Raffi Krikorian (CTO, Mozilla)

If these things are making decisions on my behalf, I kind of need to understand why. And I kind of need to understand how. And I kind of need to understand what are the boundaries around it because if it's going to be trustworthy to me, I need to know it's doing stuff with my values. Or at least we have a conversation about it, that kind of thing. Right. And if that thing is now the center intelligence of your personal computer at some point in the near future, it is an extension of you in a way. If your values do not harmonize in some way, that's a really weird cognitive dissonance that occurs. Welcome back to A New Computer, a podcast about the end of personal computers as we know them and what comes next. I'm Stephen Hood. I am co-founder of Telepath, and with me is my co-host. Rupert Manfredi Hi, I'm Rupert Manfredi. I'm also co-founder of Telepath and lead designer. Ruffi Krikorian Our guest today I'm pretty excited about is Rafi Krikorian, who's the CTO of Mozilla. I know Rafi from my own time at Mozilla when he was a board member at the time. He's a really sharp thinker and writer about AI and open source AI in particular. And so we're gonna have a great conversation. We're gonna talk about the role of open source developers in this coming next era of computing. We're gonna talk about how basically we're deciding right now if we're either gonna be the owners or the renters of our next computer we buy. Ruffi Krikorian And then we'll have a talk about our values and how our values as builders play into all this work. So that'll be a lot of fun. But before we start, we want to make it a regular thing on this podcast that we ask each other, Rupert and I, to talk about something that's happening recently in the news relevant to computing and kind of have a little debate and discussion about it. Rupert Krikorian Yeah. So the thing that one thing that popped up since our last podcast, which was, I would say an unusual new entrant into the tech discourse was Pope Leo's encyclical. And I know Steven, you did a little bit of research and reading about this. I do not know much about what was said, so I'd love to hear a bit about your take on this and what exactly is going on here. Rupert Krikorian Well, as someone who was raised Catholic, allow me to decode it for you. Rupert Krikorian Thank you. Rupert Krikorian Actually, I mean, actually, it's true I was, but this is not a subject I ever thought we'd be talking about on a tech podcast. I didn't think we'd be talking about the Pope. It's a little weird. But, you know, like, regardless of what you think of the institution, and people have many different beliefs and thoughts about it, it's this ancient institution that has a big influence. And so it's interesting. Rupert Krikorian That he would weigh in on AI, and he got a lot of coverage over the last week. I read it. And it has a lot of very, let's say, papal language. A lot of everything is from a biblical and theological perspective. But there are some very relevant lessons and opinions being shared, I thought. Rupert Krikorian I thought the most interesting thing that I was struck by was that the discourse online that I saw from certain corners of the internet was, ha, the Pope is attacking AI. See, he's validating that AI is terrible. Rupert Krikorian And I don't actually think that's what he was saying at all. I think if you read it closely, he's saying that basically it's just technology, and technology is not inherently good or bad. It's about how we use it. Rupert Krikorian I think historically, that has usually proven true. And we've often made terrible decisions about how we use technology, right? And I would certainly agree that we're making some terrible ones now about AI. Rupert Krikorian But I don't think it's inherently a bad thing. And there's this sort of discourse that I think you and I have talked about, we've noticed online, especially over the last maybe six to eight months, where there's certain folks who kind of feel like anything engaging with AI is bad and a mistake. And I just don't think that's, I personally don't think that's true. I think it's, if we have concerns about it, it's all the more reason to engage with it and try to use it in a way that we think is productive. Rupert Krikorian I like to think that's what we're trying to do. You know, we don't have a corner on that, of course. And it's complicated, but you got to try. And I think what he is saying, what the Pope is saying is that what he's worried about is this stuff being used in a way that is bad, meaning that it is diminishing people's humanity, it's centralizing systems of control and so forth. Rupert Krikorian And he uses this metaphor. I don't know if you saw this. He talks about describing between the Tower of Babel versus rebuilding ancient Jerusalem, which is like a little bit of a history lesson or a Bible lesson for some people probably. But basically what he's saying is like, are we to decide, are we building a new system of control, this centralized mechanism? That's kind of the Tower of Babel. Or are we building a new city together? Are we building a place that's for all of us, where we're all involved in the agency in it, and it sort of protects and nurtures humanity as a whole? Rupert Krikorian I think what's interesting here is there's a parallel with open source versus closed source. That immediately occurred to me as I was reading this, that open source is the means by which we can build something together that declares some sort of shared values. Closed source is inherently a system of control, is a small group of people with a profit motive typically making a decision for the rest of us. I think there's also something in here about how this concept of machine learning, which then has been like relabeled artificial intelligence or sort of that label has like risen to prominence. And it carries with it a lot of gravity and metaphorical weight and science fiction baggage. And I think it has a sense of enormity and bigness to it, to the point that you would have historical figures like the Pope weighing in, I think, in a way, not just talking about very pragmatic effects, but also this idea of like society shaking effects. And, you know, you have founders of anthropic and, you know, you have founders of anthropic and open AI talking about like inventing consciousness and there, there is, it's got a gaudy vibe, you know, the grandeur of this industry. And, you know, very conflicting feelings about that. I think, I think I'm really interested in focusing on where we can actually make it very small and tractable and useful in somebody's everyday life. But I can't help but notice that is going to have probably some very widespread effects, but also it has this imagination associated with it that can cause people to react. Lose their minds. Lose their minds a bit. Yeah. Including a lot of tech founders and, and yeah, it's, it's an enormous like metaphysical issue. Um, you and I have talked about this a lot, how we think of, we try to think of AI just as quote normal technology. Yeah. That it's not God. It's not alive. It's not awake. It's not aware. Certainly not yet. Um, it's just another tool. And I, I agree with you. I, I kind of feel like the whole God conversation, the whole, like, even the AGI conversation is kind of a weird distraction to me. Like, I'm not against it. I'm not saying it'll never happen, but I'm just sort of not all that interested in it personally. I, I think it's a technology that we've, we can use today much better than we're already using it. Yeah. I think it's, it's simultaneously just another tool and also something where the paper, the Pope is talking about it and it's kind of got a bit of both. And, uh, yeah, it's just a very interesting, weird, interesting time to be observing this weird times. Yeah. Well, let's talk about something a little, a little more mundane, which is actual technology here. I think that you've been talking about and we've been discussing has been this discourse recently about HTML artifacts. Yeah. Let's segue from the Pope to HTML, which is a great segue. Easy transition. Bring it down to earth. Um, yeah. So this is, this also happened since we recorded last, and this is a, an article on X by a guy named, uh, Tharik, uh, Shihipa, who is an engineer at, uh, Claude Code, um, at Anthropic working on Claude Code. And this was titled the unreasonable effectiveness of HTML referencing a fantastic paper on mathematics, philosophy, mathematics for anyone who, uh, is interested in that side of things, talking about his personal workflow, which has spread to a bunch of other members of the Claude Code team of relying not on markdown, which is a typical text format, um, uh, used with these models to sort of collaborate essentially as a collaboration surface, like looking at text, editing text, and sort of having this like shared interface via like text files. Uh, but instead collaborating by viewing rich HTML documents. So Tharik talks about, um, a couple of benefits that HTML have, has for him over just a plain text format, like markdown, which is more dense information. So being able to have like tables and charts and SVGs, faster comprehension. So being able to like have the visual hierarchy, for example, highlighting things, grouping things visually, also being able to have some sort of interaction with the document. And that's something I want to sort of bookmark to chat about a little bit later. So that might be like a slider or, um, a little quiz at the end so that you could retain information that the agent is pulling up for you to review. And then also easier shareability. So being able to like host and share an HTML file and share it with others. Um, and he's using this for all sorts of stuff or the Claude Code team is using it for like code review and design prototyping, visual exploration, like looking at different colors, um, stuff like that. And I think this brings us to, um, you know, something we've been thinking about a telepath and we've actually been building a product called television. This podcast isn't about our work, but this really ties in very closely to some of the work we've been doing. Um, and the whole point of this product television is to bring a visual interface to your agent and give you and the agent a shared interactive surface through web technologies, through HTML, um, in order to work with your stuff and to display information that the agent is gathering and pin that in a screen that you can then, uh, manipulate. So it was very cool to see this article pop up. I think the thing that I want to see in the thing that we're working on actively right now is how do we go beyond, you know, just viewing stuff that the agent is creating for us, like a calendar or a graph or, you know, whatever it is and go towards something that's actually doing something. And I think that's where we go from something that feels like an agent that I chat to with a text, uh, chat box and, uh, that does stuff that I can only see and to something that feels like a real computer that I can fully interact with where it's generating software on the fly, bespoke to a thing that I'm trying to do. So yeah, that's the, that's kind of the state of play at the moment. I think HTML artifacts is like a good starting point, but we've got a long way to go towards something that feels like full computer and we're actively working in that space at the moment. Yeah. Like it's our, it's our entry point into that, that space into that problem. It's probably one of several, uh, that exists, but it's the one that we're going with and it's been fun actually building this thing and then using it in our daily lives. Like it's very early where it's in an alpha state right now. Uh, people should go, if they're interested, go to television dot run and you can put in your email address to join the wait list. It's going to be open source. So it'd be free for anyone to use with any agent, any agent harness. It works with all of them, open claw, Hermes codex, cloud code, pie, whatever it's early, but we're, we've been having fun working with it because it's sort of giving us to the first time a chance. I'm speaking for myself. It's giving me the chance to experience a completely different computing experience, not just through text, but through visuals. And it changes the way I think about computing tasks, like how I, how I formulate them, how I sort of break them down, uh, and how I communicate them to the computer. And I'm just finding that to be a lot of fun. I mean, frustrating too, right? Cause this technology is so early, but a lot of fun. So. Yeah. I mean, I've been dog fooding television very hard, uh, and applying it to all sorts of parts of my life, like personal productivity, my calendar, my to-do lists are all mediated through an agent now, uh, that is forming these screens for me based on like prioritize lists and stuff to do things that have been consolidated from my various inboxes that it will form an overview for me to look at in the morning. Yeah. My calendar being generated work, even on the notes of this podcast that are all up in front of me right now in television. Uh, I've recently been using it for visual design for our website. Um, I had saved a bunch of fonts and it pulled together a dashboard of all of the fonts that might be suitable for this particular project. So I think, you know, this is starting to feel to me like the glimmers of the new kind of computer that we've been talking about for a long time in a really useful early scenario, like really key use cases that I think a lot of people are going to find really compelling. I should mention also, this was covered really well in a new podcast, which we'll give a shout out to called small talk, which is by Jeffrey lit and, uh, max, uh, max showing. They're both at notion. Um, I've been following Jeffrey's work in particular for a long time, met him a long time ago at causal islands conference. And they spoke a lot about HTML artifacts in this post. And I think one critique they had is that just because something is visually flashy doesn't mean it's useful. And what are these things actually for? And is this just like a sugar hit from the slot machine of being able to kind of throw in a prompt and have a generated artifact pop up and go like, wow, that looks beautiful. Like I made a great thing or, you know, have that dopamine hit. Uh, and I think there is a part of that that is true. Uh, and that's where it's going to be really important that we've fine-tuned these systems and also build out a series of skills to allow these agents to create artifacts and visual information for us that we can easily interpret. And that is actually doing useful work for us. And maybe that does look really simple for some use cases. Maybe it does just need to be pure text for some things, but where visual communication interactivity can enhance a use case, that's where it should be added. Well, this has been great. I think we should get to our guest who's been waiting patiently. So we're going to start our interview with Rafi Krikorian. Well, Rafi, thanks for being on the show. It's good to see you. Likewise. You know, so for folks who don't know Rafi, Rafi's, you've had quite a journey yourself. You have been a founder more than once. You were VP of engineering at Twitter. You led some of the early self-driving car work at Uber, I believe. Um, you were CTO of the Democratic National Committee, Emerson Collective, and now Mozilla. I've been very lucky. I've been a board member. Yeah. Well, you and I met about a year or so ago when you were on the board of Mozilla at the time and I was working there and now you're the CTO. Tell us about what you've been doing at Mozilla lately. Yeah, no, it's good. And thanks for having me on. I mean, a lot of my focus at Mozilla these days is around open source AI. And so like, how do we build AI technologies that are trustworthy and on the same side as humans? You know, think of it in the same way, like the natural evolution about thinking about the same way about the browser. Like Mozilla worked on the browser because we needed a user agent that goes on your behalf on the open web. Like it can protect you. We could do all the privacy stuff, all the things that you know it's on your side, not someone else's side. And like Mozilla at this moment is realizing that the web is vastly changing and we're getting to a world of agents that might not be on your side, but are navigating the web for you. And so like my work is about how do we create open systems? How do we create open software, open interfaces, open systems so that we can have that trustworthy relationship again in this new world that we're living in? I know you started a sub stack recently called Owners Not Renters. That caught my attention immediately because I instantly got what you meant from my own work at open source. But maybe you could talk a little bit about that. Yeah. I mean, the whole notion of, well, let's start with the phrase Owners Not Renters. And like what I mean by that is like this feeling of, you know, dignity that people have when they're working with technology. Like, you know, everyone feels good about with tech until you start telling them all the stuff that's really going on behind the scenes. Like you're participating in a data economy you didn't know about. Like people are actually surveilling you and that's part of the business model. You're the product. You're not actually the recipient of product. You're the product. And so in a lot of ways, all the technology that's around us, we're renting access to it. And we're being having data and all our stuff exfiltrated from ourselves. I want to get to a world where, you know, it's like the, it's what we all want the technology to be. We want the technology to be this thing that made the world awesome. Like it made us feel good. It made us do all the, allowed us to do all these things. And somewhere along the way in the past 15, 20 years, is like the internet and the web and technology has taken a left turn somewhere. And so owners, not renters is meant to be the place where we have that conversation. Like how do we get back to a community of people to the world, building stuff that's actually on our side, on our terms? We have on this show, like a thesis that these new technologies like agents and models are together like the beginning of a new type of computer, right? Which is why we're doing this whole show is to explore that in all of its various aspects. But like, it feels already very different from the original wave of personal computing, where these were things you bought, sometimes a kit, right? And you assembled it and you brought it home and you plugged it in and you decided what the hell the thing was for, right? And like you made it do stuff that no one else's did or did in your own way. And there was nobody you were asking permission for. There was nobody, right? That you had to like run it past. There was no one who controlled it in any way. And it was this weird, almost anarchy chaos of these machines that were just completely unleashed. And today, even the early days we're at with this AI stuff, it's nothing like that already. I mean, most people are running this stuff on cloud-based models. And like you said, their data is being exfiltrated. They're not really in control of their own computer. I want to talk about this with you today because I had a friend just over the weekend ping me and say he'd been using Claude for like the last year as like a creative brainstorming partner. He's a writer and he got banned and he has no idea why and he has no recourse. And he didn't just lose, he can just go to chat GPT, of course, whatever, but he lost all the memories and sort of personalization he had built up. And so he feels like his computer just got taken away from him. I mean, like, I feel like, again, we're in this world where people are not actually in control. Like, it's like we were living in someone else's bubble all of a sudden. So like, I think about this a lot with like personal context, like I'm building up this relationship with all these systems and then I get banned. And then my context is now stuck in one place and I'm forced to go to another or even just think about a different way. When, when OpenAon changed their changed the GPT model, I think it was like five, four to five, five or five, three to five, four, like the personality shifted and people were in an uproar because they built a relationship with this piece of technology. And then it was taken away from them because again, you're renting it, not owning it. Like, and like sometimes house to have a really concrete model, model of the world. I was talking to a person in Iowa city the other day and he reminded me of the John Deere situation. Like people bought John Deere tractors and then they weren't allowed to fix them. Like when something went wrong, they couldn't, but I just spent tens of thousands of dollars, a hundred thousand dollars on this thing. And in fact, he was even noting that old John Deere tractors from the seventies are selling at insane resale values right now because that's the one the farmers want because they can fix it. They can just own it. And I think that we have to think about that across our entire digital landscape right now. I wanted to go back to a word Raffi said a little while ago, which was dignity. And I think that's a value. I think about that in terms of technology a lot, because I feel like a lot of my interactions and people I love's interactions with technology doesn't feel very dignified. And, and I think part of that comes to another thing that you highlighted, another value you highlighted on your article on the Mozilla website, owners, not renters, which I implore everyone listening to read, was agency and decentralization, like a distribution of power. And these I think are values that Stephen and I really care about a lot in what we're trying to build. I'd love to hear your take a little bit further on some of those values, but also I'd love to hear your response to a criticism I often hear from people, which is, oh, like no one really cares, which I think is leveled at the privacy value a lot. But same probably can be said of like agency or dignity, like people will just do the, pick up the tool that does the job and they won't really care. My sense is that that isn't, that doesn't quite hold true anymore with a huge amount of bad will now directed at the tech industry. I'm curious how you deal with, how you think about those values and how you deal with that criticism as well. Yeah, it's a really good flag. I think that people don't care until they do, like in the grand scheme of things, like, and they care that instant shit hits the fan. And so like, I think a lot of it is, you know, it's not easy selling insurance. So like, we don't want to necessarily get in the world of like, we're going to protect you from shit hitting the fan. But I think a lot of it is, can we ship really good things, really gorgeous experiences, really beautiful devices that happen to be private by default, or happen to have the architecture in place so that you can be protected. So another way I think about this a lot is like, you know, if owners are not renters is one of the phrases. Another phrase I like these days is, I want architectures, not handshakes. So like, I want like things to be built in such a way that I have agency protections, privacy, transparency, all the values I care about. And I don't want that to be just a legal agreement that could disappear. Or I don't want to be a click through that maybe I read half of it and hit accept to it. Like, I think about signal a lot in these situations, right? Like, it is if you call up Meredith and ask her for the transcripts between Rafi and Stephen or Rafi and Ruber, she literally can't give it to you. And like, because the architecture is what protects you. And so like, we need, and I don't think it's impossible. I think it just it's hard, but we need better architectures when it comes to our day to day life, to make sure that these things are structural in the way that they're protecting us, not just handshakes. We also need the technology itself to continue to progress at a pace that, you know, if it doesn't match, at least sort of keeps up with the closed source options. And that is something that I've been concerned about. I mean, the trends historically over the last few years have been pretty good. People have said like, oh, multiple times, local models or open models are falling behind, they're dead, whatever, they've hit a plateau. And then, you know, then you get DeepSeek, right? And then you get, you know, Minimax and whatever else. So, but the thing that's interesting to me is we've seen over the last few months, this sort of withering of the US open source model work. You know, like, I don't think we're going to see another llama from our friends over at Meta. The Allen Institute has had some issues, it sounds like. All of the frontier open model work, which is so critical to everything we're talking about actually happening and being possible, is seems to be moving overseas. Yeah. Which is interesting. I mean, most of the models I'm using right now for locally on my machine are not trained in the US, certainly. And that has interesting geopolitical overtones. But I'm just curious your take on where we are at in terms of progress on open source AI. Well, I mean, I think like if we were to step back, I think you're right. Open source, open weight model systems like handle something like 90% of everyday use cases right now. I think I was reading some article that literally put it at 98%. I'm not sure I'm going to go that high, but 90% of everyday use cases where the, where the gap still remains is the frontier. But it turns out that that might not be the right battle. Anyway, the frontier is great for things around scientific discovery, long range planning, deep scenario thinking, et cetera. But when it comes to the things that I need to go do, like, you know, summarize an email or like trigger a set of actions that does something on my behalf or search the internet for something like turns out the open source models are great at that. And so like, I actually think that the open source ecosystem, again, if you're a step back, you're an alien, you're looking at earth, it's actually a pretty vibrant spot. And the questions now become less about the model layer and it's more becoming around like the harness you put around it. What does that layer eight thing look like? What are the interfaces that go and talk to it? But you know, if you zoom in and understand those geopolitical boundaries, like you're saying, yeah, there are some interesting problems that show up. Like one of my favorite papers from last year was doing, was do a study on kidney donations. Like it, you know, kidney donations is actually a very tricky thing. It's not first come first serve. Like there's a, there's a committee, there's value judgments that are being made and stuff like that. So this paper studied what would be hypothetical kidney donations by all the major LLMs. Like what would they do in those situations? And it turns out it clusters in effectively to oversimplification, but effectively to the Western models and the Chinese models. And so like, they just have very different, like it's a, for me, it's a prime example of like value systems are being encoded into these model systems just by the training, the post-training, the pre-training setups that you're feeding them. And so that's really the question that we should be asking. It's just like, if these things are making decisions on my behalf, I kind of need to understand why, and I kind of need to understand how, and I kind of need to understand what are the boundaries around it. Because if it's going to be trustworthy to me, I need to know it's doing stuff with my values. Or at least we had a conversation about it, that kind of thing. Right. And if it's, if that thing is now the center, like intelligence of your personal computer at some point in the near future, like it is, is an extension of you in a way. If your values do not harmonize in some way, that's a really weird cognitive dissonance that occurs. 100%. I just also want to add that have asking models to, to judge who should get kidney donations is like a William Gibson ready short story. Totally bonkers. Totally agreed. Good God. Yeah. I think it's the other issue if we have sort of a monoculture of models as well. And this is part of the thing that the open source movement could help prevent is that if I'm starting to delegate a lot of my thinking to a model, now we're all kind of thinking a little bit the same, or if I'm sort of communicating with a model to try and make decisions, the alignment of that model is going to be really important. No, I'm, I'm actually super fascinated these days with what local evaluations could look like, and then how that train translates to local fine tuning and local training of just like, you know, I've been thinking a lot about, you know, let's take it to code for a second. Like Git is like a really good architecture for how code is written. The thing is we're now working in a world where like you issue a prompt, which then writes code and then has bytes. So Git might not be the right abstraction layer given that just add a third layer to this. So, but if you can measure all those prompts, if I could record all those prompts, I know when I ran it, what model I ran it against, what was the context and stuff like that, then I can start evaluating on my own data, the history of my work based on all the different models are coming up to understand which ones would do things the way I would like to do them, which ones wouldn't. So you take those kinds of evaluations, and then you take one of the open model suites, I could use those evals to fine tune. Then I can actually start creating models and training models on a personal basis that are slewing more and more toward my value system. So it might not be the case that has fully aligned me on day one. But honestly, if I were to hire an assistant, my assistant is not fully aligned with me on day one, my assistant has worked with me for six years, and it's taken five and a half of them for us to get on the same page, like kind of thing. So you can maybe think of a same metaphor if we can just figure out how to do constant evals some way. I've been watching with interest what's been going on over at Mozilla AI over the last year or so. They've been putting on a lot of open source projects, a lot of sort of like supporting tooling and infrastructure for helping people and developers adopt open source AI more and more. Is there anything there you can share what's been going on there lately? Or are they doing work around evals like you're talking about? I know they've been doing work around making it easier to work with different models and different harnesses. Yeah. So right now their focused work is around something called Otari, which is basically enterprise-ready, open router-like systems. So like how do you do model routing in a way that you have full transparency, you have full accounting, but then you can also learn what's happening in all those different routing scenarios, which might make your system smarter. And on top of something like Otari, we've been working on something called CQ. And what CQ is, I jokingly call it stack overflow for agents. So like if Rafi is using an agent and it's learning how to work with me and Steven's using an agent, learning to work with him, maybe they should have a common message board that they're like literally posting things to. So my enterprise is learning faster. And so like CQ is that infrastructure layer that allows for agent to agent communication and enterprise learning in some way. So yeah, those are the two big things that they're sort of taking on. Does it have snarky comment threads too? Because that's kind of key to the whole experience. It's going to have like every problem you can imagine with social media. Like at some point, there's going to be like a disinformation thread on and all the stuff and we have to like work through it. But yeah, exactly. Yeah. I look forward to someone's paper on this, you know. I mean, you've been using agent harnesses yourself. You've been blogging about it. We certainly have been because we're using them. We're building products like interfaces on top of today's personal agent harnesses. I thought it'd be fun for us to sort of share our experiences, which I assume have been varying degrees of bonkers. I mean, I spent the whole morning trying to film a demo of our latest product work and tear my hair out because every time I issued a command to the system, it would do things a little differently than the last time. And I said to myself, hey, this is the non-deterministic computer you wanted, Steven. You asked for this as I hit myself in the face repeatedly. I mean, I do want to say, I mean, I'm glad to nerd out about my setup right now. Like I was joking to someone the other day that like, I think if you think about all the things you use the internet for, like you use it for information, you use it for education, you use it for entertainment, you use it for commerce, you use it for social stuff, some productivity stuff. Like, I think that because of these new agentic systems, you know, I've been playing with the Hermes agent from news research. Like, I think I'm taking a massive dent out of my browser time. Like for all the like actually intentful stuff I need to go do, which is like add this to a calendar or like check to see whether something's in stock at Whole Foods, stuff like that. Like, I don't touch a browser anymore. I'm just like slacking an agent that's doing half of it. So, uh, it's why, I mean, like, I'm not sure I'm in a place where I want to do rights yet, but for reads, um, you know, my own personal calendar, maybe it's fine, but for reads, like, I'm definitely in a situation where I'm just like, just gather the information for me. Like I literally, that's a huge change. No, it's 100%. Like my, my web browsing habits have fundamentally shifted. Like the place a browser fits in my information browsing hierarchy is definitely moved. Um, which has been fascinating. And I think it's really only happened in the last six months. Like it's happening at like this startling speed for alpha users and alpha users are obviously going to be the prime example of what's going to happen in the future. But to your point earlier about this non-deterministic stuff, one of the things that like really frustrates me, and I think it's just, you know, it's an example of like we engineers should do better in a grand scheme of things. I'm just like, it's unclear to me why we're like relying on these agents to do things that code would have done deterministically for me. And so like, I'm in this place where I ask an agent to do something and it writes like half Python, half of it's an LLM call. And I'm just like, finish the job. And then I wouldn't have these wacko problems that are showing up every once in a while. Yeah, absolutely. I, I think what's interesting to me is that this change you talk about in the last six months, like browser behavior, browser habits, it wasn't really that wave we had of AI browsers, quote unquote, that actually made this happen. I think it's the agents. And it makes me wonder if like the browser is even the right metaphor for some of these computing activities moving forward. I mean, I don't really fascinating to think about. I mean, I don't think it is. I mean, let's just be concrete and talk through a few examples. Right. So like, you know, I log whatever food I eat every day. Like I'm a middle-aged white guy. It's important to make sure I'm not overeating and just gaining weight the entire time. But like, I don't find the app on my phone anymore. I don't enter macros anymore. I literally, I'm just like taking a picture of my food, slacking it to my agent who on the backend has figured out what the APIs the app used and just integrates the data directly, like just bypasses the user interface. So like searching for apps on my phone has basically gone near zero, not perfectly zero integrate, like data input. It's just not something I'm going to do again. Like the other day I asked my agent, I need to do a, like a warranty replacement on a headlight battery for my car or headlight bulb for my car. And like, I was just like, just find all the stuff and fill this form out for me. And it did like, it just like read my Gmail, found the order, put it all in. Like data entry might just be not a thing anymore. Um, which is like a whole class of what you use the browser for it kind of thing. It's just amazing. Yeah. I'm noticing the exact same thing. I think I, I constantly surprised myself at situations where I would have gone searching for an app or use an existing app that I had somewhere. Um, and instead I'm doing just a really, it almost feels like very, um, ham fisted kind of input into an agent where I just like dictate some texts and like throw in an image and just like work it out. And, and it, and it can work it out. And then, you know, I've, I've set up some, um, some infrastructure around like how it should file stuff. And there's some upfront work. That's like very nerdy to get there. That's not like for the masses quite yet, but I'm constantly surprised by how well it will do the job of an app. Uh, and by the job, I mean, I'm, I mean like the actual job to be done as opposed to looking and feeling app-like, which I think is sometimes what people confuse with generative UIs and generative behavior and stuff like that. It doesn't need to necessarily like look and feel like a final polished app, but it can actually complete my job with like way less stress. Um, and, and then it can start doing things in the background for me. Like I had this experience the other day where I was like, I'm going into a meeting. Here's like, you can see it on my calendar. Give me a briefing of everyone that's in that meeting. Just pull their past emails, pull their social media stuff. So I have a conversation starter and just build me a briefing for it. It did. It was great. Very useful. I looked very smart coming into this meeting. And then it just asked me, should I do this for all your meetings? And I was like, yes. And now just my calendar is populated with all these briefings coming up for all my upcoming meetings. It does 24 hours in advance and it's fantastic. Like, so I think it's like those type of things that like, you know, that we've just never would have even contemplated because building that UI would have been yucky and like, no one's going to buy like, and now it just doesn't like kind of thing. Well, yeah. And before open claw and Hermes, like we had, we, most agents we were running were sort of like single session. You didn't leave them running all the time. They didn't really have memory. Uh, they certainly didn't have this idea of like cron jobs or heartbeats to do the kind of things you're talking about every day or every hour. Go summarize this, like give me meeting notes, whatever it is. And so what's interesting, great example of open source, I think is how that stuff came out in open claw and within like not even weeks, it felt like you saw it in cloud code. You saw it in codex. You saw them basically replicate that feature suite. Um, so to me, that was actually a really good sign that open source is still driving the conversation has the power to influence closed source trajectories. Well, and also the ethos of open source still seems viable and strong. Like I think a lot of the question I get asked these days is what happens to open source in the world where coding is free? And I don't actually know the answer to that question. Like I get asked all the time. Another podcast episode. But like the number of people who are sharing their vibe coded stuff on the, on like a GitHub repo, like the ethos of wanting to share still seems to be vibrant. And it's just like, I will, I mean, I didn't write it. I like paid for some tokens and it wrote it for me, but I will share it with you. Or here is like an entire task library to my agent, just downloaded and imported into yours. Like that ethos still seems to be there and strong, which is one of the reasons I remain optimistic that we're going to figure this all out is because we still want to behave in the same way. I think there are also two very like lucky coincidences here. And one is that early on, when these models started bearing fruit, like the large language models, um, they were released first as completely open, like by open AI with GPT two and stuff. Uh, but then also as APIs primarily before the product side with chat GPT took off. And then as such, like the API surface just kind of stuck around because open AI couldn't get rid of it because then Anthropic had one and they would have out competed. I think it's actually really fortunate that it worked out that way, as opposed to all the intelligence being wrapped up in products that you couldn't get an API call too easily. So now they're proliferated to all these other services and developers can do whatever, but also, um, now the killer use case is coding agents and programmers, uh, notoriously hate being locked into anything. You know, they have all their code locally. It's all through Git, you know, they can hack their own tools together. Like this is a group of people that do not like subscribing to a thing and having to do something someone else's way. And I think that again, kind of can proliferate this idea. Like I think people who are using coding agents a lot as these first users and extending it to knowledge work and other areas of their life do not want to just handle that over to a really nice finished polished product. They're interested in how they can own their computer and their stuff. And I think that's, that could be very fortunate for the ecosystem. I totally agree. And I think an interesting thing that spawned from this, which I kind of love, but I hope we don't all go down this way is like the proliferation of terminal UIs. Like it's been like wild to watch. It's been so great. I'm like, I totally, I'm reliving my like 9,600 baud days. Like this was so great. Yep. V2, 220. I saw the Hermes agent actually sort of apes the 320 terminals that would be like in the library in the eighties with like these amber screens. And, uh, it's bizarre. Of course you can run doom on it as well. So I did want to say is just really, uh, grateful for the work that you and Mozilla are doing around open source AI. It's an important voice that this is what I worked on when I was at Mozilla. I have a lot of affinity for what y'all are doing. Um, and I think it's really important that we push these values and these technologies. Otherwise, we're not going to own whatever the new computer looks like. We're going to be renting it. We're going to be leasing it and we're going to have experiences that my friend had where all of a sudden it's taken away from us. And that's not why I got excited about computers in the first place. I don't know about you guys, but I think it was really excited by that independence and mystery, you know, that was sort of mine to solve and mine to explore and discover and not something that I had to ask an authority figure for. And we're not heading that direction right now. And I think open source is probably a powerful way to fight that. No, I mean like it's, it's all about mastery, autonomy and purpose. I was just like, that's like, that's the thing that we're all chasing in some way. I'm just like mastery over these tools, mastery over these technologies and having a purpose and doing it. Totally agree. I mean, like, thank you for the work that you guys have been doing because like someone needs to be the crazy people exploring that next frontier. Like we have, like I've, I've always, you know, well, I always believe that like every, every wave of technology is a local maxima in some way, shape or form. Like we are, we will never reach the global. We were always at the local and it takes someone or it takes a group of people to do that kind of like exploration outside the local maxima to figure out how to get on the ascent curve to the next one. So thanks for doing it. Well, thanks for saying that we'll, we'll see, we'll see how it goes. There's plenty still to be discovered and to be done by all of us. So, well, Rafi, thanks for your time today. It was really good talking to you. It was a blast. Thanks so much, Rafi. Well, that was fun. Was that fun? We should do this more often.