Episode 522: What Was Your Eureka Moment with AI? Smarter Workflows for BC Development

Heat Wave Egg Experiment Opener

SPEAKER_01 0:00

Uh actually some of their um safeguards are triggered by my simple AL tests.

SPEAKER_00 0:07

Like, oh no, this is uh Welcome back to another episode of Dynamics Corner. How do you manage with all the different models in the AI space? I'm your co-host Chris.

SPEAKER_02 0:20

And this is Brad. This episode was coded on July 8th, 2026. Chris, Chris, Chris. AI models, workflow, development. So many questions, so many things to talk about. And with us today, we had the opportunity to speak with Torbin Leth.

SPEAKER_01 0:53

Good afternoon.

SPEAKER_02 0:54

It's nice to speak with you again. Welcome back. Welcome back. Hi. Yeah, I was just thinking as Chris and I were talking. I have many questions for you, as always, but I have we were just talking about it's now summertime. And everyone complains about the heat. And over in the northeast United States, they're saying it's the hottest it's been, or it's we're having a heat wave or whatever. And uh this recording, I'm not up north, I'm down south. And the feels like temperature is 112 degrees Fahrenheit, which is 45 degrees Celsius. So I'm going to do an experiment after this. I'm going to take a pan, and everyone always talks about this, and I've never seen this. I honestly haven't. I've heard about it, but I want to try it. I'm going to take a pan, put it outside in the sun because it's uh lunchtime. I'll let it sit for a little bit and go see if I can cook an egg with the heat of the sun. Uh, do you think it will cook? I think it's slow.

SPEAKER_01 1:55

It's low. But it's gonna be possible. If not, then put it in the car and then try again. I think uh put it in the windshield and let it heat up, then I think it's gonna be there.

SPEAKER_02 2:07

I would do it on the car, but I'd be afraid of the paint. But I will let you know how the experiment goes. It's sunny and uh we'll have to

Tech Days Stories And Stage Ideas

SPEAKER_02 2:18

see. The other thing is I was I had the opportunity to watch uh some of the BC Tech Days videos, and uh how was that car?

SPEAKER_01 2:26

It was fun. Uh quite it looked like a lot of fun. Um when Dimitri pitched it, it was like, okay, well, let's try and see how it goes. But whether it fit it fitted the team, the theme. So uh I'm just uh I've been poking a bit like uh what's gonna be next year? It's gonna be a blimp or something, uh what I it hasn't told me yet.

SPEAKER_02 2:52

You do I'll have to reach out to him as well because I think you do have to top it because you have already come in with bringing the car into the presentation. Uh you have to come up with something better next time. Like you said, uh a blimp, maybe go to a flying on an airplane.

SPEAKER_01 3:07

Yeah. Or maybe have an animal. Let's see what happens.

SPEAKER_02 3:11

Or maybe at that point you can have uh an Optimus robot or another type of robot come on stage with you and uh do a little dance or something, I think would be creative. Yeah.

SPEAKER_01 3:21

Uh well I know I a certain conference that uh wanted to do that, but uh skipped it. So uh it's I know it's in the works for some of them, so you have to tell me which one so I can make sure that uh I know I do my best to get there. They directions uh wanted to do it, but uh they had a few hiccups, so uh I think it's better they tell them, but uh it was it sounded fun.

SPEAKER_02 3:50

Oh, that does sound fun. It's uh I think it was

Meet Torben And His Tooling Focus

SPEAKER_02 3:53

fun. Uh before we get into the conversation, would you mind telling us a little bit about yourself?

SPEAKER_01 3:57

Yeah, my name is uh Tom. I'm uh one of the new MVPs in the PC space. Work at uh continua for the last what's six years or something like that. Yeah, and I love fiddling with developer tool tools and all kinds of stuff. I'm more not that I don't do AL coding, but I love doing the the tooling. So make everyone else a better developer. That's like uh what I want to do, or at least make you better at using AI to make yourself be a better developer. So a lot of tooling, that's why I love to do, and all the other sort of stuff I do on the side.

SPEAKER_02 4:39

You do a lot of stuff, and we had you on several months ago, and we scheduled a follow-up because we wanted to talk about maybe some of the changes, some of the things that have happened over the past couple of months, and how there is a lot, and how developers again with you talk about some of the tooling and how developers, AL developers, can use some of this tooling, or now with some of the improvements. I don't want to say the words improvement, some of the progressions with AI and AI harnesses, how they can become more efficient and maybe more productive with the use of AI. So I'd say we could have a rundown of the things that have changed, but we'd probably be here for a lot longer than uh we'd all uh have time for with the things of changes. But um just a quick recap. What have uh you've been doing for the past several months, or what changes were significant

Why Benchmarking Models Got Hard

SPEAKER_02 5:35

to you uh in the AI world?

SPEAKER_01 5:39

Well, the models are getting so good now that I need to redo my entire test set for like uh the central gate thing where I measure all the AI models. They're simply too good now, so they're saturated by test set. So it's like it's minuscule the amount of test they can't fix now, so I have to be even more creative of how to doing it. But it's also given gotten to a point where um now it's more about the harness. The models are even not the top models, are like good enough. Um but also some quirks like Fable, for example. Uh everyone knows about that one. Um the issue with that one is when I test it, uh actually some of their um safeguards are triggered by my simple AL tests, like oh no, this is uh dangerous. I am not allowed to work on this on that on scale tasks. So I'm having issues benchmarking it because should I allow it to then retry or should I have it allow it to do fallback to opers? Or so there's a lot of new things I have to think about now where they're doing this fallback thing. So um because what I want to know is how good Fable is at a given task. And well, I I was one of the few guys that I did do a benchmark in those two days in the early start where we were actually allowed to use Fable, like almost all of it. And yeah, of course it went through all my tasks and set a new high for yeah, all time. But um now it's a bit more like it fails a bit, and that's because I haven't enabled the Opus fallback. But that is again because I want to know when it does the fallback so I know when I can't trust the score. Um so I have to do more retries and retries on fable than I have to do for the other ones. But it could be like uh depending on how Sol or the next opening I model, if they are gonna have safeguards too, it's gonna be a common thing that they're when when you ask it for something about biohacking, it will use that model in the background. When you ask something about coding, it will use that model in the background. So it's gonna be a bit harder now to like benchmark models.

SPEAKER_02 8:14

Because it can decide. Yeah, it would have an auto mode. It's oh wow. I have so many things to talk about with you, but that's uh uh your central gauge and engaging the AL, engaging the models for AL development is a great tool. And as you you had mentioned that the models are getting far better at solving AL tasks that it's difficult to benchmark them. With your benchmarking and with your experience, because I I've had geez, as many changes as they've been to AI models, I've had conversations with people about using AI with AL. So it's a lot of different viewpoints, a lot of different experiences, a lot of results, different results as well. When you're feeding the models, do you have any context that you feed it ahead of time, or are you doing straight model to AL task?

SPEAKER_01 9:13

What I'm doing is that I'm giving it a straight up model. Um I'm giving it a task and some existing code where it needs to fill in some gaps, solve a task of some sort in this object. And then I already have a test code unit ready in the background where it's gonna poke at these points and make sure it returns the correct result. So it's like um here you need to do some kind of um fetch some kind of something in the background, make for example a flow field, uh filter or something, something, and then I already have the test ready to check check it afterwards. Um I'm also fiddling a bit with the harness stuff like uh Microsoft has. Uh I'm on good terms with the Microsoft guy that creates the tool and uh been speaking a bit with them on the back end. And um I've created my own version. Right now, Microsoft don't accept um pull requests on their repo, so I'm making my own fork. But yeah, we're gonna I'm testing a lot of new stuff on their framework with a lot of extra metrics that I don't have. So um I'm trying to uh expand it so now it's not only models but also the the harness. I just don't want to overstep too much.

SPEAKER_02 10:40

No, I understand it. It's it there's a there's a lot to it, and it's it's uh uh everyone will say this model's great, this model's garbage, this model's it's the same

Harness Choices Safeguards And Fair Scores

SPEAKER_02 10:54

model with the same type of uh space AL development with people yielding different results. That's why I was asking if you've given it any context before and what the what your results are and what you've noticed is if you preload it or pre-feed it some background context, whether it would be like the BC quality uh repo or some other existing coding rules. Um back when we had the AL dev guidelines, you know, AL dev uh website.

SPEAKER_01 11:27

I don't do that because what I want to test is the model's own training data. It's like the baseline. I want to know what the model knows about the AL language. So I'm also like asking it uh as soon as the AL language gets any new features, for example. I'm also making tests for those just to see how it knows about that concept. Uh I know uh a common thing I also mentioned before, like the GPT models have a sort of an issue with interfaces in AL. There's not always they know about that concept in AL languages. They're fine at creating code and filling in, but if you say you need something about an interface, it doesn't know how to how the format of an interface looks in AL. It I think it's TypeScript, I can't remember. So um I don't use like PC quality or anything like that. That's more the the um uh PC bench that where you do the entire harness where you can add like additional information. I though I create some I don't know if it's public on my site, but I'm creating a sort of an extra rule file so that when I benchmark my do run for all my tests, I will, for example, find out that GPG has an issue with this and that, and then I create my own like uh guideline file where oh by the way, if you're working on Ale, there's these few concepts that you don't know about, so remember those, so you can add them as in your Claude MD file or whatever you want to use and as a harness, then it knows about that also. Um so you can like raise all models to sort of the same baseline.

SPEAKER_02 13:16

Yeah, that's was the other question, and you you started to talk about with the interfaces and not having uh uh the model not having knowledge to understand, I guess. I I I hate to use the words because it's a computer, it's not like it understands, it's but I think we all now talk about we call it an it, with he, uh she. We everyone starts to personalize it, but we have these models that haven't been trained on the latest AEL syntax because that language is also it's not an old language like C sharp and as you'd mentioned, TypeScript and Python, where there was has been a lot of code available, and the number of changes with AL development in the language is greater than some of those languages. Yes, they've improved and they've added things to those languages, but or syntax or commands. Is there a way to determine how far a model is with the AL language to know when you need to augment it with that additional knowledge? Uh, for example, is uh the interfaces, for example, how would I know? Because if I give it a task and I want to do a task and it could be benchmarking or I want to review some code or write some code. If it has training up to a certain point, it's not aware of the new syntax. How would I know that it's not aware of the new syntax if I was a developer?

SPEAKER_01 14:42

I could I already show that information on my side, but it's hidden and you need to well, it's not hidden. It is there, but it's just a lot of information. So you need to know what you need to look for. There's already tasks like divided into subsections of like this is tasks about interfaces, this is tasks about tables, and this is about rec reps, field reps, and so on. And then you can see which model uh succeeds or that um completes each task in each group. So you can actually go in there and see which is good at what. Um but actually it's a bit of an issue, like I have all my tests public, and that is good because I want people to see what it is I'm testing. But the issue is that um they're using the same data for training. So I'm a bit of a pickle. I want people to know what my training data uh what my tests are, but they're on a public repo, so they're gonna be in the test data at some point. So I'm thinking of another place to host the tests itself and and not in GitHub simply because I can make something and then in half a year, suddenly all models know a lot about how to solve all my tasks because they're open. So uh that's I think something I'm struggling with right now. It's like I want to be open, but uh don't want it to be too easy for the models. So uh bit of a pickle.

SPEAKER_02 16:08

I chuckle about that because I remember conversations about even like the Cronus data in Business Central, because the Cronus data is part of the sample and it's it's public in essence. The AL models know about the Cronus data. It's it's something to really think about as far as what it knows and what it learns. Um I saw something, I could go uh many

Testing AL Knowledge And New Syntax

SPEAKER_02 16:32

different angles here, but talking about the different models uh and then somebody using it. Uh I do want to talk about token usage at a point and maybe how to manage the token usage because I see some of the stuff that you're doing about uh the models and the tests, but I read something about models and you talk about Fable, where you can have Fable be an advisor now, right? Or I think before we had auto mode for models where the harness would try to choose the best model based upon the prompt, right? Simply put, in some of those cases. Is there a way to structure and set up your environment so that you could have that type of advisor set up with different models? Whereas if GPT-5 is better for documentation or having knowledge of documentation and Opus 4.8, maybe the most recent for AL developments and such, is there a way to gauge and set up it could use different models for different tasks, which would relate to one results and two uh credit slash token usage?

SPEAKER_01 17:45

Yeah, there is. Um, it's a bit of a tough question because I, as an ISV, is gonna have a different workflow than a partner, for example, is gonna have. Um lot of my workings is in cloud code. I know a lot of people use this VS Code. I haven't opened VS Code in like weeks. Uh mostly just open it to check like structure of a JSON file or something. Um so what I'm gonna say probably not gonna work for everyone. In Cloud Code, I have Fable set up as a I uh normally work in Opus, and then I say when you I'm using also the superpower skill set that is in there, and then I say now let's create let's brainstorm about something, and then it creates a spec, and I say, Oh, ask Fable or review on the spec. And you don't have to write anything more than that. And then we'll start up a sub agent and just go through the specs and say, Oh, there's some feedback here. And I also installed a MCP called PAL. In there you can add other uh providers like OpenAI, also Open Brouter. So I actually have it set up in one of my skills that when I want uh consensus, it will ask um TG55 Pro and Gemini 31 Pro Preview for additional feedback because they're so different compared to the open a uh the um anthropic models that they often see things that the other ones don't. And I just use those for review, and then when it's all done, I then tell it, oh now the you don't have to think anymore, so you just use sonnet for all the coding because the spec is so good now that it's only minor things you have to think about, and then that sonnet can easily do, and then the cost is way lower than using Opus for all the coding.

SPEAKER_00 19:45

So it's a it's a plugin that you're using with CLI. So Cloud is your main and you have a plugin for called PAL.

SPEAKER_01 19:55

Yeah, it's just an HP server, so it can be added to any one of your coding clients, VS Code and so on. So um yeah, and oh and then uh again my entire own flow, uh which are the one uh shown at uh Tech Days. Where I have my own entire pipeline where I have different stages and then in a planning stage is the OPES in the coding states and so on, and there's like 13 stages where I've handpicked which model to use when and where. So um that's again to try and keep the cost down. Um but that's for entire flows, like from when a bug is reported to it has been uh documented and sent out. So that's uh sort of a semi-fire and forget flow for simple coding tasks. So when we go into our uh that's uh we use Azure DevOps there. If you see some tasks that's like what a junior could do or something, then I would simply just sit attack and then eight hours later there's a pull request, here it's done. Oh, thanks.

SPEAKER_02 21:07

I don't have as many steps, I have fewer steps, but uh it's reassuring to listen to that because I have a similar flow with you where you you tag it and then the agents go through the workflow, and you sit back and you watch it write the implementation plan, you watch it review it, you watch it code it, and then you get a pull request. It's times have changed. It's uh it's it's a lot of words come to my mind when I see how all these processes work. And uh it's the the coding is no longer the bottleneck in a lot of cases. It's the review, actually. Yes, it's the review, and then also coming up with the idea uh of some of the things that need to be done as well, too. So to go back, Chris had talked about so from the flow management, you may use OPUS. To start a spec, and then within Claude Code, you tell it you know, have Fable review the spec, and then you also say again.

SPEAKER_01 22:11

Yeah, that's also a new feature. I don't know how many knows about in clawed code. You can set an advice advisory model. That is like uh we can set that to Fable so when you need a talk to an adult, you can always uh talk to Fable.

SPEAKER_00 22:26

That's a good way to put it, actually.

SPEAKER_02 22:28

I like that. Well, that's what I was reading on about the uh advisor models, and uh you know to help uh I look at it to more or less help manage Fable because I've uh unfortunately or fortunately, I live my life in this stuff now and working an experiment and Fable when they had it out the first couple times, and then even now it's it's back out. Uh I think I burned through my usage, my session, my session usage in minutes, and then even going back and having to do some tasks, it just runs. It doesn't walk, it doesn't crawl, it runs at a fast sprint to get things done, and it's almost sometimes difficult to interrupt it, or I feel bad interrupting it.

SPEAKER_00 23:11

Well, sometimes it sometimes it continues, like if you don't respond in time. I had one the other day where I had to walk away and get some water, and I guess it asked me a question, but I was away from the desk, and it's like, well, the user's away. User's away. I just wanted to go ahead and do the default.

SPEAKER_01 23:26

I'm like, wait, there was a buck in called code that I fixed that now. Uh it was actually an an opt-in. They uh by mistake has had an opt-out. So uh for a day or two, they had the uh the 60 seconds uh uh away thing where if you didn't answer, uh then I'll pick for you. And yeah, yeah.

SPEAKER_02 23:48

It continues. It's um yeah, it's uh I'm just I'm I just think about some of the scenarios that I've had and and uh to bring them up, but I'm more into um uh from the AL development point of view and from the development point of view being able to, as you said, help developers be able to use these tools to complete and finish up their solutions as well. Um as far as the other there's so many tools available. We talk about BC quality, there's so many repos, there's so many MCP servers. You have the AL MCP, you have the AL tool, even Claude can use the ALC.exe to be able to do things. Um there's a lot for developers to take in. And I I'd like to ask these questions to various individuals that I speak with. What should they focus on? What should they begin with to be able to set up a workflow, right? So we have an agentic workflow of maybe getting an issue or a task to be able to create documentation, to be able to develop and publish.

SPEAKER_01 24:59

I would say the first time you have your setup set up so that it can automatically compile, deploy, and run tests. That's when you get your Eureka moment because it will just simply make the iteration of creating code way faster. Of course, you can always add uh thing about documentation and so on, the things developers don't like. Uh but when the when you get the the iteration going, where you could just say, yeah, now create a page, add that, and so on, and yeah, let's test it. And then it just compiles, deploys, runs, so you already tests, and then oh there are these two are failing, and I'll fix that for you. Just hold on, and then it iterates again, compiles, deploys, and tests. Yeah, now it's good done. Okay. Uh and that's normally things where you have to sit through and oh yeah, and oh yeah, what's the page again up, quick search, find it again, and things like that. Where when you get that far, that's when you see how easy it's gonna be from now on.

SPEAKER_00 26:03

That is true. It does correct itself several times and even does opens a dev environment for you if you'd like, depending on what you're gonna say.

SPEAKER_02 26:12

It's it's it's it is true, and I'm glad that you hit that point because I was intimidated at first, I'll admit, when I first started working with this, and I'm sure many developers are, but even with the AL MCP, I basically said to Claude, here's the AL MCP server. What do we do with it? And it went through and set up the MCP server. It set it up so now when I'm done, part of my tasks, when it my part of my workflow is it will publish it to a sandbox, it will run the tests. Um it runs the tests a couple different ways, and it even looks at the AL warning results, you know, the AL errors, the AL warnings from the source, you know, the it will excuse me, it will use the source code analyzers and look at the results of the source code analyzers as well. So it will tell me how many warnings there are and it will review the warnings if there are any errors and it can't if it can't package or compile or build or whatever you have. It will go back and try to fix it and fix the problem and then um adjust the tests. It's it sounded so overwhelming at first, but when you just say, hey Claude, here's an AL MCP server, set it up. Yeah. And then you do it, it's it's amazing. It's simply amazing at how it walks you through it with the questions that it comes back with.

SPEAKER_01 27:31

And I heard a lot of people like, how do I set this MCP up and that often you don't need to know? Just uh here is a GitHub repo with an MCP server I want to use. Make it so, and then it thinks for a couple minutes, and then you have a new MCP server that works for that. So the the ease of entrance is not that bad anymore, and I think I think that scared a lot of people before because what is the format of the JSON file? How do I add it? Add a new entry. Is this an array? And uh don't need to do anything,

Advisor Models Consensus Reviews And Stages

SPEAKER_01 28:02

need to restart anything, and you don't need to think about that anymore, just that the AI model do it for you. Um and of course I'm not in I don't help because I create new MP servers uh often. I always had my small side projects, and uh I've added like my own MCP server for debugging um performance and uh analysis of PC service and so on. I created my own um when you do an instrumental profiling, uh the it creates an MTC file, and I created my own converter for that file, so I can use it in like the Firefox profiler uh website and things like that, and add SQL statements and have a lot more stuff for debugging and performance analysis. So um and again, I didn't I don't know how to do it. I just asked the model uh I have this MTC file. What is that? That's uh flap buffer file. Okay, then what can that be turned into? Uh what is the best pro what is the best file format for showing uh performance profiling? Oh it's probably the Firefox profiler. Oh, then let's make that. And then okay, then I had that work for that in eight hours, and then no, now it works. So now I can convert it without uh being dependent on any Microsoft files. So okay.

SPEAKER_00 29:27

It it is so fascinating that you just have to really just talk it through. It's it's almost like um you know the the the the biggest part of building something, it's understanding what you're trying to accomplish and just helping not only it's you're helping it guide uh to get there, but it's also guiding you to say, hey, how about these things? Uh so it's a nice conversation back and forth. It's it's really um and I felt like it's if you're if you're a startup company, it's your your co-founder in a sense. Like if you're building something, you're having a conversation with a co-founder with you. It's pretty fascinating.

SPEAKER_01 30:01

Because I love solving problems, and that is what it is now. It's well, it's it's also been always been fun to write code and make the the best line of code and so on, but I also like to solve problems, and that's what I'm doing more often now. It's like, oh, I have an issue here. I could create something for it. Like I'm starting to switch my laptop into Linux, uh, my work laptop, and I have a lot of extra peripherals there, like and an extra fingerprint sensor. Uh and that doesn't work in my Linux kernel. Well, maybe I could just create a driver for it. So I gave it uh access to Gitra and Wireshark and so on. Started up Windows. Here I want to create a driver for this in Linux, record everything you need to do, and then create me a spec for how to build a Linux driver. And it's been working now since last night about 10, and I think it's sorta done now. So now I can use my fingerprint sensor in Linux. So it's create a new Linux driver for me. Simply just by finding a problem and saying, how could we solve this? I know we should probably yeah, capture the the traffic, and I know it works in Windows, then so let's do that first. And then it said, yeah, sure, and I can also see that this is like using AESC encryption and so on, so on. So yeah, it probably is. I don't care. Fix it.

SPEAKER_02 31:33

It's almost it's taken the value of software, as people have talked about, to zero, in essence, because it the before it was the software and the creation of the software that took the time so that you could solve the problem, and that cycle took a period of time, depending upon what it was. It's as you talked about, it's now you can create things that you may not understand the workings of, but you understand the problem that you're trying to solve, and the results of that problem uh the results of that being a successful solution for uh the problem that you're trying to solve. It takes a little bit to think about that because it's uh it's it's scary, I sense, in a sense, from someone who has been developing for their entire lives. But your point that you had made, and I think uh TNA Starts in a lot of his presentations, he or early on in some of his presentations, he put up the uh tweet that someone made about that is like how developers uh really try to solve problems. Uh I have to try to find it. I have it saved somewhere as well. But you know, developers really are just trying to solve problems, and the code was the tool that they used, but you're still trying to solve the problem, and you almost have to take a step back is every time you get into a motor vehicle, you know you need to go somewhere and how to turn on the car to get there, but you really don't pay attention to all of the workings of the car, right? You just know the car was built, it should work, and you move forward. So um it takes a little bit to reframe your thinking uh in in this this way, uh as well.

SPEAKER_00 33:12

But I think I think the building is very different now. Um because you like like you know, Terminator, you mentioned you had a problem. Yeah, you could search for other solutions out there. You could maybe download an app or someone already built it, but you know, the amount of effort to do in that, you're still having to do a lot of research where you can just pull up Claude and say, I have this problem, help me fix it. So it'll do all the work for you and you're just conversing at that point, where it's a lot easier to do, just having a conversation versus doing the effort of research, you know, will this work? Will this not work? It may work a portion of that, what you're trying to accomplish. But what if you want to you know iterate that solution? So when you you know when you when you're building that out, you can add more stuff, and it's easier to do that than using someone else's stuff. So it it definitely changed the way you solve problems now because you can just perhaps just build it at that point.

SPEAKER_02 34:14

So within from again to go back to the AL development point of view, are there any things that you have come across that are more of like gotchas with AI that to look out for? Because it's uh we're we're talking that it simplifies the development, simplifies the publishing process.

SPEAKER_01 34:33

But if you come across anything that uh anything that so someone should look out for, I have mostly only have good experiences, I would say, uh, but I don't know if that is uh just my way of thought through working with AI that I have my way of keeping the reins in and steering where how I want it to do. Um for me it works just fine. And if not, I'll fix it. Uh so that's good.

SPEAKER_02 35:15

It's it's positive. I think as time has been progressing with these and these models have been improving, the results have been getting better. And uh you still can review it, you still have to review the code, you still go through the code, you still have the tests. Uh um I think tests are easier now as well because the tests deeper. Yes. And uh I spent uh I spend sometimes I spend more time looking at the tests just to make sure it's testing properly than I do with the code, if I should you know admit that I do review tests.

SPEAKER_01 35:50

Only almost. I rarely look at the code because well I do skim it still, but what I'm what I'm reading through is the tests. If those are the scenarios I wanted to work in and it succeeds, then I'm glad. Um that is well, depends on which kind of tasks they gave it. Mostly uh for the automated stuff, it is just still just simple tasks, it's no like new features. Um and with I have like there was a bug, and that bug is now fixed, and I can also see we have also added tests for the eight other test cases that was wasn't there before, and all of them succeed, and I can read the tests. I don't care about the code, uh then it's fine. But as long as it works. Uh code is cheap now, so if you need to add a new feature next week, so well, let's add that and fix it. As long as it's still performant, we have some checks for how well the code performs uh determine performance. And if those are fine as no regression, then that's fine with me. Um I don't care. But as I also like said at uh check taste. Um in the early days we did punch cards, then we did assembly, and then we did C. And in the beginning everyone looked how the compiler which kind of assembly it made. Now it's only the few guys that actually want to micro optimize their assembly code. Everyone else that writes C, well, they do a compile and just run it. They don't care about how their assembly looks. As long as it's performant, it's fine with them. And at some point the the our compiler is gonna be it's gonna be the same thing as long as it's good enough. We don't care how the Yale code looks.

SPEAKER_02 37:48

It's I'll ask you a qu i that's a a big point, as you had mentioned, is as the years progressed, we stopped worrying about the code, right? As we advanced. We went from punch cards, we worried about the punches, then we had assembly language, then we had uh languages that were written for human. And then you know we didn't look at the assembly language any longer. I d I did start off in college learning assembly language, by the way, uh, just to put that out there. I remember the registers and the pop and all that other stuff, uh vaguely, but uh but now we have languages. I'll ask you this question. I think we might have mentioned it on the podcast before. I've had this conversation with many people. Uh, if I asked you this before, it would just pretend and just put up to me being old and forgetting. Do you ever see a like we have AI models learning code from the languages that we know? And Chris knows the question. Do you think that it will ever change to be where we just work with AI and AI creates the quote executable with a code that runs on its own without us seeing the language at all? We just know that it works. It's almost where AI will just talk when work with AI using what it needs to be able to create output.

SPEAKER_01 39:14

I can't see why not. It's still gonna be for the next foreseeable future, we will still look at the inner workings of what it creates. But at some point when I tell it uh create something that give me uh the daily weather report each morning, I don't care how you do it. As long as I get my daily weather report, that's fine with me. And it's gonna be the same thing with our there's gonna be some things about some regulatory things about we have the thing about finance is that there are some things about how postings there has to be concurrent or there can't be any polls in like number series in certain countries and things like that where you mm I don't know again you might just add tests to make sure they are fine and then you don't need to look at the code. But there might be some places where you would still But again it's the same thing about now. There are still some few guys that still look at the assembly code. It's not gone, but it's like this many compared to all the other people that create assembly at some point down the line through the code they create now. Um we don't look at it at all. I think it's just gonna be a new abstraction layer up where it's gonna be a question and an answer, and we don't care how the answer was created, as long as it answers the correct question, then it's fine with us.

SPEAKER_02 40:46

Oh, to be somebody that from the past to fast forward to today, to see the technology and to see the coding or to see what's going on, it would be a major time warp for them. And it's uh it's it no one knows what tomorrow will bring, as they say. It does seem a lot of it has leveled out. I remember at first when all the AI models were released and introduced, you know, you go back to chat BT, Chat GPT was great, but then Anthropa came out, Microsoft has the their AI um harness. Um you have all of these uh companies. There was a big rush and big feel of pressure, like learn AI, learn this, learn that. To me, it feels like it's sort of leveled out. I mean, not saying you don't have to learn it, but the fast height pressure has seemed to slow down. Is that the same for you guys?

SPEAKER_01 41:42

I don't know if it's slowed down. Um I would say the advances in AI are still there, but it's the leaps are not that big anymore because the issue was they were so dumb, so just a bit better model, like a big increase of what it could do suddenly. Now they're so good they're all good. So it's like minuscule things. Now it's that better at solving PhD math in that area. Yeah, it's fine with me.

SPEAKER_00 42:13

The public that's a great question, Brad. The the public perception, perhaps I feel like has echoed um slower. Like um it it's it's not as big of a news, maybe for us that lives in it, or you know, still feel like there's so many, um there's so many still new features and things you can do. But from a public perspective, I do feel that it has it's slowed down a little bit. But again, I think a lot of them still feel like AI is just a chat. You know, it's all the people I spoke to, oh, it's just like I can ask all these questions and they're just getting into it, which is wild to me because we've been working on it for quite some time. Like I had a I had a my neighbor who was building a uh a shelf and he just identified chat GPT and how chat GPT gave him all this information he needs, how to measure and so forth. And still the chat, but the way we use it is very, very different, drastically different than what I've heard from many. So it certainly felt like it slowed down, but for us, uh I still feel like there's a whole lot more. Because then to your point earlier earlier, Turbin, so you have all these different tools. I think

MCP Setup Fast Iteration Tests First

SPEAKER_00 43:30

for my perspective in the future is that we can we don't care about the code, but it'll be, you know, how do you make them more efficient or how do you make it more uh how does it how can it be cheaper? And I think eventually AI is gonna talk to each other or different models can talk to each other and in and it's gonna have its own language to kind of compress the cost of tokens, hopefully. Uh, because we've already heard it in the past where two agents can have their own or made up their own language to be more efficient, to be faster. So I think that's gonna be the sort of like the next five years.

SPEAKER_01 44:02

But I think also we are in a different industry compared to the rest of the world. We are in our own bubble where we know if we don't if we don't uh evolve then we are gonna be behind in a year or half a year. The the partners that are still just using the free version of Chat TPG to just ask a question here and there they're gonna be gonna have a hard time in a year because there's gonna be a guy on the other side of the street that can create the same solution in half an hour when you need 10 hours. So you need to be aware and need to know what happens. Where the rest of the world well it's gonna be better at creating your German essay uh but that's like small things here and there.

SPEAKER_02 44:55

You you I do have another question about some models for you, but at the point that you're talking about it's it's you said the person across the street's going to be able to produce the the solution quicker than you. It's not only the the velocity at which output can be created it's the amount of output that could be created because before as you had mentioned nobody liked you know documentation. Now you need documentation in a sense for requirements so that the AI knows what to write. So even now if you do make a solution you can generate a document saying here's what it is, here's how it works. And again it's it's on the flip side that sometimes it creates a lot of documentation and you can't give somebody a 38 page document and say here read this this is about your solution but now you can deliver more so not only can someone deliver a solution faster for a customer again in our world they can give them more they can give them the automated tests to make sure that it works every time that they add something new to their system. It's not just your extension for in our world it's any other extension that gets installed you can run the test to make sure that all of those solutions still or extensions work together. And if one of them one of the tests fail, you can determine the course of action doesn't make mean either one of them is incorrect. It just means you may need to make an improvement to work with another extension whereas before everyone would just install it. Ah, this one task works then a week later they find out oh this doesn't work anymore. So you you're generous with a year from what I'm seeing I think if you haven't started doing that I think you're late at this point. It's because there's so much that's being produced by some of these partners but again I think it's a a balancing act of where you need to control what your output is don't say you know it's you don't have to tell everybody, oh, I'm using AI to create all this documentation. It's cool. It's almost like saying I use the keyboard to type my code but you need to give them the documentation that's relevant for the task that you're giving them and not overdoing it.

SPEAKER_01 47:03

Just to switch a little bit uh Chris you had mentioned cost or when we talked about cost and being able to switch to other models uh I listened to an interesting conversation on a podcast about the open source models uh open source models for uh cost management open source models for privacy or depending on how you want to look at it have you worked with any of the open source models and how do you see them fitting in a workflow today and in the future I haven't used them actively I've tested them uh like GLM and Kimi and so on and they're really good really really close at being like a replacement for Sonnet not for Opus but for the coding part and that is actually one of the next things I'm gonna add um on my website here in a month or so is that I'm gonna add a lot of local models both uh open models hosted and like OpenRouter but also local models that people can use by themselves and also see how good they are at coding and I'm I don't know yet how I'm gonna do this because some of my tasks still need reasoning so I don't know if I need to split my tasks even more so I have like coding tasks and like AL reasoning tasks even more work for me. Um but simply because again we are not using the same model for everything anymore. So I need to be able like what is the best model for coding in a performance price calculation and then Kimi X might actually be the best because you can do it just as well as Sonic to like a tenth of the price. Um but I haven't tested it yet in in real Alex programming it's only been in small tests here and there nothing concrete yet um but that is also because when I started doing these benchmarks and like just in the start of the year and compared to now the open models are way better. Uh start of the year it was like that we had was it the DeepSeek I think it was that was like the one that updated it all and now there's a lot of models that is good.

SPEAKER_02 49:36

So it's uh I'm interested to see if there's going to be a pivot to using local models for some things even though I talked with someone about this and they said well you need a lot of hardware to run it. My thought maybe even if it takes a little bit longer to run, if you don't have that added cost, okay. You know if it takes three hours instead of one hour or you know four hours instead of one does it really matter?

SPEAKER_01 50:05

For mine like my pipeline I don't care. Okay it can work while I sleep. Yes. If I set a task and instead of it saying eight hours it takes two days it's fine with me. That's probably because there's gotta be 20 agents working anyway so I don't have time to review them all in eight hours. So it's gonna be fine.

SPEAKER_02 50:27

Yes I like that I like the way that you think of it because I think that's what ends up happening with it as well is it doesn't matter if it's done in an hour. It would have taken two weeks anyway and if you can reduce the cost by using a local model for some tasks. At this point it almost goes back to what we talked about before with having that advisor to be able to manage and orchestrate okay use fable uh maybe use Opus to come up with a spec I'm just saying these hypothetically not saying this is what people should do. Everyone should review their workflow use Opus to come up with a spec fable to review the spec and then offload the development to like you said you could do to Kimmy, Deep Seek or one of the other ones and then turn that back to say okay we'll now have another one of the models review the code to how it matches a spec. And that way you can minimize the token costs. I really do think that that's where the uh what that's where there will be some emphasis of what we'll see because with Microsoft going with the per use for GitHub copilot or for copilot and a lot of the different copilot you know they have how many different copilots they have 128 or something yeah uh it it just it shows everybody that that day is coming and I'm sure anthropic will go there.

SPEAKER_01 51:50

I mean right now you're paying to H2 it's I feel they're just waiting for the IPO and they won't do anything now to mess it up and just for now just keep a lot of people get them hooked on the anthropic models and then when they're public then switch to API users by one.

SPEAKER_02 52:10

I agree with you because once you have that IPO then you have a fiduciary responsibility to in essence generate revenue. It's it's not hundred percent you can still have you can still have funding you can still have people seeding but once you have that IPO uh I mean people aren't looking at a company to take you know investment money and not have a return on that investment. So I I do think my prediction for 2027 will be you'll see a lot of these local models getting or these open source local models getting a little more attention for some tasks. And again to offload.

SPEAKER_01 52:46

As we talked about before as long as they're just good enough for their given tasks it doesn't matter. So we will pay anthropic a grand sum for the plan and then we have some kind of cheap local thing or free ish depending on what your electricity costs for doing the coding. And yeah I think that's how it's gonna be I I agree with you. It might be that local models will be so good in four years and we will all have our local opus home AI yeah that's what I think it will go with what you were talking about.

SPEAKER_02 53:25

I think it's going to come up with the harness if you can come up with a harness yourself and again not saying you have to develop but if we can come up if there's someone can come up with a harness to manage the use of these models because that's the difference. The models are the same across every harness it's how they work with it and how they feed it and how they um process the information back and forth. If you can have a local harness that you don't need to pay that you don't need to pay for in essence again you're going to pay for it with hardware you'll pay for it with something else. Maybe it will be good enough for you as you said not to say that everybody would need that use case. There may be someone who just is fine with the subscription to one of the the big what is it big four now uh and whatever you get in your subscription is adequate. Again Chris you talked about someone who may want to just use it for search to how do I do something or to hey rewrite my resume or something maybe those plans are sufficient for that. But for those that need to do actual not I'd say those that um those that need to spend a lot of time doing work because you had said like eight hour uh workflow or 16 hour workflow you'll definitely burn that up in a minute. So oh yeah it's uh it's a strange and scary world. Well Tobin said thank you very much for taking the time to speak with us again. I could talk with you about this stuff all day I think I'll have to you know just call you and just sit and talk while we work on some of the stuff just so I can hear about some of the things that you're doing. I see all the stuff that you share and I I sit back and I scratch my head and I say how does he have all the time to do all this and uh to like so you're doing some great things and really he was so quick he was so quick with the fable and I was like you know I was like how much did that cost you you know it's crazy I have the the big subscriptions so uh I think I had the uh benchmark ready like an hour after Fable was released I was like sitting there now it's ready let's benchmark it that's crazy I think I saw they pushed it out of July 12th so you think you have a few more days of yeah the issue was that they announced yesterday as the the cutoff so I was like uh using it all I could I think I'm at uh my Fable limit already so when they said like oh you're gonna have uh until the 12th that won't help me it won't reset until uh Friday so um so I can just sit

Open Models Token Costs And What’s Next

SPEAKER_02 55:52

here fiddle my fingers um I have more than one account so I get my fill that way uh which helps you have two dealers yes well you need a backup you know if if uh if that dealer is not available the dealer's not available if he's out of if he's out of supply or he's locked up uh you can go to the backup uh and then get some work done uh no but to uh thank you for uh the time that you spent with us today I know that uh you could have been doing some other uh great things with uh AI and for the business central community and you are doing some great things thank you again for central gauge I do keep up with it it's nice to see the models and the results it gives me a little insight of which ones I want to try uh to to complete some of the tasks that I'm working on. If anyone would like to contact you to see some of the great things that you're doing or if they had any questions about some of uh some of the neat things going with AI what's the best way to get in contact contact with you a little tongue tight there.

SPEAKER_01 56:45

Cat be on LinkedIn or uh on X then I'll reply back in a very short while um Torben lit uh or Shadows as I'm also known as my old gamertag so then we'll be able to reach out and I'll return immediately.

SPEAKER_02 57:04

Listen to the previous episode. Yes. No I'll have to uh uh make sure I communicate with you on X as well too. So it's my new favorite place to hang out. Uh thank you again I look forward to speaking with you soon and I'm interested to see what you and the team come up with next year for tech days as your entrance for the conference you you you you all have to do a session together.

SPEAKER_01 57:26

There might be something in the works for directions.

SPEAKER_02 57:28

So uh Amia in Paris. Nice I made a commitment to go under certain conditions. If those conditions are met I'm obligated to go. So we'll see if those come through and maybe you'll have to catch me over the pond.

SPEAKER_01 57:42

There might be a reason why why I'm switching my works laptop OS so um let's see what happens.

SPEAKER_02 57:51

Oh I have an idea if you I know the the the the the the I know the circle so I'm uh excited for that as well something that I think so far now so we can switch uh finally so yeah I have a lot of rumors a lot of rumors and uh I I I hope it gets there trust me I will be right there with you as a uh I know you you may not call it but as a an avid Mac user uh I wish I could do everything on there but to get everything on Linux would be great too. So thank you again look forward to speaking with you again soon likewise thank you Chris for your time for another episode of in the dynamics corner chair and thank you to our guests for participating thank you Brad for your time it is a wonderful episode of Dynamics Corner Chair.

SPEAKER_00 58:40

I would also like to thank our guests for join joining us thank you for all of our listeners tuning in as well you can find Brad at developerlife com that is d v l-e dot com and you can interact with them via Twitter D V L P R L I F E you can also find me at mattalino and my Twitter handle is Mattalino16 and see you can see those links down below in the show notes. Again thank you everyone thank you and take care