บทถอดความฉบับเต็มของ Grok Bot, Robotics, Cybersecurity, and AI Infrastructure, w/ Rob Carpenter

plumber. And so it went out to five different locations, pinged them on my behalf, gave them my information, answered questions on my behalf, and then went ahead and proactively scheduled it. And so 20 minutes later, my bot comes back and says, Hey, good news, Rotor is coming out to your house at 11 a.m. Like perfect. Right. And so now it's being proactive and it's doing things on my behalf. And so we're gonna see its ability to do things on our behalf, whether it's Open claw, whether it's clawed bot or groc bot. Whether it's, you know, an anthropic product or a codex open AI product really doesn't matter. They're giving it the tools, the infrastructure, the scaffolding, and the resources so that it really can act. You know, for anybody maybe that's been in the kind of small to medium-sized business world, a lot of us have used like outsourced, you know, executive assistants, right? BPOs, things like that. So imagine it it's a BPO. Jason, you and I could be sitting in an office together. It's not quite like that because it's this virtual person. Quote unquote, that you're interacting with, but they're doing things on your behalf. And so that's really where we're going to see this sort of proactivity going into the future. Jason Griffing: Mm-hmm. Yeah, I like that. So let let's talk about Grokbot a little more. Let's jump in there. I am I I'd love to hear your perspective on this. So as somebody like I'm fairly deep in the AI space. I'm a big claw user, chat GPT, I have both of those desktop apps, paid accounts. I'm in there all day, every day. Rob Carpenter: Same. Yeah. Jason Griffing: and then I hear Grokbot comes out and there's a lot of buzz and excitement about it. And the first question that comes to mind is like, what is fundamentally different? About Grokbot, why are people excited? What makes it differentiated? There's a lot of different things that we could explore there, but fundamentally, I think one of the things is this idea of like you're building little agents that are like these persistent, almost personalities that are armed Rob Carpenter: Mm-hmm. Jason Griffing: with different tools and can help you with different jobs. Separately, you might have one that's like kind of a chief of staff or orchestrator amongst them. And this would be in contrast to tools that use just like threads where each thread is kind of a throwaway conversation. You have more of this persistent kind of agent personality. Do you see that Rob Carpenter: Mm-hmm. Jason Griffing: as a really fundamental difference? And I'd love to hear you riff on like as somebody who's jumped in and started using the software, what is your experience with that and what makes Grockbot a differentiated offering? Rob Carpenter: Yeah, it's a great question. And let me try to untangle it for people too a little bit. So Cursor was a standalone company. And at one point, they were just really a wrapper on any of the sort of foundational LLMs that existed behind the scenes. And they just served it up to you based on what you needed, what you wanted to pay for. Cursor then started to Build, my understanding, is some of their own true kind of like foundational AI capabilities. And then SpaceX signed an agreement with them to give Cursor access to the colossal data centers, but it also included a right of acquisition. And then once SpaceX went public, they acquired Cursor for, I believe it was $60 billion. Cursor had already been developing this Grokbot product, but with the acquisition by SpaceX Don who owns Grok, then they put on the Grok, you know, kind of Branding, if you will. But for all intents and purposes, at launch and still kind of up to now, it's been out for about a week and a half. It's still sort of running as a separate entity, although it's slowly starting to get more integrated into Grok. So that just kind of sets the foundation of kind of how we got to where we are at with this technology. I think one of the things that's really interesting and kind of differentiated about this Grok bot product is that yes, it has a chief of staff. Yes, it can spin up any agents that you need. Yes, it can turn those agents off if they're not needed. You can be Running, you know, one, two, five, ten, if you got the processing in capital, 10,000 agents simultaneously. So instead of having just a single thread in Claude or in OpenAI, ChatGPT, that's sort of doing one thing at a time for you, and it could take a long time to get through 10,000 tasks. 10,000 agents could ever do every single task simultaneously. And so it allows you to speed up that process. My own experience is I have not yet found it. beneficial to have multiple sub agents versus just the chief of staff kind of doing everything. So I don't know if I'm not using it correctly or if the tech is not quite there yet. I've tried to train the sub agents like, hey, you need to be really good at this one thing. And they're Kind of average. They're really not much better than just the chief of staff. And I do tend to find the chief of staff just sort of doing everything. So I don't know that there's a huge amount of benefit I've seen yet in having these sub agents. And in fact, more often than not, the sub agent like glitches out on something, and then the Jason Griffing: Mm. Rob Carpenter: chief of staff kind of gets stuck. So I've not seen a lot of value there. What What is really cool is that every one of these agents, the chief of staff included, have their own virtual computers that they're running back at Cursor, which I guess is probably SpaceX's colossal data center, but they're spooling up their own virtual computer. And what's cool about that is that they can then download software applications. So let me give you an example. Like I am you know deep into the AI space, but I also like video games, and I'm a hobbyist across a variety of different things, and video games one of them. So Godot is a game engine that you can download and you can build inside of and

you can release video games. So I told my chief of staff, hey, download Godot, and I was like, hey, we need these different types of libraries. It looked them all up on GitHub, it downloaded all of them, and now in its virtual machine, it has Godot, it has all these libraries, and then in a verbal conversation with With the AI, just the chief of staff, it then goes into Godot and it builds everything out inside of its virtual machine for me. And then it says, okay, here's the link where you can now download this .exe file or you can access it through the web. To play this game as just sort of a simple example. And then if it runs into a situation where it's like, hey, you asked me to buy a bunch of stuff on Amazon for you, I need you to log in. Then it pulls up a screen. I can see it's a virtual terminal. I can see all the tabs that it has open on Chrome. I can enter in my username and password. It then can get into the site and it can order and buy stuff on my behalf if I want it to. So it really, again, the easiest way to think of it is think of it as like an outsourced BPL sort. Of virtual private assistant. So instead of having somebody in India or the Philippines that's doing work and completing projects, research tasks on your behand, you now can start to hand that over to virtual agents. And how I kind of think about this, because then we very quickly get into a conversation on labor automation, you know, it's almost like a pyramid. And at your bottom, you can have a very distributed workforce all over the world, you know, and then above that, you get into like entry-level employees in your office, and then mid-level employees in Senior level employees and executives, and there's fewer and fewer people as you go up the pyramid. And what we're doing is we're starting to see AI kind of inch up that pyramid in terms of the work that it can complete. The AI, in my opinion, when you're on these high-end paid tiers, is very much at like a mid-level employee today. And so, Jason Griffing: Mm-hmm. Rob Carpenter: as the software and infrastructure and your comfort exists, you really can give tasks up to that mid-level employee and feel pretty confident to 80 to 90 percent that it's gonna. do it correctly. And so we're just going to kind of see that capability move further and further up the stack. And these virtual computers, all of their abilities to download software, access websites is just unshackling them and allowing them to do more and more work on your behalf. Jason Griffing: Yep. Yeah, that's interesting. I so on the agents and subagents topic, you know, I've had similar like I've I've been looking at the product. I have not dove into GrockBot myself yet, but scratching my head a little bit on it, saying like, okay, well, I know that like Claude, for example, can also do this thing where you can have like this one orchestrator agent that spawns off multiple sub agents and candidly I've never found much utility in it. Like I've tried and I've just always kind of gravitated back towards like just jumping into a thread and having like this one agent just kind of help me out with whatever it is that I'm working on. So I think that's like an interesting space that I'm keeping an eye on is like this idea of Rob Carpenter: Mm-hmm. Jason Griffing: agent to agent collaboration and orchestration. And there's all sorts of crazy implications and security concerns and stuff that comes with that. So it's really, you know, Rob Carpenter: Yeah. Jason Griffing: Fascinating space. The other thing that I'd love your perspective on, like d to dig a little deeper on is this architecture that Grockbot has where you're actually tying into like a cloud computer. The idea that it can like spin up a web browser. I've been doing more and more of this stuff myself. So we're doing a remodel at my house right now, as one example. And just last weekend I had Chat GPT open up a web browser and I gave it very specific criteria. For some furniture that I was looking for. And then I Rob Carpenter: Yeah. Jason Griffing: watched it just browse across the internet and find a bunch of stuff that was really great options. It was a very collaborative process. So this idea of like AI being able to s open up software or web browsers and use those tools in basically the same way that a human would is really fascinating to me and opens up all sorts of interesting implications for, especially for tools that don't have APIs or MCP servers. So I'd love to hear you riff on that a little bit more and and why this architecture of cloud computers is fundamentally different and important thing to for people to be paying attention to. Rob Carpenter: I don't know that that's the case. I think that underlying these systems, it really still just comes down to the quality of the LLM. And I've tried having the GrockBot do some like software development and stuff, and it's just still not As good as Claude is. And so I think what Grockbot does the best right now is the cleanest, uniformed architecture that's sort of the simplest for people to get access to and use, but still has 80 to 90 percent of the core sort of foundational LLM intelligence. One of the things I saw the other day, for example, any of us that are heavy users of LLM, we get the experience when we've maxed out memory and it tries to compress. Old memories, but it also starts like making mistakes, and it's like, okay, we're just gonna start a new fresh thread. Well, one of these conversations with the Grokbot, it's like, hey, hold on just a minute. I've exhausted my memory. I'm spinning up a new agent and I'm transferring all the relevant details over. And it handled everything, transferred everything, but it it told me it was happening, but then everything existed in the background, but my like unified chat interface never changed and it remained sort of the same interface. So it could be sort of like Linux. Versus, you know, Mac OS X, right? Where Mac OS X is super,

super simple. And I think Crockpot has a bit of that. Whereas for a lot of people, they want the Linux so they have complete and total control over everything. And so then there's elements, and I think that's where maybe an open claw exists. And maybe Anthropic is kind of in that Microsoft world where it's like, well, they give you a lot more control than the Mac OS X does, but they don't give you as much control as a Linux. And so I see Anthropic as being kind of in that space right now. I really think it's Just kind of at the end of the day, comes down to personal opinion of sort of the interfaces you like to use. I very much believe that over the next six months, all of these platforms are going to end up at sort of like a unified front-end interface architecture. And so it's really just kind of picking your front-end flavor between now and then. I have Jason Griffing: Yeah. Rob Carpenter: found myself migrating a little bit more towards the Grokbot lately. but Before that, you know, it was ninety five percent clawed. And so now I'm like forty five to fifty percent clawed. Jason Griffing: Yeah. Yeah. And I think that is an important thing to point out. The Grok bot is, as the name would imply, tied to the Grok models. So Rob Carpenter: Mm. Jason Griffing: depending on the type of work you're doing, you may or may not find like that you're getting good mileage out of that. But I have heard a lot of people talk about just the elegance and simplicity of the interface and it being probably the best tool out there for people who may not be super comfortable with these tools but want to get in and experience, you know, what it's like to work with agents. You alluded to something there that I've been curious about is the context window. So when you have tools like I would say I guess more traditional, things are moving so quickly it feels weird to use that word. But like, you know, you working in Claude or Chat GPT, you have these threads that you basically start, and when the context window gets full and things start acting squirrely, you basically throw it away and you start a new thread. This shift to these more persistent agents that just Stay around and stay in your interface over time. If I heard you correctly, it sounds like Claude is just managing or sorry, Grokbot is managing that on the back end. And basically as that context window starts to get full, it's doing some housekeeping on the back end. But as a user, you're not really seeing any of that. Did I follow that correctly? Rob Carpenter: Yeah, that's what was my experience with it. So you just don't have to manage that process of starting a new thread, giving it all of the relevant context, and then picking up where you left off. It just sort of migrates itself in the background and you just never see any of that happen. And I think this also gets to like where things are going. So the most recent sort of rumor and report that I heard is that OpenAI is on the verge of releasing a new model that's got a context window of two and a half million tokens. And right now, roughly speaking, the average is about a million tokens, you know, so a 2.5X increase in the amount of memory, which will be really nice. I think memory is going to be one of those critical battlegrounds over the next 12 to 18 months. And I think we'll just continue to see those context windows improve. And I think within 12 Months we'll get to a point where you won't have to manage the different threads, and the systems will be smart enough that they'll just be able to manage all of that on your Jason Griffing: Mm-hmm. Rob Carpenter: behalf. Whether it's switching in the background, or whether it just doesn't even need to switch anymore, and it could be, you know, 20 years worth of conversation and it can still, you know, intelligently sort of search through all of that. I think that's the direction that we're going. Jason Griffing: Yeah. Rob Carpenter: And then Time had an article that came out, I think it was like Monday or Tuesday, where Sam Altman is quoted as saying that, like towards the end of the year, start of next. next year, they'll have an architecture in place that will allow for just sort of unlimited persistence use. And so I think we're Jason Griffing: Mm-hmm. Rob Carpenter: really, you know, kind of moving in that direction. Jason Griffing: Yeah, I mean, you you I can already feel things moving in that direction. Like I've noticed in the latest Chat GPT desktop app that there's no like prominent indication of how full the context window is getting in a given thread. And I Rob Carpenter: Mm-hmm. Yeah. Jason Griffing: was curious about that. And so I've been experimenting with just I have a couple of threads going in that app that are Rob Carpenter: How cool. Jason Griffing: super long running. and I haven't noticed any discernible degradation in in performance. And so I it's not clear to me whether they're doing something already behind the scenes to kind of manage that. But I I did notice that like c contrasting with like the Claude app, you have the little circle that's kind of showing you how full the context window is getting as you progress down the thread. And I'm not seeing that on the latest version of the Chat GBT app. So it does feel I I agree with you. It does feel like something that over Rob Carpenter: Mm-hmm. Jason Griffing: time You don't wanna have to think about as a user, especially as they're trying to get less technical users into these systems, having to think Rob Carpenter: Totally. Jason Griffing: about this abstract concept of a context window is just cognitive noise that that I'm sure they don't want people dealing with. Yeah. Rob Carpenter: It's just a short term growth pain. You know, in five years none of us will even remember that we had to deal with that, you know. Jason Griffing: Right, right. I think that's true. I'd love to hear your perspectives on security. So how do you encourage how would you encourage people to

think about this if you're using a tool like a GrockBot and then it's like, hey, I want to Rob Carpenter: Mm-hmm. Jason Griffing: log into your, let's say, something relatively benign like Amazon, although although, Rob Carpenter: Mm-hmm. Jason Griffing: you know, that could be fears about it going and buying a bunch of stuff, right? But like certainly Rob Carpenter: Yeah, yeah. Jason Griffing: then when you get into like if you're a software developer and it's asking to log into your AWS instance or something like that. How do you sort of think about what's your mental model for thinking about balancing risk with upside when it comes to allowing these tools to log into your various accounts, whether those be personal or business? Rob Carpenter: I start with worst case scenario, how much will this cost me? And if it's a cost I'm not willing to pay, I don't do it. so I think of I haven't done this yet, but I am planning to set up there's three or four providers now that have come out with basically virtual credit cards that you can give to AIs so they can spend money on your behalf. And basically putting in like 50 to 100 bucks, you know, because that's where my comfort level is. And so it's like, okay, if this AI goes rogue and buys a hundred dollars in paper clips, it'll be annoying, but it'll be a good story and it's not the end of the world. I'm absolutely not giving it access to my, you know, checking account. my retirement savings, any of those kinds of things. And so, you know, in some ways putting what I would almost consider like an air gap where it's like, yes, you have credentials, but the credentials you have access to have $50 to $100 of sort of exposure on there. you know, things like Amazon are possible, full disclosure, that was an example. Like I have not logged in the, you know, Grok agent to Amazon and let it just sort of have carte blanche to buy things on my behalf because I also don't understand fully how all those security credentials are handled and things like that. And so because I don't fully understand it and I don't fully trust it, I don't give it access to that kind of stuff. so I'm pretty protective from that perspective. Grockbot, as of this recording, I'm not aware Having access to like LastPass or anything like that. I do believe OpenClaw has access to that. So I would assume this is part of just this being a super, super new technology. You know, Jason Griffing: Mm-hmm. Rob Carpenter: within the next few weeks, I assume they'll have a password manager. Then I start to feel a little more comfortable in terms of what I give it access to. And so, and then just kind of monitoring it closely, because you're sort of trying to build up trust. I had previously used OpenClaw and built basically what Grockbot is. Is for myself personally, as did a lot of kind of people sort of hacking and playing around. and in that case, you know, I did hook it up to LastPass. I gave it access to a few credentials, but I still kept it pretty contained. And I don't think I even gave it access to any sites where it had the ability to like spend any money or do anything like that. So I just really start from a place of like assume the worst and am I willing to take that risk? And if I'm not, then I don't do it. Jason Griffing: Yeah. Yeah. Just I think that's good advice, kind of putting your toes in the water. And Rob Carpenter: Mm-hmm. Jason Griffing: I've been taking a similar approach, you know, like using it, logging it into certain sites. Like I have a a personal project I'm working on that has a Vercell deployment. was doing some work in like Google Cloud Console. Those are both environments. Like I'm pretty technical, but I'm not like super savvy with those things. And so having the browser built into something like ChatGPT or Claude. Where you can open it up, it can see everything you're doing and kind of walk you through. But I'll be very specific about saying things like, don't go change a bunch of stuff, don't change anything without letting me know what you're doing. But I'm more using it in those cases as like a tool to basically walk me Rob Carpenter: Mm-hmm. Jason Griffing: through and kind of guide me on what I'm doing. And so I I think there's also that element of I'm sitting there watching it work and I'm having a dialogue with it versus. giving it some credentials and spinning off this huge task and then walking away for two hours and hoping that it doesn't like go rogue and and do Yeah. So Rob Carpenter: Yeah. This is gonna go great. I have no concerns. Everything's gonna work out perfect. Jason Griffing: so there did different ways to think about it, but it really is like a fascinating paradigm to me to sit and watch AI work in a browser. And I think it's it's just so wild. Rob Carpenter: it's so exciting. I mean In some ways, it's kind of in the browser in particular, right? Because vision to language is still an area that's under undergoing a lot of development. And so it's kind of like a child with a cursor where it's still a little hesitant, it'll Jason Griffing: Mm-hmm. Rob Carpenter: click something, it'll get stuck, it'll have to stop and think, it'll try again. It's not a rapidly intuitive process. And I think the world and internet companies and businesses in general are still trying to figure out how they interface with these agents. And so there's definitely still a lot of sites where. Where you wanna go and log in, and even if you're entering your credentials, you're hitting up against Cloudflare filters and things like that, that's still making it pretty clunky. And so I don't know, in some ways it's kind of like the AOL world for any of us that remember it, where it's like, hey, here's your sites you can get access to. It's almost like a curated type of kind of internet portal. And I mean you I'm referring to like MCPs and APIs, right? And so when you can give it sort of that headless data access to a lot of other sites, it actually works pretty

darn good. trying to do it visually by clicking through websites is still sort of a clunky experience. Jason Griffing: Mm-hmm. Rob Carpenter: I will say one thing that is substantially better in Grockbot, and probably the reason I've been using it more lately than Claude is that I've just been hobbying on some of these game projects. And Claude, in my opinion, is still awful from a design perspective. Even when I give it access via an MCP to design services, what it generates with the design service is still pretty bad. I don't know where grokbot's getting its images. I assume it's grok and their image capabilities, sorry, are actually pretty decent. but what grokbot generates from a design perspective are Ten times better than what Claude designs. And so I think that's part of why because I've just been using it a lot for some of these different kind of hobby projects lately. And and Grockbot is substantially better on design. Jason Griffing: Interesting. And I I think that's like just a good takeaway for people generally is like having having that understanding of different models and kind of their core c capabilities and what they're good at and what they're not so good at can be really helpful because I do think we're moving to a world where like there's just gonna be agents and different models that you are gonna kind of pick and choose from. I think it's clear that like there's gonna be a few kind of dominant models out there. We all know that. Rob Carpenter: Mm-hmm. Jason Griffing: But like there's always gonna be this sort of horse race effect where like this model's really good at image generation now, but maybe some other model gets ahead. And this is one of the things that I really encourage people to be paying attention to because there is no like one model that's just definitively better than the others. It depends on what you're what you're using them for. Rob Carpenter: I agree. Jason Griffing: before we shift off of Grokbot, I'd love to know like what what's your long term outlook on this product based on your experience so far? Do you think it's meaningfully differentiated? Like the AI market has this sort of effect, right? Where something comes out, we all sort of as a as a collective look at it and go, this is the next big thing. It's gonna change everything. And then sometimes in six months we've completely forgotten about it. Do you think this product is durable? Do you think it has legs? Or do you think it it feels to you like more of a flash in the pan and something that we're gonna move on from here pretty quickly? Rob Carpenter: I mean the exact form factor I'm not totally sure, but going back to a very like reductive perspective on reactive versus proactive, from a proactive perspective, this is the future. This is where everything is going. And so I think Grokbot is very, very important for anybody that wants to see the future. Jason Griffing: Mm-hmm. Rob Carpenter: and see how these agents and how these systems are gonna work. I think as we've talked about around things like memory, that's gonna fade into the background over the next year or so and we won't even worry about that. I'm not sure the idea of sub agents is gonna be super valuable. I mean, if if models get specifically like very intellectually spiky in a specific area, and you truly can have five different domain experts negotiating and working together, I think that's where you come up with novelty. So I could see that working. Short of that, I think it's just the idea of multiple little virtual overseas employees working together simultaneously. And right now with GrockBot, not based on anything scientific, you know, I think it's about eighty percent as smart as Claude, but five to ten times better on design. And so if you're touching anything where you want images generated and stuff like that, then it's pretty impressive from that perspective. But what it does tell us is this is where technology is going. And we are going to arrive at a place in the future. You probably honestly just have a single point of like interaction and you don't even mess with the agents, you don't mess with all the threads, you don't mess with any of that. It's just like communicating with a virtual employee, but it's doing stuff on your behalf. And you know, my wife and I talk about this. I feel like in society right now, we've reached peak complex. complexity. Like I don't know if you have kids, but I've got two kids, one in elementary school and one in middle school. And my God, I must have 20 portals that I have to manage for these kids between their classes, all their different schools, all their different clubs and sports and activities they're in. Like it's just too much. So having a virtual agent that can just be managing all of that stuff for you in the background and then only surfacing stuff for you that really needs your attention, I think we have a very real opportunity in the next two or three years that complexity will actually start to go down. And we'll have these agents working on our behalf doing things for us that will make our lives simpler. And that is what I'm looking forward to. And that's why I enjoy playing with GrockBot to kind of see where things are going. Jason Griffing: Yeah, get a glimpse of the future. I did see actually, I think just today, it might have been yesterday, as we're recording, that Grockbot is available now, apparently on the $20 a month cursor subscription. So, Rob Carpenter: sweet. Excellent. Jason Griffing: so that is intriguing to me. I was not quite ready to pull the trigger at whatever it was before $200 a month or $120, because I've already got multiple, yeah, I've already got multiple Rob Carpenter: Two fifty. Yeah. Jason Griffing: AI subscriptions and I just couldn't quite justify it. But at $20 a month, I would say it's a no-brainer to at least. Turn it on for a month or two, if for nothing Rob Carpenter: Mm-hmm. Jason Griffing: else, like you said, get a

glimpse of the future. And to your point on kids, totally agree. I've got a I've got two kids, one's in middle school, one's in elementary. And I've already got like Rob Carpenter: same place. Yeah. Jason Griffing: like I'm seeing sort of glimpses of this future because I already have like very simple use case, but a chat GPT routine that runs every morning at you know, eight o'clock and tr triages my personal email inbox. And I have in the instructions put everything school related in a separate heading and just distill it down to what I need to know. Cause these emails that come Rob Carpenter: Yeah. Jason Griffing: from schools are like way too long, use way too many words to, you know, get to a point that could have been made very quickly. And so just helping me like distill all that down and make sure that as a parent, my wife and I are paying attention to the things that we need Rob Carpenter: So Jason Griffing: to be paying attention to. Rob Carpenter: that's in place out of the box. And that's one thing that Grockbot really pushes, like when you're first getting started. So it's like a two button click. To hook up Gmail and your Google Calendar and your Google Drive, and then it'll give me a daily rundown every morning. It's like marginally helpful, but I do have like consistent types of spam that I get, probably like everybody. And so it runs at 5:30 in the morning and cleans out any spam and Jason Griffing: Mm-hmm. Rob Carpenter: stuff like that that's come in from the night before, and then just surfaces the things that are really important for me to review. In the example where I was like, hey, you know, give me a plumber to my house at 11, it used my. email to email out and like reach out to people, stuff like that. It reads my calendar, gives me a summary of the day, give me a heads up on the things I need to pay attention to, reminded me that we had our podcast recording this morning, you know, that kind of stuff. So in terms of that use case with the school, that I would do that now. Do your $20 trial. It does it out of the box and just see how that type of an experience works for you. Jason Griffing: Yeah. Yeah. And I I agree with that. In my experience so far, this email triage is not a replacement. It does not have me to a place where I never have to look at my email. But what it has done is removed a lot of the noise before I get in there. So I'm not spending energy Rob Carpenter: Mm-hmm. Jason Griffing: like cutting through all of that. And then also like I'm not always the greatest about staying on top of that particular email account. And so it really helps me just see like, there is something. Rob Carpenter: Mm-hmm. Jason Griffing: in my inbox that's actually time sensitive and urgent. But dive in there, pay attention to that. The rest of this stuff can wait. So I think I think like I said at $20 a month it feels like a no-brainer to just jump in there, spin it up, give it a try and see, you know, see what it looks like. I'd love to shift gears. We talked a little bit about cybersecurity. There's of course stuff in the news that anyone following AI is going to be familiar with. Right now both OpenAI and Anthropic are being a little bit noisy about cybersecurity concerns. And so you've got Chad G or OpenAI who has paused training on their frontier models. You've got Anthropic saying they've got models that are now even better than Mythos that they're holding back and not releasing. What is your take on these stories? Do you think these are real security concerns? Is it just sort of a marketing ploy because they want people to, you know, be paying attention to them. Like where do you come down on these concerns that these companies are voicing right now? Rob Carpenter: Think they're real. You know, I would say it's 70% real and 30% hype. You know, I mean, you can go back. Like, we can't release our models, like their impact on the world would be too catastrophic, you know. So this is a common playbook for them. Anthropic is the same way. And they did hack sites. Like nothing I have read has said Hugging Face had a heads up, Hugging Face knew what was going on. They, you know, put Open AI, put their AI in a box and said, get this information. And they made it impossible for it to get it within the box. And they really wanted to see if it could break out. And not only did it break out, it found another site that had the information and it broke into that site to get the information to come back to the first box and give them the information that they requested. It's real. it's very, very real. And I think it's something that people need to take very seriously. I think the companies themselves are taking it seriously. Again, there's still some hype. Both have IPOs coming up. Both are trying to juice, you know, what a potential stock offering would be. The most recent rumor I've heard is that Anthropic is now talking a $2 trillion IPO. SpaceX, previously at 1.75, was the largest in human history. So now we're gonna jump over that with with Anthropic. and they have every reason to kind of hype these types of things up. And They're hacking websites. I think that's true. Jason Griffing: Yeah. Rob Carpenter: I listened to a really interesting podcast where a tech CEO who had a very, very hot startup billion dollar plus valuation in the private markets, they run their own agent every night at 2 a.m. Because software companies are now getting into a place of just rapid, constant, iterative deployment. And so in any given day, they may have pushed 20 to 30 new capabilities live. And so They do attacks on their own system every single night at 2 a.m. They run whatever their own current sort of best publicly available models are, and then they attempt to hack their own system, and then it reports back to them at 8 a.m. of I wasn't successful or I was successful, and here's all the stuff that I had. And so I think for the CIOs out there, they really need to start looking at

these types of systems where they're just constantly hacking their own systems, looking for Jason Griffing: Mm-hmm. Rob Carpenter: the flaws and looking for the issues because it's real. And if these AIs can write code, then they can pick apart code. They can find problems in code, they can hack in code. And I think that's been sort of the general consensus is anthropic and open AI are like, we never taught it to hack websites or to be malicious. But just by the very nature of teaching it to be very, very good at coding, it can find flaws in other systems. And these things are going kind of hand in hand together. And Jason Griffing: Yeah. Rob Carpenter: so having a focus on that, I think, is going to be important going into the future because Hugging Face won't be the last. And we're going to see other big companies getting hacked. And people need to be aware that the Chinese models that are coming out, they're only a couple of months behind. And there is some belief that they have protections in place. But they're gonna start opening up all of their weights, which means any malicious actor could drap could download an open weight model, can look for the places where it's you know, you know, specifically blocked on, you know, harmful attacks, strip those out and re-release it, right? And so we will get to a world, I think very quickly, where anybody can hack anybody, and if the sites aren't protected, there will be issues. Jason Griffing: Mm-hmm. Yeah, I think that's a good answer. When I hear this question out there of is this real or is it a you know a marketing play, I come down on the same side as you. Well, it's it's both. I mean, the reality Rob Carpenter: Yeah. Jason Griffing: is like that there are very clear dangers with these associated with these models. I think you illuminated all of that nicely. I do think as a sidebar on this, that companies like OpenAI and Anthropic aren't necessarily pausing or holding back on these things from altruistic you know reasons. They know that if they put a model out like this and it causes some sort of catastrophic incident, that's pretty bad for business. So ultimately that's a bit of a sidebar though. I think the cybersecurity Rob Carpenter: Yeah. Jason Griffing: concerns are real. I think it's also a bit of a marketing play because there's nothing stopping a company like OpenAI from saying internally, hey, this model is dangerous. Let's just keep quiet about it and keep working on it. But they're going out there, they're making a lot of noise about it. They're getting people like you and me to sit down and talk about it. It's so good. We can't even release it. Yeah. So Rob Carpenter: Ooh, we got the scary model that we can't release to you mere humans. Yeah. Jason Griffing: so I do I do think it's I I do think it's both. And Rob Carpenter: Look in in nuance, sometimes two things are true simultaneously, right? They're not mutually exclusive from that perspective. Jason Griffing: Exactly. You you mentioned China there. So shifting gears again, robotics. I don't know if you've been watching the I think World Humanoid Robotic Games or whatever they're called. Rob Carpenter: There's been some good clips coming out of that. Jason Griffing: Unb unbelievable and so just kind of like creepy, in my opinion, to watch, but really gives you Rob Carpenter: Uncanny Valley. Jason Griffing: yeah, totally gives you a glimpse of of where things are going. The US has decided recently to restrict the sale of Chinese made humanoid robots. What's your take on this? You have some experience in robotics. Maybe tee that up for us so people understand some of your perspectives and what they're grounded in. And do you see this as being a good thing for the robotics industry overall and here in the US? Or is this just protectionism gone too far and is actually going to backfire on us? Where do you come down on this issue? Rob Carpenter: Yeah, I mean, I think everybody's got their own strong kind of personalities and opinions on this type of stuff when they're deep into the space. I think overall I am on the side of supporting it. the ban. I think it is helpful to force American industry to build their muscle, you know, whether it's physical or robotic, right? And if we're just constantly relying on China for everything, then we have no domestic supply. And anybody that's reading the news understands, you know, we are potentially headed for a conflict with China. And so there's been a lot that has been written about, should we have been, you know, banning our GPUs going to China? Should we not have been banning our GPUs going to China? And I think it there is a lot of evidence that by banning the GPUs that would have gone into China, we have slowed China down and we've allowed. out our frontier labs to kind of stay sort of ahead and continue to develop this technology. I think with the robotics, if there is a decoupling between the US and the Chinese industry, being able to build stuff on our own is super important. You know, going a little beyond, you know, consumer and politics and getting a little bit into sort of conflict and military, you know, we're seeing what's going on in Ukraine. And we've seen what a massive difference Ukraine has made by building out this whole Drone infrastructure, and it's not going to be limited to drones. It's going to be anything, right? Like you have a missile or you have a tank and you take all human, you know, potential loss out of the picture from the person that's putting that stuff on the field, and you put intelligence in all of those things, ish to the level of a human and let it go. That changes the nature of warfare. And warfare is not going to be constrained to drones. And I don't think, you know, warfare will be constrained, you know, when it comes to humanoid robotics either. I think there's a lot of people that want them out. I think that's the right call. And we live in the real world. And I don't think that's gonna

happen. And so we need to be able to like build this stuff up ourselves and not be reliant on somebody else to be creating these types of things for us. Jason Griffing: Mm-hmm. It does conjure up images of Terminator, you know, the the cyborgs out there on the battlefield. But I do I do agree with you. I I hear in the narrative out there, people say things that are very, you know, anti like weaponizing AI and using it in military cases. But the problem with Rob Carpenter: Mm-hmm. Totally. Jason Griffing: that is, like I understand where they're coming from. I also understand that if we stop doing it, that does not mean other countries are going to stop doing it. And so this is the kind of wicked game we get caught in. Rob Carpenter: Totally. Jason Griffing: with these sort of technologies and I I think we just have to be clear eyed about it and have a sober perspective about the reality that AI is going to influence warfare. And it's something that as a country we have a vested interest Rob Carpenter: Yeah. Jason Griffing: in making sure that we, you know, stay stay competitive there. Rob Carpenter: I I agree with you. It'd be great if it didn't and I totally understand where people are coming from. And we live in the real world. And Jason Griffing: Yeah. Rob Carpenter: you cannot control what the enemy does. And if the enemy can field a million You know, humanoid drones, robotic subs, robotic bombers, and you have to put that up against humans in every one of those situations. Not only are your costs substantially higher, but obviously your loss of life is substantially higher. And so you're gonna get into this sort of tit for tat thing where you just have to match what the other side is doing, or you're gonna end up in a situation where you're not gonna be able to compete in future conflicts. And I think the military Jason Griffing: Mm. Rob Carpenter: knows that. And I think that's why we see the Department of War being one of the biggest acquirers and procurers of drones and AI and robotics because they're very much viewing that as the entire kind of future of warfare. Jason Griffing: Yeah, definitely. Well, speaking about, you know, global geopolitics, staying competitive, there's a few stories out there right now. We've got SpaceX planning this new launch facility in Louisiana, trying to launch thirty plus starship flights a day. They've got this Terrafab project, expected to be the largest building on Earth, to produce chips for things like robots, self driving vehicles, and space-based AI data centers, which is fascinating. It could be a whole discussion on its own. you've got all sorts of crazy power demand going on in Texas right now for data center permitting. So there's this whole AI infrastructure build out going on right now. How do you think about some of these big projects? How do you see the US staying competitive? What do we need to be thinking about as a country in order to come about, you know, come at this challenge the right

way? Rob Carpenter: Yeah, I mean obviously it's a very big topic but covers a lot of different things that are kind of related and a lot that are I think in general it's exciting, you know, and I'll probably qualify like everybody does, you know, like Elon has a tremendous amount of controversy around him. And also the things that he is doing is super interesting and exciting from a civilization perspective and a human perspective. I think for a lot of us that grew up interested in space and interested in technology and things like that, the ability to truly be able to get that much launch material into space, you know, with this hundred billion dollar investment that SpaceX is talking about in Louisiana to be able to do all of these launches is fascinating. And I think it's gonna push a lot of different high cutting edge technology fields forward. And hopefully it's something that kind of brings and creates sort of momentum across the spectrum to move all of us in that direction. I think it's interesting the space-based data centers. I think there was very much an element of that that was sort of marketing and publicity. Cause some of this stuff's been talked about now for about a year. And about a year ago you really didn't need it. So I think a lot of it was just sort of marketing. And I think now we're getting to a point where there's a lot of pushback, some reasonable, some unreasonable, around data centers. And so a lot of these companies are like, all right, fine. Then we'll just put our data centers in space and nobody can get mad about that. So I think that we'll start to see things going in that direction. They're still probably, you know, five to 10 years out before we see it in any sort of a meaningful way. but that could be a huge unlock because there's an element of what we're seeing right now with AI, where like, yes, there have been innovation in AI, and also the very core base architecture. Is an idea that Jeffrey Hinton first proposed in the 80s. And really the takeoff that we're seeing right now in intelligence, in my opinion, has been heavily, heavily driven by just having more processing capability. And so if you Jason Griffing: Mm-hmm. Rob Carpenter: could have an exponential increase in processing capability in space, then I see, I think you just continue to see that intelligence going further and further. And what AI and what the foundational companies really need to do. And Dario said this himself is like, hey, we're getting a bad rap. Like, we need to do something big like cure cancer, you know, to kind of turn this narrative around. And so I'm glad they're thinking that way because I agree. And so whether it's, you know, terrestrial or space-based data centers, if just really having more computation is what will allow some of these breakthroughs, then I think something like SpaceX's launch cadence and getting these data centers into space, if It unlocks things like new material science, innovations into cancer research

and stuff like that, then it's a huge win for us as a society. And to some degree, we're on the cusp now of finding out if that's going to be possible or not, which is why the next, I think, 12 months are just going to be so Jason Griffing: Yeah. Rob Carpenter: exciting and fascinating to watch all of these developments. Jason Griffing: Yeah, the the implications of it all are just mind boggling and it does really feel like we're at this very bimodal point in human history where things moving Rob Carpenter: Yeah. Jason Griffing: forward could either be really, really good or really bad. And that's one of the reasons I'm motivated to do this show is to just continue explore this to explore this space and try to make sense of, you know, Rob Carpenter: Same. Jason Griffing: where all of this is going and how we as, you know, technologists and business leaders and so forth Can play our small part in helping to try to push it in a in a positive direction. speaking of that, as we're coming up against our time here, I'd love to close out with the question: how optimistic or pessimistic are you when it comes to your belief in humanity's ability to navigate this AI transformation successfully? And by successfully, I mean in a way that enables human flourishing. Rob Carpenter: in the long term I'm very optimistic. Will the next five years potentially be pretty disruptive? Yes. I hope that they are disruptive in a way that we handle Well, and kind of work through this process in a way that benefits everybody. I do think that there is a future five plus years out where every home has a humanoid robot doing whatever you need it to do. You have a digital intelligence that is doing anything digitally on your behalf that you need it to do. And so I think there's a real opportunity for the first time in history for technology to really free humans up. And, you know, maybe it's my own personal like sort of science fiction. Fiction bent, because obviously there's very dystopian ways that all of this could go. But, you know, the long arc of history, you know, bends, you know, towards justice. And so I hope as humans we navigate this process well. Because if we do, then I think we have, you know, a Garden of Eden in front of us. And maybe that's too optimistic, but that's what I choose to believe in for my own life and for young kids, you know, that I've brought into this world. I think there's just so much complexity in life and in society. today and it's exhausting and draining and everybody that I talk to, you know, seems to be having a pretty tough go of it right now in this market. And so I'd love to see this technology start to really make a difference and make people people's lives easier and better. And I think that that is very much a possibility. Jason Griffing: Yeah. There are lots of very thoughtful people out there who believe deeply in this future where AI can really help solve some of our most vexing challenges as a as a human race. And I I think it's exciting to think about those possibilities. And I think one of the ways that we can help push things in that direction is just to have good, thoughtful people in the space working to understand this technology deeply and all of its implications and be working to apply it in ways that, you know, move us towards that goal. So Rob, this has been a really fun Rob Carpenter: I agree. Jason Griffing: conversation. I really appreciate you coming on. If there's anybody listening who wanted to connect with you, maybe learn more about what what you're up to or ask a follow up question or two, what would be the best way for them to do that? Rob Carpenter: yeah, feel free to shoot me an email. It's Rob at frostriver capital. Jason Griffing: All right. Well Rob, thanks again. This has been fun. Rob Carpenter: Yeah, thank you. Jason Griffing: All right, take care.