396 10 Recent AI Announcements ACCOUNTANTS Should Be Excited For

My friends, it has been a wild last 30 days in AI, and I try to keep things super, super practical on this podcast. Accountants are nothing if not wildly efficient with their time and attention spans. But today, allow me, allow me to indulge myself. Indulge yourself. Let us indulge ourselves. We're going to run through 10 recent AI announcements that are kind of like defining the future of knowledge work and within that, how accounting gets done as well. It has not gotten to the level where it's like, this stuff is productized. I do know, like, the folks behind the tools that you use every day, they are, they are scrambling the interceptors and cranking on this new tech that they have that is kind of magic. And that's the stuff that we're talking about today. It's kind of like the tools they now have in their hands that they can put to work for us. The kind of stuff that we should be seeing around the corner. 2025 stuff. If 2024 was the year of the AI chat assistant that nobody wanted, where's tech taking us in? 25, let's talk about. I'm going to give you kind of 10 updates that I think are important to keep you in the know on all things AI. Let's do it. All right, number one, and it is kind of the biggest trend that is defining how AI continues to develop and get smarter and be more useful. It is the shift to quote unquote reasoning. So language models, historically there was this big push to make them bigger that is trained on more data, faster. And so we threw a whole ton of compute at them, train these absurdly large models that just cost astronomical amounts of money to train. These models are so far beyond what you can run on your local hardware, like on your local machine. There are definitely models that can be run on their local machine and they're getting more efficient every day. But the frontier models, the biggest, most powerful models were these big, just massive things that were running on humongous servers initially with the goal of, like, getting you as high quality of answers as fast as you possibly could. And when OpenAI launched their O1 model, which we first got 01 preview a couple months ago, we talked about it on the podcast quite a bit that, like, unlocked an entirely new paradigm of AI that is now taking us a huge step beyond what the old paradigm was capable of doing. There's been some talk that, like, at a certain size, models don't get any better. And there's probably some truth to this because there's only so much training data out there that these models can gobble up. Initially people thought that like you could just scale up a model forever and it would continue to get more and more intelligent. The reality is if it's not trained on something that's more intelligent, doesn't necessarily get more more intelligent. But how we are now getting greater intelligence is with, with this shift to reasoning. And so this is basically the amount of time and compute that goes into the model thinking before it responds to you. And a really basic version of this is, you know, if you learn very basic prompt engineering, people would talk about chain of thought prompting that is giving it a sequence of steps to follow and that yielding a higher result or a better result. In fact, over a year ago on this podcast, maybe even a year and a half, year and a half ago now, Accounting today did a, a test where they tested the GPT 3.5 model on the CPA exam. And they said, neener, neener, it's not smart enough, can't pass CPA exam, will never be as good as us. And then GPT4 came out. Well actually at this point in time GPT4 was already out. But they ran the test with the old model and I understand like it takes time to do these things. And so the, the report didn't come out right away, but I was like, this is pretty easy to test. Why did they not use the new model? I use the new model and it, it like it passed. It was pretty much at a passing score. Well, shortly after that there was another study using quote, chain of thought prompting with the new model. And basically how it works, it's kind of amazing is with no chain of thought prompt prompting and basically just copying and pasting the multiple choice questions in from the CPA exam or from a CPA exam test bank, the model just barely didn't pass. But they did two other approaches that literally the question was the exact same, but they just changed how the prompt was written, namely putting examples of multiple choice questions and the correct answers above it, which were not at all related to the question it was actually asking. Just by simply putting example questions and answers above and then putting a question and then asking for an answer, it yielded way higher scores. In fact it, in the end the study found, and this was over a year ago, the study found that it, it passed CPA exam at like an 87 to 89 or something like that. But then it went on and passed the enrolled agent exam and the CIA and the cma like every, every certification that there is, which prompted ICPA And Barry Mellon to be like, it's never, it hasn't actually passed the CPA exam because it can't see the real questions. And what they're getting at there is test banks, like the public questions we have to do this testing on. In theory, you go to like a surgeon or a Becker or something like that, their questions ought to be private. But the reality is people share that in corners of, of the Internet. Maybe people are talking about it on Reddit or something like that. For the actual CPA exam, those questions theoretically, theoretically ought to be nowhere. And so the thinking is, well, somewhere in their training data they found in a Reddit comment or something like that the answer for this thing. It's a problem in benchmarking. When they take a new language model and they give it kind of this set of questions that's maybe a common benchmark. People are like, well, that that benchmark was trained into the model. It already had those answers. It's true to an extent, but also largely not because language models don't work that way where they like, retrieve one very specific thing. So what the ICPA says to that, it is true to a degree, but largely not true. I, I genuinely believe if you put an AI model to like to the test on a CBA exam right now, it would have no problem. But, but I think the bigger issue with the response is it's not about us versus the computer. Like, AI doesn't give a crap about you. Like, that's, that is. I, I think if more people could let that set in, we would focus on the more important things, which are, where can this be useful to me? Not it'll never do this or that. It doesn't matter. In fact, you have no freaking clue. If anybody is telling you that AI will not replace accountants, they're just pandering. They really are. Because nobody knows. Like, the reality is nobody knows. I would argue, like, it's pretty much a guarantee that it will, because it already has. I've already talked with accountants who have been replaced in like an industry. And so the biggest thing right now is AR processing, AP processing. But you could have said the same before about rpa. Robotic Process automation. You have a floor of accountants somewhere doing this incredibly repetitive task in SAP. Some somebody builds a bot that will do that, and you lay off that whole floor of accountants. Technology replaces people. It just does. It doesn't mean that those people can't be repurposed and reskilled and go, go do other things, but just the whole debate of whether or not it's going to replace people is completely misses the point. Like what we would love to do is plug AI into the stuff that we do today so that we can do more valuable stuff in that, that sort of stair climber to the next better thing. That ought to be the goal of what we're, we're always doing. And, and too often the conversation goes to this primal self preservation of ha ha, look at that dumb thing that it did. It's never gonna be as smart as me. And it's those people, ironically, who are like most susceptible to being, to being displaced. Now this shift to reasoning, it's the same concept that yielded that much higher score on the CPA exam. The notion that by prompting it specific ways you'll get better, better outputs. And this reasoning approach has basically been built into how these models now serve up responses like a, like think of it as chain of prompt reasoning, like to the nth degree. And so basically it will spend a bunch of time thinking before it gives you a response, kind of mulling over its own responses and being critical of it before ultimately giving you a concise response that has already been thought about and through multiple iterations that you never see. Kind of fascinating. So the easiest way to play with this right now is if you have a paid chat GPT subscription. You hop into ChatGPT and the model selector at the top select the 01 model. This is a reasoning model. Give it a question and it'll take anywhere from 5 to 40 seconds to think about it before giving you a response. And a really basic example of how we kind of already knew that it was capable of this sort of thing is like, you know, so playing Scrabble happened to me. You take a picture of the Scrabble board and you say, give me a few moves that I can do with the pieces that I have in the state of the board. And it gives me like five words and one of them is a word that I've never even heard before. And I'm like, is that, is that actually a word though? Like what does that mean? And the language model is like, he ha, yeah, you got me. That's not really a word. Made that up. Sorry about that. Here's some other ideas. And if you literally just in the course of, of conversing with the AI, would say, here's a response. No, now double check your response and see if there's anything that's made up. The next response would be higher quality, right? Like just that's how they are right now. That's what These reasoning models are doing behind the scenes, they're having like a debate with themselves. And this gives you a vastly, vastly better response. Like to the degree that, that the O1 model, which is the best example of this reasoning model that's out in the world now, is doing like PhD and beyond human PhD level work like that is. That is wild to think I literally have a PhD in my pocket right now. It kind of shines a light on the fact that I don't quite have like PhD level problems for it to solve, but where the rules are known and the outcome can be provable and all of that, it's unbelievably powerful already. And the wild thing is, is under the hood, it's using the same GPT4 model that we've had for over a year. And so it's like the underlying model, it's technically, it's the 4O model that we haven't quite had for a year, but it's the same model. And if you just had that same model reason through that problem for a bunch of time, the outcome is way, way better without the underlying model having to be any better. And so if we ever do get a GPT5 like a next. Hang on, my wife's printing something like a next generation model. When we plug reasoning into that next generation model that's supposed to be way better than GPT4, holy smokes. Is that going to be a really, really smart AI? So a lot of the developments we're seeing right now are all around reasoning and the quote unquote inference time. That is kind of the amount of, of compute that's put into letting the model think before it gets back to you. That is the, the biggest like big picture shift happening in AI right now. Google's head of AI product, he said the other day that he actually thinks it is a straight shot to ASI now with reasoning. ASI is artificial super intelligence. So you may hear a lot about AGI, not adjusted gross income, but artificial general intelligence. That is the level of intelligence where the computer can basically do what most humans can do. But ASI is when artificial intelligence is beyond humans in, in every way. It could totally take our lunch. Kind of a hard thing to get your head around with AI stuff is that human intelligence is a plane on a vast spectrum of possible intelligences. So if you think about like an X and a Y sort of thing and the progression of intelligence over time, I don't know if, I don't actually, I don't know if humans have fundamentally gotten More intelligent over time. Actually, I think there was some science around this where like, IQs have gone up or something like that. But you got this line that represents human intelligence over time, right? And then you got this line that represents goldfish intelligence over time. And I don't know if goldfish have gotten any smarter or not, but I think most would agree they're probably way below the human level now. Artificial intelligence, it's on a. It's on a line with a different slope than us right now. Right. Like, ours is pretty flat. Artificial intelligence seems to be going up at a pretty rapid rate. And we're super fixated on the intersection point when artificial intelligence meets human intelligence. But we don't think about the fact that, well, there's kind of nothing to make us think that it can't keep going up. That in fact, you know, take the smartest human out there who has an IQ of 200, AI can pass that and continue to go up and up and be smarter and smarter. And one problem we're already having with how they benchmark models that is like, throw a bunch of tests at a model to see how smart it is, is we kind of don't have tests smart enough for them yet. Like, they're able to do all of the tests and benchmarks that have been out there. So we're having to create new tests to challenge them more. But if we go back to, like, the example of, oh, one, doing things at a PhD level and in some cases beyond human PhD level, who's going to make the benchmarks to, like, test, okay, how, how beyond PhD level are you? But intelligence, like, in terms of, like, the natural world, there is nothing to say that intelligence can't continue way past what humans are capable of. This got way more sci fi than I was expecting it to, but. But then you have this whole existential, like, is that terrifying or is that humanity's greatest achievement? The notion that we made this thing that's incredibly powerful and surely will make lives better like it ought to. And so if you think about the car, we made a car that can go faster than humans can run, and it can just go and go and go and go and go. The train, man, that sucker could go so far. The plane, are you kidding me? We made planes. But now this is our brain. This is, this is, this is challenging our kind of mental dominance over the rest of the solar system. And that's a different ball of wax. This episode is sponsored in part by Live Flow. Do your clients just throw up in their mouths a Little bit. When you send them a PDF of their financials every month. I bet they do, because what is even that? What. What are you supposed to do? A set of financial statements. Okay, most people, they don't know how to make sense of this. What if you started delivering something, Something better? Okay, think about it. Live Flow, kind of the OG what they used to do. They let you connect your qbo, your Xero to Excel and Google Sheets so you can, you can make your wildest, wildest spreadsheet dreams come true. Now they have these simple little financial dashboards you can build, have all the same connection points to your. Your accounting ledgers, but you can build, like, a more intelligent report that your client will actually understand. You can give it to them on a monthly basis instead of financials, give it to them on a weekly basis even because it can connect to the ledger in real time, still means you got to do the bookkeeping every week. But if you're looking for something, your clients will actually engage with a little more and get more value out of that ongoing accounting that you're doing for them, Oftentimes the answer is to change the deliverable to. To something besides that stanky, stanky old PDF financial statement package. Okay, I'm not the boss of you. I'm not here to tell you what to do besides mosey on down to the show notes and learn more about Live Flow via the link down there. This episode is sponsored in part by Cloud Accountant staffing. Hey, scenario. What if. What if you had this buddy that was running another accounting firm and they stumbled into this thing and they, like, they start hiring these really good offshore accountants in the Philippines, and. And you're talking with them one day, and your buddy's like, hey, I got some really good people here. You want me to hook you up? You want me to find you a few. Find you a few accountants? What would you say? Would you be like, nah, I'm good. I can find accountants all day long. I don't need more accountants in my firm. Poppycock. You'd be like, tell me more. That's what working with cloud accountant staffing is. Because the team behind cloud account staffing, they were literally just an accounting firm that started hiring people in the Philippines, had a really good luck, like I did, with the quality of who they picked up, and then they just started doing it for other firms. It's not. It's kind of as simple as that. If you're looking for somebody who understands you just works with accounting firms, not some rando sending you cold dms on LinkedIn, a group that actually does this stuff just with accountants. Check out cloud accountant staffing link down in the show notes. This episode is sponsored in part by Corb Tax firms Please do not go headlong into tax planning season without the right tool for the job. You are flying blind. You are wasting time time and you're probably underselling the value of your work. Core V to help you find like a bunch of tax saving strategies for clients. Like not just maximizing retirement contributions, but a whole bunch of opportunities. But the problem for most firms is we just undersell the value of our tax advisory work across the board. You look at small tax firms, the ones that are absolutely crushing it. People we've talked to on this podcast that are working like half time making several hundred grand a year, they are making that on the back of charging top dollar for tax advisory. And I honestly didn't know how to do it until I saw a really good tool. Prove the value of your work and we kind of forget about it. Like we understand like yes, the stuff that we do is very valuable, but we forget about just how much money we can put like we regularly put back in our clients pockets through tax savings and tax deferrals. And then we go out and you like bill the time. You, you know it took you an hour and a half to do that. Here's this $400 invoice for that 40k in tax that we just deferred. That mindset shift of how to better capture that value, it's actually made way easier with good tax planning tools. Corvee is my tippy top recommendation. Other tools in the space that do this. Steph Corby is actually the number one one that I recommend. Check out the link down in the show notes to learn more. Okay, we got 10 things to go through. That was the first thing. Thing. Number two, we'll rapid fire these. Peter's got to get off the exercise bike. Number two. Copilot Vision Copilot Alert. This is Microsoft's AI product and historically all things Microsoft AI has just really really crapped the bed in a shocking way. Still to this day I don't recommend you even use Microsoft Copilot because ChatGPT is so much better, but something that they are putting out into the wild that I suspect in practice is not amazing yet, but is really the future of work is is Copilot Vision. And this, this initial version of it that is now available in preview. It can see your browser and you can talk with it like, like Voice, using your voice as you do stuff in the browser. And I've been very bullish on voice assistance, like speech assistance. If you've ever seen the movie her, it's amazing how accurate that was. In fact, the AirPod was not even a thing when the movie her came out. Yet Joaquin Phoenix, he's got like a little, little earpiece. If you haven't seen the movie here, it's a great sci fi flick, but it's basically about like having these incredibly compelling like human interactions with an AI and how helpful and how messy that can be. But one thing we learned with like chat assistance as we work on our computer is nobody wants to have to go out to another thing to ask it a question. That's just kind of annoying, right? That's, I would say that's the greatest blocker to people using AI assistants more is they have to remember to go over to this other thing and then explain the problem. And oh no, you're not prompting it right now. Like, that's just irritating. As opposed to something I can talk with with my voice that doesn't tie up my hands. I can keep typing on the keyboard, I can keep using the mouse. If I could just say, hey, assistant, and ask it this question and immediately understand all the context of what I'm working on and go out to like fetch an answer for me. Buddy, that is awesome. I mean that's like what you do with humans. Like when the, the intern is walking by in the doorway and you're like, mark, get that croissant out of the break room for me. That, that's easy to do because it doesn't require you having to stop what you're doing. And I kind of think that's the future of work is this, this running conversation with an assistant, probably with your voice murmuring to yourself in your cubicle as you're doing work. Because the alternative of having to go and like type out a message or something like that, or interact with a text chat, that's just annoying. Right now. The real version of that that's interesting is the one that lives on your desktop. And this is why Microsoft is still in the driver's seat to be the one to nail this. And it's are OpenAI's biggest investor yet they have vastly inferior AI products to OpenAI right now. Copilot Vision just works in the browser, but the really exciting version of this is the desktop version. Loads of privacy considerations with this. Obviously you can see everything that you're working with. A lot of that needs to be navigated still. But people said the same thing. When you plugged an ethernet port into the back of your computer for the first time, you're like, whoa, whoa, whoa, whoa, whoa, whoa. I'm not connecting my, my personal computer to the entire world with this little cable. It's going to be weird, it's going to be spooky, and that's going to be the norm. And if it makes me more productive, I'm excited for the path where it becomes the norm. Number three, this is just a huge quality of life improvement. ChatGPT got projects. Oh baby. Like little folders on the left hand side where you would have your conversations. Now you have projects and you can click and drag any conversation into a project. You can put like source documents into a project that any conversation inside of that project can see into. You can even create what are called custom instructions that only exist within the scope of that project. So hypothetically, and I totally don't have this, I could have a project called Pirate Stuff with a custom instruction to say, always talk like a pirate to me. And then I have like different types of content or whatever I'm doing writing in there. And anytime I'm working on a conversation within that project, it'll always talk like a pirate. It could always reference specific materials that only exist within that project. Super nice in that capacity, but also super nice in the capacity to literally just be a folder structure. If all you want to do is just drag conversations into named projects, you can also do that now, which is amazing because historically conversations in every chat assistant are just this never ending timeline where you just got to scroll down and down and down and down and down and you're like, oh, I did that thing like six months ago and it was this Australian lady that was talking about who I don't even remember value, what even was that she was arguing with like an old man. How do I find that? Well, now you can like put those conversations into folders so you don't lose them. Development number four, buddy. Google's, Google's been cooking. Google's actually launched some pretty cool AI stuff lately, especially on the creative side. I don't talk about AI video stuff here. Just know that there's a big part of my heart that is so geeked on AI video stuff, AI image stuff, if you follow me. On social media, you see all the stupid visual things I do with AI. We're increasingly using AI for B roll in my videos. In fact, we used to pay for a big expensive stock video subscription. And for the last Nine months haven't had any because we just created all ourselves and it's not that time consuming and it's funnier and like more true to the channel when we make it ourselves because oftentimes I'm the main character of that B roll. Google ships some amazing AI video stuff. But most relevant to us in accounting, they now have Gemini. They're, they're releasing their Gemini 2 generation of models which is a quote unquote next gen model, similar to how GPT5 would be a next gen model for OpenAI. But most interesting to me is what they call deep search. And right now you can actually use this with Gemini 1.5 if you have. I think it requires a paid Gemini account which who the heck's paying for Gemini right now? So maybe I just cut the legs out on, on what I just said. But what I love is they now have a deep search tool where it will go and it will reason and it will crawl the web for three to four minutes before coming back to you with a solution. I talk about how I get excited about tax research and that sort of thing. Imagine cutting an AI loose for minutes on end. They say oftentimes it will crawl over a thousand sites to get you a higher quality answer, a version of this that you can point to authoritative material. I mean it's 20 beginning of 2025 right now and like the compute available to us will continue to increase exponentially. Do you really want to pick a fight against that in the future? Like no, it's, it's just a matter of time before AI tools do a vastly, vastly more comprehensive version of research than any of us can do. Honestly, like you can make this argument for almost anything from, from, from tax prep to accounting. Like once we have the tools that can be applied in very, very targeted ways, our job will be understanding the tools enough to know like what, like what are the things this isn't going to be considering? Like what are the things specific to the client's application where it's like oh no, it would have cruised right past that. What are the questions that have to be asked to ensure that it gets to a correct result? But boy, are those going to be some powerful tools in our hands. I Google deep search. That's something that you could try right now. It's frankly why I haven't really put my endorsement behind any AI tax research platforms yet. The two main ones are Blue J and Tax GPT. We've turned out a bunch of sponsorship money with various platforms that do this stuff because I don't know that they're any better than the general tools out there. I know there's people trying them that are getting okay results, but you can also get pretty darn good results with a custom GPT with tools like Perplexity. You know, Google Learn about is an incredible tool. I've talked about it before, but if you haven't played with that yet, just go Google Search Google Learn About. It's free for anybody to use. Type a technical subject in, and it populates this unbelievably cool learning experience about absolutely anything. I'm not convinced right now that the companies in our space that have limited resources to build novel research platforms will, will keep up with how good the general tools are. If we can point the general tools to specific information, which we can do now. And so like a Perplexity where you tell it, only use authoritative sources or upload the sources to it and say, only use this stuff, it's really freaking good. And it's like 10 bucks a month. Do they have an entry level $10 a month plan? They used to. I know that the pro one that I'm on is 20 bucks a month, but I think they even had a plan that was cheaper than that. Okay, number five, ChatGPT Pro mode, baby. You ready to shell out $200 a month? That seems astronomical, right? To pay $200 a month for ChatGPT Pro. Here's the thing is that is table stakes. Once it can do the kind of things that we're talking about, because you know what? You pay humans every single month. Whole lot more than that. Whole lot more than that. So right now, ChatGPT Pro, basically what that unlocks is no usage limits right now, especially on the 01 model, which is the most recent resource intensive model in Chat GPT, you will be blocked on the number of times you can basically use that model in a given week, ChatGPT Pro, all those usage limits go away and you also get a slightly better version of that 01 model where in certain situations where it needs to, it will think up to a couple of minutes on a problem rather than a max of about 30 seconds on the normal Chat GPT subscription. This is, you know, like all things AI. It's just a huge barbell reaction from people who are like, this is ridiculous. Why would anybody pay for this? To the other end of the spectrum where people are like, this was an instant buy. Didn't even have to think about it. And this is going to become more and more the norm. It feels like we Pay a lot for software right now, buddy. It is about to absolutely, like, go to a whole nother level because our software vendors will begin framing these AI tools as agents, as human replacements. You know, think about your practice management system right now. In a recurring task that comes through every single month, there will be a day where your software platform will ship Jarvis, or whatever stupid name they give to this AI assistant. And rather than assigning that task to a human user, you will assign that task to Jarvis and it will go out and it will do that task. And the platform will say, hmm, we just, we just saved you 73 minutes this week. And you'll be like, how did you get to that number? And it'll be like, don't worry about it. Just read this email summary of how much time I just saved you. And that is how they will justify charging the replacement cost of a human. So, like, let's be real, this whole, like, AI doing human tasks thing, like, software companies are licking their chops because if I can't hire another accountant to do this work, would I spend as much as an accountant? And we've talked about this before on the podcast, would you pay software the same cost as a human? If a software could do the job just as well as a human, there are people who will say, no, I will pay the human, because that, that gives them a good job and a lifestyle and all that things. There's a lot of people who are like, I'm going to pay the software because I don't have to, like, manage the human. Like, software companies increasingly are going to be able to use this justification as the tools get more powerful than any other software we've had before, to come along and say, I'm going to charge you an absurd sum for this. Look no further than the early research tools we've gotten from legacy providers like Thomson Reuters, where when you think you're the only one in the market with a solution, you can kind of charge whatever you want. This is the value of, of competition. So if there's only one, one company coming along and saying, well, we can automate all this stuff for you, then they kind of charge whatever they want. If it truly is less than having a human do that work, once you have more competition there, they can kind of put some downward pressure on each other. But like, let's be real, this is rising the tides for all software companies big time because it's making our software much, much more useful. A $200 ChatGPT subscription, a lot of people roll their eyes on that when it came out. But that's absolutely nothing compared to, I think, what, what the future will bring. Number six, ChatGPT advanced voice is now available on all ChatGPT apps. So on the mobile apps, on the Windows desktop app, on the Mac desktop app. The ChatGPT Mac desktop app is so, so good right now. And Windows is, is behind it a bit, unfortunately, but it at least has advanced voices. We've used advanced voice on the podcast before. It's so, so good. A like mind bendingly impressive AI experience to have a conversation with an AI assistant that can sound and act that human, can mimic your own voice, can moo like a cow if you wanted to talk like the most convincing pirate you've ever heard. I had my most viral video of 2024 when that model first came out and I just had an interaction with it and I'm like, okay, now talk like a, now talk like a trucker. Now make up a barnyard song and make animal noises. And then we did like this whole roleplay scene where it was like we were in an igloo and it was super cold and they were shivering and there was a like a yeti outside that I had to kill and it was like, had like chattering teeth and it was like, oh my gosh. And we're like whispering back and forth. The notion you can do this with a computer, like just, it's wild. So Chat GPT Advanced Voice, it's now available basically anywhere you can get ChatGPT big update number seven. Along the same lines, Eleven Labs, who's the platform that's kind of the leader on creating like AI generated voices, like compelling sounding voices. I've used clones of my own voice in the past. Whenever we've done that, it has been with the tool. 11 labs, they launched what they call their conversational agent. And if you think about what ChatGPT advanced voice is, the ability to have a back and forth with an AI that's like with speech talking, that sounds incredibly realistic. 11 Labs is conversational agent. Is that. But you can customize it however you want so you can train it on your own voice. You can give it knowledge where it uses that knowledge to talk back and forth with you. If you're an OG on the podcast. You know, we've had this kind of Jason bot assistant that we developed with a tool called Delphi. We kind of have conversations back and forth with with it. It is still trained on every single episode of the podcast. So it's incredibly good at responding like I do. But this new Tool from eleven Labs. We'll link that down in the show notes. I'll put resources for all these in the show notes. I know I'm really bad at that sometimes. I will do that here. This tool from 11 labs, literally anybody can go and create their own voice assistant that anybody can talk with that sounds like you and can pull from whatever subject matter you want it to pull from. That's wild. And while a lot of these things like we're struggling with how do I implement that in a professional services context, like if we zoom out, like call centers are going away in droves right now because just the AI assistant of it all, like it's gotten so good that you don't need humans to do this stuff the same way that you used to. One of the trends I've talked about that I see growing in 2025 is a growing appetite to interact with like a robo advisor. And that could be a TurboTax robo advisor, but it can also be like an advisor that is kind of an avatar of of you. It's a first line of defense. Answering those like low hanging fruit questions for clients that do not require a bunch of technical expertise are the same questions we answer all the time, like send me a copy of my last year's tax return or, or an invoice or just super easy technical stuff. I do think that will increasingly become the norm. And as much as we want to distance ourselves from that saying no, I'm the small firm with the personal touch, the reality is, is if there's another firm with the personal touch that has a 24, 7 assistant that that client can use and still get the personal touch, they're going to do that because they can message it at 2 o' clock in the morning. They can get instant responses to many of their questions. I really don't think there is any hiding from this. It's just going to become more normalized in all of the services that we use. Number eight, ChatGPT Canvas got some really cool updates. If you're a ChatGPT user, I highly recommend you check out Canvas. And now you can literally just say put this into a canvas. It's basically like Microsoft Word or Google Docs inside of ChatGPT. And the important thing that it does is it puts the document on one side, but then it puts the AI conversation on the other side so that rather than in the stream of the conversation making it rewrite every single version of the document when you want to make changes. Now you have a static document that you're discussing on the side and kind of iterating on. So if I have say, a Sam writing a 500 word blog post and I've got that on one side, on the left side of ChatGPT, I'm having a conversation with the assistant about the document and I can say change this, change that. I can actually select parts of the document and right click and prompt it and point it to specifically that section. There's some really handy like quick tools like rewrite this whole thing to a specific grade level anywhere from kindergarten to graduate. You know, like function as my editor, like ask it for editing notes, suggest improvements. It's like, holy smokes, why is this stuff not built into Microsoft Word yet or Google Docs? It's, it's literally the best writing experience online right now, ironically enough, is Chat GPT. It's really stinking good. And you can augment it with all the other stuff that makes AI powerful, like pulling in a, you know, a, a transcript from a podcast recording and say turn this into a 500 word post or, or a transcript from somebody's YouTube video and say like pull this and that insight and kind of combine it with this and that point. Like all that stuff that's always made ChatGPT really powerful, but now in a really good document editor. This episode is sponsored in part by Stanford Tax Gang. Stanford Tax was one of the like original, original originators of cool new tax workforce flow companies who have built killer stuff around intake. Specifically I've been shouting from the mountaintops tax firms headed into 2025, you need a tax intake tool will save you a massive amount of time. Stanford Taxes is one of only a few that I recommend. The whole having to like sit down and go through when you've got a stack of pile of client, a pile of client information and you're like, okay, here's everything we had last year. Here's what we need this year. What have they actually given us? Stanford Tax Gang. It will generate a complete checklist based on prior year pro forma information of everything that you need and you can drag all those documents on and it automatically checks off all the items. A stack of 40 PDFs, drag them on there, it'll auto resolve all those checklist items, will follow up with your clients until they get you all the stuff that you need and just completely eliminated that entire step of doing today 1040 work where you're like, do we start the project yet or do we follow back up with the client for more information? Absolute game changer, Big fan of Stanford Tax. Check out the link down in the show Notes this episode is sponsored in part by Financial Sense. If you were plugged into my 2025 practice management system recommendations, you would have seen Financial Sense on the list. I know a bunch of really smart firms that run Financial Sense. I have. I've been on record as saying it is arguably the best small firm all in one. Er, right now it's really hard to have a PM that does both tax and accounting work well. And while I'm not a believer in all in ones that do everything from tax intake to nailing proposals and all of that, what I am a fan of is the platforms that get closest to that. Right? Because at the end of the day I do want all in one and I also want them to be investing in a more all in one path to where Maybe right now that feature isn't good enough for me to replace that other tool, but I could see how two or three years down the road it is. I want that as opposed to the platforms that aren't investing in that way. Right? Financial Cents has actually absolutely been crushing their marketing the last couple of years. They run an annual thought leadership event for accountants. It's really good that I've been a part of. They're putting out actually really good social media thought leadership as well. Not just around like buy my product, but around better ways to run small firms. And those are the brands I want to get behind. Like those are the brands I want to work with. They got a smart vision for the right way to do it, the products, and to be a reflection of that. Learn more about Financial Sense. Check out the link down in the show Notes this episode is sponsored in part by Maker's Hub. If you've tuned in to much of my stuff, you know that I'm a big fan of how Maker's Hub is using AI to crush the accounts payable headaches of accountants everywhere. Arguably the domain and accounting DOM that is being most disrupted by AI, well, they recently shipped support for purchase orders. Good golly gosh how much I hate purchase orders, but they are a necessary evil for companies of a certain size. But it turns out that purchase order matching, you know, matching the invoice to the PO to the receiving slip, whatever you want to call it. Turns out that's way, way easier when your AP tool does like rock solid line item extraction. Got my hands on with with the PO system that they just plugged in recently. Super, super impressive. If you have clients that struggle with this like I did. Oh my gosh Makers Hub is about to probably save them a whole heck and lot of time. Learn more about that one. Check out my I've got a couple of videos on makers Hub on YouTube that are like quick 5 minute demos. Otherwise I'll link their website down in the show notes. Okay, two more to go. Number nine more people, more accountants are using AI tools ChatGPT every day than ever before. Which unfortunately is just a really, really big deal. A bigger deal than I'd like it to be. Because as much as I would love for the security conversation and all that to be totally first principles and be like no, here's why it's secure or not secure due to these these points. Unfortunately most people will pick up a tool because they see their peers using the tool. Even IT groups. They will be scared to like be the first one to recommend something. Insurance companies, my gosh, there are so many people who are still Spooked to use ChatGPT or put client data into an AI tool or something like that because of what other people are saying, not because on a first principles basis it's lacking something. The whole notion of, of your prompts being trained into the model on the professional version of like all of the mainstream AI tools now that's, that's no longer the case. Nothing has been trained into those models from prompts from paid business users in like a year plus. So I point people to the Chat GPT team plan it comes with, you know, your prompts are never trained into the models. You got some cool kind of team management controls SOC 2 type 2 certification from the AICPA. The majority of the practice management tools that we use that house our tax data and all sorts of sensitive things. All of our emails, the majority of those practice management tools are not SOC 2 type 2 certified. A lot of them are type 1 certified, a couple of them are type 2 certified. But chatgpt like, like I struggle to be the guy shouting from the mountaintops. This AI assistant is more secure than the other tools that you use. Ah, but it's kind of the case now. The, the biggest source of data breaches is unauthorized access. It's somebody being able to log into your account who shouldn't be there. Multi factor authentication makes that massively more difficult. ChatGPT has that you can enable multifactor authentication. I I tell people don't take security advice from a YouTuber, but here's what I would recommend is you define or challenge your IT group to define what are all of the things that would make an app secure enough for me to use. And then let's put on a grid. The, the rows are all of the security properties I need and the columns are all of the apps that I already use. Stuff like important stuff like you sock two. Is there a certification? There is a type one. Is a type two. Is there no certification? Multifactor authentication. Can we enforce multifactor authentication or does every single user have to opt into it? Could a staff member of mine turn it off without my knowing? Like important stuff like that. What happens to our data if we want to stop using the service? Can we delete all that data? How long is data retained in the system? Can I have like ephemeral chats and ChatGPT? You can hit a toggle and have an ephemeral chat and all the data is immediately like deleted. It goes away so that if somebody logged in to that system three years from now, they would, they would never be able to get to it. I would challenge you or put this to your IT group, put all those things on a grid and then put ChatGPT up there. And I would hazard a guess that it will actually stack up pretty darn well to all the other tools that we use. But unfortunately for most people, when I say most people, I don't mean most of the folks that listen to this, the normies, trad accountants. Trad accountants. They're going to make buying decisions based on what other people are doing, which is not first principles. And it's kind of a bummer. And that takes time for that ship to steer and we can all say the right things and it makes a difference bringing people around. The unfortunate thing is everybody DMs, me, they're like, we're using this. We got absolutely no problem using it. And it's doing all these amazing things for us. And I'm like, we cool? Can you go out and say that in public? And they're like, no, I don't want to get canceled. People, people don't get it. And we've done the research and we know it's fine, but it's, it's kind of unfortunate. That's where we're still at. Okay, number 10, last one, arguably the biggest one. OpenAI announced this right before Christmas. It was the O3 model and they had to skip 02 because there's a UK telecom brand by the name O2 and so they're jumping straight from 01 to O3. But it is a new in house model that they, well, if you can call it a model that they haven't released Publicly yet it is still our understanding is built on GPT4O in the same way that 01 is. But O3 is like a better approach to reasoning to how it thinks and considers a problem before giving you an answer. And the big headline here is there's a benchmark benchmark called ARC AGI and it is a benchmark that supposedly measures how close is a model to artificial general intelligence. And the headline was for the purposes of this benchmark AGI where most humans come in as a score of about 75 and oh 3 scored a 78. So the headline was OpenAI just reached AGI internally. It's such a big headline that I don't know that anybody can really get behind it. And they're like really? Did they? And, and what's changed? Even though they did like we can't, we don't actually have our hands on the OH3 model. You read the fine print, more details came out there. Uh, we've talked about the amount of time and, and compute you put into letting a model reason about a problem and think about a problem in order to give you a higher quality response for this benchmark test. How long you think that that thought about this problem or these problems. Um, in ChatGPT right now it'll think about a problem for about 30 seconds. If you upgraded Chat GPT Pro, it'll think about a problem for about a minute and a half. O3 reasoned on this problem for 16 hours. Was 16 hours or 18 hours a long time. And they spent. Drumroll please. $1.6 million in compute on the benchmark. That is they had had a whole bunch of GPUs servers spinning, their little fans thinking their little brains out to the tune of consuming $1.6 million in presumably energy to get there to be slightly smarter than a human. And it's a whole lot more expensive and a whole lot more time consuming than a fellow human is today. But this stuff is getting exponentially better and more cost effective every single month. And so it is, if anything it kind of shone a light on the reality of what our progression into this era will look like is there's going to be an astronomically expensive out of touch version that is smarter than a human before it lives in my pocket. Right? But like the fact that it exists is, is more proof that we're on a path to that which is pretty darn wild. Like it, it, it really is. I mean ChatGPT, it just turned two years old. Think about that ChatGPT and the original version of ChatGPT was a real dummy. Like it was novel in so many ways and useful for copywriting and some stuff like that. But that was built on the GPT 3.5 model, which was not amazing. And in that amount of time we have something that people can argue is AGI, you, you and I don't have it. It's still astronomically expensive and out of reach, but that's pretty darn wild. And it's doubling down on what we started talking about here. Reasoning time or inference time. Giving the AI model time to think about that problem and kind of work through a solution before it gets back to me. Rather than immediately blurting out the very first answer, which is not the highest quality answer. It's gonna be interesting like, because most of the apps that we use are, you know, you would call it synchronous. That is, you click a button and a thing happens. It's not asynchronous. We don't cut our apps loose to go work on a problem and then get back to them. We do that with humans, but we don't do that with software. But the future of software probably looks more like working with humans. And so AI agents cutting a agent loose to work on a task for an open ended amount of time and then bring a response back. To me, it looks more like working with humans than it does software. In fact, an example of tech we have that people are already using a tool called Devin, which is a tool for software developers where you cut an AI agent loose to go and work on a software development project and it will browse the web and, and pull from, you know, like technical documentation and all that to work on a problem that you define and you give it either on an existing code base, an existing tool or, or something brand new. And the way that you work with Devin looks a whole lot like iMessages or WhatsApp. It's actually a stream of conversations because it reports back to you and you tell it what to do and it goes and does that stuff like it's like texting somebody. Like you can drill in to see what it's actually doing. But because, and Devin, this is just one example of like the user interface for these agents because you can cut any number of them loose at a time, like within reason, but theoretically cost is the blocker here because you can cut any number of them loose at a time, then what you're doing is actually orchestrating a number of, of agents like with instructions and saying it reports back with a question and you send it off to do this thing and this common talking point in AI and what the future of AI looks like is, you know, the Bill Gates of the world are like, well, AI is removing the user interface. We're just going to talk with assistance now, and they're going to get all of our work done. And you're like, you telling me that the future of chat, the future of QuickBooks is ChatGPT. Like, I'm just going to talk with this thing back and forth, like, how's it going to give me reports and how am I going to do the stuff I normally do in a spreadsheet with a conversation? And it's like, no, it still kind of feels like I'm still going to need my apps and my comfy. My comfy kind of accounting stack, because otherwise how's it going to present all this data? But I tell you what's bringing me around to that worldview, the notion that you're just going to talk with AIs to get all that work done is if I am in an office with a bunch of other accountants, my interface to all the other humans that I work with is talking about stuff like, those humans don't have a user interface. I use my mouth or I message them in Slack or something like that, and we talk back and forth. And AI agents are starting to look like that, where what you're doing is actually managing a set of conversations with different AI agents that are doing things for you. And I'm not micromanaging them and seeing what are you looking at and what are you solving here, what are you solving there? Like, unless I need to hop in and troubleshoot, I'm not micromanaging that. And so in that timeline, the only scenario where I really need to get into the weeds is when the AI can't do something or I have to troubleshoot something. And it's pretty wild to think that right now, if I ran a accounting firm of 20 people and, you know, I'm looking at my Microsoft Teams or my Slack or something like that, and I can message any of those 20 people, we're probably three to five years from having something very, very similar with things that are not people. Weird, right? Super, super weird. And what will those things be capable of doing and what will they not be capable of doing? Makes your head spin. The reality is, and this is the good news, if it's. If that amount of change is spooky to you, the reality is, is the day that some unbelievably good AI model drops, your world's not going to change all that much. But because the world takes time to course correct. In fact, if we got an AI model tomorrow that could quote unquote file a tax return. Well, first you have to ask what does that even look like? Does that mean like putting together a perfect XML file that will go to the IRS and represent an E file filing? Because that's what my tax software does, is send this structured XML file to the agency. So what would that even mean? Right? A bunch of input docs. Well, who's asking the questions? Is it going to ask the right questions before it sends the filing off to the irs? Like even if you have the perfect AI tomorrow, it takes time for the world and for human beings to adopt and to course correct. And I would argue right now we already have, and we already have a PhD level AI assistant that's available for $30 a month. But most accountants don't think it's worth it. So I would argue we already have, you know, a scalable PhD that we're not putting to work because we can't figure out what are the best ways to use it right now. And to be clear, the cool stuff our tech partners are developing for us, they're all built on the exact same models that drive ChatGPT. So tax GPT, the research tool, they got a whole bunch of research materials and you use GPT to interact with it. Well, I can go out and I can grab like the IRS publications, like there's ways that you can kind of do versions of this yourself, but it's all using the same models that we have access to in chat GPT. And so I'd argue we, we're already struggling with this. We have this really powerful thing and we're not totally sure all the ways that we can deploy it to make us more productive. And if that same thing was twice as smart, would anything change actually? Or would we still be blocked by the same struggles of what are the right and wrong ways to use it? To me, it is just one more argument to put it in people's hands. Like do the security, due diligence, make sure that for where your firm sets the bar, it ticks all those boxes. And for 90% of firms, I honestly think it does. But then put it in people's hands, share what's working and what's not working. Because you think this is going anywhere, like you think this is going to disappear in a year like crypto did. Absolutely not. What we're blocked by right now is use cases and people getting a comfort level with these tools so that it will make them more productive. We talked about this before. The high school and the college kids that are coming out of school right now, they are freaking wizards with ChatGPT. Can you imagine if somebody plopped ChatGPT in your lap when you were still in school? You better believe that thing would have been your best friend. Not because you're going to quote, unquote, cheat with it, but because it will make you more productive in just about every way. Right? And so we add these people entering the workforce who have spent hundreds and maybe thousands of hours with AI assistance in a way that we have not. And then the big bosses sit in a room and they're like, it's not useful to me. No, you can't, you can't use it. And you're like, no, no, you're missing the point. This is like a new way of working where people need to get comfortable with it. And that's going to take time and it's going to take some training and, and repetitions and sharing the journey with other people to learn how to put these things to work in a very contextual way for how you work, to solve the specific to you problems that no software platform's ever going to solve. Because there's so many workflow things. I mean, if accounting was a bunch of rote, repetitive tasks, it would have already been automated. But it hasn't been because we're applying specific client things to really messy regulations and all that. And that's where ChatGPT, like personal AI assistants, I think, will have relevance for a really long time because they can solve the more personal workflow issues that maybe one or two clients have that no platform's ever going to solve for, because it's not a problem that everyone shares. But yeah, step one, just get in people's hands. Play with it, experiment with it, share what you're learning. Find other accountants who are excited about it as well, a ton of them. Tune into this podcast. Being able to multiply your learnings with a hundred other people who are like, jazzed and trying to figure out the same things. You better believe that is going to fast track what you are learning and the ways that you are putting it to work. And then you're able to kind of trickle that down to your team and get them excited as well. Because work is changing, man, and like in the, in a really exciting way. Okay, we indulged ourselves today. We don't do this that much in the podcast. A lot of the really early episodes of the podcast were just like, oh, let's just geek on AI today. But we try to keep things practical today. It was an exception, but I truly believe the folks that are like hanging out here and tuning in like you are so far ahead of the curve on your understanding of AI and where the profession is going that when somebody like takes a cheap shot on LinkedIn or you even get a keynote from a non accountant at a conference and you're like, do I need to be listening to them or taking them seriously, the reality is, you, if you're listening to this, you know a lot more than you probably think you do and a whole lot more than like the average accountant in the profession. So good for you, man, because it is you who will be like, defining what AI and accounting looks like if we all sit back and wait for the software developers to do it. We've all seen how that works before, right? So let's go define that future and help a bunch of people along the way. Gang, thanks for coming and hanging today. If there's anything you're excited about, AI Wise, drop it down in the comments. Keep that conversation going and I'll see you in the next one.