← Back to search
What Are Lawyers Building With AI? Open-Source Agents and the Future of Legal Tech
The AI Sidebar · 2026-10-29 · 48 min
Show full episode description
Law firms don’t have to build their businesses on someone else’s AI infrastructure. Dan Schnurbusch, a probate and estate planning attorney at Schnurbusch Law in the St. Louis area and founder of the Legal Tech Collective, a Facebook and Discord community of lawyers, legal ops professionals, and builders, shares what he’s seeing in real time. He explains what an agent harness is, why he runs the open-source Hermes desktop agent, and why he believes solos and small firms have a major advantage over larger firms still working through governance and security at scale. He also makes the case that open-weights models like DeepSeek and GLM are keeping prices in check for Western providers, why he expects capable small models to run on phones within a year, and why he doesn’t think subsidized AI subscriptions are headed for a price explosion. Schnurbusch walks through his own daily workflow of supervising agent sessions on probate matters, why he sees AI as a force multiplier for most lawyers, and why he lands as a mostly optimist who stays cognizant of what could be lost along the way.
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
What attorneys and paralegals are actually building with AI, and why open-source desktop agents like
Hermes matter for owning your legal tech stack.
Benefits
- Hermes is open source: inspect, modify, and run it free on your own machine
- Install or update skills by telling the agent, no vendor connector interface
- Owning your stack avoids dependence on a vendor's harness, tools, and labor
- Desktop app shows running sessions, scheduled tasks, integrations, and skills
- Solo and small-firm lawyers can adopt without governance or scaling overhead
Use cases
- Dan Schnurrbush runs Hermes as the agent infrastructure for his probate and estate planning firm
- Legal Tech Collective grew to several thousand members, mostly attorneys, on Facebook and Discord
- Daily brief skill checks email and Slack, lists owed takeaways and hours since promised
- Legal professionals in the community adopting Hermes agent as their AI agent harness
KPIs / results
- Hermes described as roughly 15x larger than OpenClaw
- Legal Tech Collective: several thousand members
- Reported $200/mo ChatGPT plan equals ~$16,000 in API tokens when maxed; Anthropic ~$8,000
There was a report recently that ChatGPT, a sub, if you max out your session limits every session, it's something like the equivalent, like a $200 a month plan is like the equivalent of like $16,000 in API tokens. And Anthropic is like eight grand or something like that. But most of the money that these companies make is through API partners who have to pay full sticker price for everything. So it's not in the same way that like Google is able to offer like Google Docs and Sheets and Gmail and Google Calendar and all these different software for free. I'm not certain that they can't run these subscriptions at a subsidy forever. What are attorneys and paralegals building with AI today? How are they doing it? Coming up, our guest, Dan Schnurrbush, has about as good a vantage point on these questions as anyone. Dan is a practicing attorney running his own probate and estate planning firm, Schnurrbush Law, out of the St. Louis area. What's more, he's the founder and admin of the Legal Tech Collective, a Facebook and Discord community where lawyers, legal ops folks, and tech builders trade notes on exactly what's getting built, what's getting bought, and what's just hype. We're going to dig into what he's seeing across that community in real time and where he thinks Legal Tech goes from here. But we're not just staying small picture. And Dan and I get into where the AI stands right now, whether AI abundance is around the corner, what role open source has to play in the market. Then we bring it back down to earth, what any of it actually means for how legal work gets done. I'm Philip Vaden, Partner Manager at InfoTrack, and this is the AI Sidebar. Hi, Dan. Thanks for being here. Yeah, yeah, Phil. Thanks for having me on. Yeah. So, Dan, I came across the Legal Tech Collective, I guess this was maybe six weeks ago, and I've learned so much from the group. And this is such an incredible collection of folks who have sort of found their way through Facebook and now Discord to hook up. And it's all, you know, legal professionals going in and just swapping ideas on ways that they're leveraging artificial intelligence and other technology. So I wanted to start out by just asking you, what got you to start this group? How did it come about? I have a hyper fixation with AI right now and have for several years. It's my special interest, so to speak. And I have to tell everyone about it all the time. I can't turn it off. And when I go to like public things and I'm talking to, you know, parents of my kids at birthday parties, at kids' birthday parties, they're like, what is this guy talking about? Like I need, I needed a, um, a place to, uh, communicate with people who would actually care what I had to say about AI. Yeah. Because I was having so much fun with it. And so, um, I really grew up, uh, as an attorney after, after leaving my firm, I grew up as an attorney on Facebook because there are so many, uh, wonderful, uh, communities out there, maximum lawyer and lawyer on the beach. And, and, um, if you're not familiar, you gotta, you gotta get familiar folks. But, um, but that's where I started is, is in these Facebook communities and the podcasts that they were running. And, and, um, so that's where I naturally gravitated. I was like, well, maybe I'll just, I'll just start a group focused on legal technology. And, uh, you know, it kind of grew a lot faster than I was expecting. We have several thousand members now, which is, which is fun. And most of them are, um, most of them are attorneys, which has been interesting. We do get a good batch of, uh, you know, legal tech, uh, consultants and stuff like that, but it's not nearly as many as you would expect. Um, you would think what this many tech interested attorneys there would be, but, but no, it's, uh, it is mostly attorneys in the group. Um, like you said, sharing ideas. And so that's how I got started though, is I needed, I needed a diary where I could, uh, drop thoughts and ideas and what I was working on. And, and, and hence the legal tech collective was born. That's awesome. Um, yeah, I, uh, I can relate. I know that, um, my wife's a school teacher and every time I'm around the, the teachers, I, I dive into, uh, some new AI thing and talk about it until I see their eyes starting to glaze over. And I'm like, all right, I need a, need an outlet. It's not, it's not this group, I guess. It's very, it's very interesting. Like how different the responses to AI are depending on who you are and what you do and what role you play in the economy. I mean, entrepreneurs, you know, are overwhelmingly seem to be exuberant. Whereas anybody who's not an entrepreneur, like I, I run into a lot of, you know, uh, W2 employees working at companies and they're like, I don't understand why I should care about any of this stuff. Like this is, I mean, if anything, this is threatening to me, like, why do I, and you know, they'll focus on things like water and consumption and stuff like this. And we're, and over here, we're like, well, wait a minute, you can, you can automate workflows. And they're like, I don't know what, who the hell cares about that? You can 10 X yourself. Yeah. It's very, it's, it's very interesting. The incentive structures and the responses are very, very different. Yeah. Yeah. And it, it has been, you know, something that, that I've, I've tried to stress with people when I'm, when I'm visiting about new tools that are out, workflows, things you can do with work. And this is, you know, internal inside of InfoTrack and speaking with other people. People that have similar roles to me about things that I do just in my day-to-day life. Hey, you know, I have a daily brief that I set up that just runs as a skill that goes in and checks my email, checks my Slack channel, and then list out like, Hey, here are the things that you have as takeaways from meetings or any kind of communications. And these are the things you owe people. And this is how many hours it's been since you said you were going to do it. And so then, you know, in the morning when I'm fresh, I can go through and just knock this list out and, uh, and start, start fresh for the day. And just little things like that are, are huge because number one, you know, those, those things don't fall through the cracks. Cause you know, inevitably if you've got some, some project that's not urgent and maybe not important yet, it may be two months from now. Um, those are the things that I think can kind of slip through, but you know, my, my skill doesn't know what's important and what's not. It's going to list all of them out. So I know, I know which things to go through and take care of, but, um, yeah. So in, in your, one of the, one of the things I really like about it, you as the admin in the group, um, you, you're really good at facilitating the conversation, keeping the conversation moving and keeping the conversation moving in a positive direction all the time. Um, and I, uh, and I think the other thing that, that is amazing to me is how knowledgeable you are in the space so that you're, you're always kind of just at the, at the edge of, of what's being rolled out. And you've already not only like heard about it, but you've already played with it, integrated it into your, your workflows. And one of the things that's come up recently that I wanted to hear, um, more about is, is the, uh, the Hermes agent that it seems like a lot of, uh, a lot of the, uh, the legal professionals and some of the other legal tech folks are, are starting to adopt and put into their, uh, leverages there as their AI agent. So I wanted to hear about like what you feel like are the pros and cons, like what, what size of law firm may benefit from, from diving in and looking at something like that. And, um, oh, that's a big question. It's multiple, several questions. They're pretty big ones. Yeah. I like Hermes for a lot of different. So I guess it, it, it, um, warrants a little bit of contextualization for a lot of non-technical people. It's like what, um, so Hermes is, is an open source desktop agent. Um, well, more specifically, it's an open source desktop agent harness. Um, and people are like, well, what, what the hell is a harness? Um, so, and this is like a term that I think attorneys are gonna have to get familiar with cause this kind of stuff is getting, they're gonna have to get familiar with a lot of technical stuff they didn't want to before. Yeah. Um, just because that's the way everything feels like it's going. Um, so at, at desktop harness, the harness is the software around a large language model. So like you have, like a model has no sense of, you know, session state or what you have talked about before or context or different, uh, threads that you have with it or memories or skills or model context protocol or, or, or integrations. The model doesn't understand any of that. It's just a predictor of the next token in text. Um, so the software around it is the harness that actually makes it usable as a tool, right? Um, and the harnesses that most people are familiar with are things like chat GPT in the cloud. Uh, it is a, it is a harness. It doesn't live on your desktop. It's in a cloud, uh, cloud environment, uh, Claude in the chat window or Google Gemini, uh, in your Google account. These are all different harnesses. There are different software environments that are kind of wrangling the LLM and, and giving it coherence and context and access to the world. And Hermes is an open source version of these. So if you've used like, uh, uh, chat GPT codecs or Claude code or Claude co-work on your computer, those are desktop Asian harnesses that are designed to allow an LLM, uh, a place to play on your computer, not in the cloud. I mean, they do inference in the cloud, but they, they can take action on your actual computer. Um, they can write code and access your terminal and delete all your files and do it and all kinds of things. Um, and that's a different conversation, different question, but, um, they can access your computer. So, um, those are ones that more people are familiar with is tools like Claude co-work and, and, and chat GPT codecs, but Hermes is an open source version. So, you know, you can actually open and inspect and modify all the code for it. It's free. It's publicly downloadable. It's now the, the largest, I don't know if it's the largest desktop harness. I don't know if it's exceeded codecs and Claude code, but it's like, it's like 15 times larger now than open claw, which was the one that kind of came out and set the world on fire. Uh, when it, during its initial release, it's now much, much more used than open call in part because it, a perception of it being much more secure. Um, but ultimately it's kind of the, the, um, the agent infrastructure that I use for running AI agents. And, and they've got a nice desktop application now. So you can see, you know, the different sessions that are running and you can set up scheduled tasks and, uh, integrations and skills and stuff just like you can. And codecs or Claude co-work and whatnot, it's just, uh, open source. Um, and there are lots and lots of reasons that I like that. I'm a big believer that people need to own their AI stack. I think as agents get more and more powerful and more and more integrated into, um, organizational hierarchy, like we are at risk of building a company on top of somebody else's infrastructure and not just the software. Mm-hmm . But the labor. Yeah. Like you're, we're, we're at the point now with agents where if you build all of your stuff in like Claude or ChatGPT as your primary harness and you use their tools and their integration systems and their skill repositories and their, you know, everything. Um, and you know, what if tomorrow the U S government decides this application is too dangerous now? Um, we're now have to go through an approval process that might take three months to get approved before you can get your agent access back. Like that's not a good, uh, situation to be in. So I, um, anyway, so there's a lot of different rationale I've written at length on why Hermes, but, but it comes down to a lot to owning your system and owning your stack. And, um, you know, um, I like the ergonomics and experience of it a lot better than other harnesses too, just cause, um, they don't, they're not as opinionated as far as like how you install things. Like you don't have to go through their goofy little interface to install connectors and skills. You just tell your agent install this skill and it works and that's it. Um, that's the, that's the whole, or, you know, or, or update my skill and it just updates it and that's it. It works. You know, I don't have to go through their, um, scaffolding, but that has pros and cons. Cause like to your followup question was like, what size of organizations should use this and stuff like that. And, um, I think solos and small firm operators are at a major advantage right now because they don't have to deal with governance and scaling. And, um, how do you deploy this stuff at scale in a way that doesn't like screw up or cause security incidents or, um, or just compliance headache or just like headaches of like making sure that everybody's using the rights, the same stuff. Like, are we all using the same skill and all this stuff? And that's a different layer and a different set of skills. And it's something I'm actively, um, studying and, and researching right now. I've formed quite a few opinions about it, but, um, but right now I think Hermes is, um, it's, I don't know if this particular thing, I'm not sure if I have a strong opinion on like whether an organization should roll out Hermes agent. I think every CEO, every owner of a firm should run an open source agent like, like Hermes, um, whether I would roll that out to every paralegal and receptionist attorney in the firm. I don't, I would have to think about how to do that responsibly. Um, because it is, it is essentially like putting like a loaded AK 47 into the hands of every person, like, which is, you know, very powerful if you're trying to like, uh, you know, uh, render mayhem. Um, but, um, sometimes if people don't know what they're doing, they can shoot themselves in the foot too. So you have to be careful about that. Um, but yeah. So let's, let's, let's dive into a couple of, so there's been a lot of discussion about open source versus closed source and, you know, protecting open source. I can't remember if Anthropic ever did sign on to the, you know, let's, let's stay, you know, focused on keeping open source live and active. I don't think they, I think I don't think so philosophically opposed, I think. Yeah, exactly. Yeah. They, they, they want the, uh, they want the duopoly really badly with open AI. Some of it I think is just that Dario is terrified. Like he thinks we're at risk of all of us like dying or destroying our infrastructure. And he's, he's a doomer. I think he's a doomer trying to ride the wave, but. Yeah. It's an interesting, it's an interesting, uh, mindset that he's got. I, I, I've, I've been chalking it up to, and, and for those that Dario is the, uh, he, he runs Anthropic, which has Claude. Um, but he, and he's been, uh, he's been a doomer all, all along. Every model that comes out that's, that's Anthropic. He's, it's almost like, what have I done? You know, we've unleashed some monster on the country. And it's, it's, uh, it, to me, I've, I've always chalked it up to he's desperate for some regulatory capture, uh, to prevent people from being able to play catch up because they're, they're, they're ahead. I mean, they're, they're, uh, their stuff is really, really good. Um, and, uh, and they're, they're quite far ahead. So I think that, um, the, this, this open source idea, um, some of these open source models, they're coming out of China. And I know that I've talked to several people that are like, you know, I'm never going to do this. I'm never going to like install, you know, Quinn on my, on my, my hardware. I'm just not going to do it. That's crazy. Uh, let the Chinese know everything that I'm doing. And I'm like, well, that's not exactly it. So, so do you want to explain, do you want to explain how the open source products, open source models work in the real world? So I will preface this by saying that I, I, I have not run a lot of, um, open weights models myself. I have more philosophical opinions about them than I do technical, but, but I will say, um, yeah, I mean, open source models, open weights models are, in my opinion, uh, critically important, uh, to a lot of different things, but they're also dangerous. I mean, there, there, there's the reality is it's a two-sided thing. Um, but where, where, where to go with this question. So on the one hand, I think, um, uh, I'll preface by saying like, I think open weights models and open source models are, are, are super important for a couple of reasons. Like number one is I think the most important role that I think they play is, um, price constraint on Western models. That I've heard many people, I mean, this was, I think this is less common now, but like six months ago, a lot of people were like, Oh, just you wait. The, the price explosion is coming. All these subsidies are going to dry out and then we're all going to get smacked with, you know, uh, giant increases in prices. And I have never believed that that was going to happen. Um, I, I don't think, I think there are lots of ways that, that the scenario could play out where that doesn't happen. And I think the more that we have access to open weights, open source models, the less likely that is to happen. Because if you see, you know, like deep seek before launched a new flash model a couple of weeks ago, um, that is only a few points shy of the frontier. And it's a, it's not that big a model. It can be run on a, on like a, a really nice home computer. And, um, even when you buy it through the API, which most attorneys are not going to do for obvious reasons with security and stuff like that. But, um, it's, it's a cost. It's like, I can't remember what the factor was, but it's like less than a hundred times or more than a hundred times cheaper than Western models. And then yesterday GLM 5.2, one of the other major, major open, open source Chinese models dropped their price 95% as well. And it is now cheaper than deep seek and GLM 5.2 is at the frontier. It's like benchmarking, like number four, uh, behind only, uh, you know, fable and, and chat GPT 5.6. And, um, and now it's, it's incredibly cheap and you can, anybody can run it. Um, if you have the hardware. And when I say anybody, I don't mean like every law firm needs to buy like a stack of DGX sparks or sparks or something. I mean, um, there are, you know, for a few hundred thousand dollars, um, an American server farm can spin up a company and start, you know, download the weights and run that model for other people in a secure American environment where you're not worried about, you know, am I giving my data to the Chinese government? Um, because it's running through a, an American company that may have, you know, American assurances about ZDR or zero data retention or something else. Mm-hmm. So it gives, it creates an infinite well of competition against Western, um, Western providers where, you know, if you can spin up a, a competitor, um, language model delivery system for, uh, you know, less than, you know, not that much money. Um, it just creates an environment where the consumer surplus of these technologies is going to be enormous and there's not going to be nearly as much rent seeking from Western model makers as there might otherwise be. So I think that piece is critically important, but like, um, but like, you know, to your, your initial question about like, how do these actually work? Well, you're, uh, most, I think attorneys don't probably don't need to worry that much about, you know, self-hosting open source models. Like some who are really security conscious or privacy conscious, or, you know, they work with the government or with like really high profile clients and they really don't feel comfortable disclosing anything to any of these API providers. Then maybe they ought to be looking at it. Um, but for most people like running a lot of these big capable models, you're, you know, you need over a thousand terabytes. Of, uh, or excuse me, a thousand gigabytes of, um, of VRAM in most cases. So you're looking at tens of thousands of dollars to build a machine capable of running that. Um, so not that attractive for most lawyers, but, um, you know, that's a whole different conversation, but, um, that's my take on it at least. But, um, I don't know. Does that answer your question? I don't remember what your question was. I just kind of get going sometimes. Yeah, no, it was a, it was a big question. I mean, the, you know, I think that the, the, and I, I, I, I agree with you that the, the open, open source models are going to, are going to keep the prices constrained. I do, I do see, you know, with, with, um, uh, SpaceX AI and, um, meta, um, there, you know, them coming into the market, they, they own their compute centers, um, which is really helpful. And then they're able to come in and do some cost undercutting. And so I think that that's, that's also helping to kind of keep prices in check. These, I, the thing that I'm curious about is that there's, there's a different client, the different kinds of, of inflation that are happening. And one of them is just like compute inflation. So like, at some point, like, well, I don't need Chad GBT, you know, I don't need GBT 10.7 to help me draft a motion. Like, I don't, if I need to solve, you know, how to build a particle accelerator, like maybe I need that, you know, how to improve a particle. Maybe I need that, but I don't, you know, so I think that there's, there's some point at which, and I think we're in, for most of our use cases, I think we're, we're getting there pretty rapidly. Or we need to do 80% of the task that, that people, that most people need for on a day-to-day basis. And I, I remember somebody mentioned on something, something I was listening to, they were talking about how, how many of the tokens that were being burned right now were people asking stuff that was like, you know, it was just like stuff we would have gone to Google for two years ago, three years ago. And it's like, you know, what's the, how, what's the temperature outside? You know, like that's, that's, that's, that's like using a spaceship to, to fly to go get your pizza delivery. It's just unnecessary to, to do that. So, yeah, so I think we, we're also kind of approaching, approaching a point there as well. Yeah. Um, so it's, I was just gonna say that it's interesting that the, the economics around that are, have been fascinating to, to watch kind of unfurl, because I think you're going to see a couple of different trends proceeding in tandem. Cause one is that, that the price of intelligence is just, I mean, as you, you will get people saying, oh, it's, it's, it's going up, but tokens are getting so expensive. These companies, their budgets are getting blown out. But if you look at like the, the adjusted cost of, uh, of a intelligence adjusted cost per, um, task, um, you know, uh, based on how, how much that's, that's, it's in essentially free fall. Uh, the, you know, the amount of intelligence you can get, especially when you factor in open source models, but even the West is trying with like, um, chat or open AI's GPT 5.6 Luna, which is a phenomenally intelligent model. Um, it's not the sole model. It's not the smartest model, but it is, it is very, very good. It is so, and it is now so cheap that it is because of the improvements in efficiency and service delivery and GPT 5.6 soul recently had a. Um, uh, made a breakthrough in within open AI that allowed for a 20% more efficient delivery of, of tokens across the organization. Yeah. That's huge. The model figured out. And that was part of what precipitated that the 85% price reduction on the Luna model. But, um, so I think you're going to see, and at the same time, you're seeing a lot of these Chinese, uh, open source models and increasingly Western open source models with like, uh, meta and the new spark models. And stuff like that, where, um, they're going to get so good and so compact that you're going to see these models running on people's cell phones in a year. Yeah. Yeah. Um, like we're going to have local inference. That's really good, um, on small devices in, in a year. And it's going to, you know, we're going to see, um, people pulling out to APIs for complex issues. Like, you know, we, like you said, we need to upgrade the particle accelerator or something, you know, that'll query, you know, the top model at some firm and you'll have your subscription and you'll, you know, burn through your tokens. But, um, I think we're going to see a lot of that is just open source shifting to the edge where we're going to see a lot of edge deployment of small models. And then you're going to see, but I think you're going to continue to see the subscriptions like that we have where tokens are fabulously subsidized and people are like, oh, well, they're just burning through cash on investor money. And I don't think so. I think they're deploying these as loss leaders. I think they know that they're, they're, um, you know, they're, they're not making as much money on these subscriptions. They may not be making any money. Um, like there was a report recently that, uh, chat GPT, um, uh, a sub, if you max out your session limits every session, it's something like the equivalent, like a $200 a month plan is like the equivalent of like $16,000 in API charts. Yeah. And Anthropic is like eight grand or something like that. Um, but most of the money that these companies make is through API partners who have to pay full sticker price for everything. So it's not in the same way that like Google was able to offer like Google docs and X and sheets and Gmail and Google calendar and all these different software for free. I'm not certain that they can't run these subscriptions at a subsidy forever. Um, as long as they're able to offset those costs through, um, collateral business lines. So I, that's one of the reasons I'm not convinced that we're going to see a massive explosion in pricing on, on subscriptions either. Um, yeah. Which, but it was weird that you're able to get super high quality intelligence on a subscription and it may just stay subsidized. Like we may be able to continue to use awesome models that are way overkill, uh, indefinitely. I don't know. It's, uh, it's hard to say. Well, you know, and we're, we're also, you know, we've, we've got a, uh, tax structure in the United States that really, um, incents these companies to go full bore on, on, on their, their CapEx spend to build, uh, new data centers. And, and, and, and along with that, um, they're building power plants and everything else to go along with it. But, you know, it's, but you, uh, you essentially, you know, I think Google announced, uh, and this caused their stock to drop quite a bit, which I thought was, was a little, little odd, but, you know, they're going to spend, I think something like 250 billion or $350 billion on, on, on CapEx. And, um, but the, the reality is, is that they get, it's basically at a 26% discount because as soon as they spend it, they get to that that's a write-off on their taxes. And so, um, when you think of 26% off 350 billion, I don't want to do any lawyer math live on, on the podcast here, but that's a tremendous amount of money. I mean, that's enough for, you know, for a couple of, couple of additional really good sized data centers, um, to be built. And, um, and then, you know, then they're really rolling at some point. But, but like you mentioned too, the, you know, the size of things, what we've learned about hardware over our lives and what, what people should be accustomed to at this point is like, you know, think about a terabyte of memory in 1982, um, it would have taken up, you know, a warehouse. And now it's like a terabyte is like, man, I hope that's enough for my camera, my pictures, you know, It's insanity. I have a Nintendo switch, uh, that, that the kids and I play and you have, you get an SSD micro SSD extension card and it's a terabyte. And it's literally the size of my, not even my thumbnail, like my pointer fingernail. Yeah. And it fits in there and it runs at extremely high data transfer rate. It's like 800 gigs or like 800 megabytes a second or something like that. Yeah. Uh, on a, on, on a chip, the size of your finger. Yeah. Um, I know, I know I just sound like my, my grandpa when I'm talking to my kids about this. Cause I'm like, I don't think you understand how miraculous this is. They're able to put this much memory on this tiny little thing. And my kids are like, yeah, what's the, what's the big deal? I think it's, what is it? Photo photolithography or whatever it is, the process of, uh, lasering, uh, microchips and CPUs and all that stuff is, uh, the closest thing. I think we have to like sci-fi magic. Uh, yeah, that is, uh, that humans have invented. I think. Yeah. It is. Yeah. It's truly, truly extraordinary. And I, I think that, you know, and a lot of the, uh, people are very concerned about like the, the pace at which data centers are being like, are being built out in the United States. And, and, you know, Hey, water, which is like a little bit of a lie, the water concern, the, uh, power is, they're obviously going to require a lot of power. But I do think that we're going to hit this, this point. I think Mark Cuban made this, this point where he was like, everybody's talking about spending a trillion dollars on this and a trillion dollars on that. They'll never spend that much money. And that was sort of the takeaway people took from that was that they were lying. And I was like, well, it wasn't really his point. And his point was that you won't need to spend a trillion dollars on this by the time that you're actually installing some of these things, because the cost will have gone down so much that it will, it will be much less expensive. Your, uh, your, your electric outlay is going to be much lower because things will become smaller and more efficient. Um, cause it's not only a terabyte taking up a whole warehouse, but you have to power this massive warehouse in order to restore a terabyte. Now terabyte is on a tiny little micro SD card says your, you know, pinky finger, and then you can drop it into your device and it costs $10, $20. And then your, your phone can run the whole thing for a day without having to be recharged. Exactly. I, I, I, I feel a lot of, or sense a lot of Malthusianism, um, around all this or Thomas Malthus for those who are always an economist, um, from a while back who was predicting that we were all going to die of starvation, uh, around the turn of the industrial revolution. If I'm not mistaken, somebody can fact check the timeline there, but, um, and this was not an uncommon belief. Everybody thought we were going to die of starvation because population was growing much faster than the rate of, uh, growth in the, in agriculture. And then, um, advancements in agriculture resulted in a increase in, in agriculture output by like a factor of like 10 or 20 X and the whole issue just disappeared. And, um, I, I can't help, but feel like a lot of that is going is that we're going to see that story play out again in, in artificial intelligence. I, I mean, I'm, I, I don't have any, like all this data center stuff. Like I'm, I think it's important in terms of maintaining our, um, geopolitical advantage, our economic advantage and the development of, of AI. I think that deployment of data center, I don't begrudge people who have concerns about where they get located or is it actually affecting, is it actually affecting my, my electrical bill? And I have heard from insiders that there may be some in certain places, this may be true and so on. And I, I, I get all that, but, um, I think a lot of the doomsday predictions of, oh, and if we follow this trajectory in 10 years, we're going to be generating as much power as the sun or something. And we're going to need to suck up all the water in there. I'm like, I, I really don't think that's going to happen. I think we're going to see a lot of different shifts happening, but I think you're going to see the supply of energy catch up and not just that. I think you're going to see the demand for it, um, level off and probably decline at some point because I think a lot of the, a lot of the small models are going to shift to the edge and they're not going to be run out of data centers. You're not going to need a data center to run your, I mean, the, the big new hardware guy they put in, put in a place of Tim Cook at Apple. He's a hardware guy. Yeah. I think Apple's about to make a big play here. I think so too. They're about to make a big play as, um, you know, they've already got their phones in everybody's hand. They've already got, you know, hundreds of millions of, uh, consumers who are going to buy the next iPhone, but the next iPhone is probably going to have, if I had to bet a much better CPU and more VRAM. And it's going to be able to, you know, maybe not the next iPhone, but the one after that, it's probably going to have the ability to run, uh, small local models pretty reliably. Um, especially if, if you, if you take Anthropic seriously, that they've essentially solved prompt injection, which is incredible. I mean, if you look at, I don't know if it's, you know, their, their internal reports are saying they've gotten their models down to a 0% prompt injection success rate, which is. If true, um, uh, a cataclysmic, uh, you know, uh, security improvement, you know, shift that is, that is, um, you know, there was a paper out of one of these universities recently that said. That it's an unsolvable problem in LLMs. You can't solve for prompt injection, but, um, or that you can, because there's no way for them to tell the difference. And of course, these are people looking at the LLM. They're not looking at the harness. Um, and I, I agree with them that from a strict LLM standpoint, it's not possible to solve prompt injection, but that's not how LLMs are actually run in real life. They're run inside of complex harnesses that have a lot of context and a lot of cues and a lot of, you know, this package is untrusted, but this one is. And so if, if that is to be believed, I think we're going to see a lot more expansion of, of local LLMs and, and correspondingly a dramatic drop in the demand for energy. Um, but we'll see. I don't know. Anything could happen. Yeah. And yeah, that, that Malthusian trap is something people definitely fall into. And some of it's just a failure of imagination or failure to like, you know, understand that everybody's, there's a bunch of people that are highly financially motivated to figure these things out. And, um, and a lot of the stuff is like 30 there, just a matter of implementing it at scale, uh, or tweaking it slightly to make this other thing happen. And so, um, yeah, I, uh, uh, and then on the, you know, I think on the, the prompt injection side, that's one that was, was really, you know, solved by, Hey, we're going to, we're going to build our own harnesses. And then this is going to be what we incent people to do by doing the subscription model. Um, and I'm sure they worked out something where it's like, Hey, you know, 60% of people are going to barely use this, but they're, but their company's going to pay for it. And, um, and then there's guys like you and I that would just use, use maximum usage every single time that it comes out. And then, um, but, but you, when you put that paywall there with the, with the API, cause then I guess this was maybe a year ago, year and a half ago, Anthropic was like, Hey, if you have a subscription, that's great. You can't use this subscription for your API anymore. So they shut that down so that, um, so you couldn't, uh, couldn't run the $16,000 worth of tokens. You had to pay as you go. Um, but that is, uh, you know, that's in, in a lot of these models, a lot of these open source models, they're being trained in on, you know, on Anthropic and on open AI. And, um, uh, and that's, that's how they're able to, to keep up for, I mean, they're spending a lot, um, you know, funding that spending a lot of money. I can't remember if it was open AI or XAI, but I think it might've been XAI. They're, they're selling, um, access to, uh, distillation to downstream Chinese providers. They're instead of like making it, uh, into this big national security issue, they're literally selling it. They're saying, you know, pay us a certain amount and we'll allow you to distill off our models without, you know, banning your accounts or engaging in like counter, you know, espionage and trying to block you and all that stuff. Um, very different approaches. Very interesting. Um, yeah, it is, uh, it is fascinating, but yeah, I, um, I, I know that we've, we've kind of ended up, um, nerding out in the weeds here, um, on the, on the pod, but this is all stuff that, you know, I think that, um, some people are, are reading the news about this and staying up to date. Uh, about what's going on in the space. And, um, I think, I think it's easy, uh, depending on where you're looking to, to get like, oh man, this is going to be a disaster, you know, with, with like these things coming out, replacing people and blah, blah, blah. And I, I still look at it as, you know, all of this technology that we've been talking about and implementation of it. I still think it's, it's just, it's going to be something that we're augmenting people. We're not necessarily replacing people. Now, some people may get replaced that are in very, um, uh, if you're, if you're doing one task, you're literally doing one task all day. Like, look at this, click a button, look at this, click a button. And it's the same action all the time. Then you're, you know, that's probably a job that's going to get automated out, but that frees that person up to do something that's more value additive, uh, for their, for their job. And, um, uh, and I, and there's, there's a lot of stuff to do and there's always going to be more stuff to do. So, um, yeah. Yeah. I, I think, I think, I'm, I think I agree with that, that, that, um, mostly will be, uh, additive that it'll be a multiplier, a force multiplier. And I know people have been saying this for years, like, oh, you know, the attorneys who do the best are going to be the ones that use AI. They're going to outcompete the ones that can't. And I, I, I still believe that's mostly true. I still, I do think there are contexts where we will see, um, wholesale displacement of, uh, certain job categories just because, um, I've done it internally. Um, so I, I could see, uh, I could see it happening in certain contexts. And I know there's a lot of venture capital flowing into like, you know, building these AI native firms that mostly I think are probably mostly marketing puffery. Um, but, um, because I, you know, I think, um, you know, especially a lot of administrative type tasks are, we're at a point now where we can pretty reliably, um, hand that off to an agent loop. Um, and the areas where I find myself still spending time at this point is like every day. My day is I, I wake up in the morning. I, you know, make breakfast, take the kids to school or what have you. And then I come down and I have a stack of like 15 to 20 new sessions, uh, in my Hermes desktop agent. And that's my work for the day is going through each session, figuring out the work that the agent did, giving it a follow-up action or approving its proposed, like it will propose for itself. Here's the next things that I want to do on this case. And right now I make it check in with me as part of that loop so that I can approve stuff and tell it what to do. But I mean, in, you know, it's thinking through the case and what needs to happen. I'm just kind of supervising and making sure that it's judgment and decisions don't sound crazy. Um, and I have it easier than most because we do probate administration. It is very procedural and process driven. And, um, and we not only do we do probate administration, but it's non adversarial. So we don't take litigated cases. So, um, it makes it a little easier to, uh, rely on the model in the same way that, you know, you might rely on a paralegal or something like that, which we also, which I also have a wonderful paralegal. Um, but, uh, I don't know that I will need to be a check in that loop forever. Um, yeah, there are like most of the outputs. My response is I approve your next action and then I archive the session and that's it. That's the end of the movement on that case. Um, you know, so it's, um, at what point do, you know, it's almost an empirical question of at what point do I no longer need to make, um, feedback and tweaks and corrections. And most of those are a product of refining skills and connections and all that. So, um, I, I don't know. I think, I think it's mostly going to be a force, most important multiplier, but I do think there are going to be once people, more people figure out how to do this stuff at scale. Uh, the scale issue is the, is the next quandary to process through, but, um, it will have interesting effects. And I think some of it is going to be shifting us into more client focused roles. Um, like, I think it's, I think, I don't know, right? But this is something I run into all the time. Clients calling and saying, oh, I haven't heard from my attorney in three months. I don't even know if he's still alive. And three months is probably generous. Like there are some that are like a year and a half. I haven't heard from my attorney. Um, and the case is not moved. Um, so does it open us up to make sure that that kind of stuff never happens again? Yeah. Hopefully. Hopefully. Yes. Mm-hmm. Um, so I don't know. Well, I'm, I'm along for the ride. It's happening whether, whether we want it to or not. Yeah, this is true. Well, Dan, uh, this has been great, man. I really, really appreciate the time. I have one, one exit question I asked guest and, and that is, and I think I know we're going to land on this, but, um, are you a techno optimist or a techno pessimist? Mostly optimist, but, uh, cognizant of the pessimistic side. Mm-hmm. I think the technology in this situation has the chance to be the most transformative and impactful technology in the history of the species. Um, it has the chance to, uh, and I don't exaggerate when I say this. It has the chance to solve, uh, disease, um, solve mathematics, um, provide a vehicle for sending something that looks like consciousness. If we get there, uh, some people think we're getting pretty close, um, into the stars. Um, cause it's probably not going to be, you know, biological organisms that, that ever leave the planet, uh, in any serious way. Mm-hmm. Um, I, I, I'm kind of a weird techno utopia, not, not utopianist, but, um, I do believe we are building our descendants right now, uh, that will outlive our species. And, um, I don't know what that's going to look like. So, I mean, I, I, I think in that sense, I'm, I'm sort of, I'm sort of optimistic, but also, um, I think we're going to lose some stuff. Um, just like every other technology, um, you know, with the advent of, uh, air conditioning in the car and the decline of, um, you know, interpersonal connection that that has drawn and, and indoor air conditioning and how nobody spends time outside anymore. And, uh, you know, social media and cell phones and home entertainment and how nobody, uh, nobody knows their neighbors anymore. Like there are, um, there, I think there are, there are downsides to technology that are, um, often hidden and quickly forgotten. Uh, once the, the generation who didn't, the, you know, as had, didn't know the world before, um, or, or can't remember the world before, um, dies off, um, the, the, the, the way things were is going to disappear. And I think there's going to be countless ways that things are going to change and probably become worse that, um, that were, you know, people who are living through it, will probably see it. And then our, our kids and grandkids will have no idea that it ever happened. Um, and they'll just have to adapt and the world will go on and they'll find new things to be optimistic about. And, and at some point they'll stop being sad about the things they never knew and that's okay. Um, so I, I, I guess it's, it's, um, a mixture of optimist and, and pessimist and, uh, perhaps a realist is, is it's good things and bad things all, all mixed in one, but we'll see. Awesome. Yep. And, and this is, uh, uh, you're, you're a dad, right? You're a dad of four. Yeah. Yeah. That's true. Dad of four. Yeah. Yeah. I'm a dad of two. And, um, and so it is an interesting thing kind of getting the kids, like getting the kids ready for, for, for this, this new, new world. And, um, um, and, uh, it's, uh, it's an interesting, interesting time for sure. Yeah. But that's something I'm not quite sure what to do about it. But I, I thought, um, there was an interesting post by Dean Ball, who's open AI. Um, I can't remember his role there, but he's very high up at open AI. Um, and he posts a lot on X and, uh, he talked about how humanity of the past was a sculptor, but humanity of the future is a gardener and arborist. Mm-hmm. I think the point he was getting at is that, um, we used to do everything directly and with our hands and create things. And it was the application of our knowledge and capacity to problems that, that allowed things to move forward. But increasingly we're going to be, um, cultivating and pointing things in direction. We'll be planting seeds that are some of which grow on, grow on their own, except for the, the, the plants are, I guess, models, uh, agents, intelligence systems that are hopefully aligned with our goals and, um, growing into beautiful flowers, but, um, could grow into other things too. So we'll have to see. Yeah. Yeah, absolutely. Well, um, I encourage everybody who is interested in, uh, artificial intelligence and how it interplays with, uh, legal technology and practice of law. Uh, you should jump in and join the conversation at the Legal Tech Collective. Um, Dan's very fast, amazingly fast at approving, uh, or rejecting people. I guess I've, I've been approved. I don't know if you reject anybody. I don't know if you had to keep anybody out, but, um, you should, uh, you should join and join the conversation. It's an excellent group, a really positive group. If you have any questions about, uh, a setup that you're diving into on your own, um, loads of people there, including Dan, um, who can, who can help walk you through it and give you advice and guidance. Um, and, uh, there's nothing, nothing quite like it. So, um, uh, we'll, we'll post, uh, uh, links to the group and all of that in the, uh, in the podcast notes. And, uh, man, Dan, thank you so much for this discussion. This has been so fun. Um, I, uh, uh, I'm glad we have the outlet to be able to talk about, uh, about artificial intelligence, uh, with each other. And, um, I, I know my wife will be relieved that I, I got a lot of the AI talk out of my system before she gets out. There you go. There you go. I appreciate it. Thank you, Phil. Thanks for having me on. It was fun. Yep. Absolutely. Have a good one, man. You too. Bye.