001 Who Are You Working For?

Today we're getting Woo Woo. We're talking about, oh, our identities as firm runners. We're talking about using GPT and these robots for research and what works and what doesn't work. Talking about prompt engineering, what that even is. I think it's like the new Googling, but even more powerful and how to hit your wagon to the things that will take you on a path that you want to go on all this on whatever. I end up calling this Jason Daly. Do you think if I call it that people will think my name is Jason Daly Daily? Jason, Can we get Woo Woo to start things off here? So I got pretty touchy feely on social media last week and a couple of people DM me. They were concerned. They said has did your robot break? But it was because I had a couple conversations with friends that were just like such great reminders of being mindful of why we do the things that we do and like pushing back against sometimes some internal motivations that aren't healthy and ultimately don't take you to a place where you want to be. So like one example of that is, I think a lot of us are motivated by proving something to somebody. Maybe it's a past boss or that person that wouldn't give you the promotion at the other firm. So you went out and did your own thing. You're gonna, you're going to show them, you're going to, you're going to build a bigger firm than them. And that's, that's like what motivates you and it's, it like takes you down a path that's, that's dictated by somebody else. And I think even if you get there on the other side of that, first of all, you won't find anything and there's nothing to say that you'll even like find happiness or fulfillment and what's on the other side of that thing. But the big thing is that person doesn't care. Like that person you're trying to prove a point to, they could care less. Yet we're going out and we're oftentimes sacrificing the relationships that are most important to us in order to make a point to somebody that doesn't give a crap about you. Like I have so many examples of this in my life and it just got me thinking about what are all the wrong reasons why we make decisions and affirm and like all these, all these internal things we struggle with that are very, very like human things and probably stem from insecurities that we have and you know, Us grappling with our identity of being firm runners and what it means to be a successful firm runner. But I just had a couple conversations that were just such great reminders to try to be mindful of the path that you're on and why you're on that path, because it's really easy to get sucked into the sweaty mindset of growth and success and all of that. But, I mean, so many times, even when you find that, like, the fastest growth period of my firm running career was the least happy time I ever had, it was cool. It was impressive. I. I won awards, I did interviews. But, man, that stuff sucked. Like, that was not fun at all. And maybe there's an element of, like, it's easier to say that when you're on the other side of it. Or once you. Once you showed that past boss, once you showed them, now that you're on the other side of it, it's easy for you to say, ah, yeah, none of it matters. But especially through the lens of like. Like, I'm married, I have children, I have young children. And I think the more I get sucked into the trap of my identity being somebody that runs a firm, the more the people around me in my life pay the price, because that's who I am. And that's the yardstick by which success and all that stuff is measured. But a lot of times it's for all the wrong reasons. So, like, I try to be like. In one of these conversations, the person was like, you don't really. You don't talk about family, and you don't talk about not being an overachiever. In fact, you probably look like the opposite just because doing so many things. But it was just a really, like, poignant reminder of really the greatest. The greatest thing about running a firm to me is the freedom that it enables. The fact that you're not working for somebody else anymore, and you get to make up the rules. And if you put yourself right back in that same box that you were in when you were working for somebody else, then what's the point, right? Like, it's up to you. Like, that's the best and the worst part about having that agency is you got no nobody to blame but yourself. But if you need to hear this today, there's a lot of days I need to hear this, man, just chill out. Like, everything's gonna be fine. Stop pushing so hard, say no a few more times. Like, give yourself a break. Because I think everything we read and consume in many ways does the opposite. Like, it tells you, here's how to find success and here's this thing you can change in that thing and here's what's gonna get you there and help you do something really impressive. But sometimes it's also cool to just chill, you know. Like I've got days where I feel like that sweaty boy that wants to, wants to do all these things. And like there's days when you got that fire and man, there are days when you don't have that fire and you wish the day before that you didn't have it because you just got yourself on, you know, a whole day of meetings or something like that that you now don't want to do. So like if I can be the one voice that's saying you're doing great, you're doing enough, you're going to figure it out. Stuff takes time, sometimes a frustratingly long amount of time. But if I can be the voice that's like the anti, you know, be okay with what you've got. Like, hear that from me today, I guess. What a weird way to start that off. I don't normally go full woo. Like that other really cool thing we're gonna see, like what format this show ends up being. What I feel like I'm missing with the YouTube channel is like the YouTube channel is a big, big lift and those videos are a big production now. And so the stuff that goes out on them is like there's a lot of latency there. Like it's like a three week production. This stuff, man, we can just jam on. Like what's exciting, what just happened like of the discussions that are happening online, like what is the best of that and what can we take inspiration from and just get like a little bump to kick off the day, you know. So that's kind of what I'm going for here. Cool stuff that happened the last couple days. I did a roundup on using large language models. So AI stuff like GPT for tax research. So US tax research. I think it's applicable to more other technical types of tax research too. And the result. Drumroll was that's pretty bad. Like it's, it's still not something that without a lot of expertise from a pro and how you word the vernacular and some prompt engineering, like it's not something that you can really get useful stuff out of yet. So I did a roundup of. We use Google for perspective because that's what we all know, right? Used Bing, an AI search engine called Perplexity, a really cool app called Chat PDF. I'll Talk a little more about that. And then the last four OpenAI GPT models, just to kind of see how far they've come, Google actually did pretty well in this roundup. So when I prompted Google, and so the premise was, here's like a semi challenging tax research question, like, find this nuance in a rule. Most of them didn't find it. When I worded the question the way that a tax pro would, Google actually gave me the answer straight off the top in a Google search. But none of the language models really got there on their first try. A couple of them, like Bing, it took a second prompt, and with a second prompt of me saying, okay, so there's no exceptions to this rule, then it goes, oh, yeah, there are these exceptions right here. But the reality is, if you were using a tool for that, you would have peaced out after the first prompt when it was like, yeah, not really. But a couple of takeaways for me here. One, one of the apps that we've looked at, which I've talked about a little bit, because I think it's like an important paradigm shift. It's called Chat PDF. The idea of Chat PDF, so chatpdf.com is you take any PDF file, could be 900 pages, could be five pages, you chuck it in there, and in less than 5 seconds you can start chatting with it. Which is kind of weird because the way that we're used to working with PDF files is probably Control F, it's searching for a specific bit of text. But that kind of sucks because it's just text matching. So if you don't have the vernacular right, if you don't know the right word to get to what you're trying to get to, you're not going to find it. And especially in large documents, especially like IRS publications, that sort of thing, it'll take you to an explanation. But then two pages later, you may have a list of exceptions and all these other important considerations that it's not necessarily going to surface, as opposed to a service like this where you can pose a very specific question to it, and it can see the entire document and pull all the bits out of that document that are relevant to your question. And this uses embeddings, but also GPT on top of of embeddings, which is how it kind of stores all that information from the PDF. And so it's taking the best of a large language model like GPT, but limiting its context to just this document. So you're not getting all the other stuff that's in There all the general knowledge that it has, it's just speaking to the document, which is kind of like the best of both worlds, right? It can frame things, explain things in a really helpful way, but it's only looking at the stuff you want it to look at. So my main two takeaways here is I think the best version of a research tool like this for a specific domain is when you can limit the scope of the language model to just the context that you want it to look at. Maybe that's the tax code or an IRS publication or that big old long bill that just passed that you don't want to read the whole thing. I think that's the best way to use a language model to get to specific information today, rather than that language model that knows everything, but with everything, there's no specificity. So it's like, okay, which tax year are you talking about? Or are you talking about the authoritative text or the hundred other blood blogs and listicles that were written about it? Reality is, a language model is like the marriage of all of those things. But when it comes to technical research, like you probably want control over what context it's pulling from. And that makes I think the output much more useful. So chat PDF, like that's an interesting paradigm we're seeing right now, a ton of chat with X types of AI applications because that actually just got really easy to do. Check, chat with your help, documentation, all these different things. So it'll be interesting to see the different ways that we can use that to, I don't know, simplify the stuff that we're already doing. The second big takeaway for me of going through that research exercise was the value of prompt engineering. So a couple of times I framed the question in the way that a non tax pro would frame it and the model, it was just not even close. But when I framed it with the vernacular that I knew would get me to kind of, sort of what I wanted, or a couple of times I had a prefix on the prompt that said using IRS publications. I just said that at the beginning that actually gave me the correct answer because it only then used publications as kind of the authoritative source. And man, it's a reminder to me of just how much prompt engineering is now this really valuable skill. In many ways it's like the new Google, right? Like, man, I've, I've toyed with this for years. Like I would love to assess a potential employee's ability to Google things. It is such a learned skill. You know, the people in your life who don't know how to do it well and how much of a barrier that is to get to good information. But man, in many ways, prompting and being able to write for ChatGPT and other language models in a way that will get you the result that you want is becoming such a valuable skill and will only compound as these models are getting better. And many of these AI products that we see out there like all it is is actually AI prompts behind the scenes. So you give it something and it wraps it with a prompt to say, do it in this voice or take the output of this and structure it this way. So for example, apps that are summarizing meetings, you're taking chunks of that meeting and then around that you're doing prompt engineering to say summarize this part, that part into bullet points. Like that's what actually goes to the language model in the same way that I would add a prefix to say use IRS publications and that tells it the right way to do it. Many of the actual apps that we use are simply adding like wrapping around the message that you give it or, or parsing things out of that message and structuring what the request looks like when it goes to the language model. But it really to me is like the new Googling. And a lot of the anti AI stuff I've seen online has kind of come from people who don't have a great grasp on prompt engineering yet. So I saw one high profile person that was posting on LinkedIn about having it generate a landing page for their firm. It generates all the landing page copy and it was like very blah. And they're like, I don't see the value in this, it's just, it's too basic. But like, you got to understand that's what a language model is, is it's like the average of everything. And if you don't give it any specific instructions to the contrary, the result is usually going to be pretty blase. So you got to say, use this tone of voice or say, hey, here's some stuff that I've written so that you understand my own tone of voice or write this other thing in my tone of voice. And I don't know, nobody's going to just know that. And the best software is like, I don't know, intuitive and maybe makes that easier for you to understand. But it's a good example of how like small changes in how you prompt it give you much better results. Another thing that's come up a lot is people talking tax or other technical things with it without Giving the language model enough context to answer it correctly. So asking it a question about unreimbursed employee expenses, which are no longer a thing generally, with a couple exceptions, used to be a thing in the past, isn't really now. That language model is not only trained on like, well, I guess it's not trained on current year tax knowledge because the training model ends for GPT4 in September of 21, I think, but it includes everything before that. So what it's actually trained on is rules from 2012, just like 2015, just like 2020. And if you ask it a tax question or say prepare the tax return based on this or that, without that context, what's it gonna do? It's kind of, you're kind of like spinning the wheel on where it's gonna pick and choose where all that data comes from. But it's really as simple as saying like under 2021, federal tax rules or you know, whatever your tax agency is, you just have to be more specific. And it's interesting, I think it's actually kind of a reflection of the struggles of working with junior staff is you can't expect it to know what you don't tell it. And there's maybe an element of mind reading like that you expect out of it. Actually there's a study and GPT4 now has like the mind reading or quote unquote common sense capacity of like a 16 year old now and the last model was like a 7 year old or something like that, where it actually is getting better at sussing out like what you're implying by what you're asking it without you explicitly saying it. But that's a huge factor in a user's experience with the model. But I see a lot of examples where with a bit more prompting around the specifics of how you want a thing to be done, it could absolutely do it. But instead we just kind of chuck that thing at it and then it doesn't do it right and we're like, oh, that's annoying. Doesn't work. Which like I've seen happen to junior people in real life so many times, you expect them to know this thing because it is quote, unquote common sense to you with 20 years of experience, whereas common sense for somebody else is going to be fundamentally different. It's interesting to me how this is like, this is like the new Googling and it's only going to become more important as those models get better. Last thing that I've been thinking a lot about and it's Kind of a framework that I've always tried to make big decisions based on with software or with firm strategy. That sort of thing is we make decisions in kind of like this snapshot of what we see in front of us at a given time. But oftentimes the more important consideration is the fact that you're hitching your wagon to something. Where is it going? Is it going to take you to a place where you want to go? And like software is a really good example, is that company going to develop the software anymore? Are they going to go in a totally different direction that you have no interest in? For me and my firm developing our casts, our accounting practice was a big example of this. When we rolled it out, I wasn't able to make it as profitable as our tax work, but I could see that the accounting practice was on a more profitable path than tax because the price for this stuff was going up and up as private companies had a hard time finding accountants and the technology kept getting better and better. So I was kind of in this wedge where I was only gonna get more efficient at doing it and I was only gonna be able to charge clients more. And that seemed like a good long term position to kind of set myself up in. And right now, man, like thinking about that, who you're hitching your wagon to has never been more important because of the rate at which things are changing. And it's why for me personally, I'm investing like so much in learning and understanding AI and embeddings and language models because it's really hard to argue that there's anything else that is moving as quickly that you could hit your wagon to. Right. Like this stuff is moving so fast. GPT4 was finished being trained actually last summer and they spent six months doing safety training around it and all that. So GPT5, most people expect is either done training now or pretty close to being done training. The impact is like a little different in every industry. GPT4 has already like fundamentally changed software development. Like it's almost non negotiable that you need to be working with AI and how you develop code now. And that's all come in like the last 24 months. So to say where we are 24 months, 36 months from now, there's aspects of it that are honestly terrifying. But man, the reason I'm, I'm investing so much in learning this stuff right now is just because where else are you gonna hit your wagon that's moving as quickly as this is? And it's what has me talking a lot about the vendors in the accounting space too, some of whom are plugged into this, some of whom are not. And I'm worried companies like Microsoft are gonna start stealing the lunch of the practice management systems in our space unless they start moving on AI stuff. So be mindful of that as you're making software decisions, especially with all the AI stuff going around. Who are you hitching your wagon to and are they taking you to a place that you wanna go? Hey, thanks for coming and hanging out. We'll probably do this again tomorrow. I'd love to get your feedback. If you're watching on YouTube, drop a comment. Big news here is I wear glasses. Podcast listeners are gonna miss out on that. Reach out to me, Twitter, LinkedIn questions? What do you wanna talk about? Let's make this a space where we can learn a little bit every day and spread some positivity.