navarre.aiClarisTalk AI · the FileMaker AI podcast
AI Impact on work, workers and jobs
Matt and Cris review the AI focus of Engage in Austin TX, and weave in the main subject: The impact AI has on work, workers and jobs, and where it's headed. Cris referred to this video: Shopify CEO Tobi Lutke told employees in a memo that artificial intelligence is making its way across the company, including when it comes to hiring. https://www.cnbc.com/2025/04/07/shopify-ceo-prove-ai-cant-do-job
Apple PodcastsSpotifyYouTubeAll episodes
Transcript
Transcribed by machine (whisper, large model), 11,870 words. Names and product terms are sometimes misheard; the audio is the record.
Chris. Matt. Claire's Talk AI. Today our topic is going to be AI's impact on work, workers, and jobs. But before, I would like to hear the stuff that you talked about at AI, about AI at Engage. Yeah, this is the first time we've talked since Engage. My goodness. Well, first of all, I've been listening to all the other shows. If people haven't, they've been a great series of shows. Missed talking with you, Matt. Missed having you at Engage as well, too. And I'm sure you've heard that the number one theme was AI. Yes. Shocker? As a matter of fact, I'll cut to the chase on one thing. Our podcast was recognized, actually. It was mentioned in the keynote, and they bestowed upon me a... I'm here to share this with you, actually. We got it... Part of the reason for this Education Excellence Award was this podcast. Interesting. Which was cited by Claire as a great source of information for AI. So, congratulations, Matt.
Congratulations, Chris. Thank you very much, sir. And so, also part of that keynote, I actually had a few different commitments. All of them were related to AI. I think you'll find them interesting. First of all, I was a surprise guest at the keynote. Clarice showed me some interesting things that they have cooking in the Clarice Labs. Believe it or not, there are Clarice Labs. And believe it or not, they're actually cooking up some AI stuff. And I got to preview it. I spent some time with them ahead of time with the engineering group, and they let me show a... Gosh, I don't even know what the NDA... I'm not even sure there was an NDA. I don't know. I was backstage. I didn't see it. But I got to show a really cool way that AI can be used to help develop applications, or like a plan of an application, right? So, sort of like...
Like it can be done with Microsoft Copilot, stuff like that? Yeah. Well, there was definitely... So, there were copilot-type demonstrations, but in the context of FileMaker, which... Okay. Should be really excited to anybody at all levels, frankly. There were some really cool things that were shown there and promised there. There was a bit... There was a large tip of the hat of all the AI features that are coming out in the next release of FileMaker as well, too, which were very compelling and got some good applause from the crowd. For sure. Then the thing that I did was actually showing how you can take a natural language problem statement and AI can create a plan, which includes an app and potentially even some supporting automation or flows to help bring that to life. And it actually was real. It was a very real product that actually created... In this case, the demonstration was a studio view and supporting Claris Connect flows. But the point of it was that you could actually build an app within the bounds of the true capabilities of the application. It wasn't just to spit out a problem and then it comes back with a plan. So, it was really cool. I just... To me, it was like a real glimpse... Oh, go ahead.
For example, like what was it about? Yeah. Like... Oh, yeah. So, for example, like what was it about? Yeah. Like... Oh, yeah. So, for example, like what was it about? Yeah. Like... Oh, yeah. So, for example, like what was it about? Yeah. Like... Oh, yeah. So, for example, like what was it about? Yeah. So, I said, you know, let's create an app for managing devices that were distributed in a school to parents and determine which ones were damaged and collect fees for it. And so then it... Yeah. And much like in a sort of like a replet manner or one of these like text-to-app type things, it then determines whether or not it has enough information to build the thing that I need.
Mm-hmm. Mm-hmm. And in this case, it didn't. So, it asked me a couple of follow-up questions, which I provided. And really, in about a paragraph's worth of words, it then came up with like a two-view connected app within Studio. And then it had like a cool Claris Connect flow, which really the flow was just allowing data access. So, it was like giving me like device data from Jamf and, you know, some other data as well, too, that allowed, you know, this thing to actually connect to backend data sources. So, it was pretty cool. I mean, it was... I mean, it literally took like, you know, a minute or so to actually devise this. And then you can... So, it's a plan. And then you hit the little generate button on the top, and it turns it into an actual Studio view and Connect flow, which is pretty cool. And the idea was just, can you take problem statements and make them applications? And it was really compelling. So...
How do you... You can't really copy and paste Studio views or Connect flows, though. How did that work? Well, they're all just JSON, essentially, at the end of the day. At the end of the day, they get assembled as JSON. And so, if you follow the rules, understand the formatting and the syntax and what's possible, and frankly, they're loading in like every little... What they did is they broke Studio and Connect down to its most atomic level components, like the tiniest little pieces, like this little configuration that's possible. And then the model was trained on those capabilities and the assembly instructions and rules, like, for example, like a flow needs to have a trigger, right, to start with, and then it's followed by, you know, actions.
And then it actually produces the plan within those bounds. So, to me, the thing that impressed me, it wasn't even so much that it was creating a Studio view or a Connect flow, for example. Yeah, that impresses me, but... That was super impressive. And the reason you can do that... You don't have access to those JSON files. Those aren't reviewed. Yeah. No, but, like, well, but, I mean, like, just look really closely at, like, a custom connector, for example, within Claris Connect, right? Oh, yeah, yeah, yeah. That is... That's... Yeah, right. Perfect example of a structure. Okay. So, if you can mimic... If you can mimic that structure, you can assemble these things. So, to me, it was... I was moved to just sort of from an intellectual curiosity standpoint that you can assemble within the bounds of reality.
Like, it literally will say, look, this is not something we can do, but we could do this instead, right? So, really promising. Of course, there was a lot of stuff... Plus UI, right? That's developer assistance. Yeah, actually, develop the UI. As a matter of fact, you could go in, in my demonstration, and by anybody that wants to check this out, I'm sure it'll be up on... YouTube channel here shortly. They are going to release the keynote, so everybody can check it out publicly. Yeah, the other ones are, like, you pay for them to get access to it. I can't remember how much it is. A couple hundred bucks or something. But they're going to be fully releasing them over time, yeah?
Yeah, I think so. I don't know. I'm not entirely sure, but it seems historically that they... It's important for them to show the keynotes. Yeah, yeah, yeah. Which is why I think there's, like, real... Like, no real, you know, NDA thing going on there. So, in the interest of me, possibly... I don't think I have, because they didn't really say anything to me about it. But it was pretty cool. It was really promising. And to me, it just was like, wow, just imagine when these things mature and what we could do. And, like, it really got me thinking about the role of developers, right? Like, significantly. And as it turned out, that was really the theme of at least my participation at Engage.
So, aside from being part of a very AI-centric keynote, like, extremely AI-centric keynote, where, of course... They bestowed upon you and I an award for our AI advocacy. Then I had two other commitments. I'll cut to the chase of the end one, because I think this is an interesting statement. I did, in collaboration with our good friend, friend of the show, Ronnie Rios, and Pied Piper, of all things AI, within the Clare's community, he and I did a training day together, which was... Actually, it was just getting started with Clare's community. Yeah, yeah. It was the last day of the conference. That's just from a presenter standpoint. Imagine from the learners. People are just bleeding from the ears at that point. So, it has to go back.
Well, for a long conference and all the other stuff. So, as it happens, I'm going to be teaching a six-hour AI class at a conference in a couple of weeks here. In the one in Toronto, right? The one in Toronto. Yeah, yeah, yeah. Yeah, I'm super pumped about this. It's after the conference. But it's a one-day conference, so I don't think people are going to be as fried, hopefully. It's different. Yeah, in fairness, I'm going to a conference. I'm going to a lane-chain conference next week in San Francisco, and I think the... I think the training is the day before. They did the same thing. One-day conference with a one-day training the day before. I think it's before. But anyways, from...
So, you did... 68 people, which was amazing. There were, like, some heavy hitters in there, too. People who were just like, look, I just want to lap up every little piece of AI that I can, right? So, I thought that was fun. Is it hands-on demo file? Yeah, it was... We did a... There were actually several exercises that we did in hands-on, and then at the end, Ronnie came and just killed them all with his, like, cool demos and everything and found... Yeah, he talked... I had a little bit of a preview of one of the things he showed. I'm not sure if he actually did, which is all stuff possible with the current version, where he had a... I can't remember which plug-in, but the real-time conversation one, like, talking to...
Yeah, that's... You're actually having a conversation with your data, and, like, very deeply, like, having a conversation, asking about the structure and the relationships, and... Yeah, yeah, yeah. He got... What did he show that? Was it the keynote? Dude, he showed that at the keynote, with the, you know, the voice bellowing through the hall, right? It was so... It was so cool, and, like, that got the biggest pop of all the presentations, definitely. Yeah. Well, curiously, that and a bunch of the bug fixes that were announced for FileMaker were kind of even, and then I got no pop at all. Actually, I got no one even... They were like, Chris Imeloy, and it was like crickets out there.
Yeah, exactly. But that's fine. Chris Moyer? Yeah, actually, if they would have just said Chris, I maybe would have got some half of claps, maybe, before I actually got up there. Yeah. But that was a big hit. He showed that, and then he showed that again in our... He kind of dug deep. He actually did some crowd work with it in his session, too. So that was prominently presented. But then the other one that I did, which really is... It makes me think a lot about the topic of today's conversation, was the other session that I did. It was called Making a Business Case for AI. And I started off the top talking a lot about what we spoke about on the last podcast.
Like, what's the role of Claris developers in the AI world coming forward? And I told people about our story. We talked about AI solutions, where we deliberately pivoted. We didn't pivot. We just deliberately offered AI services, how we presented ourselves as experts, and, you know, what... And more importantly, like, some of the crossover soft skills that we didn't... Or even hard skills from a development standpoint that we didn't realize were going to be so valuable in the AI space. But, like, all the years of experience as file maker developers, we didn't realize it was making us... It was putting us way out in front of AI developers, too. So I talked a lot about that.
And then I brought up... Customers. So I profiled two specific use cases. And intertwined within each one of them were, like, you know, topics that we've talked about before. Prompt engineering, few-shot prompting. Like, you know, all these kind of things. Context and all that. But I did it along a storyline. So I had two customers in there. One of them was 3M, which I'm super excited about. If I remember, we'll put the link to the story that I have a blog post on about all those... What was really cool about it is, like, all of it was done in FileMaker. And the ROI on the impact of what we did was just huge. I mean, literally, it was, like, $2.5 million ROI on just, like, one integration.
Wow. And the biggest part was 3M was kind of going through it at the time. It wasn't so much an economy thing. It was more about some stuff that they had going on with some, like, settlements and stuff like that. And so... So they were bottoming out. They were, like, bottoming out from a resource and funding standpoint. And so they were, like, having to let a bunch of people go. And they just still had to get work done. And so this thing actually helped lift these groups back up to help them still produce at the same levels that they did before a bunch of the layoffs that they were forced to do. And then later, being able to even rise above that and become even more productive with the people that they had.
And then even the people that they were able to bring on, those people are operating at, you know, an X component of productivity. So I was really proud to do it. My friend and colleague over there at 3M came all the way out to Austin on his own dime to present. He did a great job. Nice. And he's an old FileMaker developer as well, too. Matt Tandy. Shout out to Matt over at 3M. And the ROI stuff was amazing. It actually ended up winning a global marketing award within 3M. It was one of the first deployed AI applications deployed within 3M globally, dude, in FileMaker. That's crazy. So if that's not enough reason to... To, you know, explore AI. And then also, that saved our account.
I don't mind telling you that that account, we started as FileMaker developers and we're now AI consultants. And frankly, if we didn't make that switch or present ourselves as experts to those guys, we probably... I don't know if we would still have the account. No, you wouldn't. I've talked to four or five partners who they were talking to. Yeah, they're all sitting in the parking lot, actually. And then the second one, just to tie a bow on the AI stuff, was also human-related in that it was a... It was a use case of another FileMaker user who I had met, who went to my session the year prior in Austin. And we engaged them to perform some experiments and then a production version that just rolled out of not using embeddings for semantic search.
This is one of my frustrations, by the way. Okay. Everyone says the new features that were released in the last version of FileMaker were semantic search. Well, no, they were embedding support. And one of the things you can do with that is do semantic search. And so our story was about... Jobs and resumes and these various different somewhat sophisticated techniques we used to use embeddings to do matching of those. And it got pretty deep. You know, people can check out... I'll figure out where to find more information on that in the show notes as well, too, in case that doesn't get out publicly. But the idea was they were able to... It was pretty amazing because what we're really doing with AI in both of these cases is not replacing human work, allowing human work at scale.
So, for example, in the job placement company, because of one of the techniques, they were able to have AI create a larger pool of candidates for them, which then meant that they could free up the expertise, the human expertise that they had, to actually take on other new lines of business, essentially, for them. So they stayed with the same amount of humans in their resource pool, and they were able to expand into new business lines and therefore generate more revenue with the same staff that they had. So two really... Kind of amazing stories of how AI embellished human work, did not replace it, very specifically, and helped scale and route the right work to the right humans, and so that those humans could perform efficiently.
And these companies experienced incredible ROI in games. I'm super curious as to what you've put into the embedding to get those things to match. Like if there was JSON-LD or something, did you use generative AI to expand it so that it would match? Would you be better? Because a lot of the stuff I'm doing, I'm getting, you know, fairly low score. I'm not even sure what to take of a 0.3, 0.4 type of... First of all, I love hearing you talk about cosine similarity as a way to score effectiveness of matching. That's one of the key stories there, that, you know, 0.3, 0.4s. And by the way, if you're working with... Is that from which Cohere that you're experiencing that, probably on the embeddings front?
No. Well, Cohere I've been using for image embedding. I haven't tried it that much for text. Yeah. And embedding models are really slow to come out. Ah, but I did learn recently that I think that some of the generative models now also support embedding. Yeah, maybe. I think I just learned about that like last week. Yeah, I mean, I guess. I mean, it wouldn't be that much of a leap. I mean, that would just be another form of multimodal, but I don't know. That seems like a solution without a problem, honestly. Like, you know, embedding models are embedding models. They're trained. They're trained to do a specific thing. So, sure, maybe you don't have to change your API call to do an embedding versus an LLM call, but...
Yeah, what was the general approach there? Yeah, the approach I'm really excited about. So, the original approach came from an idea that I presented in my session the year prior. And the idea was, what if we just take a job posting doc document that says, here's what the job is, and someone's resume, and we just embed the entire document and then do a cosine similarity? Right. I mean, it's like the most, like, brute force wrote thing. Yeah. Well, as it turned out, and that's the experiment that we... That's what I was talking about in my session, and so that's the experiment they engaged us to do. Well, as it turned out, we were like, we saw the first results, and we didn't even get back to the customer.
We were like, no, this isn't going to work. It was garbage. They were like, low 30. You know, this was cosine similarity, so low 30s, you know, that type of thing. Point is, there's just too much noise. Yeah. So, what we ended up doing, and I'm really proud of this, is we went with this, like, ontology mapping approach. This is like an old ML thing. Mm-hmm. So, what we conceived of, we realized it wasn't working, went back to the customer midstream in an experiment, which, by the way, is why, when you're working with AI, doing experiments is the thing. Because think about it. If we had engaged these guys to use that as part of the spec, now, all of a sudden, we'd be dead in the water.
But instead, we said, let's spend 15 hours doing this experiment. If it works, great. If not, we'll figure out something else. Yeah, absolutely. You really have to do that. I've got a few projects like that right now. It's a must, and it has a lot to do with the black box nature. So, what we did is we ended up breaking, it just so happened that the customer, what gave me the idea was the customer actually broke down their job descriptions into, like, six different specific categories, whatever they were. Like, they were just, like, you know, the nature of the job, what experience is necessary, the role description, whatever it is. I don't even remember.
And then, each one of those little categories, we then embedded those individually. Okay? Okay. So, they didn't have to do that. They didn't have all the noise of all the rest of them. They were just specific. Now, the challenge is, what are we matching it to on the resume side? Yeah. This is where LLMs really came in. It was kind of sophisticated. So, hopefully, I can represent this briefly. But we took a resume, and we said, look, a job posting that you're going to have to match with is broken down into these six categories. You are not a job posting. But imagine how would this resume fit those each individual six sections. Now, use the language model to, like, extract.
Review the resume, extract from it a little block of text that would be a good match. So, like, if the thing said required experiences, blah, blah, blah, it would look into the email, or I'm sorry, the resume on the resume side and populate that little experience necessary with their experience and why they would or would not be a good fit. So, then we embedded each one of those things and did individual matching. So, now we had six different scores. And what that allowed them to do then is look and see which area. Like, some of them were a .75, some of them were a .45. So, they could actually look at which of the match was more, you know, impactful.
Then the final thing we did was pure ML. Actually, that wasn't the final approach. But the final thing we did on that was pure ML. What we did was we still wanted a single score for matching. So, instead of averaging all those things together, we did a weighted system, and we did a regression model to create one score. So, what that means is the user, when they set up a job, the administrator. At the company, it can go in and set a weighting of all six of those things. Weighting is just, you know, it has to add up to 100%, right? So, they can go job by job and say, experience matters more here, so this is a 35. Or in this case, it's all about this one thing.
Sure, yeah, yeah. Totally. And it rolls up to, then it does this regression model to roll up to one single score. And that's ultimately what they ended up using. But then we did the last thing was called ideal state, which was really kind of fun. What we did was. What we did was we used the language model, and we went to this job. We go, here's the job posting. Create the ideal candidate description based on this job posting. Put that into text. Now, over on the resume, create the ideal job for this candidate. Put that into text. Embed both of those, and then match those in the ideal state. That last one was what allowed them to, without going into their business, create larger pools of analysis that they could send over to humans.
So, they used the second one. The third one allowed them to open up some. Interesting work. So, none of that had to do with semantic search. Yeah, I understand all of this, right? So, I've tried some other approaches. You've given me some really good ideas right now for some things I'm working on. But, yeah, it's the beautiful combination, in your case, actually, generative plus multiple types of embedding semantic plus ML. Yeah, plus ML, and none of it was semantic search, like literally nothing was searching in there, so. Well, yeah, embedding comparison is sort of semantic search. You know, it's just not entering a prompt or a search term. Honestly, well put.
If people hear this and they think what you just think, then they're on the right track. What I worry about is that people who have yet to get started think that embeddings are solely in service of semantic search. It just so happens it's one of the things you're doing, literally matching the question to the potential answers. Yeah, or like a photo. You can say, hey, I think there might be a duplicate of this photo. See if you can find it. And it'll find another one that's a different resolution or taking a different day, and it'll be like a .9, you know? Duplicate management. Yeah, go try to find duplicates. Duplicate photos any other way. No doubt. Well, and so to me, the thing that was important about both of those, Matt, was they were really about enhancing the human experience.
One of the things I noticed in the early years of selling AI services to customers at iSolutions AI was that there's a lot of fear around being replaced, right? So I ended up coming up with this sort of kitschy, somewhat tongue-in-cheek, but very meaning, I mean it, approach where we say, look, we don't sell. We don't sell robots to replace humans. We sell Iron Man suits because we want to make, we want to give humans superpowers. And people really got that. And in both of those cases, we absolutely were selling Iron Man suits for humans. That's for sure. To be able to level up. Yeah, yeah, yeah. But that's, you know, unfortunately, not everybody thinks that way about this, right?
I mean, if you look back at all the other things we had, you know, like, I don't know, but my favorite example from recently from composing music is. When I first learned how to compose, I did it with pencil and paper. And if you're writing a piece for an orchestra, and you had, you wrote a really nice melody for flute, and then you had, like, the cello part, and then you wanted some other stuff to be in there. And if you wanted, like, oboes and first violins to double, but then you had some other stuff, and you wanted the viola to do something with a different octave or, like, a thirds, that was all totally manual. You had to do all the math in your head.
But now you can say, oh, okay, I want to take this flute line. I'm going to copy it to the oboes. I'm going to put it on the, you know, the violins. And then take it, you know, a sixth down or whatever for the violas. And then you can push a button in the software that I use, Dorico, to fill out the chords for the rest of the string quartet. And that's just busy work, right? To a large degree. I mean, it follows all the correct rules set down, you know, a book that Bach read forever ago on how to do counterpoint. And then you fix the errors, right? You go through and you go, yeah, I don't like this line. I'm going to fix this line here. But it's an iron.
It's an iron man suit, right? It is an iron man suit. It gives you the ability to do more. And all of the stuff that you're doing is the creative part and not the make work part. Well, what's interesting about that is when Copilot first came out, before it was in Microsoft, it was just like a plug-in for VS, right? And it was the first thing that AI came out with. And ironically, all the AI developers were like, oh, we're the last ones who are going to get disrupted. It's AI. They need us. Well, the first thing out. The gate was like coding tools. And what was fascinating to me is after it was like after nine months of that, they sent out a survey. And we might have even talked about this, honestly, on the pod, but maybe ages ago.
But they asked people three questions. They said, you know, what percent more productive are you? How many more hours of, you know, code are you doing in a given day? But the last thing they said was how much more fulfilled are you? I don't remember the other percentages. They were probably what you would expect. But the fulfilled one specifically said like 78% of respondents said they were. And so I kind of dug in on that. And the logic was because now they're spending time in more of that creative zone than that repetitive zone. Right. So, you know, and I would say from a FileMaker standpoint, like I get it. Like back in the days where I would like wake up on a Saturday morning and be like, oh, I want to try this cool experiment with FileMaker.
Well, that, oh, my God, I'm so excited to do something that creative energy or that wow factor dies after you hit a bunch of errors and you're just in the weeds on dumb stuff. You know, and then after a while, I'd just be like, ah, all right, I'm over it. The energy is gone or whatever. But if I could have just accelerated and maintained that enthusiasm, that fulfillment, I think that's actually extremely meaningful. I'll even add that one of my one of my new hobbies, dude, is amazing old people with AI, like like trying to do like magic tricks with people that are older than me. I'm an old man. So I try to find people even older than me. And I had this gentleman, sweet guy who came to.
To do an estimate at our house for carpets. So he goes and he's got his pad of paper and his pencil. And, you know, he's been doing this for ages and he's really good at it. And he sketches out all the measurements for our rooms. And then he sits down at our kitchen table and he goes, look, I'll get back to you later tonight. I got to sit and do these calculations so I can really set an expectation for you on what the next step is. Right. Like we're in the flow of the engagement. Yeah. Yeah. And he has to stop it abruptly and say, I'll get back to you later tonight. Well, it's already almost. Dinner time. He's sitting in our kitchen table. Like, I kind of want this just to be wrapped up so I can think about my decisions.
You didn't invite him to dinner. God, what a jerk. Well, what I did was I think I broke him. Unfortunately, my wife, my wife says I broke him. So what I did was I took out I'm pretty sure it was my chat GPT app, you know, something with vision. And I go, I go, yeah, hold on a second. Let me take a picture of that. I go, click, click, click. And he goes, oh, what's that? And I hit enter. It comes out. With, oh, I think these images are blah, blah, blah. Here's. Yeah. Hey, by the way, here's what the calculations would be per room. You have to factor in 10 percent for waste. And, oh, do you want us to do the the diagrams, like how you cut the patterns and what sizes?
And he was like, wow, what? And then seriously, he just broke character. And he's like, wait a minute. What is happening? And so then he sits down and he does it all by hand and ends up. It was like right on. So his impression wasn't, oh, my God, I'm hosed. Like now my job's gone. His impression. His impression was, oh, my God, how do I use this so I can actually tell you the information you need right while we're in the moment still. Right. So, again, Iron Man suit for that guy. Yep. Yep. Yeah, for sure. It's going to make his job more fulfilling because it'll be it'll be so much a better experience for him with a customer to be able to give him the information right then.
Oh, dude. And then you go, wow, that's a thousand less than I thought. Maybe let's take a look at a different carpet or something, you know, or another room while he's still there. And frankly, although carpet, come on, man, that's I don't even allow that stuff in Greece, whatever. Tile, sorry, tile. Everything. Yeah, of course. Of course it is. But I mean, not the least of which is that the guy, you know, he was running late because of a previous appointment where he probably had to spend a bunch of time doing this type of stuff rather than being engaged. He probably could add a whole nother thing on there. So there's a ton of conversation about the impact on jobs.
And most of it, like, you know, one of my favorite podcasts is the All In podcast. I don't know if you listen to it, but they talk about AI all the time. It's a bunch of VC, like one of the most popular. business podcast, but they also talk a ton about AI and they should. These are VCs, the guys who started up like Uber and Facebook and, you know, huge, huge, big movers and shakers. And one of them had this quote, and I wanted to share with you. They said, in the Valley, unfortunately, we've been co-opted a little bit, somewhat with a bit of a doomer mindset in some areas. And we think of AI as, okay, it's a fixed pie. It's going to replace things. On the ground with large enterprise, the vast majority of the use cases are the ability to finally review contracts that we never got around to reviewing. So it's really the ability to finally automate an invoice process that we never did. It's the ability to go in and just create marketing campaigns in every language that we never got around to. I think this will probably be actually like 90% of the usage in the future will be things that if we look back on it, that we snap in line right now, we said, this is what knowledge work is today. And so essentially 90% of AI usage will be things that we don't do. And so I think that's a really good point. And I think that's a really good point.
didn't do today, 10% will replace what we're doing in some areas. I thought that was hugely profound. Your thoughts? Yeah, it really is. That's like not at all the thing I was talking about, about music that flips that on its head, you know? Well, I think so. I mean, if you think about just, I mean, I think the marketing campaign in different languages is a perfect example, just like you would have just not done it because it was expensive and time consuming. And if you could do it in minutes, then why not all of a sudden, right? So, so I don't know that there, I believe that I, I, I believe that there's opportunity for that, but then there's like, then there's two other things that I want to explore with you. The first of which is hiring this one. I'm not even really sure how I feel about yet, but, um, your thoughts, did you hear about the Shopify's CEO came out a couple, couple of weeks ago and he just, it went viral. He said, um, well, actually it was an internal memo that then he just owned and he released publicly because there was like a lot of chat about it.
It was the Shopify CEO. He said that staffers need to prove that jobs can't be done by AI before asking for more headcount. So before you can hire a new person, you got to prove that AI can't do that. Now, I don't know how I feel about that. That was Shopify CEO, Toby Lutke. Um, what are your thoughts there? Well, you can't prove a negative. There's that great call. Yeah. The philosopher of the show comes in with the most obvious part. Yeah. Russell's, you know, the Russell's teapot, uh, thing. Yeah. Maybe we did. It's a great one. Uh, well go ahead for those that didn't hear our Star Trek episode where we talked about it at the end. Go ahead and reveal that. Well, because people are going to wonder what the hell it is. So he basically said, imagine a teapot in, in orbit around the sun. And, uh, you don't know where it is disproved, but it's there. How can you prove that there's not a teapot orbiting the sun?
And that's how people talk about a, a GI too. That's what you're talking about. Yeah. That's kind of a thing. And I shared a podcast with you about AGI and how, what it's going to do in the medical field. Wow. Well, did you, did you hear, uh, Bill Gates? He came out and he said, gosh, I don't have that prepared, but, um, something along the lines of like within three to five years or something like all like, like the X percent of the jobs and medical advice will be AI. Right. Um, so a lot of this talk is going on out there and, um, I don't know how I feel about make sure AI can do it before you hire somebody, but I feel strongly that AI capabilities are probably the most important thing to consider in tech, you know, in tech at the very least, like we've done hiring, we've, we've added on teams. I just had two recs. I just had two things that we floated out for job openings. And in the back of my mind, it was critically important for me that I knew that these people were leveraging the tools in any manner that they could.
And here's why, because when I started, I've mentioned this in DevCon or Engage presentation before, but when I started up iSolutions AI division, I really, I just had like three different people and I was super freaked out, Matt, about like, oh my God, you know, I'm the sales guy and the resource manager. So those are usually in contrast with each other. So what would happen is I would get a big new client and then I go, oh God, like the sales guy in me would be like, yeah, let's go. We're doing it. And then the resource guy in me would be like, oh my God, we're, I feel like we're already maxed out. How are we going to do this? Yeah. So then all of a sudden, like, um, co-pilot came out, like I mentioned before, and then eventually like things like Devin, which we've talked about before on the show, and then just frankly, Cursor, you know, Windsurf. So now my whole team uses Cursor. The, I then about a year ago, I said, all right, how much more productive are you guys all? Let's just do like a, a, a, a team survey.
And we really spent like an hour talking about it. And the, the dev guys, we're like a thousand percent more. And they cited specific examples, literally like this used to take me 10 hours and it takes me an hour to write like, like literal things. And then even the project manager was like, well, I used to have to take notes and send emails. Now we just send the transcript over to our AI bot that partitions out next steps, sends an email to the customer, tells us what JIRA tickets we need, all that kind of stuff. Right. And then I, as really the product owner was like, yeah, I'm using this to vet out whether or not these things are real before I even bring it up and dah, dah, dah, dah. So anyways, we did all the math.
And it ends up that we are, and that was a year ago. And I can tell you that we've increased our revenue by like 40% since then. And I haven't hired a single person. And the, that was just since last year. And then since we started, when we first built the team, it's probably a hundred percent growth. But the point is we're operating in a 5.5 X manpower, um, uh, uh, a differential, uh, essentially meaning I, if we weren't using these tools, I would have had to fire hire at least another five people. So the point is I didn't, I don't have to stress out about it. I get a new job and our new project, and I don't have to freak out going and hiring another contractor or something. So, so one can certainly say, oh man, that sucks for the five other people I would have hired.
But then at the same time, we're just able to do more. So I don't know, where do you come in on that? And then I want to go into what the, the, the AI. Yeah, I definitely want to get to the, the main heart of our, well, we've been really kind of skirting about a lot of parts of it. So, and I've been also kind of bouncing the thing around about the Shopify CEO, which I hadn't heard that. Um, and the thing I love about that is it actually forces his people to be thinking about AI and to really think about, Hey, can this be done with AI? Cause it makes more sense to do that. Uh, and then I think some of the things you're going to get to is like a fundamental expectation that they are using AI, right? That's like him both saying, I want you to, but he's also saying, I don't want you to, but I want you to, but I want you to, but I want you to do that.
So that's kind of interesting. Yeah, exactly. And I, you know, I'm, I'm definitely one of the people who, uh, on the positive side of it, uh, God, some of the interviews I've heard with, I can't remember, believe I can't remember his name. Um, the CEO of open AI. Oh, uh, Altman. Yes. Yeah. Um, and the stuff he has to say, uh, and also Nvidia CEO, they're just phenomenally Jensen Wong. He's like all, I can't even believe listening to him. Uh, he's like, Oh, I can't even believe listening to him. Uh, he's like, Oh, he's like the spiritual guru of AI right now. I know, I know, I know. He's got like waxes kind of philosophical on stuff. Um, yeah, yeah, yeah. Uh, and then he like, doesn't understand people who are naysayers about it anyway. So, um, yeah, I understand naysayers. I just want to point that out. Like I, I get it. Like I understand the emotion, but at the same time, you know, honestly, man, I'm trying to think about how to say this. I understand the emotion, but I, but I also like, like just going, Hey, I'm going to go learn machine learning and I'm going to go teach myself AI and I'm going to take a chance that none of this will pop or that it's just the next like Bitcoin or NFTs. That was hard. And like, I personally like sacrificed months and months and months of my life. Like I could have been just doing other stuff. Uh, and I decided not to do that
to try to, you know, explore this. So, you know, I have an emotional relationship with that reality as well, too. So I don't know. It's just math. Sure. But there is a reality that the naysayers are right or a possibility I should say, right? I mean, it's possible that we get a Terminator, um, situation, uh, some other story. I'm not going to, I'm not going to acknowledge that. No, I mean, like I'm not, I'm not all gas, no brakes. Like a lot of those months and months and months of my life that I sacrificed to try to learn all this stuff involved learning risks. Like I probably spent more time in the early stages learning risks. Cause I wasn't about to be the guy that people were writing me expose their data. So, um, the reason I have so much conviction over my, my stance on data privacy and how models can't learn from your data is because I know this stuff. Like I actually go and research it and read the papers and listen to the experts and watch people testify in front of Congress on these statements. Like I'm not speculating. Like I actually know this stuff.
And so I had to go learn all of it. That's where I get my peace of mind from is the hours and hours and hours of figuring out. So I'm not saying I'm any better, smarter than anybody else. I'm just there's the emotion. Well, there's three, maybe there's the emotion of, Oh my God, this is going to take my job. Then there's the, Oh my God, this is going to take my job. Maybe I should do something about it. And then there's the other part, like, you know, um, the fantastical, wow. Like I, I, I still every day, almost I'm like, wow, this is amazing that it can do this. And I just lean in and go, I'm going to just have superpowers. That's what it's going to be from here on out. So, I mean, I get, I get a lot of parts of it. I don't know your thoughts. I don't want to be insensitive because I know there's a lot of, you know, I think it's taking all that stuff into account. And then, you know, there's, uh, there's some really, really smart people who believe that AI is going to be net negative for society.
Um, and I, you know, I can't say that they're wrong or right or whatever. I can just have my opinions and I can spend my day. I choose to spend my day helping people use AI and file maker. That's my mission. That's my thing, right? I mean, you don't give a damn about, I'm like, or I know you're long past it, but no, that's not true. That's exactly what this is for. This is literally for dedicating all that time to teaching the file maker community about AI. Like, like no one asked what interesting thing that was there at the museum. Oh, look at this. Just talk about analog, the folders. Yeah. That was, uh, these were, they were selling merch and they had a museum, which is a bit of an oxymoron with like, you know, AI emphasis and then having a museum of stuff from, from 40 years ago. But it was a lot of fun actually. And you and I are, are, are calling that out here. We should maybe, for the people who are listening, maybe we should say that I just pulled out my, uh, Mac or my file maker to Mac, draw to Mac, write to coffee mug. And then you showed your, my Dixie cup. Yeah. Folder. Oh my God. It goes back to the same era documents.
Yeah. Yeah. I mean, if they're probably supposed to be the same thing. Um, well, you talked about smart people talking about this. You want to listen to a couple of smart people talking about this and let's pontificate a little bit like specific to developers and the opportunity. So there was a really important, um, conference that just happened. It was, uh, Lama con, which was at the time of this recording, like last week where, uh, they, the, the, the ending of it was a fireside chat between Mark Zuckerberg of Meta fame and Satya Nutella, Satya Nutella of, uh, Microsoft, like two of the most informed biggest luminaries in the space. But here's what's interesting.
Matt, uh, for those that are watching, you can see it, but everybody's going to be able to hear this clip here. And then we'll talk about it. And I have a couple more clips I want to play too, but this one is about them talking about how they use AI, not that they're trying to sell it to other people. And, and, you know, this is exactly, this is about their experience. So let's check this out in terms of the coding and how it improves that. Do you have a sense of how much of the code, like what percent of the code that's being written inside Microsoft at this point is written by AI? As opposed to by, by the engineers? Yeah. So there's two sort of things we're tracking. One is the accept rates itself, right?
That is sort of whatever, 30, 40, it's going up monotonically. Uh, and it depends like one of the big challenges we had for a long time is we are, a lot of our code is still C++, um, uh, C sharp and C sharp is pretty good, but C++, it was not that great. Python, it's fantastic. So we've now gotten get better at that. So as language support has increased, the code completions have gotten good. The place where the agent code still, it's very, it's sort of nascent. For new Greenfield, it's very, way high. Uh, but as I said, it's nothing is Greenfield, uh, in many cases. And so therefore, I would say maybe at this point, the PR, oh, by the way, code reviews are very high. So in fact, the agents we have for reviewing code, uh, that, that usage is increased. And so I'd say maybe 20, 30% of the code that is inside of our repos today and some of our projects are probably all, uh, written by software. What about you guys?
Wow. Um, I actually don't have the number off the top of my head, but I mean, it's, um, I, I think, you know, we, I think a lot of the stats that people say are still effectively of this like auto complete variety, but we have a bunch of teams that are working on, um, on basically doing feed ranking experiments and ads ranking and like very contained domains where you can study the history of all of the changes that have been made and like, and, and, and make a change. And that I think is like, is kind of an interesting area for us to work in. But the big one that we're focused on is, um, building an AI and a machine learning engineer to advance the llama development itself, right?
Because I mean, our, our bet is sort of that in the next year, probably, you know, I don't know, maybe half the development is going to be done by AI as opposed to people and then that will just kind of increase from there. Dude. Okay, a couple of big things to unpack. Microsoft claims 30% of their existing code is generated today is generated by AI. And he also made an interesting comment in there about how they would use it more if everything was Greenfield, but it's not. They're literally going back to like legacy code. And then Zuckerberg comes in and goes, yeah, well, we're, our goal is that half of our code is going to be done by AI and then even more from, from there, moving on within a year, your thoughts.
So I actually picked up something else from what he said. He said, he said, Zuckerberg was talking about, um, how he wants to have a lot of, a lot of the team actually focus on the AI engines, the models, and actually improve those kind of rather than using them right now. But like, let's make this thing better so we can use it even more in the future. I thought that was really interesting. And obviously, you know, uh, a very smart investment, I think. Yeah. Well, I mean, their perspectives are fascinating. They're both, uh, companies that, you know, are talking about leveraging AI, but also very responsible for the AI that they're, that they're leveraging at the same time to competing luminaries too.
So it's not talking about these people is so hard because they're the CEOs of all these companies are so politically divisive right now. Oh my God. Yeah. There's, there's their own, um, yeah, there's their own, there's their own politics within all of this, but you know, we shouldn't forget that people felt that way about it. They felt that way about Steve Jobs too. Like, um, you know, he was kind to everyone. Everyone knows that. Yeah. I've been, I've been, uh, watching a lot of Steve Jobs stuff lately and I, and I didn't realize he wasn't so kind, but, uh, whatever. He's still, he's still, you know, my, uh, guy up on the altar for sure. Um, so now let's, let's do this though, Matt. Cause like whoever's listening is likely in a developer, you know, role. And so there was another clip in here that I want to play.
There's really a couple, one of them really hits it for me, but I want to first play their closing statement in the same Llamacon. And by the way, you can go check out Llamacon. It was like 35 minutes. It was really compelling. They talk about a really, a lot of really interesting things, but this is how they ended. I want to, um, share this play. Like, what do you do if you're a developer? What's their advice for developers? And then you and I can kind of pontificate on this and I'll play one more from a third AI luminary. Look at the chart. Yeah, let's do it. Getting from us. So here we go. It'd be a good note to end on is just, you know, when you look at everything that's going on, I mean, I'm, I'm curious, what is, what gives you like the most optimism or what are you most excited about over the next period for what developers are going to be doing over the next few years?
Yeah, I mean, look, I mean, you know, I'm, I always take my inspiration from that to whatever that Bob Dylan line, right? Which is either you're busy being born or you're busy dying. It's get better to be busy being born. And especially at a time like this. Um, I think what gives me optimism is even with all the various constraints, it turns out software in this new form of way I is still the most malleable resource we have to use to go solve these hard problems. And so that's what gives me the optimism. And I also sort of say that's the call to action, I think, for the people in the room and for all of us is to be able to sort of take the opportunity to lean into this, but then also build solutions. When I look at whether it's an I. T backlog in a company or the unsolved problems in the real world, both of them need something new.
And I think that's where I think the greatest benefit of all of this tech is, and it'll only come down to, you know, developers in particular, being able to go at it fearlessly. Get busy living or get busy dying. Your thoughts? Yeah, I mean, I definitely agree with that. I kind of like the idea of being born, reborn, whatever, right? Yeah, that's fun, right? Yeah, yeah, yeah. Well, so, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.
Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this. Yeah, I mean, we've talked about this before, Matt, like, again, possibly not a popular topic, but I think it's a good idea to talk about this.