navarre.ai

ClarisTalk AI · the FileMaker AI podcast

2025 Year in Review – What happened? And what's next?

December 31, 2025 · 88 min · Matt Navarre and Cris Ippolite

Show Notes Podcast segment that summarizes recent paper: Microsoft Copilot Usage Report 2025 It’s About Time: The Copilot Usage Report 2025 (PDF): https://microsoft.ai/wp-content/uploads/2025/12/What_people_do_with_Copilot-8.pdf Microsoft AI OpenAI State of Enterprise AI 2025 — State of Enterprise AI 2025 Report (PDF): https://cdn.openai.com/pdf/7ef17d82-96bf-4dd1-9df2-228f7f377a29/the-state-of-en

Apple PodcastsSpotifyYouTubeAll episodes

Transcript

Transcribed by machine (whisper, large model), 16,440 words. Names and product terms are sometimes misheard; the audio is the record.

Claris Talk AI! Chris Ippolite. Matt Navarre. Hey Matt, what year is it by you? Still 2025. Oh, so we still have some time for our... Well actually it's interesting that we're teetering on the brink here because we can spend time today looking back on 2025 and maybe even peeking forward into 2026 and maybe you know making some projections which I believe we did last year if I'm not mistaken. Yes we did and we'll be taking a look at that. Oh fun! Well I'll tell you what I have a well yes so you're in Greece just so in case anyone's not up to speed. 6 p.m. where I am. Okay gotcha. 8 a.m. where you are. Oh yep perfect so we have plenty of time so we don't need to get into 2026 we can still faithfully make our projections about 2026 without cheating into 2026. This is true.

You know I'll tell you an interesting place for us to start today I think might be with this sort of I don't know if this is gonna be a pattern moving forward but for those of you that are just listening I'm showing Matt a couple of actually a whole bunch of papers that have come out over the last three weeks or so and they're all really interesting like for example this one here is open AI's 2025 state of enterprise AI report this was actually from the folks at Microsoft AI it's about time it's called the copilot usage report from 2025 these are all these might be a little bit creepy and this is perplexity adoption and usage of AI agents early evidence from perplexity and to our yeah there's a a yeah Harvard University and perplexity did that one and then there's a Pew research PEW research like pews at a church on teens and technology I thought this was kind of interesting since a harbinger of where things are going some of those sets of numbers there some really interesting tables oh yeah and then and then finally the last one that we're gonna be sort of summarizing and and taking a look at is estimating AI productivity gains from cloud

conversations now what all these have in common well they're all from providers well except for the Pew Research one they're all from organizations that that make chat BOTS which is essentially you know this at least today kind of the main line of how consumers are working with AI now just to frame this up these are not necessarily just business use cases and why we generally talk about business use cases of AI it's worth noting that this is the full birth of what people are actually using AI for and it breaks it down into a bunch of different categories and I thought it would be kind of fun for us to look at these pull some really interesting stats out of this then actually maybe even find out a little bit about how you may have been using AI using these some of these same terminology yeah yeah and we'll and we'll link to these papers in the show notes by the way we'll make sure to get that over to folks as a matter of fact I'm sending this to Matt right now as a matter of fact just so I don't forget so what do you think about this Matt is it are you creeped out that they're actually tracking what they were all very careful to mention that they anonymized all these conversations for the purposes of just aggregating you know the the data and of course I mean I don't feel at all creeped out by it

because I mean you know it's you're taking hundreds of millions of of things and probably only for the people who say yes you can you know use my stuff because it's check yeah probably yeah to speak specifically to that one of the biggest data sets is actually from the Microsoft co-pilot team they analyzed thirty seven point five million they call it D identified conversations which is January through just September so they've been crunching these numbers since September and they and so really that this one is probably a little bit wider of a use case because this is co-pilot which is integrated into all their products whereas the rest of these like you know perplexities I guess you know kind of a product even though it's you know a search engine for example so so what I guess I have that I have the top six I have the tops three I'm sorry the top three conversation topics and the top three intense which is what people were trying to do yeah when they were doing this what would be what do you think your top three conversation topics would be just like one word categories well I spent a lot of time on Greek learning Greek Greek history like researching and

understanding the etymology of words which is super fascinating mmm because so much of English comes from that um and that makes it a lot easier to learn that's one huge area got another area is just like life health dogs drug interactions for the 17 diseases I was hospitalized for I'm just kidding mm-hmm like just a ton of things done I'm all good well so but a ton of things like that like yeah yeah you know I think you got your finger you got your finger on the pulse here buddy like that that's really what people seem to be everything from like cooking to health to you know just really really basic lifestyle not tech and then the third area which is actually probably not number three is everything I'm doing to learn about AI and FileMaker and integrations and API integrations and how things are evolving and how to change this code and what you know all the things I had to learn for the switching from the chat to the responses AI uh endpoint and how fabulous that change was nice um well and that's tech though that would fall under that's yeah I'd say those I'd say those are the three groups yeah so you would say health Fitness um tech and then um like

language okay so you're you're right there with folks so number one with the bullet and and again this is just the Microsoft co-pilot team so for those of you that want to check out this paper and the show notes uh you can die to be cats right this isn't YouTube yeah otherwise it would be uh number one is in fact technology so um and again think about this this is co-pilot so it's probably got some coding built in there too yeah I bet number two is work and career kind of a you know a general topic but that's number two and the number three uh you're right there health and fitness and I think it's I think there's a bit of a common theme about health and fitness and I think that's really interesting that so much of us are relying on that I certainly am I have a um a personal context and a project set up just for my health um because and I and I use a chat for that all the time and and then in addition the top intense now the topics versus intense are different intent means you know like I'm searching on a topic but this is what I'm trying to do for example uh number one kind of unsexy uh searching for information kind of a duh because it's people are you know supplanting their Googling uh habits with uh chachapeteen or whatever their flavor is I almost never use Google anymore me

neither um as a matter of fact it I I have a hard time even you know using it as a verb and when I'm even talking about chachapeteen uh that doesn't have the same verb right now but yeah no I know it's weird it's it's it definitely doesn't work uh but searching for information um getting advice which is interesting that's one that's the number two ranked uh intent and then creating and that's obviously a very vast uh large birth of an intent there so in general it looks like at least from the co-pilot folks um people are using or in 2025 it helped them find things help help them decide things and help them produce a first draft of things it's not glamorous it's not AGI um but it's kind of like a hey just help me think a little bit better uh type approach your thoughts um I I'm kind of right there with that I think those all totally ring true um with kind of with me and with what I would expect use would be uh for for all of that um I guess I'm curious you know what I would also say I think that's not going to change yeah that's probably not going to change yeah I mean yeah different forms but it's not going to change much right because so I guess we could

loosely make that a projection projection that if these same things come out next year that we we both don't expect that these categories and intents would change that much yeah well I mean just like for health stuff as we get older tracking things that didn't uh seem that important actually throw one thing out I'm going to throw my one book recommendation out that uh that I think is super important and I strongly recommend to everyone and it's called outlive by Dr Peter Attia and it's uh if there's a must read book it's that I don't want to try to be all you know not tech bro but like health bro uh you know Keto diet all that stuff but basically he's talking a lot about um not just uh aging well not longevity but like health he doesn't call it health Jevity but something like that basically um having a lot of uh strengths to be able to do the things that you want to do when you're older he talks about a veterinarian decathlon like plan so like plan for the kind of bone density and strength and blood pressure and you know all the other things that you need to have if you want to be able to like play with your grandkids when you're 90. start planning when you do financially like it's the

same advice so if you want to be able to retire then you know make decisions now that will help you there so yeah I get that and and it sounds like that's probably something that you'll be um using your chatbots for to dive into your health Jevity uh a little bit uh as you move forward well duly noted and and we should put that in the show notes as well too a nice little uh link for people yeah yeah we should um yeah yeah yeah and there's he also has a podcast that talks a lot about he gets very very deep into it but it's just basically it's balance and it's thinking about things that you hadn't really thought of before like your metabolic rates and what your vo2 Max is and hmm um that's like if you're in the top quartile of uh of of strength in really basic ways you're like 30 less likely to die of all-cause mortality when you're in your 80s I mean there's the stats are just insane um and he really goes into the data anyway I'll get off that topic but I what I have done is I I've made a custom GPT with all of his information that you can get and then talking about like what my goals are what I want to do what my numbers are what blood test results I had blah

blah blah um and you can create yeah you can just create conversations in there or create one big conversation that you keep putting me back to yep so that you have that conversational memory that is guided by that as context because you can upload the entire book if you wanted to or if it's available on PDF and then just speak against it I did something similar to uh um like with this like I'm using it mostly in voice mode I found myself kind of using a lot of voice mode for like I don't know like just therapeutic reasons I guess I guess is probably the right way to call it and and I just have it as a little desktop companion now like literally like hey how would I navigate you know this work scenario you know um wait Chris if I don't know if there's anything else I can help you with okay yeah unfortunately I gotta be able to turn that stuff off but anyway well you gotta yeah you you do have to get past like the but I think the thing is though like actually I'm not gonna say I I think if you're using your AI as a therapist you don't have to get past the fact that it's not a human I think that we have to actually get past that human therapists are human and just sort of be like all right this person's not going to go tell everybody my secrets or they're they're

probably broken too like you don't have to think about all that stuff you just kind of regurgitate it and um so and I think I believe that is what we we're talking about when or they're talking about when they talk about health and fitness as well too although although this next part actually is a little bit interesting um the the Microsoft one went a little bit deeper I think you'll find this fascinating it actually broke down um how AI behaves like two different products is what they said desktop versus phone so I've often said um like even when we're talking about building FileMaker applications in my training classes I used to talk about how important it was to have um you know to identify who your users are not just for gatekeeping but for the purposes of well actually come to think of it establishing the context of who this person is and and where they might want to go of course those are the things that were meaningful I feel the same way in in AI applications to provide context of users so that the responses from language models can be consistent but what I used to say in that is um that that I Chris uh like we have this one system that we use at iSolutions that we run everything on when I'm logging in via my phone I'm doing very different things and often have very different goals than when I log in on an iPad or let's say a desktop right absolutely and that just makes sense to us

and so this study actually addressed that from a chatbot perspective too and so Microsoft didn't just ask people what they do with AI they asked how it changes by device and by time of day actually so they found that mobile looks uh fundamentally different from desktop use probably not a surprise based on what we both agreed on a second ago but on mobile health and fitness is the dominant topic so kind of when people are in their fitness activities or health activities or about to eat something or whatever right it stays dominant and it stays dominant across every hour and every month that they observed now on the desktop the workday as you can imagine really you know impacts their data so between eight o'clock and five p.m work and career take overtake technology as the topic so technology actually maybe for you and I might be our work day you know um you know uh most common topic but for some it's work and career and then technology is after so kind of interesting huh yeah and as I think about I hadn't really thought about the distinction but totally it's there um they taught they said this interesting phrase um AI is quote a colleague at their desk and a confidant in their pocket you know kind of marketing speaky but I thought that was kind of interesting yeah I also pretty much use that I use the chat GPT app on my phone but

not necessarily the one on my computer because the website gives you a richer experience on the computer it does yeah I'm glad you mentioned that because like you know sometimes you can I use one of my pro tips is I I um sometimes if I lose the thread and the conversational memory I will intentionally go back into the timeline and I'll edit one of my um you know uh instructions and then you just hit save and then what it does is it erases everything that came after that from the conversational memory if things start to get a little noisy and so um I find that often the edit option isn't even available on the desktop app and if I just simply go into the web app for example it is so uh kind of a pro tip two pro tips there maybe for folks I'll give you a pro tip um one of the tools one of the tools that you get on the iPhone version is the Apple music tool oh I saw that I was trying to figure out how to actually I saw something online about that um it allows you to query the whole world of songs and say I want a playlist by such and such of an artist that's this or I'm just trying to find some is vastly vastly better than what you can get like in well I think I accidentally used that just yesterday because yesterday was my uh one year anniversary of when I we first had brought our dog packs into our house um and so I

had a one-year retrospective that I sent to the uh Foster family and the shelter people to just you know make them happy and look at how much he's blossomed and all that but I needed the right song and so I went into chachi BT and I said well look here's my situation this is what I'm doing I don't need it to be about a dog you know and it just gave me just layers and layers of all these songs and then I could click on it it actually went to a preview and then from that preview I could go into Apple music and it would launch it on my computer and listen so it was perfect and and you know what we're talking about stuff that Siri should be doing yes without question I mean what are we doing come on Siri I used to it's so easy I'm a nerd about classical music choral music especially and I'm I just want to test it out the first time I tried it I said give me the top five amazing choral pieces of all time and one or two that I knew about and then like the number one was a piece that I've known and sung but it was a version from a choir I've never heard of and oh my gosh absolutely perfect recording just you're gonna be so proud of me because I just bought tickets yesterday to go see um a uh organ performance like uh like uh uh at the Disney uh

Walt Disney you know not proud of you but go ahead and I'm very excited about it mostly because I'm obsessed with the movie Interstellar and that that Hans Zimmer theme just like is just experience for me when I listen to it and I thought what the heck I'm gonna go listen to something like this live um and uh I've never heard actual like pipe organs really going live and so at least it's getting me to a concert hall but maybe AI didn't help me you know uh having the experience of having an organ in a concert hall or in a church uh is pretty amazing because there's so much I've never experienced that there's a direction of every single pipe they're all in different places the bass shakes the room with those big 32-foot you know well that's the thing with those big 32-foot you know that I will bring it into AI because I said to AI hey look these are the seats that are available so obviously I want to be down on the floor where all the VIP tickets are that are more expensive and they're like no pipe pipe organs have nothing to do with directional they're not coming at you right you know they literally said go on the side go about halfway up into the cheap seats and sit there because it's actually filling it in from the like the roof down and and like creating this thickness of of sound and I'm like oh perfect so it literally gave me the exact section and what like range of rows and saved me like 100 bucks frankly so um

thanks AI and you know maybe another thing that Siri could do so here we are let's see we've just talked about our aging and uh how we want to be healthy when we're 80 and church stuff and uh orchestral music how about we we can do that but I have one more nerd thing to show about pipe organs right okay so it's a kind of fugue and D in D minor right like a super famous okay it's a D the the uh is that the um the the Halloween it's for everything is it okay yeah it's just been ever so it's in D and the lowest note on the foot pedals is a C pretty much on a lot of keyboards just like that and if you play the D and the C together the two lowest notes which wouldn't be a good sound because they're next to each other at the very very low end of the uh foot pedals you can't really hear the dissonance but the but this undertone beneath that mathematically is the D an octave below interesting the room shakes man really yeah so basically it's creating a note that the organ can't play by using a mathematic wave guide type thing not a wave guy type thing but anyway oh who knows if they'll even do that but this is the kind of stuff that's amazing dude I'm seriously I'm I'm so I'm I'm into this I get it I'm a tourist here in this this type of

hearing yeah my experience with you and recorded organ music I just can't it's just you lose so much compared to yeah and it's still moving uh to me so um yeah I'm gonna go see a particular pipe organist who will in fact play the interstellar theme and I know that's very like commercial of me but I can't wait and do it whatever gets you in there misty-eyed now here here we are sounding like a couple of serious old dudes right now uh why don't we try to figure out what the kids are up to when it comes to AI as a matter of fact one of those papers yeah that paper that I showed you from the Pew Research Center actually surveyed about 1400 U.S teens uh specifically between the ages of 13 and 17. this was um cheers indeed this was uh between September 25th and October 9th and um and really the idea is what are the kids doing this is a bit of a harbinger of you know what young adults will do and then adults and so on and here's what they found 64 percent of teens say that they use an AI chat bot so 64 two-thirds over two-thirds it's interesting and when you look at frequency you get this really interesting it's under two-thirds two-thirds or 66 we'll let that oh it's just under two percent damn um and teens would probably know that as well uh and so when

you look at frequency four percent of all teens say that they use chat bots quote-unquote almost constantly 12 say several times a day and 12 also 12 another 12 percent say about once a day so it's like 28 percent of these uh kids teens are using AI daily so um it's like this would I assume get into you know the uh the new normal type uh range here and chachi BT is the top brand among teams uh saying that 59 of teens use chachi BT so um you know when adults say kids are doing everything with AI well I guess in fact they are is this surprising to you it wasn't too surprising to me I mean U.S wide or where's the oh good call yeah U.S teens I'm sorry just U.S teens yeah because it's got to be different in different parts of the world I would assume just I mean it has to be South Korea probably a lot higher you know yep Greece probably a lot lower other places um yeah if you want to check out the paper it breaks down how they use this versus like Instagram or X or Facebook or Snapchat or YouTube Reddit yeah yeah yeah well that kind of stuff it's kind of that sounds fascinating does it say what they use it for is it different no it would it just went

into frequency of usage not into the what like like all these other papers do but um but you know that's what the kids are doing kind of getting out of you know in pure consumer probably school work stuff you know I would have assumed a lot of that is but um but yeah the some of the other papers uh talked you know get more into the business use cases so we're kind of swinging the pendulum around here a little bit uh there was one of them the open AI one I think is probably the most interesting they actually um yeah it's called the state of Enterprise AI but that went up it had a really amazing graphic on the front but I'm curious what that means yeah you know they do all sorts of um um here it is they have yeah they do this kind of funky like this pattern thing they do on all their blog posts so I think maybe they went with like it doesn't really look good it you know what I'm sharing but if you check it out um it's pretty cool uh yeah so they're um you know the you see what some of the key findings that they say stand out here are but in this particular one they these are some of the things that I thought was interesting more than one million business customer users now use open ai's tools what you know the tools meaning chat and some of the other you know anything you hit on the plus one one there and look at the first one man oh

yeah volume increased by 8x yeah that's the other thing too weekly chat GPT Enterprise message volume grew about eight times since just November 2024. eight times and the reasoning token consumption per organization increased 320 times year over year and chat GPT Enterprise seats increased about 9x year year over year since they introduced it so people are just gobbling this up yeah yeah and the more interesting use case it's not going to catch on Chris yeah I think it's like rock and roll or designer jeans um it's definitely going to go away because it was the same thing I mean from one year to designer jeans to the next there was a 320x adoption rate it's insanity and actually this is kind of I think it's maybe even more advanced and uh snowballing than we think because um the the the highest growing or fastest growing use case signal in this paper was what they described as configurable AI workflows are becoming uh normal and so that's like you and I have already talked about it creating a custom GPT falls into that category or um you know just tools where you're you're I I would dare say doing some context engineering so um and then the last really

interesting stat to really freak us out is they reported that weekly users of custom GPTs which you mentioned you were using and projects which I mentioned today I'm using have increased 19 times year to date since this time since the publishing of this uh paper you know what I really want to know is how what percentage of people still aren't using them because there's still so many people that I know that don't know that you can customize are aware of chat GPT but are not aware of the customization right yeah I'm not aware of the yeah yeah yeah anyway or we can ask chat GPT yeah I have a conversation of this somewhere so um look at this codex engagement uh 2x in weekly active users 50 increase um oh here's some uh workers uh yeah 87 percent of IT workers work fast 85 percent of marketing 75 percent of HR man it's just yeah definitely yeah that's a dressing paper yeah yeah it's got some good insights as you can imagine and you know they're you got to take it for what it's worth there's well I mean they use tragedy to write the paper so 100 of that's just made up telling the research it's all hallucinations um they also

did do a survey of 9 000 workers across almost a hundred different enterprises in this paper and a couple of interesting tidbits there were that seventy-five percent of the surveyed workers say AI improve the speed or quality of their output so that's encouraging yeah uh they also attributed uh 40 to 60 minutes saved per active day because of AI usage um I would definitely concur there I think the in the little shorthand data that I'm gathering on my own it's significantly higher than that um 75 say that they can complete tasks that they previously simply could not now that I think is really interesting at the worker level Matt because I'm hopeful that that from a entire business and Oregon org standpoint we've talked about this on previous pods I I think that uh businesses shouldn't look at like oh I can use AI and I don't need to have humans that's I don't prescribe I I'm not yeah that's not my message at all but I do believe that businesses can now do things they previously couldn't do with the same amount of individuals that they have and I think there's something really profound there yeah I really I kind of break it down into two categories I still see the whole world everything that I think about with AI is how I can empower FileMaker customers and people

that I know to use these tools in an easy way to do their job better yep that's it and and and those two areas really specifically it can either save them time during the day by doing things that are rote or boring that they could otherwise do it was just to be manual and time consuming and then the second category is things that they just couldn't do before it was just not possible multilingual search you know like a multilingual totally ridiculous you know fuzzy logic is what we used to call it search for names where you can make whatever kind of misspelling you want for whatever any language that was just not possible before and now it's trivial I you I know the the the speed at which things went from not possible to trivial is probably the the craziest of all these um the another one of the papers I'm showing the the cover page not quite as sexy as the open AI one this is the one that you pointed out earlier the Harvard University and perplexity do you use perplexity no I feel like that's one of the tools that the cool kids are using that I don't use so I have it like open um I did steal something from uh I'm blanking on the CEO of perplexity's name but I I've heard him multiple times on podcasts and he mentioned

one thing that I then stole the idea of and built into an actual business AI application and that is that in perplexity behind the scenes when you type in whatever your question or your query is they then refactor that five times and you know try to ask the question better before they go do searches based on that question I thought that was so simple and so brilliant and if you combine that with like personal context if you're gathering that and you're doing it in a sophisticated manner you can now rephrase or refactor questions more accurately for what the intent of the user was than the user could even do in a lot of cases right depending on their environment factor so that was kind of cool but this paper there's a there's a important word here agents we've talked about this before um and of course yeah we're in the alleged year of of the agent so we'll have to figure out if that came true or not but the term agent I get a little prickly uh with the misuse of the term agent I hear it a lot um and I'm not like the you know the judge and jury of agent usage but they did

execution maybe the executioner though I think what happens with agents is that um well first of all um you could ask chatbots things and you could ask chatbots to go do web searches for you before language models had the capability of reflection and what why is that important to the definition of agents well agents at having it act in your agency or doing something on your behalf is not what an agent is an agent is actually um something that uh creates its own plan executes against that plan decides what tools it needs and then it reflects back to determine whether or not that was done right which gives it autonomy right whereas the first example is merely tool use it's tool use and and that's fine agents is one big part of tool use but or vice versa or tool use is one big part of yeah yeah no doubt about it and there you know there was a paper that came out that I think did a decent job of providing some clarity on the industry and it was from uh Anthropic and they defined they I I don't think they did an amazing job of defining what agents were but they did an amazing job of of saying what uh LLM workflows were and I think what that allowed people to realize is oh that's what I'm building I'm not building agents I'm building LLM workflows

I I have you know LLM calls at X number of times throughout a somewhat deterministic workflow maybe I'm using it to make the workflow somewhat dynamic but I'm the one as the developer who determines whether it was done and what the plan is what the steps are and what you know maybe a couple of the steps happen to be language models but in true agents they're the ones determining what steps need to be taken and what the plan is or what tools are necessary or what data needs to be gathered and then they determine whether or not that was executed properly and in some cases which we've talked about in previous discussions there's multi-agent systems which may or may not be evolving even further in 2026 where agents are telling other agents what to do so anyways in this paper they say an agent takes actions in external environments not just using tools like web search but other tools as well too so so in this paper uh essentially what they're doing in the in this was supposed to be the you know the the year of the agent so harmony yeah um but that might be a little scary um no I mean that's actually really important yeah to the environment that the user is in that session the user is in which is the

definition of what it is right it's you know anyway yeah super helpful um and so what they did here is they published this paper this was just in December they analyzed hundreds of millions of anonymized interactions with Comet you know perplexity had that AI browser come out earlier this year called Comet and they scraped all that that super valuable information and this is what they came up with they said that uh this is what people use agents or agentic experiences I don't think that people know their agentic experiences I think this is just another what people use chatbots for but in the case of uh perplexity productivity and workflow were 36 percent of the queries learning and research were 20 percent and together that's about 57 percent of all agent use uh usage productivity workflow no surprise and learning and research those are pretty wide categories but probably not a lot of surprises there I'm guessing and uh let's see and then you can actually see and within research that breaks down into summarize and analyze research information create and edit documents search and filter products and search and filter research information so yep yep um you know a little bit of that extends the help me think that we saw from uh the earlier paper now that leaves us with just

like two really big problems with that paper please let's go one of them is that they use two spaces after every sentence rather than one which is a complete error and second of all it was right it was right and left justified come on amateur they had a minimum number of pages for this assignment so they yeah whatever they could have just increased the font size by 0.2 percent no one would have noticed anyway well they did say something about environments and I'm trying to check that a little bit it showed that environments are where agents operate and the top environments help me make sense of this they define the top environments as Google Docs was the number one environment at 11 email all email services combined tied it just at about 11 okay LinkedIn was an environment YouTube was an environment and Amazon so yeah I'm not sure what to make of that does it well well I mean I use a lot of those right so um uh like in Gmail you can I think you and I talked about this one time um I was trying to remember a name of someone and uh I couldn't so I put in to uh mail like there's this guy I knew from a long time ago he was really into shoes he

was all into uh value building at the time and he was a solo developer shop but I can't remember his name for the life of me three seconds later it came up with Jonathan Stark and I'm like that's now that was just there's no possible way yeah that's Gmail with its AI agent in there so it's it's great at something yeah yeah yeah so and I think the same thing I don't actually use it with LinkedIn that would be interesting um they have like the little rewrite this in AI and and I'll tell you what LinkedIn clearly has although I haven't used it myself but most comments I get written so it has like a couple of the things that I posted about and I I get I think you can just turn that on to in increase engagement but it's just really obvious that the human people aren't actually writing it when they're commenting so I would say maybe write your own holiday cards by hand no AI there and uh and actually make your comments on people's stuff genuine in human um uh the the last thing on this paper by the way was just the there was a usage breakdown 55 personal usage 30 professional and 16 educational so you know professional still not yeah yeah you know cranking it out but but um okay one last one for you man and then we can

uh move on to making our own predictions this one's got a pretty cover to it too estimating AI productivity gains from Claude conversations so uh as you can imagine Anthropic did the same thing here um they um you see now hold on here this one is using the correct formatting it's uh it's jagged justification on the right there's only one space between anywhere let's go there was there was no I don't want to give you PTSD I don't want to make you nervous about this margin usage here you guys can check it out on um on YouTube but uh they analyzed in a privacy preserving analysis system I love everybody's word usage here they analyzed a hundred thousand conversations you know yeah of course naturally um so uh they uh they generated two core estimates for each task time estimate without AI and a time estimate with AI so they're really just trying to see you know again productivity gains right so um the headline claim is pretty aggressive they said that the time spent is reduced by about 80 on average if they you if you use AI you're getting an 80 increase they estimate that the average task would take about 90 minutes without AI so whatever they determine their average task out of those 100

000 transcripts uh would take 90 minutes with without AI and I'm assuming it didn't take 90 minutes with AI I was unable to find out the actual piece of data as to how long it took so was this well I guess the one thing it says time estimate without time estimate which is I mean the estimated time after the fact or the estimated time that the person thought it would take I think yeah yeah that's it so there they were determining what they thought it would take so for example someone might say hey help me write an email to my colleague with blah blah blah right they're saying they think a human written email would take X amount of time and they can determine how long it actually well they can actually do a with a large enough sample size they can actually do a pretty good uh they can get it they can get a pretty accurate result of that for different types of things yeah and I and I'm assuming and they go very deep into the science of how they determine that actually so deep that they've already lost me uh if you can see some of the um yeah so some of the example AI accelerated task vocational educational post-secondary makes sense library and science often these are these are roles so this is like research right um really interesting paper uh Claude's estimates by occupant occupation category sorry uh time

without AI hours so yeah you can see what they're basically doing for those that are not looking at this they seem to be saying if you're in management you can save two hours a day on AI tasks if you use AI because it would take you a two two hours extra without AI and that is uh an increase in 68 dollars an hour in a wage which that's an interesting way to look at it and 133 dollars in cost and a time savings of 85 I mean it's that's insane I mean that's the top column of time savings there's not a single one of them below the lowest one I see is 61 most of them hover around 80. that's that's crazy well actually let's look at the lowest one transportation and material moving yeah that's the lowest 79.4 yeah that's right yeah and there's still an increase of 23 an hour I mean this is this is undeniable money in your pocket especially if you're a small business right so yeah um yeah it's crazy so so essentially if you pull all these different reports together you know people are using AI for probably some no not too shocking things like searching but advice uh creating drafts um which everyone

should create drafts desktop looks like work mobile looks like life team teens are building habits um which is interesting for the future and uh in Enterprise they're not just using it for chatting the custom GPTs projects and embedded sort of context engineering systems have a 19x growth which is amazing yeah um agents uh are moving into the next step of delegation which is often described as multi-agent systems and product every gains are real well but measuring them some way or another certainly matters so crazy stuff uh check out those papers here I posted something on LinkedIn too if you guys like it give me a non-ai generated comment um and uh um but you know summarize some of these things that I thought was interesting as well too so yeah still never see all right so you're now you want to frame up how you uh muted oh I think I know why um I think when I shared my screen I turned to share sound on um you want to talk about your process for putting these together yeah so I took I did a um a

um transcript of the entire podcast we did for the year in review last year cool um and then I put that into chat GPT and had it analyze it um and so the predictions that each of us made so let's see I thought token Windows actually this one was maybe more you you talked about token Windows yeah getting larger I think maybe we both did the average token window at the beginning of 2025 was 128 000 that was like one of the biggest ones you can get and now it's a million two million yeah I I think though like I I'll you can take the a on that I'll take the C because I openly made predictions about 10 million token Windows uh by end of 2025 and that didn't happen not in the podcast you didn't yeah no not on the podcast so I'll take I'll take the win when I can but you know in this in the podcast according to the actual text you get the total a on that one awesome yeah I'll take it I'm not going to talk you out of that for sure yeah I talked about not predicting AGI in 2025 and yeah no AGI I'm going to give you even more credit for that not only did AGI not happen in 2025 and by the way people were legit talking about it happening in that short of

a term but I'd go so far as to say kind of people aren't talking about AGI anywhere near as much as a matter of fact they've introduced the new term of um it's not super intelligence yeah like ASI or something like that yeah yeah yeah so now the goal posts are moving yep so as they always were I mean the Turing test has moved and moved and moved you know yeah we've long passed it um I not to jump ahead too far but I actually predict there's going to be a few organizations claiming to have reached AGI in 2026 but that it will not be widely accepted well and that's and therein lies exactly why I'm of the AGI is not upon us because there's no accepted measure of of AGI like there's no one out there that says this is what AGI looks like and how we determine when it's here and also it's not much like human cognition works and learning works it's not a light switch it's not all of a sudden one day oh my God we've achieved general intelligence it it's a frog boiling exercise on intelligence and knowledge so I'm with you on that one um I do see one of my predictions for 2026 here in your next line so I guess maybe well you know small language models I didn't actually talk about so I think this kind of put stuff together because I was talking about that

in a later conversation in this whole thing okay um so that one that one is actually one of my things for this year and I'm not so sure how big they're going to be but anyway and then the last one is Apple which is really bad it was it's so essentially we were saying it's just not even there yet and oh yeah we were talking about how we completely turned it off I guess that was but it's just so frustratingly bad it's I mean we already talked about it here what it it should be the music experience we just described that someone else is now doing besides them it should be super easy and I mean yeah I travel with this thing all the time I mean just so much of the stuff that I'm I'm offloading to other AI tools that could just happen right on my phone it's so frustrating I'll say get directions to blah like there's a store called practicate which is like a German store that sells hardware like at Home Depot um and it will say getting directions to practical North Carolina located blah blah blah it's like dude I'm in Crete I'm not gonna drive to North Carolina the like you know where I am you know where I'm going you know what I've previously searched

before you know the kind of things that I've looked at you know the previous conversations I've had you know that I've been here before work really hard at figuring out what I said if you don't understand it as opposed to assuming it's in North Carolina it's like uh you touched on two two of the trigger things for me one is their promise of personal context could have been amazing they could have done it better than anyone else could and they still could do it better than anyone else could because of the trust that they have with users in their sandboxing all of that is what I have faith in them to actually not you know expose all my data to the world I'm not as worried about that stuff as some people are but but like you've got to be kidding me you have every text message you have every utterance you're you're constantly listening to me you know everything I'm doing the way I talk who my friends are what I talk to them about essentially where I go to eat like every app interaction and intent that's in there how dare you Apple not figure out how to do something with that because that's the that's the type of personalized consumer experience I'm going to have with AI and apples just yeah and then when it comes to health stuff it knows I mean with an Apple watch it knows my heart rate it knows all the other stuff that I'm putting in there my heart rate variability uh how much I exercise blah blah blah anyway even if

you're not using like like the aura ring it still has the app and the day I do and it is critical and sleep management for me um but like my that app is on my device and that data is stored locally as well so you know it could be part of my personal context that's true I use the Apple watch to track my sleep it's pretty good okay so I don't know if there's anything else that you want to add to that okay so your claims um agents agents agents you're yeah you got an a on that one because they're absolutely were huge a token window explosions you actually did talk about that you didn't talk about 10 million it says a plus I'm going to say C because I did I I thought in my brain it was going to be 10 million and we just never got there well I guess that's going to be coming what's going to be happening for next year um uh Clara say I wish I should have been talking about more but basically you said it was going to become much more built into the application and have a much bigger effect on users you were totally right about that you know they came up with a couple of yeah I mean think about it all we had was semantic search this time last year yeah I know and then adding the other stuff was just huge huge huge uh and now that now that MCP is you gave the F well X so you said X will blow your mind it's gonna be huge well first of all go look at you definitely gave that F uh for your own personal

reasons but no I didn't literally go look at every Benchmark out there for actual performance of models and they're and they're I'm not stumping for that no they're literally number one with a bullet in most of the uh reasoning categories and the categories that everybody else seems to care about so what I saw in a huge study was they were they were the far and away the worst for hallucination like 94 percent hallucination rate to 14 percent hallucination rate on some of the other 74 percent hallucination rate that's absolute insanity what I saw that's a hallucination dude yeah um they move at the time of this prediction they were only it was only available inside of X now it's a completely different tool uh they went from uh a player who was in their second generation after only a year to now their fourth generation after two years and they have the largest data centers and they're completely crushing people on compute I think yeah but they have 0.1 percent corporate uptake you mean oh you're talking about like adoption rate yeah yeah the adoption rate is tiny so yeah I don't know maybe that will change oh my prediction of uh breakout player absolutely happened they they were not even considered and now they're yeah in the top tier of

all Benchmarks and they are there as a matter of fact what happened earlier this year where um you know how like one model provider you know a foundation model provider comes out with something they're number one and then two weeks later somebody else does it yeah that happened an entire cycle of open AI and Google uh and Anthropic and they still didn't usurp the X model at the time but I'm not going to sit here and stump for X because well I think it's worth checking out because it's been off people's radar for a long time and I think because of the way you feel about it it's it should be and the tool is actually really good the super Grok tool uh has way more like one of the things I get frustrated with when I'm trying to build uh documents in a canvas and open AI is the throttle the amount of tokens that the doc like you get to a certain point of the document and it just stops building it properly for you or editing it properly because it has its own internal limit the super Grok tool doesn't do that some and I've been actually very frustrated with uh 5.1 and I'm I'm not so sure if it's fixed in 5.2 from open AI with a lot of model slippage and like you know context memory stuff issues I was running into and I would find out in the context window got smaller than 4.1 was a million tokens and it's like 600 000 and something like that that's maybe part of what I'm experiencing too and it's much larger over on super Grok so those are just you know I I don't

have a single daily driver I use chetchpt Claude super Grok and various other tools like perplexity you know I'm gonna have to move into it because we talked about how we need but okay I'm not just one you know one tool anyway we both did great you got dragged down by X but I don't know maybe it'll happen um so uh I don't know if these are some of the other predictions this part I didn't I didn't do that much work on but basically like what what thing's gonna happen so kind of my predictions are the AGI prediction oh yeah this one I am this is dismal but here's my prediction I think that in the FileMaker world there's so many apps and I think there's gonna just not be overwhelming adoption by clients to use it I think there's so many thousands of small databases out there that are just chugging away for their company that by the end of 2026 they're gonna go this is great but I haven't seen anything that my workers really need so I'm not going to put any AI into it this is the thing that I'm fighting most hard against but you're talking just to frame it up specifically adoption of AI tools in the FileMaker native yeah natively adding native tool okay yeah so that so that there's some API call inside of

FileMaker to run an AI function summarize translate uh sir I'm not gonna I'm not gonna counter that at all I'm gonna I'm gonna throw another log on your fire I I think I am actually going to predict that FileMaker will not keep adding AI tools but instead focus on um mapping outcomes to the existing tool sets that they have right now I think um it's pretty evident that some of the uh product decision making on the FileMaker AI side has not been validated by the marketplace desires and needs and they've gone now two and a half cycles on that and it's quite obvious because I think uh there's probably single digit adoption of the tools that they have out there right now single digit percentage adoption is what I'm talking about yeah uh evidenced by some of your own research that you did when you were talking to even presenters on AI topics at some of the big conferences who said they don't even have anybody and and I'm not even counting I don't care if somebody says they use it and they tinker around with it my question is is anybody paying you to do this like like first of all do you have the ability to do it so that you can confidently charge people for it number two are people actually transacting on AI and I think that's in the zero point whatever percent adoption rate within Claire so I

think it stays well below 25 but Claris I'm going to be hopeful for them will change their focus from just you know t-shirt canoning uh acronym AI acronyms out into the world because I don't know why how they're deciding this they lick their finger and put it up to the sky and you know just throw something out there they'll actually be more deliberate with this and will now communicate uh in plain English to people and talk about what outcomes can be achieved with these tools and that will drive it's just not gonna be enough time to get above 25 though yeah I totally I totally get that so my mission like my goal for 2026 is to actually uh offer services so that if companies want to do like a proof of concept or even a fully integrated really really small small thing to be able to in a few hours meet with them take a look at their application I've got a big toolbox of ones that I've already done that are small figure one out that's the lowest hanging fruit for their application integrate it let some users test it and then see what happens you know it's a foot on the door kind of a thing because it's so hard and people are thinking way too big about what they're going to do yeah it's like oh yeah yeah we're going to implement an eye it's going to take over the entire thing and you just push one button and the whole entire thing happens it creates a project and assigns

it to people and blah blah blah blah no you start with something small and then kind of get to that yeah I'll validate validate that by saying um our process on the AI side not the FileMaker side but just the AI business is we deliberately start with experiments we just say what small little thing even if they go we want this big huge thing we actually work flow it all out and then what we do is we do we like circle little areas like well we got to prove that AI can do this one piece we got to prove that AI can do this and so on and so on then we do a bunch of different uh experiments some of them will prove fruitful some will surprise us some won't and then we stitch together what we learned from all those into workflows and they usually aren't big huge swings they're just little things and I believe so strongly that you know wrote wrote tasks can be eliminated with AI and users like what's going on in the world of AI yeah I think that's a good point I think that's I would love it that that's where every organization needs to start and I and I I like your methodology we do something similar and I think everybody should do the same thing um let's see by the way is a generic and probably unmeasurable um prediction of mine I said it in the frame I framed it up as far as uh Claris was concerned I think ROI driven AI deployments like

where businesses shift from experimental pilots like I've talked about to actual projects with clear ROI um I think we will see growth of that but it's going to take where people identify specific tasks that can be made more efficient with AI I think it's just a shift in thinking the tech doesn't have to change it's just how people are going from I think you actually articulated this perfectly stop swinging for home runs like with you know going big just go with the little stuff and yeah little tiny ROI pennies into nickels and the dimes first yeah yep get a guy on for oh look at that I mean sports I mean not even that right like uh hit hit a ball that's coming at you and don't let it don't let it be an out you know yep get on base this is money ball it's literally just get on base or at least at least not have it be a not have it be a strike that goes past you make contact with the ball even if it's you know it goes out of bounds because it's a fly I mean keep with the baseball metaphor um no I think it's I think it's actually perfect and so yeah that's what I'm hopeful for I I think that that mindset shift will change I think um well I'm going to go back to one more that we already took the W on but I want to revisit this and that's small language

models and there's a couple of interesting things in here uh first of all this is going to become a problem because like I felt a little icky when we when we went we were talking about like language models and all of a sudden everyone just started calling them llms it was a little tough for me to adopt that but it was so clear that everyone was saying this that I just went ahead and just started calling them lms as well but the first L is for large like it's not and the assumption was that they were just going to keep getting bigger and bigger and bigger and pretty soon they were just going to take over the entire universe um so much so that uh in the the brilliant and extremely well run state of California where I live and have my business uh these geniuses uh came up with this um luckily it didn't get voted in but they wanted to completely shut down and restrict all AI and AI related Commerce and one of the measures was that if your models were certain parameters in size like they were so sure that this was just going to keep getting bigger and bigger and bigger is that they wrote um what was going to supposed to come into law probably would have been a federal law but it didn't end up getting uh didn't end up taking in California but they were assuming that

was to get their thing was just gonna get bigger the interesting thing about it is the the metric that they used in AB 1047 in case people want to look into it was never achieved as a matter of fact what happened instead was and it was actually this time in 2024 when that was trying to get passed so what happened in 2025 is model a lot of this was the um God what was the China one again damn it with the whale uh deep seek so deep sea came out and they and you know distilling of models came uh you know these processes training of models with less compute that which is what deep seek really changed the game on all of these actually were able to produce larger models with smaller processes and less compute it doesn't mean they're using less computer just going to keep you doing more compute yeah but the models are getting small yeah and then all of a sudden we started seeing model capabilities uh I think oh four mini was the one that really was profound for me and that was where the first time you know they were coming out with the big model and the little one you know 400 billion parameters and then seven billion parameters or something but the seven billion parameter ones were kind of hot garbage you really couldn't do much with them but then the 04 mini came out maybe it was maybe they're doing reasoning or I think these were distilling uh the larger models they were really capable and so not only do I think that this is actually going to

continue it is and I know it is but I think the awareness of models like llama 1B Mistral 3B Gemma 3B all these are out there and available and that's moving uh reason are those in the same are those in the class similar to like uh 41 uh like uh 41 Nano 50 Nano from opening on um I don't know what the billion parameters are but the Nano concept is is indeed consistent with this class they're essentially trying to move the processing requirements for reasoning off the the compute into onto the edge and we've in in one of the things that Apple has done very well somewhat ironically is they build excellent silicon for uh for inference like our laptops are some of the most powerful inference uh capable machines available to consumers that aren't you know like Nvidia class chips or something like that so bringing that onto the device and I in and the reason I wanted to bring it up as a prediction is I I hope to see it becoming more profoundly moving to the edge but also this is how people should think about building AI systems um one of the things that I see working with clients or teams that are building products

um one of the most common things I see people doing wrong is they think that the model has to do everything like the most the the biggest example I can give you is let's say you're creating an AI app and what people will do smart people too and large teams of smart people will try to make it all happen with one big prompt and they just go well the model is smart like you know 5.2 just came out it'll figure it out no stop doing that first of all sure maybe the model is smart break your one prompt into as many different calls to an LLM that you need to that match your workflow um you know one that I looked at for example I was doing an analysis on somebody's monolithic prompt which is a really poor short-sighted approach and instead it could have been broken down into about 14 different calls and then when you start looking at those calls you go the way you should do it is you should try to make these calls work first of all as many calls as you need break them down the more the better and then look at those individual calls and see if you can accomplish what that one tiny task is maybe this is just classifying a chunk of text or extracting the the city from something or redacting some PII I think you'll find that these less than 10 billion parameter mini or nano models

can totally do all that as a matter of fact look no further than on our devices back in June WWDC at WWDC Apple actually announced that the foundation models were going to be available through app intents on device so Swift application developers could develop it could actually load those models locally onto people's amazing inference machines on those cool mchips and perform little things now the problem is the token windows are only 4096 tokens for the windows for those models so most people who think it's all about t-shirt canoning AI acronyms into the into Foundation models is the way to do it they would think well I can't do anything with that oh no I think if you break your things down your processes down enough it can absolutely handle most things that you're doing and then you leave the big calls to the cloud for the Foundation model so anyways your thoughts on that no that makes sense and that's one of the things again that Apple I guess they were talking about doing it but um they're really good in a good position to do and I think really it's the whole commitment to privacy just um was such a huge limiting factor because Google and other ones just don't care about it yeah there's definitely some yellowing

going on and it's the reason to stick with Apple in a lot of ways because of that um and then just run the chat GPT app on your phone which works perfectly you know I I am um I am uh uh holding out um uh uh that they will be able to leverage that somehow um so yeah you know what it might be like tool use if they open up um if they open up a bunch of the tools on the phone for that like for example the music when I was talking about it can create a playlist so you can have a conversation and about a bunch of stuff and and then at the end you can say okay make a playlist and then you push a button and poof you have a new playlist in Apple music that's from then on there that you can then use for whatever you want but what it cannot do is edit an existing playlist because privacy so um I think we're going to see you know some other stuff like that but what it's not going to be able to do is like read all my emails um and find something maybe well I don't know if it's on device in the sandbox I don't care if it's accessing that information I just don't want to expose that information to the outside

world where it's unnecessary but if it goes and finds out like it reads every text message I've ever sent to my wife since I got an iPhone sure cool that'd be so great here God if I could say there's this thing that I sent a couple years ago it was a picture of a blah I find it I can't find it because you know if you're going to find it um and like the text search thing like you have that right now it's just useless to try to find oh my God it'll find a word but then if the conversation was above or below that forget it you can't just at least employ semantic search across the device on the OS level yeah so but I mean I don't know so you know these things have to be coming um have you tried the language translate function the real-time language translation yeah yeah yeah I haven't um but I I'm only because I haven't had opportunity to but yeah it's cool I mean it's it's cool for sure great it's not one of the languages though I think it supports like just three like Portuguese Spanish and Italian or something very very few well check any box Apple keep doing what you're doing um I'll throw one last one out for predictions only because I feel like why hasn't this um happened already voice I I just can't believe that voice is not just completely

dominated the models are there the tech is there the the support is there from a silicon standpoint and a device standpoint I I think there's the the one profound limitation for voice adoption at work is the cubicle effect it's the the like you don't want to have people hear you say this but I think that's probably offset by the fact that like fewer people are working in offices than ever before so one would think this is the right time for talking out loud to your device but um you know I found it as a user I I on dog walks I just ramble on for 45 minutes or something you know and and and I it's easier I'm not I'm not a good typist I never overcame typing um so hunting and pecking my thoughts is not as easy as just saying them and the the transcription technology is where it needs to be I just like where is it right so um I don't know a big a big change will have to happen somewhere for that but yeah totally maybe that's 26. it's almost impossible to type as fast as you can talk unless you're using a court reporter machine and those can easily and you're a court reporter well yeah trained on it of course yeah

exactly yeah so yeah I'm a very good typist and fast and there's no way I could keep up with something but at least it's not frustrating for me um yeah it is for me and and I and yeah I'm here for voice uh certainly yeah I get that and you know when I don't know if you knew this when you read like if you just pick up text and read it you're sub vocalizing what you're reading it's just a built-in part of your brain oh parts of your mouth are moving as you read words on the some people's more I think everyone yeah but I think it's a thing so why couldn't there be an AI thing a little sensor on your neck or whatever well who knows maybe that uh open AI device that Johnny I've and yeah yeah so I mean this means like if you're just thinking like if I think out loud you know I just thought a sentence to you and I could feel my mouth moving so I could be sitting in my cubicle and if there was a device that could actually track that you know we have this I mean you see the crazy Tech that's coming out with the brain implants and people are able to like send messages and play video games like there's there's there's a connection there I mean that's the fastest way to get it and that's with hardly any resolution like the number of the number

of electrodes to the brain like 1024 but imagine you know millions yeah yeah yeah I'm uh definitely here for that as well uh so do you have another prediction you want to throw out there gosh predictions voice was my last one yeah that's a big one um what I really am interested in uh the two categories that I'm that I'm interested in are the ones that are going to be surprising and they're going to hit us and the ones that we really want that are just not going to happen yet okay a big one is the dishwasher robot so it's it's it's a turns out to be really really hard for a robot to take dishes from a sink put them in the dishwasher wash them and then unload them without breaking them you would think that would be like just an arm but it is absolutely not there they're not uh if you do actually a big search on it you're not going to find a bunch of stuff as physical AI by the way there's I heard I've been hearing people talk about the distinction so yeah so so that's coming because almost everyone would identify that uh I think I heard a thing today like number one on the list was a dishwasher robot and the very very last

one on the wrist list was a present opener oh a Christmas a Christmas package unwrapper like nobody knows what happened no that's the one time you do want to do labor exactly that's precisely what I'm saying but nobody wants to load and unload dishes and so if I'm and there will be that will happen but I don't think it's going to be 2026. no but I'm here for it you know then of course just the the they have to get the price points down so that everybody can have it and you know it's a whole like when the tech is in a rental car then everybody's got it like you know when when unloading self-loading and unloading dishwasher washers can happen inside rental apartments then it'll actually have an impact on society otherwise it'll be like a bougie rich people thing um but whatever I would like to be one of those people someday it's kind of shocking I was a couple of weeks a couple minutes later I mean today because I was put to get putting together an exercise bike that I bought for oh the house and and it's it's actually fairly high tech and cool and it was 180 euros oh so I mean if you think about if you could think about something with that much technology a screen a magnetic uh you know device

for setting attention um and all of the other hardware yeah obviously a dishwashing robot's gonna have a lot of other moving parts but why would it you know with with manufacturing that could be under a thousand dollars yeah I agree and I think you know look no further than the fascinating example of Flippy the burger flipping robot which is of course just an automated arm Flippy actually has a remember that yeah Flippy is the one where um it when it first came out I don't know the actual dates but if you look them up it's actually even more fascinating when you look at the timeline of this but they've they said hey uh you know franchise burger joints reduce your cost by X firing insane amounts by having them back there just flipping burgers right and none of them wanted it because it was all just such bad PR and it was a bad look and everything and then fascinatingly like in some um jurisdictions and municipalities uh minimum wage like uh rates went from four to 15 I don't know yeah yeah to like 25 20 and 25 right and so and it hits those organizations the hardest and so then um in California uh they did like a 20

something dollar one and Flippy the burger flipping robot just like took over all over all the franchises in California so they finally were able to say well look this is the only way I can still make a burger for 4.99 right like I have to figure this out and and it just it flips burgers and so when it economically makes sense then I think people the factors have to make it make sense and politically make sense and just you know not just be trying to eliminate humans I guess like for all those factors have to kind of come together for that for the tech to actually take place so I think it's it's fascinating to see that it goes beyond just the the development of the hardware but the environment has to be good for it too well I mean it's eliminating a job of flipping burgers but I'm not sure who really wants that job like even if it's in food I think there's more important things like maybe there'll be less staff in the restaurant but the staff that's there is going to be doing things that are more core to the to the function of the restaurant into the restaurant they're going to be able to spend more time taking care of guests you know yeah that I mean I think there's certainly some some truth to that and then and then also it does keep burger prices down and you know unfortunately the communities that are most impacted by uh wages

are also the ones that will turn to places like that for to feed their families and so you know there's 12.99 right um so flippy's here to save the day flippy's feeding families what about if there was a if there was a like a gas station that used robots to fill your car with gas assuming you still have a gas cousin car what do you do you drive out you drive a Suburban right with a with a I drive a I drive a gas guzzling car I'm holding on tight to my gas guzzling card um but I I don't know you know I'm I'm fond of all the electric car Technologies I test drove a Rivian this year before going back to a gas guzzling you know baby dinosaur killer um but I I think on the horizon I can't imagine even though I'm stubborn and nostalgic with cars that I'm gonna not be in it like I totally get that over here they charge a significant tax by the size of the motor so if the engine is like 1.1 liters or less it's almost free 1.2 1.3 1.4 it kind of goes up and up and up our cars are 1.3 liter diesel because like 54 miles to the gallon um well and if you get if you get like a VA Range Rover it's like three or four thousand a year but

if you get a really small car it's like 300 a year for registration well if that was working like we said model development and training was um it should go the other direction it should like force the the model creators to be able to create the same torque with less um leaders of uh a smaller leader engine isn't there like a so isn't there like a really small Formula One car I don't even know like a less than two liters with crazy turbocharger and 600 horsepower or something like that I mean you can get the power out of it so yeah it runs on um you know pure grain alcohol or something like that uh is part of the challenge but uh anyway but I think all all this goes into our physical AI prediction I like what you're talking about there maybe we start to see this happen on a consumer level this by the way is an area where I think Apple could come to smart home with tiny little devices and maybe a better make Hardware you know yeah but that's it that's the thing like integrated they make integrated Hardware is what they make integrated Hardware with their great silicon and someone else's AI you know actually creating you know improve HomeKit and uh you know kind of boil in some of the the promise of HomeKit

with actual like really good AI grade uh Hardware I'm going to make I'm going to make an app for Flippy like an Apple silicon-based burger Flipper I'm sure it can be done with iot and some level of AI at the present but I just think it's now I'm getting hungry now I need an omelette flipping uh Flippy so uh here we are are we is it is it what are yours what are your predictions no it's not quite it's just oh no those are my predictions I I think there's another level of slms uh and I believe that actually um it's going to get really awkward because this whole MLM terminology and slms are going to be so profound that we're going to not be able to use either term and we're just going to back to lms uh and then voice was the other part that I was predicting there too and then Claris I'm saying no more AI stuff actually that's a big swing that they're actually going to just try to learn how to actually message and listen to their customers more about what outcomes they want and then have the stuff they already have to those outcomes I'm going to say you're wrong about that one I think we're actually we're going to see a bunch of stuff and I have by the way I'm not on the inside track for beta for any of this stuff I just know kind of what I want there's no currently no beta right now for anything that hasn't been released so this is

a good time to make predictions I think we're going to see some AI that helps for creating the database like some AI tools for the developer to write scripts write calculation codes check type stuff like that it's a pretty big issue for script writing because scripts are under the scene like a really specific XML and you can't it's not text-based like all the other programming languages I'm factoring those I'm factoring in those limitations yeah but I know they're working on it and since it is under the hood XML and XML is something that you know can be dealt with in a really effective way I think they're going to be able to put some I think they're going to be able to construct and uh uh and some sort of an AI tool that would allow it to work where you can say take this section of my script code and make rules and poof the new strips the steps will appear or check this script and find out what logic are or you know something like that or even write a script that does this and it will look at and understand all of the other scripts in the solution and I think I think those things are coming I don't know if they're going to be in 2026. I also think we're going to see um some other tools like uh rag for example is hard right now it doesn't you can't really do it like directly built into the application

easier by their rag tool either yeah so but I mean it's hard but there's no reason that they can't actually store some of that stuff directly in the database like so so you don't have to go out to another service provider where that can be a different way to do it I don't think they've um I've been thinking a lot um gosh I'm trying to think of some of the other things that for prediction I think those are both areas they should have started in yeah actually I think focusing on dev assist would have been the move yeah yeah okay MCP oh okay maybe we talked about this before well we had an entire episode if people haven't seen it yet we had Louis de la par on here talking giving us all the details big listenership I'm not sure how much I got into it maybe I did a little bit um about some of my questions MCP doesn't really have much impact for any FileMaker developer who just works in FileMaker like if they just write code MCP is actually huge for all the other stuff coming in well only actually it's even narrower than that yeah it's profoundly impactful only for those that are communicating with language models outside of their FileMaker application and that need to reach into FileMaker well it's it could go both yeah no

I'm sorry right the thing that's currently out there right now is the ability to reach into FileMaker that's correct but you know you you could also leverage MCP if you were connecting to a in-market in beta by the way we should note that what we showed and Louis came on to talk about is simply beta it's not like the the last of the MCP well shipping changes right yeah it's not a big push you have to go looking on their website for it yeah well that I think is intentional someone tapped me on the shoulder and said hey by the way that's not actually it's beta is what that is and I'm like well maybe fix that messaging no it doesn't actually yeah so I predict there will be some I okay there's my predictions for January I think they're going to be tweaking their messaging on MCP uh for filing anyway my point is that to make it useful for a regular FileMaker developer they could they could open up some kind of a flavor of MCP to a database that then you're scripting or your own AI queries can use that tool in that standard so so in other words you could have because it actually understands the context of your database right well so I'll I'll pull I'll zoom in a little bit on that exact example because you're totally onto something

the first thing is MCP outgoing MCP meaning applications an MCP client connecting to MCP servers right isn't just to connect to data it's not even just to connect to data and or actions it can be anything as a matter of fact you gave a great example of an MCP based uh context management or memory system that you can then integrate in using the standard protocol and provide memory or statefulness to your own applications like let's say FileMaker um and and like a million other things you can do like if if you if you just go search some of these MCP directories you'll be blown away by some stuff that you can actually connect to and bring into your applications it's it's I would say it's a level above API endpoints because of the the context and awareness yeah different yeah it's different it's a little bit harder maybe but it's vastly more powerful definitely yeah um and we're I think still that directly affects the FileMaker developer in terms of what they want to reach out to for all these other thousands of things you can do yep but I still think that there has to be a path where a FileMaker developer can need to can solve some problem in their own scripting and their own work on their own database to make an MCP call to another FileMaker database

in the organization basically kind of a look back in I think there's a lot of power there yeah that they haven't I haven't thought about it I thought about it well quite a bit but I think there's still more well that would really affect so it doesn't need to be there right it doesn't necessarily have to do that that's the thing like MCP without the model the context isn't as meaningful then it's just a connection using a protocol so there's definitely value in the C and the connection part of it but the C being context and M being model yeah you need both well the mp-server the mcp server though that you connect to FileMaker and you publish a set of tools and data and stuff like that yeah you technically go through your FileMaker databases and mcp server using this toolkit and then it can actually then query and understand not only the data but also structure about the data like yeah well you can ask questions about what you know tables and fields and well and but that's you have the opportunity to also provide your own description of what that context is and how it is meaningful and to help make those um interactions more meaningful to your users so

that that I think is and so like when we talked to Louie you know they mentioned that like the standard MCP protocol has little description field and how that's fine it's certainly better than just connecting to an API endpoint without a description field but if you really want to you can load an entire like template or like a schema description of that data source as part of the MCP connection into that context and now you get like super relevant yeah very hyper focused connections to data and that could certainly benefit and there's not a FileMaker script step that says tell me about this database yeah interesting and so I think that's actually one of the big things you can do okay and there's not a there's not a FileMaker script step that says call this other tool in the database to do this action that's an AI type action and of course you do need an LLM for that but I think that part it's like it's reaching out and back in and taking advantage of this really amazing feature to do things that are are that are now really difficult to do in FileMaker for just regular old FileMaker scripting I don't know if that's if this is there there but anyway it's something I fantasize about okay well we won't add that to the predictions that's just the wish list

how's that there's different things there well we certainly did it I think it's probably almost midnight now in Greece it's only 20 after seven man oh okay um my math is off a little but uh another uh we did it another year of talking about AI um it's always fun uh for me to do this with you I I look forward to this being sort of the one thing I was doing this week that was productive um but yeah we'll check back uh in a year on these predictions uh we'll certainly kick off the new year we've got some things we're thinking about doing nothing's set in stone we'll have two or three podcasts between now and the year end 2026. yeah at least two of them uh you know whenever something relevant comes out we'll have a podcast so maybe two podcasts but uh um I look forward to it maybe we'll be able to call back to some of these predictions as they unfold uh throughout the course of the year so as always it's uh been my pleasure to sit and talk uh one of my with one of my favorite people about my favorite topic um and I look forward to many more of these moving forward all right yeah hi Alexi bye Alexi Happy New Year Alexi all right cheers happy new year to everyone bye