139 What OpenAI's Latest Announcements mean for Accountants

Mom, get the camera. We just got a bunch of new OpenAI stuff. OpenAI is having their, their first like conference right now. Sam Altman just did the keynote. I'm recording this on. What is this Monday? I think this goes out Wednesday. They just dropped a whole ton of stuff, like way more than I was expecting. New AI model, new build your own GPTs with like natural language, like normies can do it. Come on in, friends, let's talk about it. What does it mean for what we do? Any short term impacts, let's do. So just a refresher. The last time OpenAI had a live stream, they prepared somebody's taxes for over a million viewers live. And that was with the announcement of GPT4. So anytime there's going to do like a big live stream or announcement, I try to tune in because it feels like I should. And so this was the second one they did this year. It's not very often that they do events like this, so people were excited. There were some leaks before the event and some of the stuff that they had announced had kind of come out, some of it hadn't. I was shocked at just the volume of stuff they cruised through like it was rapid fire. So I'll run you through that kind of the rapid fire list of notes I have, but then we'll spend the meat of this on the, the big release. What they're calling, quote, GPTs or your ability to build your own, they're calling it GPT right now. But they're also saying it's a precursor to agents, which if you tune into my stuff, you know, agents are like the holy grail kind of where we're going, these autonomous agents that go out and do work for you. And so a big part of the talk was about that and how those are on the way and how to roll those out responsibly. But first, the rapid fire kind of roundup of stuff. So some updates to ChatGPT. Big one, buddy. We got a brand new AI model, GPT4 Vision. So it's got the eyeballs, you can throw images at it. Turbo. So it's faster and it has knowledge up to April 2023. The previous knowledge cutoff was September of 2021, which feels like a long time ago at this point. And that was like eighth grade for me. So we got a new faster model and it knows stuff up to, I don't know, five or six months ago. And they said they will never let a model get that old again. Now the other big change to ChatGPT which there had been some murmurs about already, is when you're in ChatGPT and this is a bigger thing on plus the paid version you have, you like pick from a dropdown which model you want to use and that's now totally gone. So you know the data analysis model you had like the Dall E3 image generation model, all of that stuff has been rolled up into one user interface, which is either awesome because it just kind of knows what you want to do from a single chat interface, or it may be a pain. Some early testers of that had shared that it didn't always exactly know how best to do or like what model to use to achieve the task that you gave it. Long term, this is great. Like it's the way the way AI models are going long term is actually the models we interact with are more like a controller model and it's that model's job to then hand those tasks off to other AI models. Depending on what it is. Am I asking it to add a number? Am I asking it to write a book? You have this controller model that then decides, okay, this, this model and that model are actually best suited to do this job for you. It's actually part of the secret sauce of GPT4, most people believe is it's actually a collection of quote unquote expert models, not just one giant model, but it actually carves up tasks and has has them be done by various different models. Which may be why, like there's a just a fundamentally better user experience with GPT4 and ChatGPT than any other language model tool out there right now. So big updates to ChatGPT, that turbo model, GPT4 is going to be faster now. And then the model picker going away. They also announced that there's going to be a version of GPT4 with over 4x the context limits of what it currently has. It can now basically contain like upwards of 300 pages of a book. Is effectively how much information that is now. Now historically they haven't opened up those larger context models in ChatGPT. So if you use ChatGPT, I don't know if we'll actually see that, but for developers who are building stuff on top of large volumes of information, that's great. A cheeky one that some people like just kind of cruise past that was almost this throwaway is they're launching version three of Whisper soon. And so Whisper is their speech to text model. So if I chuck like a zoom recording or an audio file at it, it returns the text transcript and let's not forget that I think it was late last year, maybe even the beginning of this year, when we first got Whisper, I think it was last year. This was a massive step forward in automated captions. And it's actually the secret sauce behind why AI meeting assistants are so good these days is before, when the automated transcripts were just so inaccurate, like that thing wasn't gonna be worthwhile. Now those automated transcripts are really good and they're about to get much better with version three. So that is awesome. And another really big thing there is the notion of how that like eliminates language barriers. It's something we don't talk about enough with this current crop of AI. As we're getting more and more chatbots that we can like talk to in real time. Like they have like speech processing, you can do that back and forth in any language. And like we are super close to a reality where you've got like that universal translator in your ear. And that's going to be absolutely wild. Three things that are available in the API for developers today. That new turbo model for GPT4 Dall E3, the image generation model, for the first time, you can use that in the API. Before it was just available in ChatGPT and then their new text to speech model. And this is what's used for the speech, the talking back and forth. So now the people who build your practice management system or any other software that you use can now chuck some text at the API and actually get speech back. Like the same way that we do this with ChatGPT. So a few updates around fine tuning that I won't bore you with, basically it's gotten better and there's now experimental access to GPT4 fine tuning, which is pretty interesting. Big misnomer around fine tuning. This phrase gets misused a lot. Fine tuning is for directing a model on the form of what it returns, not what it knows. So if you want to like build knowledge into a model, it's not done with fine tuning. Fine tuning is more like if I always want it to respond with an account from the chart of accounts, you know, like an accounting category or always with in this very specific format or something like that. That is something that you would use fine tuning for to control the form of what it returns. And interestingly, you can actually, literally anybody can go out and fine tune their own model without code on the OpenAI website if you have the data to do it. Like it's totally accessible for non developers. So if you have something where you like you want the language model to maybe classify a group of data or something like that. Fine tuning is used to get it to adhere to a specific form of like, what it returns. They said they're going to work with a few select companies to train custom models from scratch. But basically by what he said, he's like, we're only going to be able to do it with a few people and if you have to ask how much it's going to cost, don't bother. They announced copyright shield. Number of companies have announced something like this. Now, basically, if you're sued under a copyright claim for anything you generate with their tools, they will come in and like pay your legal costs, basically, Gang. This episode is sponsored in part by Copilot. Are you sick of Copilot ads on this podcast yet? Because you shouldn't be. Because bless their hearts, they're making all of this magic come together and all of this possible, you know, Okay, I know we got some hyper nerds in the audience out there. If you saw our practice management report that we released recently, Copilot was on the list because they've got this hyper flexible approach to building a client portal and they're not getting in all the other stuff, the workflow and all that. All they want to make is a super flexible portal platform for you. And if you are a mega nerd, boy, wouldn't it be nice if all these tools in our space had killer APIs so that we could automate all this stuff automatically? Wouldn't that be nice? Not a thing though, right? Wrong, It's a thing. If you use Copilot, they've actually got a killer API that lets you work with pretty much everything in the entire platform. So the notion of building some like fully automated stuff where before that wouldn't have been possible. It is. It's possible with Copilot, it's possible. But let's say you're just a little bit nerd, not mega nerd. Your little nerd. Through stuff like Zapier and make, you can actually build a bunch of your own automations. In fact, I got this app, your page up here and there's what the cheese. There's an absolutely shocking number of triggers and actions. Hang on. Pause. Quick count. Break 35 triggers and actions in Zapier for Copilot. If you like building your own automations. Pretty cool, right? Hey, if that is exciting to you, check out the link to Copilot in the show notes. And thank you to Copilot for your continued support. It's that time of Year again, gang. Halloween? Nope. Thanksgiving? I guess it's soon. But worse than all that, it is time to fulfill your annual continuing education requirements. That old chestnut. This episode is sponsored in part by Pronto Tax. We know how that CPE deadline usually feels, right? More boring than watching paint dry. CPE really do be like that, especially virtual cpe. Man, you usually can't even see the person that's talking. It's like this awful, like, set of slides with these fade animations between each of the slides. And then you hear this like super corporate voice, which frankly, might even be AI Actually, if it isn't, it absolutely could be. Well, are you ready for better cpe? Me too. If you're ready for a continuing education experience that's different than anything else out there and not incredibly boring, yet still highly educational, you could be ready to give Pronto Tax School a try. The people at Pronto are a little crazy. What? Their training's not for everyone. Things can get spicy. But the people at Pronto really know their stuff when it comes to taxes. And they make a valiant effort to make completing your annual continuing ed at least halfway entertaining. They're approved by irs, CTEC and nasba, so that you know you're getting valid credits that match your credit credential. And over 4,000 tax pros choose Pronto's CE Slashpe solutions every single year. Visit Pronto tax school.com or check out the link in the show notes to learn more. The big Announcement GPTs so big GPT Little S the notion that you can build your own GPT with natural language, which is kind of the holy grail, right? Like, okay, cool. I see that this chatgpt, like, it does this cool thing in kind of a general way. But what I really want it for is maybe this specific domain or this body of knowledge and to always talk like a pirate in this way. Like the notion of being able to build your own version of that. And there's kind of hacky ways of doing that with prompt engineering and third party stuff like vector databases and chunking and stuffing and like all this developer Y stuff where you can kind of sort of build knowledge into it. But amazingly, like we're. I mean, we're not even a year from when ChatGPT launched. The idea behind this is that anybody can build their own version of a GPT and either keep it private or even share it publicly. Like you share a link with somebody else and they can use that GPT. Pretty interesting, right? And the framing they provided for this was basically that you Know, agents, A G, E, N T S. Agents are, you know, this notion of this autonomous, autonomous agent that will go out and perform tasks for you using the systems that you use and all of that. If you've been around my channels, the podcast, the main YouTube channel, we've talked about agents a lot and that this is kind of where everything's going. It's probably the fastest moving thing in AI and is ultimately like, if you've ever looked at like a chat bot and you're like, that's cool. And maybe sometimes feel like, feels like a party trick and sometimes it's useful. Agents is where the rubber meets the road in a big time way because they are actually doing tasks the same way that a human user does by communicating with a large language model and reasoning through like, what should I do next, given the task that I was instructed to do? And so that could be ad hoc tasks of me just saying from QuickBooks, you know, go out and download a 2022 balance sheet and P and L cash basis. I've shown before that we already have browser agents that can do that stuff reliably. They aren't particularly cost effective, they're pretty slow right now, but it will open a Chrome browser in the background and do that thing for you. So like we're already getting there. Like this isn't a sci fi thing. But they were very careful in how they talked about agents and saying, you know, this is something that we have to do slowly and be really careful with just because of the reality that regardless of how you feel about AI and you know, the Terminator stuff versus the job displacement stuff, the reality is as soon as, you know, to use an example in our space, as soon as somebody drops an agent that can go through, you know, a PDF of a bank statement and a QuickBooks reconciliation and can reason between those two tabs and mark up a PDF and go through the QBO UI and reason through that. Like, and we're really darn close to that. As soon as something like that happens, that is a huge scale task that happens in bookkeeping. That changes a lot. Like it changes a lot really fast. So we're kind of at this weird tipping point where like, as soon as these agents are good enough, a lot of stuff is going to start changing really fast. And so OpenAI didn't go as far as to call these things agents, they're calling them GPTs. And they framed it as kind of a step in the direction of agents because that is where we're headed. But to like baby Step into that responsibly. Let's start with this thing that we're calling GPTs. And so the notion behind what a GPT is, or building your own custom GPT is, you give it a set of instructions and you can give it expanded knowledge, so you can build knowledge into it and you can set up like, custom actions that it can do. So they ran through an example of how you can actually have a chatbot conversation with the thing that will build your GPT for you. So it starts off by asking you, like, tell me about what you want to build and who it's for. And so like, he went through an example of being kind of an advisor for startup founders and he said he wants to make it to help startup founders think through business ideas and get advice. After they've gotten advice, grill them on why they aren't growing faster. And then it responded, how does he feel about the name startup mentor? Then it generated an image for the GPT and then asked some clarifying questions, like what should be emphasized or avoided when interacting with users. Now, the biggie here, I think the most significant here is use. You can upload files that will serve at it as its body of knowledge. And we've talked about this a lot. The idea that GPT and any language model, it is not a knowledge base. Like, we shouldn't really trust factual things that it gives back to us. It's amazing for ideation and creative things and all of that. But it honestly doesn't really matter if the answer it gave you back is right or wrong. Because what I'm looking for is not an answer on technical things from AI, but like the source material. So we don't, we shouldn't be coming to AI for answers. We should be coming to AI and showing it how to research the same way a human does. So I can point it to an authoritative body of context and say, give me verbatim, word for word, the most relevant quotes from this thing to the question I'm trying to answer. Pulls all that stuff for you, gives you citations like, like a research assistant would like instantly. That's what I want from it. I don't want it to say, yeah, no, this, it looks like this or that is the answer. Because it's irresponsible for me to blindly trust that as the answer. And ultimately I don't think that got me any closer to the answer. And maybe down the road it's going to be super mega smart. I can do that stuff. And we could just trust the black box aspect of it in the same way that when we first got software to prepare tax returns, we were like, well, I can't trust the output of this stuff. And it probably did a lot of stuff wrong. And frankly, it still gets stuff wrong. But in the beginning there we were like, I'm sure we were like, I don't, I wasn't there, but I'm sure we were like preparing it on paper alongside and then comparing it side by side with the computer output. And then at a certain point you just kind of start to trust it. And there'll probably be a day where like the black box of AI, after a certain amount of time, sure, maybe we just trust it, but we're definitely not there yet. Right? Like, we can't just blindly trust the output. So I need it to draw information from the specific knowledge that I pointed to. And to date we've been able to do this with like the browsing model where I can link, you know, an IRS publication and say, just give me the answer according to this via, you know, a list of direct quotes. Like, that's been the best way to use it so far. But these GPTs, now you can like build into them your own custom body of knowledge. And in the example that they built, he loaded in some transcripts of talks that he had done, like giving advice to founders. So the age old, like, Jason GPT thing that we've talked about for a long time, where we pile the transcripts from this podcast into a GPT like that just got really easy. What I don't know is if there's like an upper limit to the volume of knowledge that you can give it. Like all it was was literally like attach files and it would let you attach files. So I'm not sure what the ceiling is on that in other applications. And it'll depend on kind of what the technology is that behind it that's actually servicing that. Because the language model behind the scenes is still just the language model. What they're doing is like injecting external additional information to the prompt. And in other applications of this, basically how you had to hack this together in the past, it couldn't handle, like, theoretically no limit on the volume of knowledge that you could give it. So I'm not sure if these OpenAI GPTs will have a limitation to that knowledge or not. But obviously, like my mind goes immediately to what we've talked about, which is like, what does an avatar of your expertise look like? It's not a replacement for like sitting down and talking with you, but the reason that we do stuff like write blog posts and that sort of thing isn't to get people to not come to you and isn't to get them to take and make investment decisions and tax positions and all that just based on the stuff that you put in your blog, but as a way of getting a degree of access to your expertise while you sleep without you having to do any extra work. And like being able to build a chatbot around your documented expertise, man, that's a big upgrade over some blog posts as far as I'm concerned. Like, that is a big step towards you, the real deal. And this podcast is an easy example. If you got a like, very specific question for me, like, what is Jason's thought on this sort of thing? How the heck are you ever gonna find that? And like going back in podcast recordings, right? But the notion that you could chuck that thing into a chat pod and immediately get answers based on the most relevant things I've ever said across 100 and whew, how many 120 some 30 minute podcast episodes, that's pretty compelling. Right now, most of us don't have that big of a body of context. You don't have like that much that many transcripts, that volume just sitting on the shelf. But what these GPTs will do is like further embed chat experiences into everything. So you're going to have more and more accounting firms where maybe that's just a little assistant that's on the landing page and anybody can use it, or maybe it is part of a package that is like a more limited access sort of offering. Again, not as good as the real deal and like, not as good as sitting down with you one on one, but. But depending on how documented you are, could still be pretty good and changes the mental math. Or it ought to change the mental math of how documented you are going forward. We talk about, you know, creating a podcast and stuff like that. What we're talking about right now was a big driver in the reason that I started this podcast was that ultimately the ability to extend yourself with AI is going to be a product of how well documented you are. And what better way than a daily podcast, right, that covers pretty much every corner of running accounting firms that you could imagine, Gang. This episode is sponsored in part by Live Flow. Live Flow is the easiest way to sync that QuickBooks data back and forth to your spreadsheets. You may have seen this actually had a big announcement lately. So this fall, G2 gave them the top spot in their fall 2023 report as the leader in the financial analysis category. That's right. They won number one. Nice work. If you've been around my channels for a while, you've seen Live Flow kindly. They have sponsored quite a bit of stuff. And I'm not, I mean, I'm not saying I'm taking credit for it, but that was probably why I have nothing to light with Live Flow. Super easy way to sync that stuff. Sync your QuickBooks data back and forth with Google Sheets. They got a whole pile of templates too to make the process of building that stuff for the first time as easy as possible for you. Stuff for managing cache, AP, KPIs, like everything you can imagine. Sync that data into your existing sheets to make them smarter, get it to auto sync or build your like custom new sheets that talk with QuickBooks totally from scratch. Pretty cool tool. Check that one out at Liflow IO. This episode is sponsored in part by by the fine folks at Cloud Accountant Staffing. Do you hire accountants? Bless your little heart. Not the best part of the job in my opinion. Not something I ever enjoyed. Well, listen, you can build your accounting dream team with talented offshore accountants in the Philippines that work 100% full time for your firm. Their accountants aren't freelancing or contracting for multiple firms. They're all yours. They work exclusively for you and are incentivized to stay with you and your team long term. They're not gonna get swiped. Cloud account staffing is 100% dedicated to the accounting industry and founded by a former accounting firm owner that understands your business, knows your pain points. They had to hire some accountants and they said, you know what, we're gonna build our own pipeline in the Philippines, gonna pull in some super talented people and then open that up to other firms. Basically that's the story I've been talking about a lot about staffing, building more resilient staffing pipelines for your firms. I had staff in the Philippines like totally red pilled me to like oh geez. Like we need to globalize the way that we get our work done. Check these folks out. Link in the show description Cloud accountantstaffing.com and so today I think like the people that have the advantage are the ones that are really, really well documented. But it's absolutely a part of running professional service firms. If you know you've got two firms that otherwise look the same and one will give you access to a chat assistant that you can ask questions of 24, 7. When you think about that thing in the middle of the night, that's really compelling. Now the big thing this doesn't solve for is specific to a person context. So it can answer general questions, but it's still like, you know, your blog posts on steroids. It can't see Steve's tax return, it can't see Steve's financial statements. And we talked about this maybe a month ago, but I see a huge opportunity right now for the people who have that context to create some really interesting client facing applications. And in the accounting space, that's probably practice management systems, maybe it's document management systems as well. But if you think of who today are the biggest holders of context, from email communications to files to projects and all that, there's an interesting opportunity there where we already have these one to one client portals where the clients could then have like isolated conversations that can see into the data that is already in their client portal. I don't know. As the practice management software ecosystem I think has gotten more mature, man, it's grown so much in the last like three years now you've got a whole bunch of tools that do like really compelling things. Five years ago when I was picking a task management system, like I went to ClickUp because the space was so immature. And that's absolutely not the case. Now if you got that practice management report we did a couple weeks back, like even the little guys on the list, like have a ton of features. But as that, you know what it means to be a practice management system gets a little more crowded and eventually more commoditized. A really interesting differentiator I could see right now is, okay, you already connect to my client's ledger, you already have all of my client files, you already see all the emails I've ever had with this client. At the very least, give me an assistant that can see into all that stuff, specifically the client's context. Maybe even expose a chat experience to the client in that mobile app, in that browser. That's pretty darn compelling. That's something frankly that I could probably charge for in my accounting firm if I can enable that access for my clients. You know, a really frustrating aspect of being in public accounting. And I'm a US tax guy. So like this, that's about as bad as it gets. But a really frustrating aspect of all this has always been being subject to the limitations of the tools that we have to do this work. And so often like you work with people and they're like, goodness sakes, why does this have to be so hard? And you're like, I know, I'm sorry, my tax software is 25 years old, I, I realize how dumb it is, and trust me, nobody wants it to be better more than me. But that's, that's just the way things are right now. But the notion of having an AI chatbot that could see into your stuff in a secure way, where clients could like chat with it and have questions and you can see the transcript history, that's really interesting. Now the, the biggest stopper and the biggest, you know, fear people have is, is that akin to giving somebody advice? Because obviously somebody acting on the advice of this avatar of your expertise, like, your mind immediately goes to like, well, shoot, if that's exposure for me, then I don't want anything to do with it. And it's still really early days on all of this stuff. And it's a little annoying to me how skittish, you know, your carrier is and IT groups and stuff like that. And you know, we did that AI usage policy episode the other day, and it's easy for all those groups to just say no. Like, you just, you can't do it, can't use it. We don't know enough about it. And that's, that's really frustrating because that's, that's a lazy answer. But I suspect where this, like, stuff goes is, and like kind of everybody's thinking around this is still kind of developing and getting there. But when you put a blog post on your website, is that you giving tax advice to a specific person? Like, I mean, I think we kind of have the understanding that, that no, it isn't. If somebody's talking to Jason GPT or, or your firm GPT and they ask it a question, is that tax advice? I think what we have to get an understanding of is like, no, it fundamentally can't be because you're not talking to a human being. And I don't think people quite have that understanding yet. But that's, that's where we need to get to AI ultimately is not a replacement for the accountability of a human giving you that advice. And if we can get to that level of understanding, that's really valuable because you know what, that chat betters, that chatbot of yours is going to be a whole lot better than Google WebMD. Like, the fact that people are gonna get these answers somewhere one way or another. But the more we can contextualize those answers because it can see into their tax filing history, it can see into their ledger, we can pull in context from quick finder. Pick a company, somebody who can like, provide a knowledge base for common questions like, oh, what's my IRA limit going to look like this year? And they're like, well here's the stipulations that matter and what that could be. You start rolling up all that context in a place that can see into client specific circumstances. That's going to be really helpful for the client. It's also going to be really helpful for me because even if I don't have a system that will just give a client answers, that is a hecking great assistant for me who can then like fetch all that information for me along with citations to like where, where the source is to help me answer that question in a much more efficient way. Right? So like if, if by the time I get to that email it's already gone out to, you know, the language model that has context from, from, from Quick finder and their QuickBooks file and their past tax filings and says according to this, this and this, like here are the different details that ought to inform your response. Even if that's not a chatbot that's just responding to them, that's still really helpful to me. Right? So I the biggest, most exciting thing to me here is their launch of quote GPTs. I love that they launched the API for this that developers can use at the same time as they launched a version of it in ChatGPT that anyone can use. Literally just by talking with a chatbot they can develop their own GPT, which I love. It's not something that's like, you know, gated for developers. The ability to build custom knowledge in there is a big deal. The other, frankly the other big thing here, which is more of a developer thing is custom actions and function calling. The notion that these bots can also like make a call out to an API. And so two examples they showed. One was with Canva, the notion of having a conversation back and forth with the chatbot to like help generate some visual designs for promotional image or something like that. The second one which is, which is really interesting was a zapier example and Zapier, if you don't use, it's an app that connects. It connects with like something like 8,000 different apps or something like that. And it's kind of a, a way to customize automations between different apps that you use. Couple examples they showed for a GPT that was connected to that person's zapier who's and that zapier would then be like connected to other apps that they use. So it's kind of a bridge to get ChatGPT to be able to see into, you know, your email and your QuickBooks and these other apps you use. The Zapier example they used was the person just typed into the Chatbot what's on my calendar today? It went out to Google Calendar, which Zapier has a connection into, pulled off today's events and then there was like a double booking there and they just put in the chat box, let Sam know I gotta go chasing GPUs. And from that it worked out okay. I'm going to, because I have a Slack connection. Slack, the team chat app. Because I have a Slack connection in Zapier, I'm going to go out and send a DM to Sam and Slack that says that message that they just said. And that was all from a single ChatGPT interface that like that was no hopping around a different apps. That was because ChatGPT could connect to all those other apps via Zapier that in a single chat box they could do all those things and all those different systems. And we've talked a bit about how like ChatGPT may end up rolling up a whole bunch of different types of software and the future of software might just be chat and voice and all that. And that feels pretty sci fi and still pretty abstract right now. But we just got a big step closer. I'm probably most excited for that Zapier GPT. They had a plugin before and if you ever use it, it kind of stunk. But things as simple as being able to like chat with a QuickBooks file and get it to actually do functions like pull up any outstanding bills for this or that customer. If I can enable that for one of my clients to have like mobile access to like or even to be able to talk with it, which you can also do with these. Talk with it. And I go and fetch that stuff from QuickBooks like what are my outstanding bills right now? What's the last payment I got from this customer or last payment I sent to that vendor? Zapier actually might be able to enable that for us. Like, like very near term. And that would be interesting and could potentially change the, the interfaces we present our clients with to work with their QuickBooks data. Right. I like the notion that I could help one of my clients set that up on their phone on a mobile app. That'd be pretty cool. So big day. Honestly, like there was a lot more in those announcements than I was expecting. No news on future like GPT 4.5 or GPT 5 or anything like that, but some cool stuff, especially around customizing GPT and getting a version that's more useful for specifically what you do. Anything that you saw that got you especially excited, drop a comment. Obviously, we'll keep talking about this stuff as we start seeing it more and more in the wild as the tech partners that we have start leveraging this stuff, like in the apps that we use. And I'll see you in the next one.