or three jobs into each. I queued up a bunch of things. I looked at the output and I was like, oh, my God, I was considering spending a quarter million dollars building this piece of software or hiring two people, which would also be a quarter million dollars. And I just built it all. And it's running in the background. Whoa. And it's actually working. And when we tried this with Open Claw in January, February, March, it sometimes worked, but was brittle and broke. Something's happened, I think, in this year, very clearly in the first half of this year, where this stuff went from brittle to brilliant. Oh, yeah. Boris Charny talks about this, the cloud code crater. You know, like it really crossed the first chasm was sometime in the summer of 2025. But then when 4.5, Opus 4.5 hit is like when he started to write significantly less code. And then I think, you know, 4.6, I don't think he's writing any code anymore at all. Like he's been doing these interviews where he's been talking about that. But it's interesting, you know, like with your pedal, if you think about it, you kind of have this early version of Jarvis. That's kind of how you're interacting with the computer now, where you're just like telling the thing, literally like speaking to it and saying like, hey, build me this thing. And it's just like generating the thing after a few minutes. And then you go in and you can refine it, right? Like you can fine tune it. You could even look at the code that it generates if you want to. But this is like, yeah, the very early version of Jarvis in 10 years. Just imagine what it's going to feel like to create things. It's going to be this incredible experience. Here's the tweet from Boris. I use Opus 4.5 with thinking for everything. It's the best coding model I've ever used. And even though it's bigger and slower than Sonnet, since you have to steer it less and it's better at tool use, it's almost always faster than using a small model. That was January 2nd. 4.5, like before you even said it, I was going to jump in and say 4.5 was it. Like that was the line where suddenly, you know, it's funny because if it's even like 2% or 3% less good than me, well, why would I have it make something not? But like once it's even like 1% better than what me, then like why would I ever do anything myself at all? So, and it just kind of 4.5 is when they cross that line and it's just been gas. And I think now we have maybe Fable out now today. We'll see. So it's going to get even crazier. I was going to say, Jeff, I think it'd be interesting like, you know, with Fable out, like how your harness evolves too. Just because, you know, one of the other things that Anthropic talks about with harness design is that as the models get better, you just keep kicking stuff out of the harness. Like it just all kind of like, you know, becomes more and more of the purview of the model versus the harness itself. Yeah. I mean, I think obviously like there's going to be places for the harness to be useful in the sense of connecting it to the quote unquote real world. The real world being like, you know, the digital actual world. You know, I think there certainly there is a future. Like if you asked me 15 years from now, like there is no reason why we couldn't have full model architectures that can like speak and read TCP IP packets. Like they're multimodal. Like, you know, they literally could be the entire brain. That's probably we're still like one few generations away from when we get to there. But obviously, yes, everything else was sort of like the scaffolding to like pick up where the models emergent capabilities left off. And you can just kind of like delete more. Let me give you a per for example. Suppose now Fable is still a million context link. Suppose we got like 10 million context link and it was like 10 times cheaper. Right. You had 10 million context link and it's like 10 times cheaper. Then you'd be like, well, why even why even like just record everything that's ever been like you said, you get into this record everything that's ever been said. And just that is your context. And, you know, we don't even need to like have sessions and, you know, be clearing all and doing compaction and all this other stuff that we still have to kind of, you know, pick up a little bit. So certainly it's only going to get less and less handholding as the models get better. And like I said, 10 years from now, I imagine there will actually be like the final the final boss to be defeated of the old guard is the operating system. Right. Like that's an interesting. Yeah. Can you get down to the operating system level and just take a hardware platform? Yeah. Yeah. And it would need to be fully multimodal. Like I said, it'd have to be able to speak TCP packets, render video at real time. You know, all of these, you know, 60 frames per second, 4K video, real times. All these things sound crazy, except for the fact that we have, you know, quality, we have human level hyperintelligence of coding. When eight years ago, you know, you had the XKCD comic about, we can't tell what bird, you know, what bird colors are. So it's like, there's not, it's not crazy to project that far out 10 years and be like, it might actually. As a, as a way to frame this, Steven, if you took everything we did in web 1.0, you know, just take the top 200 websites and you gave it to one person with unlimited tokens right now. I would say every hour they could rebuild every one of those sites, which means in but four or five weeks, they could have rebuilt every innovation for the first, I don't know, five, six, seven years of the internet. Then you fast forward a year from now, what would it be capable of doing to Jeffrey's point? Maybe you could just build the operating system, take out an old, you know, laptop and say, hey, make me a new operating system here on this laptop. And I'm going to compete with Windows and Linux. I mean, it's kind of a weird moment to frame it that way. Well, also on the UI level too, Jason, I would say like, you know, yeah, the underlying operating system, but like, what is the kind of basic visual metaphor? Like it's something I've been obsessed with forever. Like I've like wrote a whole book about computer interfaces, you know, and you know, we've, we've kind of been like, okay, we're going to standardize around these particular models. There's a browser and hypertext model. There's a like Windows model and stuff like that. But maybe the future AI kind of will just spin up the perfect UI for you, given the project you're working on. That is, looks like nothing that's ever come before, but because it knows that you are actually working on this kind of project with these competing needs and you have these tasks that you're trying to do. Like it will invent a whole new surface for you on the fly. Like that's, it's, it's, you know, to Jeffrey's point, that's imaginable now. You know, whether we want that, whether people want to have standards and a traditional way that they use a computer so that they feel comfortable or whether they want that open-ended space, probably they want both. Wow. Yeah. And the interface is adapting in real time to your need. All right. Let's talk a little bit about token maxing. Jeffrey, people are going wild. Uber decided, uh, Andrew McDonald, I think their CFO president was like, oh no, president was like, you know what? You used up all your tokens in month four. We're going to ration them now. You blew out the budget. Supposedly there was some report of Amazon potentially blowing through like hundreds of millions of tokens and somebody forgot to close the faucet in this metaphor. Um, and then some people just saying, hey, the amount we're spending is not equaling the actual output. What's your take Jeffrey on the value of token maxing? And if you should just YOLO it or people need to be more thoughtful about it. Well, I think, you know, in a corporate environment, especially at a large corporate environment, you know, you, there is sort of a failure case. And I think we're starting to see this failure case, which is if you have a, you know, you have an employee and now you give them an unlimited budget, they'll do their entire job with that budget. Cause it's just easier to say to the computer, Hey, do the thing. Hey, do the thing. Hey, do the thing. Keep saying, Hey, do the thing until it does the thing. And if the output you get is just replacing what that worker previously was going to be doing, you know, like you're getting the same output, but now it costs twice as much because you have to pay their salary for them to sit and turn on the computer and say, Hey, do the thing. Hey, do the thing. Like, what did we actually got here? You know? So there's a failure case where if people can just be lazy, like, you know, water goes downhill. Right. And like, there's a lot of that, that if you enable it in an irresponsible way, it will go downhill and people will, you'll get maybe like the same output for twice the cost. And that's certainly not what you want. So I think what's important is to be identifying the, you know, what we do at Noose is we have budgets for people, but there are people who are worthy of 10 X the budget. You know what I mean? Like, and how do you determine that? How do you determine that? Yeah. I mean, I don't know how you do it at scale, to be honest with you. We do it at Noose because we have 40 employees. I know all of them, you know, I can, I can, I can do it. So, you know, a couple of like, like Technium, who's like the lead maintainer on, on Hermes agent spends 10 times the salary, you know, on tokens and it's worth everything. Millions of dollars on tokens a year. What? Well, we have, we pair, we're, we're, we're pretty good lean shop. So I wouldn't say millions, but like, let's say, you know, around, around the million of dollars on, on tokens, maybe, um, annualized. Uh, and so, um, and it's totally worth it. Totally worth it. Would, would spend it again in a heartbeat. Right. Um, but now if we have a junior developer who is, um, you know, spending, you know, three days to just, you know, center a div on a, on a thing. And they're just telling, you know, Opus move the file and, you know, instead of copy, you know, opening up a browser and copying it themselves, then like, we got a problem here. Right. So how you tell, I think is still an open question, um, at scale, but it's certainly something that I think is a problem that will need to be solved. Russ and Romance. Jason, one thing to just quickly jump in, kind of brings us back in a way to your original question about the booing and the, and the AI backlash and the booing at the graduations in the sense that the reason I think these things are connected is that in that backlash, there is a conception largely among people who haven't been using the technology recently that the tech companies are foisting this unwanted AI technology on the public and spending all this money on data centers and all this kind of stuff when nobody actually wants to use this thing. And that's, you know, this is not something you would hear in this podcast, but like, if you go outside of the tech bubble, you hear that all the time from people. I hear that from my New York friends all the time. And I think what hasn't been explained enough is that the reason data centers need to be built is largely because there is so much demand. Like there is so much demand for these things, whether it's the extreme of token maxing or just people like we at notebook or just, I don't know, we have so much more demand. And so people want to use these tools and that's why there is this incredible, you know, increase in spend. It's not because we're, we're inventing a fictitious market that doesn't exist. You know, I, in some ways I look at this like water, Steven, you know, if in America, I invested in a company that did like water monitoring and it failed. And then I had another one and it failed. And I was like, but water is this incredibly precious resource, you know, resource. And everybody wants to save water. And we're constantly talking about it and hand wringing. And then I realized it's a tragedy of the commons type situation. We actually don't, nobody even knows what their water bill is. If you get to the point where you know what your water bill is, because you're a golf course or you're planting almonds, you then go bribe some politician to get a special dispensation. So you no longer have to worry about the cost of the water. And what we really need to be doing, if it, and I think it's the, whoever had the water runs downhill. I think that was you, Jeffrey analogy. It reminds me exactly of that. If you go to somebody who owns a golf course, they just let the sprinklers run because they're not, they got some dispensation for their golf course. If you talk to anybody in the almond business or like, well, we don't pay what you pay for water. So F it. We're going to make almonds. Is it almonds that use all of the water? Yeah. By the way, almonds are a completely mid nut, like cashews. Very much better. I'll take a peanut. I love a good peanut in a, in dark chocolate. I mean, I could give you 20 nuts. What macadamia? Let's just rank our favorite nuts. I think that'd be a good piece of the rest of the time. What do you got? Macadamia, cashew? Where are you at? I'm a pistachio guy. Pistachio's great. Macadamia is up there for me. I mean, there's so many better nuts that are more water efficient. And these assholes with the f*** almonds, sorry, bleep that out. They're just so selfish. They're like, yeah, we're just going to use all the water. And then we're going to complain about a closed loop data center water. It's just absurd. Russ, your thoughts? And first, I need to know your favorite nuts. Yeah, macadamia is probably my number one, but super fatty. So I don't eat them very much. You know, I think like similar thoughts with Jeffrey. I mean, we have budgets, but then there are folks on the team who we kind of let them do whatever they want. And I think the general feeling is this. I think for the line of work that we do, the value of the token is so high, right? Like just the output that, you know, an amazing engineer can have. They, you know, they have that whole like 10X engineer. It's like it turns them into a thousand X engineer or whatever when they have AI kind of out their back. And so I think that it's one of those things where we don't really, for the best engineers on our team, we don't really care. Like they can, you know, spend whatever they want. What's the max somebody spent in a month? Be honest, Russ. Because you got the bill. Yeah. One person max spend. It's not that bad. Maybe like 10 to 15K. It's not that bad. Oh, that's it. Okay. So yeah, that's like, you know, whatever, 150K a year. Yeah. It's not terrible. It's not terrible, but I think this is, have any of you experimented with running local models and just saying, you know what, we're going to spend $5,000, $10,000 on a workstation, which is like a fancy way of saying a beefy PC. I was having this conversation with Michael Dell and with Jensen from NVIDIA, dropped two names at once in the same sentence. But I was talking to both of them about this concept of like, isn't this your opportunity to sell everybody a $20,000 desktop running a local model so they can token max without looking at the register? How close are we to that possibility, Russ? Well, I think like maybe, and maybe Jeffrey has more in-depth thoughts on this just because I think he's probably closer to having played with all of these models. But my general take is that the local or like smaller parameter models are just not that good in their quality. Like, I mean, let's talk about for coding agents, right? Like, I just don't think that they generate the same value of token for these smaller models. I do think that they have their place. Like, I'll give you an example as it pertains to voice. All of the frontier labs now, you know, the coding agents is the most lucrative type of agent. Voice is actually second after that, but coding is the number one. And all the frontier lab models are in favor of like having better, higher quality tokens through more thinking. They're kind of sacrificing time to first bite. So the speed at which you can generate that token. And for a voice-based interaction, you actually want something that favors speed to a degree as well as quality. And so these smaller models like a Gemma 4 31B have like pretty solid quality on the conversational side, but also a much better time to first bite, you know, which is important for speech. But then again, for coding agents, I don't know if that's true for these local models. Yeah. I think for coding things, people kind of, I look at it, I want my engineers using the best, right? And, you know, the second people come out, I'm telling our best guys to switch to that, right? So like, it's already like expensive as it is. It hasn't like sucked out. Like I still would be willing to spend more if it could be better. So you spent a million with one person. You spent a hundred thousand in a month with one developer. Yeah. And you're willing to spend more. I'd be willing to spend more for him. Yeah. I'd be willing to spend more for him. And so for that, you know, I think it's the local story is a little bit different. You know, it's about finding who's that person who would spend $20,000, but wouldn't spend, you know, more than that. It's a little bit difficult. And the real thing is we're still scaling. That's the problem. You know, is that like, we're still making the models, we make the models bigger and they just still get better. The bigger we make them and we throw more debt. So like if we'd hit like some sort of plateau and we could have a compression event, maybe. But like we're talking in the era of, you know, the next, the Verirubin chips that are coming out, you know, they're designed for 10 trillion parameter models, basically. Just like the current gen were designed for trillion parameter models. You know, we're setting it up for 10 trillion parameters for the next. And I'm guessing Feynman will be for 100 trillion parameter models. So like at that scale, until we hit some scaling wall, it's a little bit difficult to see how this train doesn't keep going in the direction it is. Yeah. It feels like it's heading there. All right. Let's wrap with bad VC stories. This is like the crazy trending topic on Twitter. And I don't know. I think it started with Greg Eisenberg describing a GP falling asleep. And then Mark Pincus had a VC fall asleep. I had John Dorff fall asleep famously in a meeting with me. But it was actually when I was doing Mahalo, human powered search. Uh, he had gotten in a, a, a biking accident in Woodside that morning, flew off the front of his handlebars. He had his arm in a sling, scraped down his face with a bandage. And I guess he had taken some kind of painkiller and he was, he nodded off during the meeting. His partner subsequently came to me and said, John fell off his bike. That's a, you know, and he's just very sorry. And I'm like, why did John Dorff show up for this meeting? If he fell off a bike, he's like, he's going to the emergency room after the meeting. I want to insult you by not showing up for the meeting. I was like, wow, that's a great one. But I'm curious, you know, you've run companies before Steven and obviously Jeffrey and Russ, you're in the game right now. Any crazy VC stories, uh, that could top the ones we're seeing right now? My, most of my VC interactions, my main VC at, uh, outside in the company was Fred Wilson. So I, I have only great stories about Fred Wilson. He's awesome. He can be a little spicy though. He can be spicy, but, but he was. He was a spicy Fred story. No, no, no. He fly off the handle. You got it. I just was reminded of another story that Biz Stone told me that I love, which, you know, for co-founder of Twitter. And he was raising money for this thing he did called Jelly. And it was this dream where like Bono from U2 was like affiliated with like Elevate or some kind of partners. Like, so he was like briefly a VC as well as an international. And Al Gore was a VC. This was the age of the celebrity Ted VC. So Biz told me this story about how he was pitching Bono like on a voice call and Bono's like, okay, Biz, lay it on me, man. Lay it on me. And he's like, wait, wait, wait, Biz. I want to lie down so I can listen to it while I'm lying on the floor so that your idea will wash over me. He's like, I'm lying on the floor, Biz. I'm lying on the floor. Lay it on me now. Just pour it over me. Like, oh man. Wow. Yeah. The age of the like A-list celebrity VC, kind of good that it's over, I think in some ways. Now a B-level, no disrespect to Ashton Kutcher. He's not A-list. I'm D-list. He's probably B-plus. Like he's been in a couple of movies, but he's not a leading man. But on a TV show, maybe would be the leading man. That's perfect because they're hungry and they're going to do the work. But if you're Bono or like Al Gore, you're like, hey, let it wash over me. Russ, you can abstract these stories. Yeah. Stephen, I want a spicy Fred Wilson story. He can get spicy, but. I definitely won't name any names other than I've pitched Ashton before and he was great. You know, he took the meeting, taking his kid to a summer camp and his kid was wearing like a bear suit. I don't remember why, but that was fun. But he was great. He rallied, joined that meeting. So nothing, nothing bad to say about Ashton. But yeah, one, I got a few of these, but. No, the one, the one that you're just, I could see when you look down as a poker player, you're considering, should I go all in with this story or not? No, I'm not going to. I'm going to give you permission to go all in. Don't say the name. Abstractly. Yeah, I'm not going to say the name, but you know, all top funds. But one, one that I pitched, I remember it was just the worst. This was for a previous company I worked on maybe 10, 15 years ago. And so I go in there and sit down. Um, they asked for a meeting with us and, uh, the, the investor, you know, shows up late. Um, and then he shows up wearing like, uh, this, like. Super tight, like tank top, like wife beater thing. And then he's got like a blender bottle in his hand. And so like, he's like, I'm so sorry. I'm late was doing a workout. He's kind of sweaty. And then like sits down and like, you know, get the deck going, start the pitching. And the dude, he's got like the blender bottle with like the metal thing inside. And he just starts shaking this thing while. Shaking his protein shake while pitching. Yeah. And then like popping it back, drinking, shaking it up. I'm like doing the whole pitch. He's still shaking. It's like just the weirdest experience I've ever had. Um, well, that was a good one. You definitely know the person, but I won't say the name. I mean, I'm sure I do, but that's a kind of, I, that would be Keith. Where boy is a workout addict, but it's not cute. It's not Keith, not Keith. No, Keith, Keith would say something spicy, but he wouldn't be late being late for the meeting rule off. And I, uh, rule of had a rule when he was running Sequoia, where if you were late, it was a hundred dollar bill and you had to keep a hundred dollar bills on you. I was taping all in at the Sequoia offices last week. And I brought my latest accelerator class there to meet rule off. I'm trying to get off the call. Tremont's got to, and Sax has got to get his one more thing. One more thing. Yeah. I'm going to finish. I got to get the last word. And I'm like, and rule off walks up to the conference room. I'm in smiling. Going like this. Cause it's, you know, it's two Oh five. And he knows you. I got to pull out a hundred dollar bill. I give him the hundred. I have to autograph it. I have an autographed hundred dollar bill from him when he was late one time. Uh, but that is just the table side matter that you just have to have. If you want to be in these deals, Jeffrey, are you based in France by the way? What's the whole thing with the French everything new research? Yeah. You're from Detroit, but you're a Franken file. Explain to me what's going on here with the word news is actually not the French word. It's the Greek word news, which means mind or intellect. So there's Greek have like, it's same thing with like the word love. They've got like four words that all get translated. The French is all derivative from the Greeks. We know that. Yeah. Yeah. So it's, it's, it's the word, it's the Greek word. Uh, I'm, you know, I've pitched a bunch of people too. Um, I don't have any like crazy stories other than, uh, I just love the, uh, the pretense of like the, uh, the come sit you down in the boardroom and we're all sitting and talking to you in my experience and having pitched a lot of people, the best ones are the people who are just real and honest. The best calls I've ever taken are with one guy who's just on his iPhone, just talking to you and being real versus, you know, we're going to fly you to New York and set you in the boardroom and do all that. There's a place. You don't like the pompous circumstances. I think you represent circumstances. Yeah. Is just, uh, uh, come on, what are we doing? It's performative. It's intended to either intimidate or impress. And, uh, a real founder will see straight through that. And Peter Thiel, or I should say, yeah, see that was me. It was okay. Truth be told. This is why you did take a picture. I'm sure I haven't rust. This is why I didn't want to say anything. I passed on investing. It's true. This is when I was a little bit thinner and I was more like, I was less bulked up. You ripped there. Really? Yeah. This is when I was thinning out. Actually, I was going to a competition. I did show up a little late in the wife beater. It was- Yeah. I forgive you, Jason. It's all right, man. Ignore it. Are we, are we good? We're good? Yeah, we're still buddies. Yeah. You had me on this show twice. It's, it's all, it's all good. Yeah. I'm just trying to make up for my mistakes. Right. Yeah. God, you guys. So what I did in constructing this, having been on the other side of this, I told another one where more David, I named them, found out I got a term sheet from Sequoia. They used like two or three people around me to beg me to do a meeting with them because they had famously, I believe, passed on Google and they saw Mahalo in the Valley did as like, well, maybe Jake Al could figure this out because he's good with content. And, you know, he did blogging. And so this feels like that a human powered search engine would be like Wiki plus blogs plus a search engine. Right. And then even Marissa Mayer and Sergey and Larry were like, this is a pretty good idea that, you know, this is before AI could do what it does. And it was 10 blue links. So like one of my really good friends begs me to take this meeting, even though I have the term sheet, I'm like, fine, I got to go to Sequoia anyway, more David. I was next door. Guy cancels the meeting on me. I'm flying up from LA. I get up at five. I'd get the six o'clock flight. I'm coming up for four or five hours. I rent the car on the way there. My phone goes ding, ding, ding, like messages. The guy sends me a voicemail. Hey, the meeting's been canceled. I couldn't get my partners around it. Now he knows I'm flying up. He leaves the message while I'm on the plane. So I said, you know me, Steven, I was full contact in my younger days for better or worse. Indeed you were. I was full contact. I'm like, fuck this guy. I'm going to his office and I'm presenting. So I show up. The partner meeting's occurring in a glass fishbowl. He sees me. He turns white. And I look at him. I give him one of those. He comes out. Did you get my message? Oh, I got your fucking message. Wow. I got your message after I got my rent a car and came here. He's like, well, I'll make it up to you or whatever. I said, let me explain something to you and how unsuccessful you're going to be in your career as a VC. You're the stupidest human being I've met. Not just the stupidest VC, but the stupidest human being I've ever met. You're going to be the biggest failure in the history of venture capital because you could have let me pitch, told me it was a brilliant idea, and then said, oh, we had a conflict in our portfolio or I couldn't get the partnership around it, but we want to look at the series B. You had a hundred options of how not to be a dipshit and salt and insane founder. And you know I'm insane because you could type Calacanis asshole into Google and find innumerable stories of me acting like a lunatic. I'm going to say bad things about more David Dow for the next 20 years. And here we are 20 years later, and I'm still not over it. Funny because he has an insane founder story, and it's the same story, weirdly. It's the same story. So we're all crazy. Exactly. But I am of the school that these stories all become legendary because there's so little at stake. You know, that's the reason this is so contentious is because there's so little at stake. Like, just move on, folks. Get the next VC. If your founder is acting insane, just move on, right? But you have to have great bedside manner and be of service as a VC. So this is my public service announcement. I'm not going to end on this. I studied this, and I realized founders don't want their time wasted. So we're going to offer them a first call, an introductory call, where they share with us for 10 minutes what they're working on. We ask two or three thoughtful questions for five minutes, and then they ask whatever are the questions they want for five minutes. If they want to extend the call, they can. Or if they want to schedule another call for a second call, they can do that as well. And at the end of every call, I came up with a sentence that I say to every founder. Let me try it on you, Russ. Russ, thanks so much for sharing your vision for LiveKit. May I repeat it back to you so I make sure that I didn't miss anything? You say yes. Love it. I repeat back to you my understanding of your business. And then after I repeat it back, I hope I got that right. Is there anything I missed? And then you go, holy cow. He paid attention. He wants to get it right. He's human, so he knows he's fallible. He wants to just have that last moment to understand the business. We started doing that. Our scores went up with founders. I told my team, we're going to score every call. They kind of fought me on it. They would do it every three months. We would send something. Now I have it programmatic. Like 72 hours after somebody on my team meets with you, you get this email from me, from my personal account. This is an automated message from Jason. You met with this member of my team. I would be really helpful if you could score how the call went so we can get better at these. And of course, if you hit reply, you'll get me and I'll reply back as a human. Man, our scores went off the charts. And then anytime anybody has a seven or less, we then do some relationship maintenance and we re-watched the call because we record the call now that it's on Zoom. And we re-watched. We have the person who did the call re-watched the call if it's a seven or less. And man, our scores went off the charts in terms of... So many... Yeah, I just don't understand. VCs, they should know that all the founders talk, right? And we are always like, oh, should I raise from this guy? What do you think of this firm? Can I get an intro to these guys? And for some where I've had a really terrible experience, it's like, don't raise from them. Here's why. You don't want to do it. And so I just think the reputation matters so much. How you treat people matters so much. Even if it's not... If it's a no, that's fine, right? But you got to be respectful and set the expectation properly. Okay. Okay. We've learned a lot here today. To summarize, Stephen wants a dynamic interface. And here's the... This is the AI slop roundup. Stephen wants a beautiful interface. There he is. Minority Report. Looking good. Looking good. You're well-fed in this photo. I see you gave it a couple of points. Yeah. You got a little punch there, maybe. You know, a little extra ramen. Never hurts. Jeffrey, we learned that you're not in competition with open claw. You're not in... You don't consider yourself in competition with open claw. But perhaps, Jeffrey, you're not in competition with open claw because you cooked them. There it is. You said it. Not me. Okay. I actually had to do that once, by the way. I actually had to put a lobster into boiling water when I was 12 years old. My mom dropped a live lobster on the floor in the kitchen and didn't know what to do. Yeah. Anyway. There's a much better thing to do. What you do is a lot of chefs who, you know, know how terrorizing it is to do this to a lobster. You just take a knife. Bang. One shot on the skull. The big sleep without the boiling twitcher. And yeah. Ross, we know you want your Apple human robot. So you picked yours up, I see. There it is. This is the Apple App Store experience with this new CEO, the engineer CEO. All right, everybody. Another amazing episode of This Week in AI. If you haven't used Hermes Agent, where should they go, Jeff, to get started? Noosresearch.com. N-O-U-S-Research.com. Okay. You couldn't have Hermes.ai. You just got to make people. Yes. Just here's a piece of advice. Rename the company Hermes. Yeah. Yeah. That's a good one. Unfortunately, someone has that. Doesn't matter. You just put get or go Hermes in the front of it. And then you negotiate with them. And when you negotiate with them for the domain, all you have to do is say, Jeffrey, we don't believe domains are important. We have go Hermes. And like everybody just finds us by searching Google. But we'd love to have the domain just for simplicity and make sure if somebody else happens, I'd be happy to give you 250K for it. That's the way you present it. You just put get or go in front of it. Russ, where can people, developers, obviously, and people building products find out more? Do you have a startup? Startup program? We do have a startup program. Hit me up for that. I'm on Twitter, LiveKit. Jeez. X.com forward slash DSA. If you want to get in the startup program, we can hook you up with a discount there for some credit. But yeah, you can find out more at livekit.com. All right. Well done, Stephen. Where can people find more about the latest agentic version? Yeah. Yeah. Notebooklm.google.com. That's obviously the site. And yeah, a bunch of folks. The Notebook LM handle at X has a bunch of tweets about it as of yesterday and links to some of the other folks talking about what's possible with it. Lots more to come on that one too. I wanted to ask you, can you set up like a custom domain and publish your notebook for mass consumption yet? You can create public notebooks. People are doing, we have a crazy number of public notebooks have been created. They're still all at Google domain so far. Got it. But we would like to explore that. Yeah. I think that's like a killer thing that Substack did. And I think you could do where like if I had a knowledge base about something I do, angel investing, whatever, love to publish it to like angelthebook.com and just make a notebook my website. So anyway, put that in your feature request. We will come back to you for that. Okay. Sounds good. We'll see you all next time. Bye-bye. Bye-bye.