navarre.aiClarisTalk AI · the FileMaker AI podcast
What is Chain of Thought?
Show Notes for Episode 17: Links: OpenAi o1: https://openai.com/o1/ Reflection: https://arxiv.org/pdf/2303.11366 Chain of Thought Prompting Guide: https://www.promptingguide.ai/techniques/cot Devin: https://github.com/Significant-Gravitas/AutoGPTdevin Cursor: https://www.cursor.com OpenAI01 Ten tests: https://youtu.be/iINPTCQWy9E AutoGPT: https://github.com/Significant-Gravitas/AutoGPT BabyAGI: ht
Apple PodcastsSpotifyYouTubeAll episodes
Transcript
Transcribed by machine (whisper, large model), 14,312 words. Names and product terms are sometimes misheard; the audio is the record.
Claris Talk AI. What's up, man? How are you, man? Long time. I'm doing well, and apparently I read a grammar book, and the answer to that can either be good or well, and they're both correct. Oh, well, good. Yes. You see what I did there? Let's do it. So we've got some exciting stuff that we're going to be talking about. Obviously, the new chapter T, things that have come out, which actually is wide-ranging, not just like a new one thing. Before that, though, little plugs I've got. I'm going to be doing a class at Rome FileMaker Week, October 9th, FileMaker AI integration class. So that's the class you can buy on their website. And in Malmo, Sweden, I'm going to be there in person and teaching a class, also semantic search and AI integration, on November 17th.
And all my classes are on navar.training. The other thing that I'm kind of excited about, which is pretty nerdy, is I started a YouTube video series and blog called the FileMaker Opinion Page. Oh, interesting. You can search that on YouTube or see it on my website. And the most recent video I made is making a strong opinion case on why I think people should stop using the button tool and switch over to the single-segment button bar tool 100%. So anyway. Okay. Yeah, no, I don't know a bigger champion for calculated buttons. Using the calculator. Calculated feature on buttons than you. So a compelling dialogue. Well, it sounds like we're doing a lot of the same stuff, just not at the same time.
I have a Claris webinar, live stream, whatever, community live coming up on October 24th, which is going to be an intro. It's online. It's an intro to FileMaker and AI with special guest Ronnie Rios, friend of the podcast. And we are going to be doing literally like, where do I start? Like, I keep hearing all this stuff about semantic search and all these new features. Just where do I begin? How do I do this? And then just step through it. It's like a 90-minute presentation. It's free. And, you know, we get as far as you can reasonably in 90 minutes and still have an interactive dialogue. And Ronnie will be part of it as well, too. So looking forward to that on the 24th.
Yeah. Yeah, yeah, continue. Well, and then, you know, they announced Claris Engage, right? And so I'll be doing a training day at the U.S. That's Claris Engage in Austin in March of 2025. It seems like it's very far away, but registration is now open. And they anticipate that the FileMaker and AI training day might go fast. So if you're interested in this and you're going to be in Austin, you should probably sign up. I'm also going to have Ronnie Rios be a special guest for that. And that's like a day-long version where we do all sorts of stuff. We, like, get into prompting techniques and we'll start off with vocabulary and then some prompting techniques.
And then we dive into the FileMaker stuff. Then I show some cool, inspiring, real-world use cases of AI. And then Ronnie actually comes in with some stuff he's not going to be showing during the actual session, like the what's next. So really, the way to see the most AI stuff is to probably go to training day this year. And it's interesting because training day is at the end, not at the beginning. So I think that has something to do with the format. Yeah, that could make sense. Yeah. The classes I'm going to be teaching are probably going to be covering less AI stuff. But they're going to be totally online. Totally hands-on. I'm going to share my screen as little as possible and have the students take turns actually integrating AI into their systems.
So I'll give a, during the class, I'll have an API key that people can use if they don't have an account, blah, blah, blah. Okay. Let's talk about our subjects today. We have 37 things on the agenda. As always. Maybe just three. Yep. We could chain them all together, though. Chain. See what I did there? So I want to talk about this concept of chain of thought. And this absolutely has, this is going to be the through line of like all of our topics here today. But when I say chain of thought to you, Matt, what does that mean to you with your background and thinking and philosophy? My philosophy classes from 30 years ago? Yeah. I certainly didn't come up then. But I think from some of the other science books and stuff that I like to nerd out on, it's actually not a phrase that's really, that I have used or thought of much in life.
But chain of thought. Chain. How about train of thought or tree of thought? Like these are all the same concept, really. Yeah. It's basically sort of taking different types of things and making sure that when you stack them, they're correctly linked to the thing before, which is very much Aristotelian logic. Yep. Or Claristotelian logic. Oh, hey-o. This is going to get way too punny if we keep doing this. This is a bad start. Yeah. So basically, yeah, like taking the input of one thing. Or that. The output of one thing and to the input of another. Yeah. Double checking your results. Like making sure. Yeah. That's what chain of thought means. Yeah. I think people probably are probably familiar with the concept when it comes to train of thought.
I think that might be one of those, like, you know how in the U.S. we mess stuff up. Well, it's actually what you just described. The train of thought means that I have this consistent through line. I don't have these, like, weird non sequiturs or right edges in my discussion points. Train of thought to me means just one thought. Like. But that's the whole point. As long as it remains one thought. So I think what happened was somehow we started referring to train of thought and then we actually said chain of thought. But the reality of it is chain of thought actually means something in A.I. And we're going to talk about that today. Tell me. So chain of thought is really.
It's actually kind of a simple idea. It's where instead of just interacting with one of these large language models with just a prompt and a completion. We refer to that as like zero shotting, you know, in a somewhat pejorative manner on the podcast here. What if you could force. The model to go through a reasoning step or a series of reasoning steps where it comes up with an explanation and then arrives in an answer. And it's almost like intermediate steps before arriving at the final answer. So you're kind of intercepting the thought process, so to speak, in between. And so by the concept that this is rooted in is by forcing models to go through reasoning, sort of a reasoning pathway.
You're ensuring that they do not overfit on extraneous patterns. That's where hallucinations come from, is it's reaching out to try to fill in gaps. With these weird extraneous things. Right. And so this way you can answer new questions that it's not seen before. You can train it on how to think and address other questions that might be totally new and novel to it. Maybe. And of course, not even in its training set. And since we know models are not databases, this approach should be quite important. So in summary, at a really high level, because of the way they're trained, the language models are trained and they're not databases. Right. Remember that. Apparently can't say that enough.
I just. One of my family. I was just texting me some some like thing where there was a massive misunderstanding about data, privacy and language models. But this way they seem to perform. And actually, in my own experience, they do. And this is widely accepted in the industry. They seem to perform way better at NLP tasks if you force them to do this kind of chain of thought concept up front. So really, you're thinking you're having to think step by step. And there's a bit of a history here. But first, I want to see if this sounds familiar to you before I jump into some of these. And this is all part of the new ChachiBT 1.0 model that came out. Yeah, actually, what we're doing in a somewhat coy manner here is you and I are going to like move into the whole what is ChachiBT's 0.1 model.
Not 1.0, yeah. Yeah. So 0.1 has been released since we did our pod. It's probably only a couple of weeks old. And we're going to explain what's actually happening in it because it's not just like GPT-5. It's like a whole nother thing. And really to understand it. How we arrived at this is probably we kind of have to look at how we got here in the first place. So what I'm sharing for the folks that are watching this on YouTube is a paper that came out. Let's see. In October 2023. And it's called Reflection with an X. Language agents with verbal reinforcement learning. And basically what this was, was when GPT-4 came out, we might have even talked about it. Like one of the sort of hidden features, it's sort of leveled up to this.
The ability to reflect upon itself. So this paper goes into very, you know, very specific detail. We'll have a link in the show notes here, of course. But really from a really small bite-sized version of this, one of my biggest pro tips when it comes to interacting with chatbots like ChatGPT or Cloud or something is always ask it after you've given it a task. Like simply say, can you verify the results? Or did you do a good job at this task? And a lot of times they'll just like give you a checklist or if there's a problem with it, it'll go back and figure out what that problem is. That actually wasn't possible in GPT-3 and it really wasn't consistent or a thing in GPT-3.5 when 4, and then of course 4.0, came out.
Reflection was actually a deliberate feature or interaction with the level of intelligence these models have achieved. You've probably done this before yourself, right? And it can also go out and like fact check the web, right? So I actually, I had an argument with ChatGPT earlier today about something that I was sure was true. It was about... Scooter models. I just bought a Piaggio scooter. And I was sure that one particular model only existed in a certain engine size. And I was wrong. It's like, no, no, you're wrong. It's like, this is the fact. I thought it was great. Yeah, no. And I love being wrong. Just ask my wife. No, which is not true. I actually played for her our last podcast where I admitted that I was wrong about something because ChatGPT corrected me.
So that has really helped my marriage. But I think this is an important technique. So it's not about not trusting. And the answer, it's just about like, it's kind of like just closing the loop, I guess, is a good way to think about this. Like, hey, I asked you to do something. Did you do that thing? And most of the time, it just tells me, yeah, here's what you told me to do. And this is proof that I did it and so on and so on. Or you can just be like, check to see if you did this correctly. And if you did it correctly, just say yes. Right. And it'll come back and say that. Now, you can actually... So really what that reflection concept allowed is some new techniques came out.
So. Specifically, there were some techniques in prompting. And one of the ones that, and frankly, I use this all the time, and this is built into our intuition environment as well, too. I strongly recommend people look into this, is this idea of chain of thought prompting. Now, that's not exactly what we're talking about with O1, but I think this is important for people to understand to really get what's happening with O1. And on screen for the YouTube folks that can see this, but for the folks that can't, Matt, you want to describe what's happening? We got the standard prompt. On one side and the chain of thought on the other. So these are great. These are like math questions, like one of them.
Roger has five tennis balls. He buys two more cans of tennis balls. Each can has three. How many tennis balls does he have? So these are kind of questions that are sort of word patterns that a human frequently will get wrong because you look at something that you look at. It's like, you know, five plus two is seven. They're almost like a sleight of hand or something. Like it's trying to make you think it's one answer, but then it's really something else. Right. Yeah. And so there's really two different ways you can do this. You can just do the straight zero shot, which is what we see on the left standard prompting. Just simply saying, here's the question.
And then so in this case, what they're doing is they're creating like an old school version of providing examples of the output that they're looking for. So you see here, you know, it gives a question, then it gives an answer. Then it goes with the second question. The cafeteria has 23 apples. They use 20 to make lunch. The question doesn't matter for the listeners, but then it gives an answer. But the point is the answer. The answer is wrong. Right. So on the right hand side. And humans also get it wrong frequently, but yes. And honestly, that is the, I, so actually, I guess the question does matter. The cafeteria had 23 apples. If they use 20 to make lunch and bought six more, how many apples do they have?
And I came up with 27 when I first read this, and that is the wrong answer. So, so this has to. What's weird and cool is that you get it wrong and Chachabiki gets it wrong for the same reason. Exactly. Because you're looking for a pattern. Right. But what we have to, but what is a better way for humans to do this and for us to train in our prompt, how to do this is to provide chain of thought. So the alternative approach to asking this question is you see on the right hand side, for those that are on YouTube here, we ask the tennis ball question. And instead of saying just the answer is 11, we, so I'm sorry, the question of the tennis ball, the answer was 11.
But instead of just giving the answer as an example of how we want it to respond to us, we give the answer of. Roger started with five balls, two cans of three tennis balls. Each is six tennis balls. So five plus six equals 11. So therefore the answer is 11. So we're literally putting those steps in there and telling it how your work, go your work. That's literally what chain of thought is all about. And so, um, so then when we ask the same question about the 23 apples, now it had, now it's been prompted. Hey, this is how you're supposed to approach this process. Right. So, um, so this whole chain of thought prompting became very popular and, and. And it should, and for those that are hearing this concept for the first time, you should absolutely go find everything you can about chain of thought prompting.
As a matter of fact, the link to the paper that we just shared a second ago is in the show notes. And so then what ended up happening, Matt was the, um, there, these, these new kind of approaches came about. Did we ever, we might have talked in the past about, um, baby AGI and, um, auto GPT to any of those. Sounds familiar from our previous conversations. It doesn't ring a bell. Well, the, the links are in, um, oh, well, the good links are going to be in, um, uh, in the show notes and, uh, and, uh, when you log into your GitHub accounts, you'll be able to see both the baby AGI and, um, auto GPT. What those are was that some developers out there said, Hey, if language models can reflect upon themselves, why don't we use that to our.
Advantage and why don't we ask the language model to come up with the plan and then it's essentially teaching itself a chain of thought technique. And so if we have the language models come up with the chain of thought, instead of the humans doing it in the chain of thought prompting, and then maybe if we give it some memory, like a database or something where it can store in an intermediate, even variables or something like that within a web application, then really kind of amazing things can happen. So if you research, um, uh, baby AGI. Or, uh, auto GPT, what you'll end up finding out is that these became revolutionary within the, uh, the AI development space, uh, specifically, um, I'll show you this little example here.
What we're looking at on screen are the different steps that are taken, uh, when a prompting is actually happening. We're essentially creating a loop and, um, what, what happens here is that you're saying, okay, I want you to, what was the thing that you were actually doing with the scooters? Let's use that as an example. You were saying what different motor size. I don't know if that's really a good example. Uh, but there's, there's a medley model and a Liberty model, and they're available like in one 25 and one 50 CC motors. And I was sure that there were, there were the, the name Liberty and, uh, whatever are only available in one size and the other, that the name of the thing was actually what told you the CCS.
Oh, I see. Interesting. So, so it had to actually do some web searching and cross-reference. That's the part. So those are steps, right? So like go use your tools to perform some steps. So in this case, use your web searching tool, go search for this, that, and the other thing. And then maybe you have to do some math tools if you're actually comparing dimensions or, you know, whatever, maybe not in the exact example that you gave, but the idea is come up with the steps, have the language model, figure out what the steps are, and then deploy those individual steps using some memory. Now, what actually started to emerge at that point was the concept of agents, like apps.
And that's what we're going to talk about in a little bit, but we're going to talk about that in a little bit. So what we're going to talk about in a little bit, but we're going to talk about that in a little bit. maybe not in the exact example that you gave, but the idea is come up with the steps, have the language model figure out what the steps are, and then deploy those individual steps using some memory. Now, what actually started to emerge at that point was the concept of agents, like actual agents. There's a little bit of confusion with agents out there. I think Copilot just announced agents. That's not what I'm talking about. Those are just GPTs that they put a super confusing name on.
Agents are actually tasks that language models come up with the steps of the task, and then they use memory, like databases, and they use tools like searching and like Python for mathematics and calculations, and they come up with the results, and then they use reflection to determine whether or not they've completed the task. That's what an agent actually is. And so in this baby AGI, in this auto GPT, they actually ended up coming up, you know, essentially coming up with that concept, and then people just went bonkers. Well, to take a step back, AGI is general intelligence. That's like, you know, human type stuff, not just, you know, language models. Is that what they mean by baby AGI, like the level of a baby or something?
Yeah, I think they were getting a little cheeky with that at the time, you know, because it was, the discovery was so profound and significant that they were saying it was AGI, but it's really not. And nor is OpenAI's O1 model, right? Which we're building up to here. But yes, that's what they're talking about with AGI. So if that helps you guys remember the repo to go find it, then it's the baby version, right? Yeah. But, and you can see my screen here. Yeah. Up on screen, you actually see this is like a tool that's using this process. You see how it says embarking on a new goal, plan the perfect date. And you see where it says thinking, Matt? And then it said a task, plan a week-long vacation in Paris, book a flight.
Let me actually go back. So, so for those that are- Perfect date, a week in Paris? Damn. You live. You live well. It's actually, yeah, it's like an episode of The Bachelor or something. I saw that movie. So, so, so the, the key of an, of an agent is that you give it a task. And so the task here is plan the perfect date. Matt and I are planning the perfect date for each other. Now notice the part here that says thinking. That is actually where the eight, an LLM is now tasked with coming up with what are the steps to figure out the perfect date in this case, right? And so it goes out and it does a search and it comes back and you see there's a little progress bar.
And then when it comes back, it comes up with the following tasks. So Matt, these tasks say plan a week-long vacation of Paris. The next task is book a flight and hotel reservation. And the final one is reserve a restaurant for a romantic dinner. Sounds great, right? So now then what these agents do is go deploy another string of agents, a whole nother plan with tools and such, and deploys those against the whole task. They then come up with their results and they put all those results together and the managing agent, sort of like the project manager, agent, so to speak, compiles those all together and comes up with this description here or actions. Like if some of the tools you're providing are APIs.
So let me take a breath and say, does it make sense how like reflection in this concept of chain of thought prompt prompting led us to agents at least in the first place? It does to me. Yeah. And I think, I think we're even maybe oversimplifying a little bit because it's a, and you showed it a little bit in that one because of the type of tasks that it does, but what we actually get to, in one Oh is PhD level math. So four and three, five sucked at math, but now it's full on PhD level math. It's going to help discover new things in math that humans can't do. Right. All that's changed is the approach. There's not more intelligence in there. All it did was in that, that example that we showed with the apples thing, where you just kind of ask it to answer a question.
And it's a terrible math instead of encouraging it to show its work and plan. So that's literally all the change. That's crazy. So it's, it's very crazy. But, but I think the takeaway for people is that, Hey, your prompts can be better too. Cause remember when people complain about, you know, getting bad stuff out of language models, it's really their, their bad question or lack of the, you know, the approach. One of the things you could do is like, help me write a prompt to get this answer, you know, and they'll actually help you with that. No, you absolutely, no. I mean, that's literally what agents are doing. It's sending, it's creating a prompt to a language model, to come up with the right approach.
Right. So if you're a web, so here, so that's not what we wanted, but so other products have come out of this whole thing. And these products are profound, man. So we briefly talked about one. Let me just share with you. You'll probably recognize this. I'm not going to play the audio, but I just want you to see what this was. This was something that was called Devin. Do you remember the whole Devin thing? I guess not to be confused with anything in the file maker space, but Devin, this is the one where it went on to Upwork and it read the description, and then it went and created the app that the description. Remember this whole thing? I do remember that. These are agents.
And if you've ever used Devin, you can actually see these agents at work and you can actually see the steps and it kind of interacts with you a little bit and it's building code on the outside. So this whole concept is really important. I will say that since we've done a pod last time and people should really check this out. Is that we have Yeah, there we go. So we there's a new thing also that's called Cursor. Have you heard of Cursor? This is just another agent. It's sort of like a evolution of the Devin thing. It's a it's specifically for code and you can go download it for free, but it's this kind of like next level copilot thing that everybody's going bonkers about.
And so and it's got a bunch of seed round funding and it's, it's very, very popular. But from a code standpoint, this allows you to actually plan out your your coding process. So this evolution is that what that is? Yeah. Sorry, let me figure out. It's actually Cursor dot com. Oh, Cursor then. Yeah. And so, you know, you talk about math and why this is, you know why this works in the first place, I'm going to show you a little peek behind the curtain into our, our intuition environment. So that's something that we use as sort of a starting point, like a platform that has all sorts of these layers to it here at iSolutions. And this one little part here that I'm showing you, Matt, is one of the things that we do is we have the environment load all of the data.
We do this little twinning process where you might have a FileMaker database, for example, but we have a twin of your FileMaker database that lives over in this other like, Google based like tool or environment that has access to all these tools. It's a security thing. And then also mostly a performance thing, because with all due respect to FileMaker, it's not super performant on these high level tasks. But what we have to do is we have to interpret what the user's request is like, go create a report and give me some insights based on blah, blah, blah, blah, blah. And it has to do something with the data. And so if you notice over here, Matt, for those that are watching on YouTube, you can see there's just some steps, load the data set, convert, the date column to this, verify that the data set is not empty, do this, that.
We did not come up with those tasks. We're using this agent approach and reflection and memory and the LLMs to help us write the right Python code on the fly, giving it access to all these different libraries. But it has to first start with what are all the things that need to be accomplished. And when it figures all that stuff out, it goes and then finds the right tools and the right data and follows all our instruction to put that all together. And that is why, this approach is better at math too, because it's saying what would be the most optimal way to do this? What tools do I need and what order do I need to do this? What dependencies are there? And then deploy those things, right?
Yeah. So that is the show your work on steroids basically, right? So I'm thinking, and it's not much better for the generative text thing. It doesn't write any emails any better, but as complex tasks, writing code, checking code, math problems and actually one of the things we do ask it to do, probably report generation and getting facts from something is what it's much better at. And when it first shipped, it didn't have the API and now it does. So you can actually make an API call with 01. Yeah. So this all leads up to OpenAI's 01. So for those of you that are not familiar with this, but you have ChatGPT Plus, you can go into ChatGPT Plus and check it out. You just have to go to ChatGPT Plus, have to pick the right model.
It would be the 01 preview or the 01 mini preview model and just simply ask it to do stuff. And one of the things that you're going to notice is that it comes up and it says, I'm thinking, and then what it does is it comes up with a series of different tasks that it needs to do to figure this whole thing out. And it tells you how long it took. So that's the first thing is there being somewhat apologetic that this takes a while, but you and I have talked about system one versus system two thinking a bunch of times before. And the main difference between that is take your time and just go through it. Go through all the steps that you need to. That's the system two approach rather than system one is use heuristics and your immediate memory and just come up with what you think the answer is.
All that we have known up until GPT and up until OpenAI's 01 model has been system one. And you and I have talked about it on previous podcast episodes that the future is system two thinking with tools and memory and models that can reason, and all these steps. So this is the first way that we're seeing at the tool level. - It must just take a huge amount of compute power. - Well, yeah, that's OpenAI's problem, but not only does it take a huge amount of compute power, but it also takes a huge amount of tokens. And I think a lot of people are talking about this as sort of a downside to using this if you're in ChatGPT. First of all, you're fully rate limited in the ChatGPT+ tool because of the compute power that you're talking about.
But also all those little steps are burning tokens and you're essentially creating all these little offshoot conversations that are going out and doing things and you're paying the token cost. So I think unless you really need this functionality and you're not already doing it and I'll explain what that means in a second, then you probably aren't going to need to pay the extra tokens. So I did a video that is going to be in the show notes that it was, I used to do these, like when all the models kept coming out all the time and I frankly had more time on my hands, I do these 10 tests. So like every time a, I don't know if you can see this one, I'm not going to play it or anything, but I just want to get to this last little part where I'm basically doing a side-by-side where I did these logic puzzles.
I had these 10 logic puzzles and I asked on the left-hand side of the screen, ChatGPT 4.0 to figure out the logic puzzle. And it was mostly accurate in what it came up with for answers. But then I also asked on the right-hand side, the 0.1 preview. And the difference in the 0.1 preview is that it showed its work and it actually shows you all the different steps that it takes. And so I think it's just really fascinating if you want to write your own prompting so that you don't have to use this tool, you can actually go in and see what the, you know, what it came up with, like what its actual steps were. And then maybe that's a good way for you to actually learn how you might be able to create your own prompting tools, actually.
Yeah. So like just watching what it does is education, right? Totally. That's, that's what my, that's what my thought process is. So, so what a lot of people are talking about is that OpenAI 01 is introducing reasoning. And what I'm trying to explain here today is that its technique for reasoning is really just this chain of thought, concept. And so the takeaway for all this is, Hey, it's cool and you should check it out, but use it sparingly because you are rate limited. And if you're using the API, for example, it's going to cost you a ton more than you think. Yeah. Four times as much or something. Well, I mean not only is the per token cost higher, but the amount of tokens to accomplish the same task will be much higher.
So you really got to use the right weapon for the right job. But for some tasks, this is the only way you can do it. Now, when every every time a model comes out, we evaluate the model in our intuition environment. And like, for example, when GPT 4.0 came out within 45 minutes, we pushed it out and all our clients had access to that model. Right. We we did the same thing when 01 came out and we did not push it out to our intuition environment. And the reason for that is because what I just showed you earlier, we are already doing the chain of thought. We are already doing the offshoots. We already have agents in the background that are doing. This we already have all the tools and the memory built into it.
So it would be redundant. We would literally be saying we've already done this work. So now you go do the same work basically. Right. So it's really just an example of it's awesome. It's amazing. But maybe you don't even need it. Like if you're already doing chain of thought, you're already using memory, you're using reflection and tools, you might not even need it. But if you don't, and all those things and everything I just talked about for the last 15 minutes just sounds like blah, blah, blah, blah, blah. Then good news. You've got a, a model that introduces reasoning. And so this is why you're what you're identifying is how they're pitching this as you know, for stem like science and engineering and math.
Why this is such a big deal because the reality of it is success at math is showing your work, not memorizing formulas and what their answer would be right. So this is really just giving it a new thinking process. And that's kind of what we're showing here on screen. And some call that reasoning. So your thoughts on whether you really think this is reasoning as the philosopher, of the pod. I mean, this is the same process that a human would use for reasoning. And that's what it's a little bit like what the graph that you're showing here where the input kind of keeps getting bigger at every channel, which is something that I, you know, I don't know if that's actually the way that it works in the human brain, but maybe it does.
Maybe just kind of on a different level, right? Like, like you learn, you learn math in grade school. So your input is that level of math and then all of that. So you learn high school math, and then now your input is bigger and then you go to college math and now your inputs even bigger. That's kind of like what you're showing if, if that's a fair comparison. No, it absolutely is. And I think Matt, what we're going to do right now is just a quick little one minute demo. I don't know if people are aware of the fact, and I want to encourage people to do this, including you, my friend, you can actually go do this yourself. So if you have a open eye, open AI subscription, you can go in, go in and use this tool, which is the get rid of that.
It's the playground. And the playground has a, an option where you can actually compare. So you'll notice over here that I've, I got GPT 4.0 on the left side and I got a 0.1 preview. And so what I'm going to do is I'm going to go in and I'm going to ask it a logic puzzle. Okay. So I'm going to put it into the system instructions over here. This is just the deities on the mountain top. You meet three deities. One, always tells the truth. One always lies. And one only tells the truth or lies randomly. We can call them truth, false and random. They understand English, but answer only in their own language with either yuppers for yes or, or nope for no, but you don't know which is, which you can ask three questions to any of these deities.
And you can ask the same deity more than one question. And then they will answer with yuppers or nope. What three questions will you ask to figure that out? So, um, so if we go ahead and, and then that, if we just put that in as our, um, oh, actually, so do you know the answer? It's kind of like, it's kind of like if I asked the other two deities, you know, whether the earth is around, what would they say? You know? Yeah. Well, and that's kind of the idea, right? So, so I now notice on the left-hand side, it already gave me an answer. What's happening on the right-hand side for those that are just listening, Matt, it's still thinking, right? So O one preview is still thinking while GPT four O gave the answer.
So it goes, it goes through and it says, well, here's the first question you ask. Uh, if I were to ask, you know, are you the truth? Would you say? Yep. Or so on and so on. So you see, it came back with this system. One approach. Yeah. Um, on the left-hand side gave an explanation. So there's the three questions right there. Um, and then on the right-hand side, it's still thinking, right? So that's the big thing. So you can see how much more time this takes. And by the way, I think it's, you see, the latency is like 11,000 milliseconds, um, at 684 tokens. So I think that's something that's useful for us, uh, to look at as well, too. So, first of all, think about if you're integrating, okay, now it, it stopped.
So the difference was 11,000 milliseconds versus 50,000 milliseconds. And 684 tokens to almost 5,000 tokens. Okay. So not only is the, even if the per token prices were exactly the same, um, this would be, uh, meaningful. Um, so in the first questions, uh, Okay. So, so, um, so, so, so, so, To any, oh, let's see. So God, are you random? And then the interpretation. So it actually, I mean, the answer is pretty significant too. So really what I wanted to show you more so than the answers was how much time one took and what the different token costs was and how you can go play with these yourself. Well, I guess I want to know if they both got it right or if only one got it right.
Yeah, there is. Well, I can tell you that in my video that'll be in the show notes here that I went through 10 of these and I, and I, and I, it's like a 20, 30 minute long video if you want to check it out. And I talk about the answers and we compare each one of the answers too, if you want to, if you guys want to check that out. So I, I would say have fun with it in, in the YouTube video. I even include the 10 questions with the correct logic answer. So you can actually play with them in the playground yourself if you'd like to. So, and then see if they even come up with the same thing. So, but the main point of the takeaway here is the approach that's being taken and when you're actually.
Doing this inside ChachiBT and you pick one of the O1 models, it does the thinking, but then when it's done, you can hit the little drop down button and expand it out and see all the different steps that it came up with. And that's what we were talking about earlier is probably a great tool to use to kind of figure out how all this stuff works and do your own chain of thought. Yeah, yeah, yeah. Right. I guess one thing I would throw out there, if you're doing, if you have something in FileMaker that uses this approach and the user doesn't really think about which model it's using, they're going to think it's broken. If it takes 50 seconds to do something.
So you have to sort of think about that too. Sure. Well, let's, let's, let's, let's explore that concept there. So first of all, you said user and the user interface expectation is not going to account for legacy. They're not going to have the patience for legacy or latency. I'm sorry. Like you're talking about, like no one's going to want to do that. So then really these, this is meant for tasks that just sort of happen on their own schedule or they go and do all these things and present it to the user. When they're done. So already it's a fundamental change in, in the way you're thinking about these things and approaching these. Right. Yeah. So that's, that's the first part.
So think of this more as like, if we were to bring it down to FileMaker level, think of this more as a FileMaker script that a user is running in front of them versus one that you would rather choose to do as PSOS or something instead. Or just a scheduled script. Yeah. A scheduled script. And, and, and I think the, the PSOS one might also introduce the fact that you're giving it more memory. And, you know, the user doesn't actually have to have a session in order for those things to run. And that's where the set, the scheduled one comes into play too. So I think that's probably how you should look at these things, you know, as far as how to implement these into, into your process.
Also agents, this is, it's basically reasoning and these whole agent concepts. Agents are going to take forever. And, you know, you hear a lot of people talk about agents right now. And the kind of things they're saying is like within a couple of years, I mean, there's some tons of these. There's tons of these out right now, but within a couple of years, the expectation, the user expectation is going to be different. It's not going to be, I hit a button and my answer comes back. It's going to be, Hey, you know, that every Monday I need to do this thing. So go do the thing before, you know, I'm going to be logging into this system and just tell me what you found out.
And I don't care if it takes three days. Right. Do this. Right. So in advance. Yeah. So I think that's kind of fascinating, which, you know, is kind of lead, you know, sort of like the conversation. That we had on our last podcast episode where we were talking about UI and UX. And if folks have not seen that episode, I, I, I suggest that you do. It was a great one. We got some really good feedback from a bunch of our friends on that. So if you haven't seen it, please check it out. But man, I have a couple of quick thoughts on, on our discussion from last time around. First of all, you enjoyed that, didn't you? Talking with our friends, Matt and Alexis. Well, I mean, how can you not enjoy talking with, with Alexis and Matt?
Because they're both awesome. They're brilliant. Alexis is on this crazy world tour right now. She's taking pictures of awesome things and wherever she is in the world is pretty cool. I think you'll probably run into her at a couple of the conferences. Well, I know she's going to Rome and I won't be there, but she, I'll, I'll see her in Malmo, hopefully. Oh, and by the way, to clear up any confusion, I guess I'm on the speaker list for Rome. Um, I'm sending my bot in my, in my stead, so I will not actually be there in person. Just a, uh, AI facsimile of me will be there. So, um, uh, pleased to enjoy and make sure to buy him drinks. We'll get to those AI facsimiles a little later, won't we?
Yeah. Um, so first of all, a quick update, um, uh, gen AI first UX so long point and click how generative AI will redefine the user interface. This link is going to be in the show notes. I just wanted people to, and this, I think is a good sort of, you know, follow up to the discussion that we had, uh, when we were chatting. But one of the big things that we talked about in our last presentation, um, was, well, first of all, if anybody, if anybody follows me on, um, on, uh, what is it called? LinkedIn. Yes. That, that one. Um, you've probably seen me, uh, talk about the, um, the, like the, uh, uh, the artifacts from, um, hold on. Let me just share this screen.
So yeah, artifacts. That was on our list to talk about today too. Yeah. So we did talk about artifacts on our last podcast, right. Um, where we were, where we were saying that this is, um, this is sort of a new, uh, uh, paradigm for UI. That's the Claude AI thing, actually, if you remember that, um, let's just see here. And for those that, excuse me, did not, uh, see our last episodes, first of all, do you have Claude AI as well? I'm curious if you've been playing around. I do. I haven't been playing with that much though. Um, and, and now it has like a huge, much, much larger token window too. Uh, yeah, no, I haven't spent as much time with it as I should have Chris.
Okay. No problem for, but just a quick, um, 32nd update, because that'll frame up what we're looking at here is write me a webpage. Yeah. So on the left-hand side, those of that saw the episode probably recognize this now on the left-hand side, we have the chat where we're interacting with the chat. And this is important because this is a different style of chatting with the language model. This isn't the question answer machine thing that most of the world thinks we're supposed to do with this, but we're actually doing is saying, write me a webpage. And then on the left-hand side, it's actually writing code, but it's got this like preview page. And so this is actually the page that it's writing.
And we went into some discussion in our last episode about whether this is production ready, or if it even needs to be and all that kind of stuff. Right. But the idea is. As a coding partner, you can then go in and say, um, I like red white and, and, uh, blue. You can just give it, you know, new updates. And then what it does is it goes and changes the code. You see the code being written. And then when it's done with the code, it shows us the thing that we're writing. So for me, yeah, your thoughts, first of all, I know we talked about this. So this is a really visual version of something that I do use chat GPT for all the time, which is to create a. Like, if you're thinking about something to create a table and you can, you can say things like delete that column, add another column for miles per gallon, you know, go out on the internet and find other ones.
I use this for a lot for scooter comparison. Um, you know, go find any other models that are available in Greece that are within that same price range, compare and contrast, which is better for a tall person at a column for seat height. You know, you can totally scope a spreadsheet over a long conversation. This is a webpage or a God. I wish I would do it for, for FileMaker or layouts scripts. And yeah, thank you for giving a much better spread, creating, crafting a spreadsheet example, because I'm using this web thing again, only because we showed it on the last episode, but this is, I've never used it for creating a webpage either. I do it the exact same way you're talking about.
I say, um, give me this dataset or I upload something and I go extract just this data from it. Same kind of thing as you do add another column, do this, do that. Okay. Now give me a visualization of all this and then write some. You know, um, synopsis of what we're looking at or something. The whole point to kind of step back away from the process, the whole point of the visual here is that we've got instruction and conversation on the left and this sort of somewhat ephemeral UI on the right. Like whatever it is we're doing, if you're writing code and you want to preview it, then show the preview. If you're creating a table, if you want to write a report, the whole point is it keeps showing, you notice on the bottom, it shows version two of two and all that.
Yeah. Yeah. So I think in this version, you probably. Can not drag and drop, um, visual elements around. Um, yeah, I haven't had that, you know, so I think we're, that's where we're going though. Right. It's going to get there. This is great, but I want to put this over there. I want to make this to the other size. I saw an ad on one of the other pages you were looking at for Wix studio. I use Wix for my webpage. Wix studio actually has some of those AI features. Um, it doesn't have the, the both direction stuff yet, but it has quite a bit, um, to automatically. Generate, uh, things like you're showing here. Well, essentially every thing where you're building stuff, not that includes FileMaker should have this sort of, let's call it.
I mean, I know I'm going to, if I listen back to this or somebody listens back to us a year from now, it'll seem dated what I'm about to say, but left-hand side of the screen is where you're interacting and coding and having a conversation. And the right-hand side of the screen is the thing you're seeing that you're actually interacting with, we would call it layouts and FileMaker on the right and the thing that the user wants. So they're having a conversational experience rather than a UI clicking through a whole navigation tree to go through various different layouts. It's creating the layout that that user wants at that moment in time. And then it disappears.
You could recall it if you want, but that's the whole ephemeral part of it. And just to show you that this is becoming even more of a thing, you can see this page too, right? Or did I just share the Claude? I just saw it. Okay, great. So then just a week or so ago, um, Microsoft announced something called. Co-pilot pages, and they also said for multiplayer collaboration. So basically they did the exact, um, I'll mute this, but they did the exact same thing, Matt it's code on the left-hand side and on the right-hand side is the thing that you're seeing. But what they added was multiplayer, which means that, um, you and I could be looking at the same spreadsheet at the same time and adding our own data.
And because they're Microsoft, they are adding another layer where you can integrate with like SharePoint or something like that. Right. That's cool. Or, you know, I'm sure Google will be right there with this too, right? I use, I use Google docs and Google sheets with other people all the time and you can see where people's cursor is. It's cool. Yeah, it's that, so there's the share thing. It's that, but with access to all the tools, Google has one, the Google suite in this case, the Microsoft suite, but the point is more traction on this interact with it on the left and the thing that you're looking at and now sharing on the right. So the, the thing that you and I were just speculating on, even just a month ago.
Was that this might be the direction that user interfaces for data and applications may be going. And now we're seeing validation from some of the biggest players in the space. So I don't know, Matt, this video is cool. What, what could possibly be, yeah, there's your drag and dropping right there. So what, what are your thoughts on the impact this has in, for our listeners who are custom application developers for business? Huge. I think this is profound. And extremely important. I don't know what it means within our stack. We obviously can't do anything like this in FileMaker. We can do it in the web. No, but we can, if we can use something like Figma or something like that, right.
To come up, if we have a client meeting and we're using a tool like this, you can, you can get a really quick proof of concept and understand what you're going to be building for them and then go build it in FileMaker. So right now, obviously this can't write the FileMaker code and all this, you know, perfectly. To do all that, it can do some of it. So that's probably the right answer for this conversation. Whatever you're looking for. Well, no, no, that's the right answer. I'm saying, I think the real answer is that's the app, not the prototype for the app. Yes, yes, yes. I think the important takeaway is this could be where things are going. Yeah, yeah, yeah.
There's, there's a traction behind that. Tremendous amount of traction that is completely enveloping all developers. And if you combine that together with the things that we talked about before, agency, writing code in the background, memory and reflection, and these new types of reasoning models, and then you combine that together with the user experience and connecting to the database, data sources and the apps on the backend. This fundamentally changes the user experience forever and permanently. Yep. And it's not just AI. I mean, like, you know, question answer machine, it's all these tools working in conjunction with each other. All Google is going to do when they come up, they're offering and all Microsoft has done is just take all their apps and feed it into this whole paradigm.
But they're still using the user interacts with something and the things they want ephemerally appear on screen. That is what applications are going to be and arguably already are. That's what I'm saying, I guess here. Yep. And as a reward for those that have stuck out to this point of the podcast candy now, terror, just plain terror for you. But I feel strongly about this. And I know that these are conversations that are happening at Claris as well, too. So what do you mean terror? This is purely exciting to me. Yeah. But well, okay, well, I'm just going to say it. We make a living. I bought this house. You bought that house off of doing what? Yeah. Charging people hundreds of hours to do something that shouldn't take hundreds of hours to do if they can just get the thing that they want to do another way.
All right. Okay. You and I have made our living by using a tool that's much simpler and faster than other ways to do the same thing to deliver for a certain client base. You and I are going to be able to use a different set of tools to do the exact same thing. This is a great conversation. To be able to build things better and faster for people that we can do now. So we'll be able to do cooler things, new things that we couldn't do before. Yep. So let's do the math on that. But all the time we've been doing what we've been doing, the customers that we have have been able to do it as well, just at a lower level because they haven't put in the hours to understand the tools.
Okay. Yeah. First of all, you just said hours. Okay. So let's do the math on this. Okay. So I totally agree with you. That's exactly what we've been doing. So let's say that these tools allow us to do it 10 times. 10 times faster. Let's just use that as simple math. So that means that let's just say you have a business and you build a thousand hours a year. I don't want to get too many zeros on this or I'll lose the math. You have built a thousand hours a year and that's how you make your living and everything is great. Now your tools allow you to produce the same thing in 10% of that. So now you're doing the same output and you're only billing a hundred hours a year for that.
And we could get into a whole value build conversation, but I think value build, the air comes out of the balloon on that just as much because of what you just said. Our customers are aware of these tools and these efficiencies already. Okay. So now we go back to our 10%. So that's all great. If we can go back, become salespeople and go increase our sales channel 9x. But your math is wrong. No, it's not, Matt. It is because we've built a thousand hours a year or whatever. Yeah. And now you're building a hundred. No, we're not. We're still building a thousand, but we're doing more. We're doing more stuff with it. I know you're not listening to this part of the math.
I'm agreeing with that. The point is in order to get to the build more stuff, you now got to put your sales guy hat on and go find nine times more work than you had before. And what I'm trying to say is, Matt, explain yourself because you make no sense. When I set aside work to do four hours for a particular customer, I can do as much as I can to get a really, really cool product for them or whatever I'm specifically trying to do in that block of time. They've asked you to do a thing. They're not asking you to show off and do your- No, they're asking me to solve a specific problem. Yes. So you solve that specific problem. So they're saying the customer, like two years ago, the customer would say, "Hey, I need a sales system so that I can find a prospect that's a good prospect for this particular thing."
Whatever it is. And so I can write an email to them and say, "Hey, we've got a new piece of inventory." Whatever it is. The task is not relevant. The fact is that two years ago- Two years ago, it took you how many hours to do that and how long does it take you to do now? That's not what I'm trying to say. It's not what I'm trying to say. Two years ago, I would have spent four hours to do it and I would only have been able to do so much because the tools could only do what they could do. Now when they say to do that, I'll spend four hours and I'll make the same layouts and stuff that I did. But now I'm going to do a generative AI call and I'm actually going to spend more time because I'm going to write some custom DJs.
Jason to collect all the information about the customer into one block that I can then send to generative AI and say, "This is a summary of this customer." And then another generative AI call that says, "This is information about this particular product. Marry those up and give me a really tailored email for this customer and why they want this product." I could not do that. There was no way to do that before two years ago. Yeah. So you're saying two things that are important to our conversation. The task that you're being asked to do and the deliverable have changed. Yes. In two years because of the capabilities have changed. I'm suggesting that that thing you just described, you'll be able to do way more in considerably less time is what I'm trying to say.
But the customer is still, that's still only one part of the thing that they want me to build. Okay. I couldn't do more. Now the next thing they're saying is, "Yes, now I need a way to do export of images to the web. Now I need another way to do all the other things that they still need the application to do." Yes. Still take those hours to do it. It's just, I couldn't do as much as I could. And so the math that I would use is this, because I go back even further than you- Well, I just want to summarize what you're saying is that, I just want to summarize that you're saying that because you can do more in a thousand hours in a year, your customers are getting more basically, right?
So that of course- It always has been. Every year as I have become a better developer and as FileMaker has become a better platform, I have been able to do more for customers in the same amount of time, every year. For 66 years. Okay. Stop talking for two seconds. It's not that long. But dude, that assumes your customers keep asking you to do things. I'm just trying to get out this one very important point, my friend, is that we would need a market that is expanding, demand that continues to expand, customers that keep coming to us for stuff endlessly without us ever having to do any lead gen or demand gen or anything. And guess what? Our market is not expanding.
I'm a person of faith. Okay. Keep your faith. I'm just saying our market is not expanding. It's contracting rapidly. And the expansion, the market for the tools that we're showing people here today is expanding rapidly. But the expectation for those users is that should not take a hundred hours to do something. As a matter of fact, a lot of these low-hanging- You don't have enough variables. Your equation is lacking variables. The number one problem every single Claris partner is telling me they have right now is finding developers to hire that can do the work to fill what the clients are asking them to do. I'm sure. That's a variable that kind of disproves what you're talking about.
No. Who knows what it's going to be like two years from now? I don't know. I'm talking about the future, Matt. It's adorable. It's cute that everybody is demand. I do too. I run on business as well too. I'm not retired. I have these same problems. We literally just had these conversations yesterday. On the AI side, guess what, Matt? On the AI side, I have the same five people on my A-team and we have been able to increase. We literally did a study. I did it in my presentation. At Engage, 5x manpower or person power or whatever we call it. That means that I'm able to do 5.5x more output with the same people, meaning I'm accomplishing way more things that customers want and making way more money without having to add headcount.
The reason that we can't do that in FileMaker is because we don't have the tools, dude. FileMaker's not building these. All right. You can wait around. Well, I'm not going to... That's not what I'm saying. Yeah. Everybody can go sit over there and wait for FileMaker to do this. FileMaker only has the tools that it has. But FileMaker can connect to all the other things. That's where it's been going, right? We can build a JavaScript thing. We can have a JavaScript and stick it in a web viewer or whatever. And so all of the complexity for that thing-- Or just build it in the web. Or you can build it in the web, right? But customers have been able to do that this whole time.
But it's been more complex and difficult, so they haven't. So there still has been a space for FileMaker this whole time, whether it's going to continue to increase as it has continued to increase year over year, including last year, including... No? Yeah. OK. No. It has not at all. But anyways, I'm not going to sit here on a podcast and freak everybody out. Then why are all my partner friends trying to hire more employees to do FileMaker development? You're absolutely right. The universe is expanding. The demand is there. We cannot get enough people. All we need to do is get this into universities and teach people FileMaker, and we're going to take over the world.
I totally agree with you. No, I don't think that's the case. I don't think... Totally agree with you. No, you don't. Yeah. Everybody's fine. That's not my thing. I don't think that's where the FileMaker market is. I don't think that's ever going to happen with the whole university thing. No, that was just me being silly. Yeah. We've talked about that before. Yeah. No, you know what? You're absolutely right, Matt. You've convinced me. I think all the partners in the Clare space should stay the course and double down, triple down on FileMaker. Don't... Just forget all these silly fads, like crypto and NFTs and stuff like that. Ignore all that stuff and keep doing what you're doing because there's not enough room on the arc for everybody anyway.
So how about some candy, actually, Matt? You sent me something really interesting last night for the ears. We got... This is one of the most heated, cool discussions we've ever had. I know. I love it. I love it. I got about 25 people in my head right now that I can't wait to hear their thoughts on this whole discussion. So, hell yeah. So, anyways, fun times. We've got something really cool. I want to tell you about a podcast in the FileMaker space that I heard about. Talk about, like, expanding. So before that, though, you've got something for people's ears that I really can't wait for you to share. I think it's so clever. All right, FileMaker fans. Oh, I'm sorry.
Go ahead. Hey. Hey. I don't know what that was, but... Okay. So maybe I'll do it and then... And then talk about... Talk about what it is, yeah? Yeah. That'd be perfect. Hello, everyone. This is Matt Navarre 2.0. Multilingual. For example, Greek. Yes, all well and good. But what about Cantonese? Let's try Italian. Okay, let's move on. How about French? And finally Ukrainian. Hello to all three of you who don't understand me. I support you. So you've been busy. I've been spending some time with Duolingo. Since our last pod, you've been, like, just crushing learning new languages. Hey, buddy, I am impressed. I've always thought you're one of the smartest people I've ever met.
And you just proved that you're even smarter than I think. How in the world did you learn that many languages that quickly? Yeah, I didn't. Oh, come on, man. So what that was, was... It was a website service, probably something you told me about a while ago. It was a website service called naturalreaders.com. Oh, yeah. Which, by default, you can have it read... You can just give it text, and it will read it to you as a book, and you can save it as audio. But it has a feature where you can sample your voice. So I made a half an hour recording of my voice. Some of them say they want five hours of audio of your voice. And then it captured all of my pitch and speech and everything about that.
And then I can... Then I just put text in. So the thing that I just played took me about 10 minutes to make earlier today. Really? Yeah. I went into ChatGPT, and I said, "Hey, I've got a thing I wanted to do. Let's pick some languages and write a simple sentence." And by the way, the thing I said in each of those languages may be entertaining to people on the podcast who can listen to them. Oh, no. Are we going to get flagged for anything that was in there? I don't think so. There's nothing like that. But... Well, so, you know, I'm not sure if I'm going to be able to do that. I'm not sure if I'm going to be able to do that. I'm not sure if I'm going to be able to do that.
well so you know first of all that's amazing and then i think when it comes to tools like that like when when you and i are are educators in this space right and uh um one of the big things has been localization as a matter of fact uh filemaker roam uh i'm really i really just submitted a video presentation and they spent a bunch of time um translating it into the different languages for the attendees for filemaker roam um i suggested that a tool like this would make that process a lot easier as a matter of fact like arguably every content you know like linkedin learning uh probably is doing that since they're microsoft but like localization used to be a big deal as a matter of fact like many many moons ago when i did my filemaker uh essential training uh clarice wanted or clarice or linda or whatever wanted to see it in japan or some other geo that i didn't speak the language and so i had to meet with somebody go over my school and i had to meet with somebody go over my school and i had to meet with somebody go over like they had to translate it into whatever language it was and then explain a bunch of stuff to me the whole point is you could have just run it through that thing and done dubbing on it and then be able to deploy in a bunch of different languages arguably we could do this podcast in that same manner frankly right and open up our reach um i think that's amazing i'm so impressed that
you learned all those languages because um and by the way in the interface i should i could maybe show it um oh yeah go ahead yeah i will um okay in the interface to do that every little bit of the speech you can choose the language and the voice that it goes in that's how you did that right so like this can be 1x speed and then when i'm speaking in um this one here let's see if it plays can you hear it wait yep okay so so if that's two two fast you call me make the speed much slower oh and so like if i'm trying to learn a language i can use this to like listen right much much slower beautiful and then so in other words i can actually have your voice and my voice and so i could pick like a different voice in here if i have you know you and me and then each paragraph can be a different person so you could you could build a whole audio book in whatever number of languages this thing speaks paragraph by paragraph speed by speed
it even gives you the ability um to like change emphasis but only in the um the voices that it includes it includes a ton of voices when it comes with it by the way this is a hundred bucks a month at the level that i'm at right now and and and just let you know like the thing you were thinking maybe i showed you there was a 11 labs as a tool that i use for that's a direct competitor yep now 11 labs has a dubbing tool and i can tell you that and i and i've played with it before it doesn't have that it doesn't let you go in line by line because like you listen to your five minute video and it's dubbed and it's fine but you go well i would rather change this or the speed on this isn't matching the click or whatever so that level of control i think is is better frankly so um you know these are we talk about like like deep fakes like you literally just showed a deep fake basically like i know and there's a lot of people who are like oh i don't know i don't know i don't know i don't they're terrible it's all bad and no one should represent people the wrong way but there are actual reasons why these types of technologies can be useful and you just demonstrated one of those yeah that it's it's pretty much indistinguishable from me you know agreed i spoke without a lot of inflection and stuff like that and then what i love is it doesn't make mistakes right so if i'm reading something really long it just doesn't make mistakes and that's pretty that's because
in other words if you try to read an audiobook you're gonna read you're gonna have to do a lot of editing and a lot oh my god um but this one i can just throw text at it and say okay here's a book do it well you know i was telling some people at engage that i did the um micro certification course for clarice for clarice connect and it's like uh 32 videos on the three different i think 36 videos for the three different courses yeah and if you haven't done it yet when you go do it you should be aware that um five of those are my fake voice oh and uh i did it for reasons like you're probably hearing right now i'm getting a little hoarse i just i don't have the stamina voice-wise that i used to when i record yeah so you know to fill in the blanks it was me and i would challenge anybody that hasn't done that micro certification yet to go through there and try to figure out if they can tell when that was me or not i do all the micro certifications but i just go right to the test well i've watched so many videos the top level you'd have to watch the videos unless unless you just connect a lot yeah because that's the one where what was it again oh the creating the custom connectors right like there are some like things you got to follow um all right so my ear candy at the end of the show is look we do a
podcast in the the ever-expanding universe of file maker opportunities and um i just wanted to give a shout out to a new podcast in the file maker world we don't usually plug podcasts without having the host on like we were actually plugging uh matt and alexis's uh podcast right this one i think is really compelling there's an episode i'll just do a quick little snippet of this episode on working with language models in file maker 2024 see what you think all right file maker fans buckle up because we're diving into the future file maker 2024 is here and it's brought some seriously cool tech along for the ride it's a game changer no seriously ai and machine learning are officially in file maker we're talking about teaching your database to think well kind of like you do understanding what you mean not just the words you type yeah it's a whole new level of database intelligence we're going to break it all down today and we're going straight to the source the claris engineering blog their july 2024 post to be exact all right what do you think you're the you're the sultry tones uh audio on this podcast so what do you think of what you just heard there um that is a new tool for those of you that have maybe not heard about this yet but it's called notebook lm from google
and it's basically a gpt see these popping up all over the place this is what applications are about now you you upload a bunch of your data sources like a chat gpt gpt would be for those that are just listening and in this case all i did was i took a pdf of the working with lms and file maker 2024 engineering blog and i uploaded it and of course like anything else it can do summaries and all that kind of stuff but this cool new feature if you click on notebook guide it allows you to generate on the fly the data that you're going to be able to generate on the fly and it allows you to generate on the fly a podcast recording so um it you heard there it was two different people interacting laughing with each other cracking jokes you can even hear the microphone pop sometimes i'm taking breaths it really is very listenable um this is sort of an extension of i can't remember what the apple is anymore but there's this thing where you could like feed a website to it and it would scrape it and then come up with something similar like this now you can actually feed it the content so um we actually have clients that um like we have a client that they do like a shift change and so they had us build a solution that goes and gathers all the data from uh you know this is what happened and this who did what and all the transactions that occurred and then we feed it into a language model to create a script and then
it it through an api generates um an audio like this that they send out via email so when the next shift is driving into work or they're just got their ipod their headphones in they can listen to all this stuff and consume it in a different manner um it's similar to the nfl custom podcast uh demo that i've showed before where it's just personalized this could be personalized just for you but it's all about like the listening being the new intake method uh instead of um reading for example so uh i thought that was a pretty cool thing so that again is notebook lm we'll put the link to that in the show notes here so um another little tidbit listening instead of reading i uh have been thinking all along that people learn in one way or the other but i've been thinking a little bit more about that and so i'm going to go ahead and show you a little bit more about that and i'm going to go ahead and show you a little bit more about that so let's go ahead and go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a
little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go
ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a-
let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and share this with you. let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and show you a little bit more about that so let's go ahead and share this with you. let's go ahead and show you a little bit more about that so let's go ahead and turn this off.