Episode Transcript
[00:00:00] Speaker A: Welcome to the Relay, the legal show for personal injury law firm owners presented by Lexamica, the number one attorney referral network. I'm your host, Gabriel Steeritz. Joining me today is Matthew Kirbis, founder of the subscription attorney and legendary podcast host. Matthew is a champion for new billing models for attorneys and according to LinkedIn, is now AI enhanced. I'm going to ask you about that, Matthew. Today we're going to wrap up 2024 and look into 2025 together. Matthew, welcome to the show.
[00:00:30] Speaker B: Well, Gabriel, than for returning the favor and having me on your podcast because you gave some great insight to the listeners of the Lost subscribe podcast and happy to be here.
[00:00:40] Speaker A: Thanks so much for coming on. So before anything else, your LinkedIn does say that you're AI enhanced. Does that mean that I'm talking to a real person anymore or is this just your avatar?
[00:00:48] Speaker B: Beep boop beep boop. I am a robot. So I've actually toyed around a lot with like what to call myself in that regard because I've been using AI since before it was cool when I started my practice in March of 2022, even though I've been practicing law around 11 years now, that's when I started my own solo practice. I was using machine learning based AI logic if this then that sort of based automation and AI back then. And so I happened to be there day one when Gen AI hit the scene in November 2022. I should say like useful Gen AI because I was playing around with GPT3 and it was just like not useful.
[00:01:23] Speaker A: And so I just for clarification because look, I don't know if our listeners are like that much of a nerd, Matthew. I am, but G 3.5 was the first release of ChatGPT to the wilds. Like the stuff that everyone that blew up was 3.5 and 3.0 is pretty early. That was not something that most people were aware of. So yeah, that, that's pretty early days.
[00:01:42] Speaker B: Yeah, yeah. And even like spellbook was integrating GPT3 and I remember playing with Spellbook at the time and being like this is a fun trick, but it's like not like useful. And now they've, they've come a long way as a company because the, the backend tech has gotten so much better and even now we as consumers or as practitioners can play with that backend tech directly from those companies, which is unusual and very cool. And so I was calling myself an AI expert there for a while, but then I realized I was learning from the Experts like Ethan Moloch, Daza Greenwood, and like other like, and like Damien Real, like true experts in the field of AI. And I was just more an expert user than like an AI expert or gen AI expert. So I played around with what I call it, and so I just, I landed on AI Enhanced. And I like it because it's succinct. Other than AI expert user, like, doesn't exactly roll off the tongue. It's a hyphenated wor. AI Enhanced. So like, it's like kind of like one word. So, yeah, I'm, I'm very much leveraging AI as much as I can, and all of my work product, and not a work product is AI enhanced.
[00:02:43] Speaker A: That. That's awesome. Yeah, I think we should talk a lot more about that. And I. Look, I respect that you're not calling yourself a guru, an expert. Everyone out there is doing that. That's a level of credibility that a lot of people don't have in the space. But yeah, look, we're, we're talking about the 2024. What changed? What were the big things? I think the goal with the show is to give actionable, practical takeaways to law firm owners in the plaintiff side space. I think the best way to do that is talk about what you've, what you've started using this year. Like, what tools have you found that have added the most leverage to your practice?
[00:03:16] Speaker B: Yeah. And look, even though I'm not like a plaintiff's personal injury attorney, I'm the subscription attorney and I'm not billing by the hours, so there actually happens to be a lot of overlap between contingency fee based lawyers and fixed fee or subscription based fee lawyers. Because we. It's all about efficiency.
[00:03:32] Speaker A: That's right.
[00:03:33] Speaker B: Faster we get something done, the more valuable it is for our clients, and the more money we make as a law firm, the less time something takes. So there's a lot of synergy there. So every PI attorney should be thinking about how could I leverage AI to have a more efficient practice. Now, I actually have given multiple cles on this and they're out there. It's on Lawline. One version's on my podcast. If you actually want cle, look for the Lawline ones. If you just want to listen to the content, you check out my podcast. But what it really comes down to is there's, there's a handful of tools that you should be leveraging. And I think every single attorney, period, could benefit from one of these legal AI assistants. And there's all different kinds of legal AI assistants out there. The one that I've been playing around with and using the most and that I use more than any other legal AI tool. Legal AI tool. I could give an hour long presentation on what makes AI legal specific is Paxton. What's unique about Paxton compared to some of these other tools is they have their own proprietary large language model that's legal specific. So if you think about AI in terms of being useful for lawyers, the simplest way I could put it, and Gabriel, I could give you a, like a little graphic I made that you're welcome to use with this and publish with this episode is there's legal, there's industry specific AI and general use AI, and then there's like even GPT wrappers where they wrap around the general use to try to be more specific. Then it's either using retrieval, augmented generation or rag, or it's not. If you have even a legal specific AI that's not using RAG or a general AI that's not using rag, do not use it for anything for substantive legal work, period. You need it to be leveraging some form of retrieval augmented generation. And if it's not legal specific, it might even not still be that good. So like.
[00:05:10] Speaker A: So let's break, let's break that down a little bit. And I appreciate that because I actually did a, a webinar for a group of lawyers very similar. I have a hierarchy of. There's. There's like specific legal LLMs, then there's general LLMs with legal guardrails on them that have been essentially like you've wrapped them where you put guardrails. And then there's just general LLMs like Claude, OpenAI chat, GPT, whatever. Right. So.
[00:05:32] Speaker B: Right.
[00:05:32] Speaker A: Very much agree with that. RAG is not something that I specifically talked about. I think that's a good call out. And then. And so just for someone who's maybe not familiar with the term rag is when instead of the output directly being the text characters from an LLM, you're essentially pointing to another source and you're using an LLM to point to that source for. Is that, is that basically. That's basically the idea. Right. So instead of the LLM being the source of truth, the source of truth is some other document. And an LLM is, is essential being tied to the mast of ground truth rather than being able to just generate stuff, even if it sounds really good.
[00:06:12] Speaker B: Yes. And I've come to this concept of source of truth as well. So I really like your use of it. And I would go so Far as to say an LLM that's generating output based on its training data should never be relied on as a source of truth, period.
[00:06:27] Speaker A: That is a, that, that, that's a controversial take, and I think it's a correct take.
[00:06:31] Speaker B: And I think that the reason why lawyers and judges and other people are not understanding AI is because like gen AI, which is now just colloquial being called AI, is that exact thing. It's only when it's connected to some outside external source of truth. And you could be the source of truth. So like in Claude, I still use Claude, I upload an hour long podcast transcript and say, summarize this into three paragraphs for a podcast and focus on the guest. And I put the guest's name. So I'm the source of truth. I'm providing it with source of truth. And when I say only summarize this what I'm providing you, then that's reliable because it's only leveraging, it's only creating a summary from what I've given it. But I can't just say, here's a link to a podcast episode or here's whatever, give me a summary of it. Because it's going to also respond based on its training data. Think of it like a human. If you gave a human, a really smart human that's read the whole Internet and said, I'm doing case research, I want you to provide some research for me on this thing. But they don't actually have access to case law and they may or may not have even read it. And even if they have, would you rely on that person to give you factually accurate case law? Never. Never. They must actually have access to it. Even if they've read it. You still want to go back and be like, like if you have an associate, like, I know this case, go look up such and such a case. Like you still want them to actually go look up the case. And that's what rag is.
[00:07:48] Speaker A: Okay, so then, so let's come back to Paxton because you said that was your number one tool for the year. What is it?
[00:07:53] Speaker B: It's legal specific large language model that also plugs and plays with the other models that are out there, the big ones we've mentioned, and it leverages retrieval, augmented generation to case treatises, to documents, to everything. Right. And the reason it does better than Lexis and Westlaw is because it's also leveraging this legal specific large language model. At least that's my guess as to why it's so much better. And it's far more accurate. And the Stanford benchmarking test has shown it's significantly like by a wide margin, more accurate, like in the knees of percent. I can't remember exactly what percent. And the other ones are in like the 40s and the 60s in terms of accuracy. Right. For this independent benchmarking.
[00:08:30] Speaker A: Yeah, that's, that's, that's, that's super interesting. So you're saying. So it does sit on top of these other large language models or work, work with them, but then it's got ground truth associated with it. So it's not going to just generally go off the rails and start creating case law and facts that just aren't.
[00:08:47] Speaker B: Attached to reality in my experience and use of it. And whenever using any new AI tool, you should always start with, with a somewhat complicated question that you already know the answer to. Right. And see how the tool does to get it right and get a sen. You know, it's quirks. Just like when working with a new person, I always try to think I'm writing a new article, an upcoming article in like the January, February issue of the GPSOLO E Report, and this is what I'm talking about right in it and so I'm thinking about it right now. Like, imagine you're working with a human. It's just the smartest human who has more capacity than a human. But if you think about it as if you're working with a human, the way you're going to interface with it is going to be significantly different and get much better output than if you pretend you're working with a calculator. Because language is made up. We made up language. It's not a calculator. A calculator, when you program it to do two plus two equals four, that's an immutable truth. It will always get that right. Language, there is no immutable truth. We made it all up. So like these are statistical language machines. And so like there's always going to be potentially a different output and that's how humans work. Like it's not, it's a reflection of how humans work. It's not exactly how the human brain works. Right. But like that's how you have to be working with it. Not like it's some immutable truth like a calculator.
[00:09:56] Speaker A: Yeah, absolutely. So look, I use Claude all the time. I've loaded a whole bunch of documents into it. So I've got the project set up and that's incredibly powerful. I do completely agree with you. Like one of my typical questions for Claude, after it gives me an output is tell me what you got wrong about this. And more times than not it'll say, oh, I'm so sorry, I put a bunch of hallucinations into this long output for you. And so it's obviously just very, very. It's got a lot of holes in it, right? And it's a little bit like Swiss cheese. It's like, yeah, you try. You got to layer a bunch of these slices together, knowledge based, plus prompting plus reprompting, plus eventually like your own eyes on it before it goes out to anyone at all. So I totally agree with you. Like, if you're using a raw output from Claude, GPT Gemini, you're just asking for trouble. A thousand percent. And even with something like Paxton, it's going to make mistakes. Things can and will fall through the cracks because like you said, it's language, right? Like we're talking about statistical outputs here, we're not talking about calculations where you can go back and just check if, if the 2 plus 2 equals 4. Did it do the math? Right? Because it's math. So, right with that, like, I think that, I think that's really great. Just give me like a ballpark. Like, what level of productivity increase have you seen through Paxton and the other tools you've adopted this year? What do you think your own output level has increased by this in 2024?
[00:11:17] Speaker B: It is so hard to quantify that. I'm just going to throw out what sounds like a ridiculous number and it's like, I don't know, like a thousand percent, two thousand percent. Like it is that efficient, right? And even, like, even math. So there's this, there's this API that the AIs could plug into called WolframAlpha. And I don't have time to go into WolframAlpha, but basically, for a long time there's been these algorithms that do math and complicated math. And so if these AIs plug into WolframAlpha, it's going to do math, like really good math. And you have to pay for these tools, right? Like, Paxin isn't cheap. Like I think it used to be $99 a month and now it's maybe like 199amonth. And I still think that's way cheaper than what you'd have to pay an associate or multiple associates or staff people and associates. Like the arbitrage is still amazing, that should be taken advantage of. But now it's able to do complicated math. And I don't think the free versions are necessarily plugging into the WolframAlpha API because they, WolframAlpha charges them to do that. But like I'm doing like complicated like mortgage stuff and other transactions where there's multiple pieces and formulas and valuations and other things. And I used to just ask for the formulas from the AI to put it, plug it into sheets or Excel and now it could just do the math like right in there in the paid tools. And that's pretty amazing.
[00:12:34] Speaker A: That's incredible. So, so you're saying 10, 10, 20x productivity look, even if you, even if you discount that by 5x for a learning curve or that's incredible. Like there is nothing, there is no better dollar that you can spend in my opinion right now than AI tools that you are able to implement and adopt yourself. Like I'm paying a $20 a month cloud subscription to load in my knowledge so that I don't ever have to start on a business model. A, an email, a slideshow presentation, podcast, show, outline. It sounds crazy to say 10x more productive, but I think for, especially for a business owner where you're a generous, you have to be a generous. You have to like these are generalist tools and I think like anyone who's not using you're, you're losing out. Like the future is a hundred percent going in this direction. The tools got way better this year. I think one of my main takeaways was like last year is quite speculative. Seamlessness of the tools increased dramatically this year. Just the kind of like baseline level of usefulness. Like Paxton pulling in Rag, Claude allowing in OpenAI, allowing you to pull in documents and ingesting those and being able to do better math and searching the Internet. Like the tools just got so much better this year and so much easier to use. So I think that's, that's a, that's a huge takeaway for this year. So let's just talk about.
[00:13:53] Speaker B: We've only three more tools really quick.
[00:13:55] Speaker A: Oh yeah, yeah, yeah.
[00:13:56] Speaker B: Specific to personal injury space that I want to mention that I'm not using because it's not relevant to my practice. But since I care about learning as much about these tools because I become a source, a resource for attorneys to learn about AI. I've learned about these tools. Two of them do similar things, although they do them differently. And just like ice cream, like it's still delicious, you just got to pick the flavor you like more than the other. Parrot and Scribe are like your AI powered deposition assistants. Transcribers like everything right. Come up with a Strategy pre deposition. And it could keep you up to speed during the deposition. Right. Like, well, hey, this was an interesting question and based on the strategy that we outlined, you should ask these questions or ask this follow up question or you take a break, a bathroom break. Like you could check the AI, right? And like come back and have a more, have a better deposition. And then you've already got a summary of the depot, like right when it's done. Right. And like so like Parrot and Scribe. And I'm sure there's some other competitors in that space. Those are the two I'm aware of. Check those out. Much more efficient depositions, much more deep depositions. Even Up. Even up is going to radically change, I think, the pricing structure. Because I'm all about law firm business models and pricing. I think tools like Even up are going to radically change what lawyers can charge as personal injury attorneys because you're getting to results faster, more efficiently and more likely based on all the data that's out there in similar situated soft tissue, non soft tissue style cases in this jurisdiction. Like, I think it gets pretty granular in terms of its data that it's analyzing. And so if you could get to an end result faster for your client, like you could fight them tooth and nail the trial and then settle for what you might have just been able to get earlier based upon the data. I mean, like there's bias in its processing, but like the data is unbiased. Right. So like is that worth a third? Well, maybe because they're getting money faster, but maybe it's not. And what are some other valuable things that you can provide to clients that's not just settling or taking to trial a personal injury case? And you have to start thinking about that because we, the lawyers are actually good at a lot of things that maybe we don't even think of as like legal services per se, but are still extremely valuable that we could offer to clients who are say, subscribing. And you want to be on the cutting edge of this because this, I think this is going to happen. And there's some firms out there that are saying, well, we take 10% less and it's like 30% instead of 33%. Right. And so there are, there's already starting to be some competition in this space and if you want to win that game, you have to be more efficient and get to end results faster for clients.
[00:16:12] Speaker A: Absolutely. I could not agree more. I think the amount of leverage with the tools that you talked about, Scribe and Parrot for depositions, that's More your litigation. Even up a lot of pre litigation work there. Look, the leverage is so high that, that you can't afford to not use the tools anymore. It's just the, the landscape has shifted. Given a couple minutes left on the episode here. What's coming down the road in 2025? What are the trends that you're most excited about or and you think will be most impactful on law firm practices?
[00:16:43] Speaker B: Yeah, I mean we've kind of been talking about it this whole time. Right. And maybe I should have waited for you to ask that. Talking about the tools that I'm. That I just mentioned, because I have.
[00:16:51] Speaker A: But those are available today. Like you have to go look at those. Like do that. Like they are not future tools. So 25 will obviously bring enhancements to those. But what's, what do you think is going to be new and changed about next year?
[00:17:04] Speaker B: Yeah, I mean, agentic AI, right? If, if listeners have, you know, seen anything about agentic AI, it's all talk right now. And I really think we're going to see 2025 being where agentic AI takes off. And what does that even mean? Well, like for example, in Perplexity, if you're a pro and you're searching about like what's the best greater than 10 foot long cable that still transfers 4K video, which was the thing I was researching this year and I found one and I'm using it right now and I've only been using it for two weeks and we'll see if it's actually, if it actually holds up. But I can now buy directly within Perplexity and like that's agentic AI helping me do the research and then going out and buying the thing instead of me having to like go to Amazon or go to this company's website and like buy the thing. And that's like low level agentic AI. Like it's an agent. Agentic is like like creating a verb from agent. And so the agent is going to go and do the things and seeing that within legal specific AI, like we're going to be borderline unlicensed practice of law with something like that. But as long as a law firm's using it, you don't have to worry about that. Right. And so I think a lot of these tools, we're going to see them start with law firms and have lawyers adopt agentic AI. Now imagine if you could send that agent to task on prepping for the deposition instead of having to prompt back and forth in a conversation with the AI. Now you can just say, yeah, here's a bunch of things. Go do all this stuff and it's going to do it and it's going to make reasonable inferences based on the instructions that you give it. And it's going to go do all these things without you having to prompt back and forth and back and forth and then you'll be ready for trial or ready for your deposition or ready for the contract negotiations or whatever it is. Right. And so being able to set the agent to task like you're just going to have to have somebody review the work and it's not, it's going to be as simple as math like we talked about earlier. But you're still going to need a lawyer in the loop, at least in the foreseeable future. But agentic AI is going to take a massive amount of workload off of lawyers plates.
[00:18:53] Speaker A: I completely agree. And just to, just to drive that point home. So we've been talking for most of this episode about tools that are available now that are providing 5 to 10x leverage on individuals. And those are tools that you use to get work done. What Matthew is describing is a future which is very real. And I've, I, I've been using Devin AI, which is an agentic coding subscription service for some personal projects. I tell it, I want you to build a front end web app that does these functionalities and it just goes and does it and it'll work for hours without me even talking to it. It'll just start doing things on its own. It'll come back and ask for approval. That's what we're talking about in 2025. Like this stuff is not very far in the future and in some places it's present. Devin AI is something that I can buy and use right now and so that's coming to legal. So look, my goal with each of these episodes is to give you a key takeaway. Here's the takeaway for the 2024 wrap up. These tools are available. They're incredibly useful. Lawyers are using them right now. Matthew is one of them. Many of the people that I talk to are using AI tools and I know there's a lot of adoption generally, but you need to find ways to get this all the way down inside your organizations. Otherwise you're going to be behind because agentic models, bots that are going to go out and do stuff and come back to you and ask for your approval. They're coming in 2025. One of my big things that I'm seeing is AI voice. I think that's crazy revolutionary for lawyers. We've seen it, we've seen the chat, we've seen the document stuff. I think Voice is is next. Eleven labs, new agents that can have incredible conversations. They just released this like two weeks ago. I've been playing around with it. It's real time conversations that are absolutely phenomenal. AI voice, I think agentic models absolutely coming in 2025. Matthew, I really appreciate you being on to wrap up 2024. Look into the future a little bit. Thanks so much for being on the show today.
[00:20:42] Speaker B: Thanks again for having me. And I encourage, if you like what I'm talking about here, I do post a lot on LinkedIn about it. So if you search the subscription attorney on LinkedIn, you'll find me.
[00:20:50] Speaker A: And obviously check out the Lost subscribe podcast. Matthew's a great host and has great content. Thanks everyone.
