Distributed Dissent — Episode 7: The $500 Million Gamble, Temu Opus, and the Build-vs-Buy Reckoning

[MUSIC] >> Hi, welcome to Distributed Descent. Our podcast about law, AI, and technology. My name is Enhong, the CEO of Generous AI, and I am joined as always by my friend, Matthias Bach, CEO of Tukumalabs. >> So this has been another big week for law and AI news. A heading up is Kirkland's massive $500 million announcement that they are spending on AI over the next couple of years to build their own in-house solutions. Matthias, you have a lot of thoughts on this. >> Yeah, they're the first to publicly announce that step. I suspect that other firms are thinking about going in a similar direction. Of course, $500 million is a big number. We were just talking before for a firm like Kirkland with $10.7 billion in last year's revenue reported. It's actually not that big of an amount. If you think in terms of a corporation, that would be something like 5% in R&D investment. >> That's over multiple years. So it's even less than that. So maybe not even that substantial, but it's of course a huge number. But the thing is, of course, when you go down the list, or lay them, sits at $7 billion in revenue, DLA number three, sits at 4.2 or something like that. And as you move down to 10 and top 20, you're going further down to more $2.5 billion. So the percentage of that investment, if firms were to invest something similar, it's going to be just proportionally much bigger. >> Yeah. >> And now the problem is by and large the same. So just because you have $2 billion in revenue, it doesn't mean that the problem is smaller for you. Kirkland is not particularly large. I don't exactly remember how many lawyers they have, but probably only over $3,000 or something, right? There's other firms that are bigger, and then all of these top 10 firms are all similar size. So the kinds of investments that they all would have to make to achieve a similar outcome are probably in a similar ballpark. Now, of course, they had a question as $500 million, what are you going to do with it? >> Yeah, exactly. >> Even it was $100 million a year, and you're a law firm, would you, and you want to build? >> Yeah, I don't think they're crazy enough to try to train their own foundation model of some kind, because if they wanted to do that, $500 million wouldn't be nearly enough. But if it turned out it is, then surely, and thought big and open now are doing something wrong. >> Well, you can train a deep seat like that. >> No indication. >> So far, what I suspect though is I could imagine they would be fine-tuning some smaller models for particular use cases, and this whole mixture of experts kind of orchestration, is this something that we're going to see more of? It'd be interesting to see whether, not that they'll probably talk about it, but whether they would be relying on deep seat or Q-Wen, those models out of China, which presumably, if they ask their clients, it'd be a hard note, but they're good models, right? They're maybe four or five months behind Frontier, in some cases. >> And you could fine-tune those models. >> And you could actually fine-tune because they're open-weight and open-source and all of that, right? So that's going to be interesting to see. But I have sort of cheeky hypothesis in that sense that I wonder whether maybe the biggest portion of that money is not actually going to go into the development of something specific, but is actually earmarked to overcome internal resistance. Because the thing is, if you want to make these things really useful, you can't only give people tools. This is the thing that drives me absolutely mad when I look at a lot of the news where all of these products that we're seeing today, they're essentially tools, right? They have tools for individual lawyers, but that's not going to be the big winner. You need something that really orchestrates, that works across teams and entire teams, right? So you're going to have to really fundamental everything how you work. You're going to have to understand your workflows in detail, and that's kind of along the way what they've announced, right? We want to centralize our joint intelligence, all those things. But they'll really need to look and find ways to change the structure, how they work. And I suspect that we'll also meet internal resistance. So yeah, maybe possibly some of the money has just taken to say, well, here, here's a check. Just let's move on. Is it part of it? Is it also sent a signal from senior partners to say, we're serious about this. We're putting this big, flashy number behind it. And so you guys better get in line. And I wonder about your earlier point about the tools versus the workflow. I think that hits at this very interesting point, which is a broader question, where Harvey LaGora and all the existing Leo AI tools today, including co-council, Lexus, et cetera, they're all tools to make lawyers more efficient or more accurate or work a little bit better. But they are not tools that deliver the end product, which is law, legal services. And I wonder with Kirkland's announcement, and then at the same time, you have both Claude coming out with their legal plug-in, and you have Codex hiring high profile legal tech founder to run a Codex for legal solution. And we're going to veer more and more towards producing the solution, producing actually, producing legal work instead of just helping legal produce legal work. My personal thesis there remains that I've come to find this analogy quite useful to bring to the fore what I think is happening right now. So I think most of legal tech products are treating lawyers as gymnasts. And what I mean by that is when you, like a gymnast, they go to the Olympics, they go onto the mat, and they perform their routine from beginning to end. And they're done. And this is what most of these tools are doing, right? I would give you this thing, and you're going to be faster and better and all that. Now I think it's quite misguided because at the very least, I would argue that at least in a decently sized law firm, the whole exercise is more of a relay. You're passing the baton onto the next person who's not going to have to do something with it. But if we're actually honest with each other, at least that's my experience, practicing the whole exercise is really much more like a basketball game, where the ball is in constant motion and you're going to have to move on the court in different places. You're going to have to observe what other people are doing. And you're going to have to incorporate that feedback. And the way where I would draw at the analogy is if you're building tools to enable Olympic gold medalist gymnasts, and you put them on the basketball court with a regional team that they're probably going to wipe the floor with you, right? So you can think of the practice of law in two ways. It's a team sport, like basketball, or the individual effort. And in many ways, it's both. In popular culture, sometimes we see a lot of depictions of a hero's soul or a single lawyer, where through their sheer brilliance, they bring a case from beginning to end. But like you're saying in many cases, before corporate work or complex transactions, or any kind of complex, a lawsuit, litigation, is very much a team effort. And it's not a single brilliant mind that takes it from beginning to end. And we're going to touch on this later when we talk about unfripping. But that's kind of what you see where I see that difference, right? Of course, their tool is empowering maybe also individual lawyers. But it's fundamentally structurally different because it's really trying to reinvent how the workflow goes and how we're delivering this piece of legal analysis that we have to do. And unless you do that, you just give people more tools. That's not-- that feels more like 2024. Yeah. It's like the old way, oh, you have a task. We're going to build software to make your task more efficient rather than let's look at the practice of law and the production of legal work and think about almost from first principles. What does it mean to produce legal work? Now, the question that I still have about what they announced is they said they were going to move much more to value billing. And as I said, it's one of my hobby horses. I don't quite understand what that value is in all cases. In all cases, they do a lot of private equity work. So maybe that's quite easy to put a value on that. But that's going to be interesting to see how they structure that and how they attribute value and what the client receptionist, right? Because I am not sure-- and I think we touched on that last time in our conversation as well. Like, where is the margin shift going to accrue? Is this going to accrue on the firm level or on the client side? And there's no answer to that yet. But of course, I think they're moving in the right direction. And I suspect that other firms will come forward with similar strategies. And yeah, I'm sure everybody's thinking about it. Yeah, I think the old competition or envy in the legal world, especially for private equity lawyers and M&A lawyers is they always want to be paid bankers. Like bankers? They are now, right? They will be able to pay us as much if not more than bankers now, but in terms of the fee structure, bankers get paid a percentage of the transaction, whereas lawyers get paid only a fixed salary or a fixed hourly rate. Of course, it can be a very high hourly rate, but it's still fixed. So if you do theoretically, if you work on a billion dollar transaction versus a hundred billion dollar transaction and they're of the same legal complexity, a lawyer is going to make the same amount of money, but a banker is going to make a lot more money on the hundred million dollar transaction. I think Wachtel and Kravaff are the only firms that somehow managed to break through that, right? Yeah, Wachtel has been able to have some, they do a lot of contingency, and even on their, I'm not actually exactly sure on their corporate side where they're able to take a percentage or not, but I wonder if that's also like the move that my co-clinist is moving towards, because they do so much private, could you work? Their clients get paid a percentage of assets and a percentage of returns, and they probably are thinking, well, wouldn't it be swell if we could get paid at least a percentage of assets on deals that we advise, and then it makes a lot of sense even from business perspective because, okay, once you move to that kind of, once you decouple the hourly rate, then you do have a lot of incentive to be a lot more efficient and to increase your margins because then you can, it becomes a lot more, a mercenary, but yeah, then you can kind of start poaching the very best deal makers, the people that have the best networks, the Raymakers from all of their place to get on the biggest and the best deals and to capture more market share. The one thing that I do wonder about is, this is now like a new evolution of the buy versus build kind of paradigm, right? And, or somewhere in between, and then there's a big posture on build, even though they also said they're going to continue to invest in external tools. Now I wonder what part of that is also maybe a bit, not misguided, but we're all empowered with cloud code, like everybody can code, we can all do all these things now, right? And I wonder, it's going to be interesting to see where that comes out and whether it can hire the right people, and I think they're dedicating some like 250 people in total, I think 100 engineers, 150 lawyers, something like that. I read, and that's a big team, right? Can they bring the traction onto this, onto the street, onto the pavement and working through this and, yeah, and what are the internal hurdles that the organization is still and comfort with, right? And what's the management posture, and is the signal strong enough, are people actually on board with this, or will they have to kind of sunset some of the older partners, or just like get away from me, I don't want to care? It'll be interesting to see. But yeah, I was quite excited to hear that because there is now a firm that really puts the money where their mouth is, and trying to, yeah, just do something differently, at least, yeah, as far as they've properly announced. Yeah, that's a good thing. Yeah, I love that point where the Bivers is built because Claude in all of these AIRT was definitely amp up that Dunning Kruger effect, where it's, oh, I made a neat app, I must be it like, these engineers have no idea what they're talking about. This isn't that hard. Like, anyway, we can build an app where real software development is very difficult. And the billion app is just, and we've talked about this so many times, building an app is just the first step. See buggy and making it work at scale, making it work in all your edge cases, actually rolling it out, security, reliability. That is 95% of the work, not just the 5% I built it. Even all the work before building it, no one talks about it. Okay, thinking through your use cases, architecturally correctly, making sure your feature set is achievable, making sure you don't like, have feature blow to all of the place, making sure your UI makes sense, and it's usable, and the workflows and the data, and all these other things, and across a variety of workflows. It's one thing to build something for yourself that's, oh, this is a neat thing that I built for myself, and also to take an existing piece of software and to copy it. That's also, yeah, a clock can do that. You give it some screenshots of Logoro or whatever, and you say, you can do that really well. But software development, actually making a new product, not just a copy of an existing thing, is very difficult. And that's like the crux of the buy versus build debate, because sure, you can probably build a copy yourself, but you have to maintain it, you have to iterate often, you have to do all these other things, and you can spend, of course, for more money building it yourself, then just buying it from a vendor that can take that product, resell a thousand times. Yeah, for $20 or so much. Exactly. For $20 or so much. And so how much, unless you really have a lot of proprietary data and a lot of competitive advantage in building yourself, there is not, what is the, what is the end game for Kirkland, right? What is the end game for? If you are a law firm and you do set aside, you say two or three percent of your revenue is to build an in-house solution, one, you're not a software company. Fundamentally, you may think it's easy and you may think, oh, I can just hire a bunch of engineers, but culturally, you are a services company. And so the entire way you run your business is completely different from the way software companies run businesses. Yeah, I think that's a very good point because lawyers, especially the more senior get, they tend to have this understanding of the world that they enter the room and then everybody sits down and listens to the wisdom that sort of, that they impart to everyone, right? And it's, I found it surprisingly hard for them sometimes to part with that because, yeah, all your career, it's been like that and you've accumulated more and more understanding and you're really expert at this one thing to come into the room and now have a, and then let's also think, well, what the conversations and law firms are like, right? Like the partners are in the end, the decision makers, if not in every individual case, but directionally, they are the ones who say, this is how this goes. And now you transplant them into a product meeting, right? Like where you have an engineering head and a product head, who will bring a very different perspective to the table and to take that step back and say, okay, well, I'm now here as a lawyer and of course as a customer, but I need to really openly share how these things work without prescribing the solution and allow these people to do their job. I found that lawyers are not necessarily the most skilled at that. And because they don't need to be, it's not part of their profession. Exactly. And when you add a law firm, there is that divide between OMA fee earner and fee earners and fee earners and fee burners. Yeah, exactly. And this is true in every organization. And I think, I think they're going to struggle for a right reasons, right? Not just an AI part, but just on those, like I think one AI gives you confidence to build software that shouldn't exist. And if you think about it, if you look at it from a pure, okay, we're going to build a software product perspective, you can look at like just history. Like finances are a really good example where you have these giant banks and some banks are really good at technology. Some banks are really bad at technology. And you would think, oh, this is something that should be really important to them because just technology is so critical to the delivery of finance. So finance is not even talking about AI, but just like pure, normal software, building databases, moving data around, delivering data, ensuring it's compiled properly, like calculating risk, calculating exposure, all of this stuff. These are fundamentally not difficult things, right? It's not a hard thing to build like a database. It's not a hard thing to collect. We've known how to do these things for 50 years and yet when you go into big banks, it's a constant source of, oh god, why is our internal technology so bad? We try to build this thing and it's horrific. And it's the exception. Very few firms have built really great technology. And even really great technology is, it exists in a kind of like a financialization of that technology, right? BlackRock Aladdin, as I know, for being like amazing. But if you compare BlackRock Aladdin to amazing consumer SaaS software, it's like a piece of crap, right? It's a why is it user-in-face horrific? Why is it so clunky? Why is it so slow? And if you show some of that versus like a great piece of consumer software, I don't know like, even HubSpot, right? Like a really good piece of small business software. HubSpot, oh, this is smooth. User-in-faces amazing. It's so fast. It's so, you can tell it only thought about user-in-faces a lot. It's with banks that spend 20% of their revenue and technology. I was always blown away at Morris Stanley when I talked to tech people and they were like, "Well, yeah, no, this piece of technology, we just don't touch that." Some people in the firm who still know how, "If anybody breaks this, we're in deep trouble." Or nobody's touching that. Yeah. Yeah. Are we now, have we now talked about it long enough to come to the conclusion that they're going to struggle to solve the literature of shit show? I don't think so, to be honest. I think there's smart people, in the end, I think it will really all turn on how are the people that are building this, actually empowered by management and what's the general of the firm? Is this a flashy announcement? Or do they actually mean this? Have they actually internalized the need for this project? If they have, then I think they have a good shot. If they spend the money on the right things, and surely with that money, they can hire some smart engineers. Can you say you're saying that about banks? Right. I think we'll see. But it's one is a really interesting development and Kirkland will be kind of the poster chart for it. If they succeed, then other people will try to jump in the bank. They can either flame out horrifically, then it will also send a signal and say, "Oh, this was a, this was not a good idea." What were we thinking? should have stuck to our bottom. Yeah, we were there. We're not. We'll see, right? And maybe just to pass this on quickly. So I've recently seen people talk about some of the bigger law firms actually subscribing to the likes of Harvey and the Goro on extended pilots. So some of these 12, 18 month periods might be coming up for renewal. And of course, we don't know what's going on inside. But if some of these firms are walking away from these tools, or if they're substantially downgrading, downsizing the number of subscriptions because they've learned a lot, and they've now-- they're now taking this to build their own stuff. That would be interesting. Yeah, and it goes to-- also there's a race, right? It's a race between the law firms, either they go and do their own thing. Or on the other end, it's also a race on the Codex and Cloud side, where eventually there will be a day, probably, where a cloud will become fairly feature parity with Harvey. Or maybe not 100% feature parity, you would think, as long as they revenue. They're in-- they're in their tool, but maybe 70%. Good enough for 70% of the lawyers out there. Yeah, and surely Harvey does continue or would continue to have this advantage that they can orchestrate different models. And of course, all the models have slightly different postures. And they're a little better at this and a little bit worse at that. My friend Anna shot out to her. She just released her new benchmark on these frontier models. And yeah, when you look at it, and this actually really fits my personal subjective experience-- yeah, just different like Gemini and Opus and GPT-5. They're just good at different things. So to be able to combine those and play everyone out for their strengths, that I think is an advantage. But how long is that going to last? And if everybody kind of catches up to where it needs to be, then maybe orchestration in that way is no longer a real differentiator. So we'll see. I think that's a good segue into, on that topic, like the cost of inference. Because one thing about, of course, basing your product line off of Clawed and off Codex and Gemini is the cost of inference is huge. So there's this step that's thrown around that the $100 Clawed Pro or Clawed Max, whatever it's called, the subscription actually costs Clawed something like $4,000 or $5,000 in underlying compute. And no one is going to pay $5,000 a month for what they-- Well, I'm not saying that. So some people-- But the vast majority of their $100 are like, we wouldn't. We wouldn't pay $5,000 for what we get $100 with Clawed Pro. And I think that's a really good point. Where maybe the orchestration, the tool creators, that they are not bound to any single underlying model, to have an advantage, is if, at some point, you're the cost of what the product you're delivering and the cost that it cost you to deliver that product have to equalize. Or as you just got a business-- Not a business, yeah. You're not a business anymore. And so if you are a Clawed legal-- and it's only you say, oh, this is going to cost you $4,000 a month-- and Harvey says, oh, but we've switched to-- QN. QN, whatever on the back end, and QN6 has gone good enough, where it's not bleeding edge, but you can still pay us $500 a month for this. I think there is space there to say, oh, Clawed is not going to start suddenly switching to you QN. But it just won't be-- the economics just don't work out anymore. Yeah. I-- that's-- that's-- that's actually, I think, that's right, with the cost, though. And there's also now been rumors about-- where some of the companies have come out and Uber, I think, and Microsoft even have come out. And so my god, this is crazy. Like, we've already burnt through our whole budget for 2026 by the end of April or May or whatever. And so now we have to figure out what are we going to do. And some people have interpreted it OK. And nothing's working. There's all nonsense. There's a complete hype. I don't think it's quite like that, because you also have to see where the-- what these organizations have told people. Of course, when management comes in and says, we're now going to have a leader board that says, who's burning the most tokens and only those people who burn the most tokens are deemed efficient engineers. Well, yeah, engineers are going to find ways to burn tokens. And I've heard these stories of people basically-- yeah, just inventing things so that the machine keeps running. So they'll go home and it just-- no, let's just keep burning tokens, right, just so that we'll make it onto the leader part. I mean, yeah, OK, that's stupid. So if that's-- then there is no way you're going to recover your cost. And that's-- >> Done mesentos and you get done mesentos. >> Exactly, exactly. But do it efficiently, right? And switch in what you-- but even the frontier model makers, and you don't have to run an opus 4.8 on everything, right? There are certain tasks that Sonnet will do just fine. And that's going to be much cheaper, right? And yeah, we also need to talk about Anthropics Legal product that they released. I'd be interesting to see where they're taking that and whether that's going to be part of the measures to make that a bit more effective. At the same time, other than orchestration, I still feel that in the end-- and I'm not sure whether this is right, but my intuition has always been that at the end of the-- like today, when neither the frontier labs nor anybody else is making money of a token, because you need to burn too many in order to produce something useful. You can see that curve moving to a point where compute is cheap enough that the token-- you'll spend less-- and it's going to become cost effective. At that point, of course, the frontier lab-- nobody can give you tokens at a cheaper price than the frontier lab, because they're always going to charge a customer who they give the token to, something else, right? Today, they're not really incentivized to do it themselves, because they can charge someone and they can make money. But if they can actually provide the token and do the service, and they're in the best position to do it, right? And yeah, that's a slightly different debate. I think it's interesting that both OpenAI and Anthropic have announced these service companies that they're looking to buy, right? I don't know, like $10 billion fund each to acquire services, and whether that is to acquire knowledge or to do the business. It's going to be interesting to see. But should we talk about legal toolkit? Yes. What's the call? Yes. My understanding is that these tool kits are essentially just collection of skills, right? Skills and context. So what you're doing is you have specialized skill sets that they've built for certain NDAs, for example, for certain specific kinds of products, so that the-- when you're using this toolkit, and they've built some nice collaboration on top of that, so you can share projects and do other things across your team, it's essentially-- not to denigrate what it is, but it's essentially taking their already really robust orchestration stack and adding more legal skills on top of it and adding more specialized-- essentially playbooks, right? Here's a playbook for how to do an NDA. Here's a playbook for how you do a-- I don't know, like a will. Or here's a playbook for how you do it. I know the contract. Yeah, I think at the core that is right, I was quite surprised to hear that they're them creating that was not like some product manager sitting and saying, oh, let's create something legal and what should this look like? Let's do a bunch of interviews. But as far as I understand, they've actually created this from usage. Yeah, from their internal legal team. And that's interesting, right? I wonder whether it's purely internal or whether it's also external. I'm not sure. I don't think they talked about that. But it'd be interesting to know. And of course, when you project that forward, like they'll-- if they were to use what their customers are using it on, and they're using that as ground truth, then no vendor other than themselves can ever outcompete the quality and the direction of what is being built. Because of course, they sit at the source, and they see all this. The vendor always only sees a small part of that universe. So that's the first thing I find quite interesting. And the other thing I also find interesting is-- especially when you go Harvey, LaGora, those closed systems. And there are some other vendors who are open-source about their-- what works inside the black box. But they open-sourced, essentially, all of the-- or I made public, all of their skills, the prompts inside it all of it. So there's a little bit of a clip here I feel towards these vendors who are very protective of those, saying, well, we don't really care. Here it is, and go knock yourselves out. Also, my understanding is that you go through an interview when you start using it. So there seems to be some structured, onboarding interview to lay some of the groundwork around it, to properly basically set up the harness and give it direction, which I find quite interesting as a way. And some while ago started using a harness myself, where that's exactly the process. You go through a pretty extensive interview, a few rounds to set up the things you care about, what are you working on, what's interesting, what kind of books are you reading, those kinds of things. And that, of course, makes the whole thing so much more powerful. because you don't have to go explain again again what matters and what really doesn't matter. - Yeah, it's queued up the most relevant skill files to the top of the pack. And that is, yeah, the distillation of legal work into, I think at law firms, I think they're very protective of their playbooks because that's, they're very proprietary IP. But if you come down to it and you say, I'm sure Anthropic has a very sophisticated legal team with a lot of very talented lawyers. And if they're sitting there saying, "Oh, I can just distill my last job I did a thousand of these service provider arguments." Of course, I know all the in and out of this. And so I can sit there and work with Claude. We can build out a pretty good playbook probably in a week, maybe in less. And Claude makes money selling you tokens. Well, eventually they will make money selling you tokens. And however they can get you to use more tokens, they will. And they can give away this other stuff. - Sure. - Because they don't need it to make money as long as they've kind of hooked you into the Claude ecosystem. And I think that's where, that's where I think there's a couple points on pack. One, I think the data point is once they're not supposed to be able to see what you're doing, right? In your agreements, when you sign up for the business solutions, it's supposed to be very clear, oh, we do not retain. We do not train. And one of the things about Claude code is, yeah, it actually all of the files live on your local computer. If you install it on a new computer, you kind of have to move. You have to sync that data yourself over. Which is a little bit of comfort. It's not actually being stored on Claude's Anthropics server somewhere. But the second point of that on, oh, there are the best positions to deliver versus a vendor. It's, and why are they buying services companies? It's the ownership of the client relationship. And that is, is in the end, kind of the most important thing. Cool owns that client relationship. Is it, and how stick is that client relationship? Is it the service company? Is it the tooling company? Is it the token provider? Is it, who is it in that supply chain? That owns the client relationship. And it's important because whoever owns the client relationship is best positioned to extract the maximum value from that relationship. Maybe I see this a little bit differently, only to the extent that you, of course, by acquiring these companies and sending your forward deployment engineers, whatever it is, and then really extracting kind of the essence of what was being done there. That is all the kinds of knowledge that a company like Anthropic doesn't have today, right? Because they're not a services company. They're a computer scientist, they're mathematicians, and they're those kinds of people. So they don't really know what it is to provide a legal service or to provide some other professional service, right? So to bring that knowledge in house, of course, build better. I don't really want to use the word product, but position the year offering towards that market, but also being able to potentially just do some of that yourself. That'd be a logical conclusion or a biological direction for me. Yeah, and they've partnered with, not on the legal side, but on the finance side, they've partnered with, announced, partnerships with, I think, Morgan and some of the big banks to do kind of an Anthropic for finance. And the banks will gladly do it because they want to be underwriters in their IPO's and they want to manage all of these soon to be desi and sent to me millionaires, assets later on and sell them finance and other things. But it is not like bringing it back full circle, it is not surprising to see how someone like a Kirkland would bulk at that kind of arrangement and say, oh no, we are not going to partner with you. We're not going to have you for deploy 150 engineers to work alongside our lawyers. And extract everything we know. Yeah, because was to stop you from just one day saying, oh, okay, we can provide you Kirkland, quality service at a white label price. Fresh real students seem to have those concerns, right, because they sign exactly that we don't know what the agreement says, but that seems to be the understanding and the agreement, right? I have one more thing that I, that's just my close to my heart about that Anthropic release. And because they also integrate with a whole bunch of existing software, right? So I manage, I can access your files. And all of that makes sense. That's pretty generic. Why wouldn't you? Yeah, just all of these companies have created MCPs. So why not have it there, right? The one thing that stuck out to me and we're doing took well, it's this automated time sheets and workflow discovery. The one thing that stuck out to me is there's not a single one of these existing time sheet companies on there. And I really had to smile a little bit because, yeah, like on the one hand, it doesn't surprise me because, well, these current time sheets are useless. There's no information in there that could be, you could turn into value really. Yeah, one fourth of an hour, replying to emails. Yeah, exactly. But of course, nobody's really seems to have caught on yet. That could be the greatest source of all these details that you'd want to be looking for. I know what I try to understand, what your people do if you did it better. So come to me if you have questions about that. I think it's building the narrative underneath those is capturing the to dig down deeper into that. It's not just that the time itself, like when you look at traditional time sheet, it is kind of meaningless because the entire point of it, the entire point of it traditionally was just to deliver an invoice to a to a customer. And with just enough information where that customer is just, what is this? I don't want to pay for this. But if you'd capture the full context of what they were doing, and instead of a single sentence on, or even have us on what I did, drafting, but you could capture pages and pages of a context. Okay, they did this workflow, and then they went over here and they did this. And you could distill that data into a playbook. That's immensely, immensely valuable. And you do that, and there's multiple ways to do that. You do that by acquiring a service company and just sitting there and just watching them, and what they do, or a smarter way to do it is through software, right? And you have software that is both capturing your time, and also capturing the context of what you're actually doing. And then that data is incredibly valuable. Yeah, well, that's certainly what we think. And the last thing on the anthropic thing, what I wanted to get you to take, what do you think? I'm confused by their enthusiasm to provide some kind of MCP for Harvey, because so now I have a customer, or I have a user here, he's paying $1,000 a month for this Harvey thing, and of course it comes with a nice interface and all those things, so that's all great. But so now I'm opening that up to use whatever's inside there, and take that over to co-work. What? Yeah, I don't understand. I don't understand why anthropic would be more than happy to do it. I do not understand why Harvey would build an MCP and say, "Oh, hey, yeah, we are more than happy to connect with Claude, which is kind of an existential threat to us. I don't understand the reasoning behind that." So we don't use Harvey, of course. We do use Claude, and I don't. Like, would we let, just think of from our product, would we build an MCP for Claude to connect into, like, R? No, not Tardis. No, no, no. Right? Definitely not. Yet you want our clients and our workflows and our data, you either buy us, or you know, or you come out with us in a different way, but we're not opening up the kids for you in a short way. But I enlarge, I think, all of this is a very smart move on anthropics part. But if you. Actually, to take it one level up, if you think about the people who are most involved in lobbying, there's such a disproportionate number of lawyers to bring them on board and to make them comfortable and to show them how amazing all of this technology is. I find that a really compelling thought that you would go the extra mile to make that particular constituency particularly happy so that, yeah, they'll talk on your behalf when it's time and necessary. I mean, it's brilliant from anthropics perspective. Like, everything they're doing, they just kind of keep hitting the marks, right? There were 4.8 just came out a month after 4.7, and comes out a month after 4.5, at 6.6. And it's just. The pace of the releases is incredible. The features that they build in are incredible. I think the partnerships with Microsoft are really astute because it's like. I never think. I never thought. It doesn't make sense for anthropic code their own competitor to word or Excel or other things. But by building that plugin that just sits right inside PowerPoint and sits right inside Excel and sits in Word, that's amazing. Because then you're using it every day and it's great at it. Honestly, it's like all this great at Excel. It's pretty good at PowerPoint. And by doing that, it's two things, right? Microsoft games a lot from that too because. They head off all of these PowerPoint and a box on competitors, where I'll go to your AI startup and I'll describe what I want, and I get a nice PowerPoint presentation of it. Stay inside PowerPoint. You can use Clawed and you can get an equivalent. You're still in PowerPoint, which you are familiar with. >> I'm in with Excel and Word. That partnership is super, super astute on-- >> I think it's a real hard point. >> It's a really good point because it also, like all vendors should take note. If you're building some other interface, and another tool to throw it into the mix, no, like anthropic is showing you that, "No, no, no, everybody calm down. We're going to keep people exactly where they are, where they're comfortable, where they're spent the last 15 years. Nothing changes in that regard, and that's a very strong-- >> Yeah, and it can't believe how badly OpenA had dropped the ball in this. And you can say, "Okay, Microsoft dropped the ball too," but yeah, Microsoft tried to build it with Copilot and Copilot socks, and I'm going to hit it right. So I think it's also pretty smart on Microsoft's part, and this is definitely the new Microsoft rather than the old Microsoft, would have been like, nobody's building these plugins. Now they're like, "Are crappy plug-in or the road?" And I think it's a part of the new Microsoft, where it's like, "We are more than happy to have an ecosystem of things that plug into our stuff." We are not-- it turns out, we're not great at everything, and some of our products are truly terrible, but people love using Clonson. We're like, "Use Clon and our thing." But it's like, "Hey, you could have built this ages ago. You're already like-- they're already your closest partner. They already own a huge chunk of you." Why didn't you guys build Codex N Word first? They asked myself that too, and I'm not sure kind of what happened first, but it really seems that Anthropic was just so much better at rather than trying to anticipate what people would want and sitting around a big table and saying, "Ah, people would probably want images." Right? That's too much of an image and video. Yeah, we're going to part with Disney on Sora and all this other stuff. They didn't do any of that, right? They just seemed to have listened to what people are actually doing. Yeah. Yeah. Maybe that comes from, like, Altman is a-- he's like a B2C kind of guy, and when you sell to-- and OpenAI was a very B2C company, and when you sell to consumers, you could have to try to predict where they're going to want, because consumers are very fickle, and you kind of have to be the cool kid on the street, or the hot new thing to get that kind of consumer demand. Whereas Anthropic went business first, and they just, when you sell to businesses, it's not as-- it's harder than selling to consumers because businesses are picky buyers, but it's also-- Yeah, but in some ways, those are easier to be. You kind of just-- in some ways, you just need to listen to what they need, and you come to them and say, "Oh, I can fix that for you." Had a hopefully cheaper, better price than what you're doing. And so you build for the problem that exists today, rather than for consumers, you have this fickle market, and it's just like, you had to market to them, you had to make yourself cool, you had to be at the bleeding edge, I need to-- when you make this cool, yeah, Sora thing, and they have abandoned it now, but because, oh, this will get like eyeballs and clicks, and people-- this will go viral, and people will be like, "This is such a cool thing," but doesn't actually driver strategy at all? Does it drive like your unit? What is your goal here as a company? Yeah. But that's just some ducat. Well, man, so much is all about it. We could go way deeper into the China open source model angle. Maybe we're just a little bit, and we can pick this up at another one, but what I've been thinking about, and we touched on a little bit earlier, is this race towards ever increasing compute and inference costs, and the runaway, and you see it, you see it, a part of it may be due to poor incentives, but it is true that the newer and newer models use a lot more inference than the old ones, because thinking uses-- or reasoning used an order of magnitude, like three times more tokens, four times more tokens, just pure chat, right? And then agentic workflows, when you use skills and other things, passing packet 4, those use another order of magnitude, more tokens, than just pure chat. And this is almost like a fundamental thing, but these frontier labs open AI, and then, the topic in particular, I think not so much Google, because they're more grounded in commercial realities. They're kind of racing towards this AGI goal. And I think the models are already super capable, but in my mind, if they fall short of that AGI goal, or AGI and ASI are kind of the same thing, because it's like, oh, once we reach AGI, then it's going to invent ASI by itself. But if they never reach that goal, I wonder then if all of these open source providers come in and just say, "Oh, hey, we are 80%, 90% as good at 10% of the cost." And then the economic realities are too, like they have truly a dollar in data center commitments. They have, on the liability side, these hyperscarers are so stacked up that if the revenue falls short, they're out of business. Like literally, they have too much liabilities, stacked against their revenue, and they kind of kill themselves, and then you have all of these, I don't know how these Chinese open source makers are going to make money off of, because anyone can just download the weights and run them on their source. But I wonder if that is the more likely outcome, where if you bet on a world where actually ASI is not going to have, they despite the Silicon Valley crowd, where it's like, this is the last five years we'll have to make money, and economics are going to end, and like the world is not going to make sense anymore, but if the world does tune to turn along, they have the same kind of trajectory. And we just have AI, but better. It's going to be really good at certain things, and maybe not so good at other things, but it's not going to be a godlike intelligence that reinvents economics as a whole. Yeah, I wonder what the, so the frontier models, the US frontier models seem to be four, five, six months ahead of whatever comes out of China. So there's a capability differential, right? And the question is, what is that really, and then how useful is that? Because if that is really useful, and you can just do so many more cool things, but not only cool things, but useful things with it, and economically valuable things with it, then of course, it would make sense to continue to invest or pay for that kind of difference, right? Because you're ahead, but you could also see, and I'm also ready today, or so capable. So let's project the two years forward. So whatever, when 4.5 at that point will be more capable than Opus 4.5 today, right? So nobody could convince me to say that's a useless piece of technology that nobody wants, right? Of course, I don't know what Opus 6 will be able to do. I think if there is economic value to be extracted from running the most frontier model, that will continue to work. And I kind of grow into, is that the story? Like with OpenAI, IPO, and Anthropics IPO, that's the entire story, right? Well, and it's not, I guess it's unobloid. I think another, yeah, it is, and another thing also will really turn on just geopolitics, right? Because as long as all the US companies will basically have to say, I'd probably sign some document to some customer, we're not using Chinese models. They are kind of beholden to what is there, and there is a lot of people in the US, in the West, more largely who say, we really need to get our act together as the West to invest in these open source models. Because if we don't, then we're all going to be outcompeted by Chinese companies who are using, who have the ability to use this open source stuff for cheap, where everybody here is paying thousands of dollars per user to do approximately the same thing, right? And that's not a, it's probably not the outcome that, that, yeah, the rest of the world wants for themselves, right? I think that's exactly right. I think, I think the deeper issue is, they should be terrified of China, but for the wrong, but for the reasons that they're not terrified of China today, today they're like, oh, if we lose the AI race, China is going to get to AGI, and ASI, and it should be the opposite. It's, no, think about what happens if you never reach AGI. If you never can really truly invent vast, like, new ways of commerce that justify your spend, if AI is just good enough to augment the vast majority of commerce, which I think is actually already is today, then you guys are screwed because, oh, it's expensive. Yeah, it's like the team who is, is Asian, right? It's, oh, we got, we got T-Moo Opus 4.8, and it's one-tenth of price, and sure, it's not as fancy, it's not going to be the hacking of the NSA, but it's good enough for-- Most people don't do that on that. Yeah, yeah, most people aren't trying to hack into the NSA on a daily basis, and that's good enough for 99% of commerce. Yeah, I wonder whether there's a bit of a Pascal's wager going on, right? Like where, okay, well, if God exists, then it probably makes sense to abide by the rules and try to invent him. Then the infinite loss, yeah, of going to hell, but maybe arguably there's the infinite loss here, so. Right, nice to have any endorsements for this week. I do, I do. So I read, I'm a big fan of Sam Chris. He's, I think he's one of the finest SASs of our time. He's got a great substack called Numb at the Lodge, and he had his recent piece called something like, if you use AI to write, I'm going to find you and kill you. And it was just so perfect. And because you see it everywhere now, you see this AI application of like almost like all written language like everybody. The LinkedIn especially, it's really bad. But he was saying he's pointing out, "There's articles in the Guardian." It's all over the place now, where you see, it's just people just being lazy and using it to write for them. And that's that cadence, is that there's a thing now where you can kind of tell. And it's just, it was just so well written. - It's not this hot, it is that. - And it's like way worse, like the flowery language. And there's just this style. And I think it basically really won't. It's not that it's poorly written. It was known as being the reason why all the AIs were to sound like that because it was good writing in the past. - That's what the M-dash. - Yeah, well we consider it to be, quote unquote, good writing. And it's just that we hate it now because we see it everywhere. And a big part of it is actually, write poorly if it's your voice. If it's your voice, write poorly and people will like it more. - I think it depends a little bit on what the subject is to me honest, because if it's the, I don't, like a financial report, I don't care if it's written by I, I just, damn it. - Oh yeah, yeah. - Well, actually, for legal service, like a contract, no one cares if it's your voice, right? We want that mutual voice. - You don't want the voice. We want this mutual legalistic thing. So I think as well, so like back to our podcast, like Law is perfect for AI, because no one gives a shit about your voice, right? We just want the legalese, but for many other areas that distinct voice, wherever you can get is really valuable. And it is something that almost by definition, AI can't replace. - That's it? I have to, I read this really cool article in Harper's Bazaar about Esperanto. So Esperanto is this 150 year old universal fake voice. - It's just cool to see how that survived all this time and then they were really caught on. Yeah, go look it up if it's interesting. And then there was a cool episode on 99 PI, 99% invisible is a podcast, which I really love. And they talked about the standardization of the world, like around the Second World War, how, yeah, every screw was different and I didn't know this, but basically the allies during a Second World War, the Americans, like they could send missiles to, or bombs to Europe to their allies, but they wouldn't, like American bombs wouldn't fit on British planes, because the whole, - The clams and everything, yeah. - Yeah, different, right? And they didn't have MCP back then. - They did not have MCP. And yeah, this is a really cool episode. We'll link it in the show notes. - Yeah, go listen to it. It's really awesome, super interesting. - Well, okay, that's all the time we have today. Thank you so much and it was a pleasure as always. Good to see you, my friend. Nice, good to see you too as well. (upbeat music)