How'd you like to listen to dot net rocks with no ads?
Easy?
Become a patron For just five dollars a month you get access to a private RSS feed where all the shows have no ads. Twenty dollars a month will get you that and a special dot net Rocks patron mug. Sign up now at Patreon dot dot NetRocks dot com. Hey guess what it's dot net rocks insane heat.
Version in.
A little warm where you are, brother, one hundred and two degrees fahrenheit today.
That's no funny. That's dangerous being you'll be careful, it is dangerous. Thank God for air conditioning. Yeah, no kidding, Yeah.
Told, I've told neighbors who don't have it feel free to come over and cool off.
Yeah no, that's smart. Just get an hour cooling down? Right? Yeah?
Hey, Mark Miller's here. We're gonna have a lot of uh, some really good stuff that he's got to talk to us about. But you know, we're all we all seem to be on the agentic Ai bandwagon, or at least topic today because that's where my better know a framework is so oh.
Roll the crazy music. All right, man, tell me all about your agentic AI coding experience.
Yeah, all right, all right, Well, I haven't done a Blazer train in a while, and I figured this was the best time. So, you know, Build, we saw all sorts of stuff about the new features for developers that Microsoft is bringing out, right.
Hunter seth w O'Brian, you know, was don't stop.
And even the keynote kind of wet our appetite a little bit. But you know, our friends talk to us in terms of developers, which is great, and we all agree. I think that AI is a very real and helpful and safe thing for developers to use. You know, we have to be careful about safety, but for the most part, the tools that I've been using have been pretty good. So I came home from Build and I immediately started
messing around with this stuff. And I started working with the agent mode in visual Studio and that didn't work out so well, only because it kind of got hung up on files, couldn't access files that it had written and all that kind of stuff, But it did it was able to generate some good stuff anyway. Then I'm
at devsum in Stockholm. I met the rooftop bar the hotel and Hunter is there and I'm just kind of picking his brain and he's like, oh no, you got to use the copilot GitHub copilot coding agent in GitHub, right, that's what it's all. AB I just did this demo.
That's where the action is. Yeah. Right.
So I was like, Okay, here we go. I went home and I started playing with that, and I agree with Scott this is this is really really good. And so the whole idea is that the co pilot is a partner on your in the repos that you give it access to, and you can assign issues to it and it will go out and figure out those issues, and you can view the logs and view the session
and see what it's doing. And then first it creates a sort of a draft poll request, and then it creates commits, and you can view those commits you know, in visual studio, and test them out, and if it didn't do something you liked, you can just add a comment and it'll go off and fix it. So I and then of course when you're done, you can merge the poll request. So I thought this was just absolutely brilliant.
So I did a Blazer train about it. Episode one oh seven, get ub Copilot Coding Agent and we'll have a link to that. And also the first thing I did was I documented my experiences using it to create a whole Blazer server forms over data application using the north Wind database. Cool, and I use SQL like because you know, not everybody uses Windows. Yeah, and the video the one I did in the video kind of I kind of got stuck in the code at the end, but if you look at the results that i'd have
in the actual repo that we'll have a link to. Also, it was much better and didn't take you know, many iterations to fix the bugs and stuff.
But I do like that those fighting through moments, so I think are really useful teaching moments too. It's like when the software is not behaving the way you want, how do you push on it to get to get results?
Absolutely, and along the way, I learned about the system prompt MD file that you can put in your repo where you tell it sort of all the general stuff and how you want things to be organized in the things that the coding standards and.
All that stuff.
Yeah, overall was really good. And so there's a third part. I've created a four hour webinar called a Survey of AI Tools for Dot net Development, which goes beyond just the Microsoft tools but looks at the best of breed and AI tools across all platforms and services. So that's like a survey of these things. It'll include GitHub and all of that stuff. And guess what, it's even going to include code Rush, which I learned about today from Mark.
He's going to be talking about that. But that link is in the show notes as well, and also linked in the repo, which is in the show notes. So that's all the stuff that I have. Richard, who's talking.
To us today, grabbed a common top of Show nineteen fifty four, the one we did I had to build with w O'Brien where we talked about the play right MCP, which definitely one of those moments where like, oh man, we can get up to stuff with this. Yeah. Yeah, and John MacArthur had this great comedy said, listening to this and talk about copilot, I love using it for my get comments. It does such a good job of them,
and it's forcing me to get better. Oh no, upon help me, help me no better by checking in more often, because I don't want lots of unrelated changes grouped together. See yeah, Because the tool actually remembers everything you did and spits it all back to you, so you have to be more organized and unlike when my handwritten comment was fixed stuff and things. Yeah, yeah, seriously, the generated
comments are brilliant. Yeah, I gotta say, John, like, there's more that can be done with these tools, Like I appreciate your sentiment and certainly I like those parts too. But keep going down the path and we're just getting started. Yeah, it's early days, isn't it. Yeah, and John, thank you so much. Your comment and copy of music code By is on its way to you. And if you'd like a copy of music codbe I write a comment on the website at dot netrock com or on the facebooks
we publish every show there. Any of you comment there and I read it on the show, we'll send you copy music. O.
I just sold another flak collection of music to code By. We're up to twenty two tracks now and you can get them an MP three flak and wave music to code by dot net. All right, before we bring on mister Miller, let's talk about nineteen fifty eight, because that's our show number.
Ninety fifty eight, A very good year in some respects. I mean, you know.
Oh yeah, So where do you want to go with it? Well, let's talk about international developments. The European Economic Community was formed, which is precursor to the EU. Precursor to EU. The West Indies Federation was formed, the Soviet American Exchange Agreement. Can you imagine? Yeah, Brazil won the World Cup, defeated Sweden five to two to win their first World Cup title. The first undersea voyage to the North Pole by the USS Nautilus. Hey, I know about that submarine.
Yeah, that's that's your part of the world.
Yeah.
It was built in my hometown. Yeah, and I'm sure my father probably had something to do with that because he was in planning, probably in the United States.
NASA was established on July twenty nine. Yeah, the International Geophysical Year. Yeah, established with NASA IGY, the Bunch of Satellites. IGY is also the first track on Donald Fagan's excellent The Nightfly, which is all about fifties fantasies about the future. Sure Explorer one was launched on January thirty first detected the Van Allen BELTZ. You know all these things. It's so awesome. I do Federal Aviation Act. If you want to just forget school. Just go have a cup of
tea or glass whiskey with Richard. You don't need school.
Or if you've been in school and you want to feel like you've never been in school, do the.
Same, right, listen to a geek out.
The Federal Aviation Act was passed, which was later the FAA Federal Aviation Agency, giving it authority overall aspect of aviation in the US. Elvis was inducted into the Army on March twenty fourth, the B forty seven nuclear weapon loss incident. A B forty seven bomber accidentally dropped an atomic bomb on Mars Bluff, South Carolina.
Oops, well that's the one that went off, but we had no nuclear core. That same year they lost one out of a B forty seven and have never found it, Oh my god, somewhere.
Some other things, the first radio broadcast from Space, President Dwight D. Eisenhowers and a Christmas message from Space.
Yeah, first communication satellite too.
Yeah. Anything that you want to add in terms of computers or technology.
Well, let me do one non computer one, and I'll do a couple computer ones. On computer one And this is from my friend Mark is the first regular surve transatlantic service of the Boeing seven oh seven. They really beginning of the jet age. That the prototype flew in the end of fifty seven, but it was already in service in fifty eight eight pan am in the beginning of what they called the jet age. It wasn't the
first jet air liner. There was ones before that, but at the seven o seven, and that form factor still flies today, that four engine rig that's the KC one thirty five. Like that plane's never going away, it seems, which is a good one.
It was also mentioned in the song jet air Liner by Steve.
Miller band down Rights. They get on the seven o seven, get on the seven oh seven. There's lots of variance of it. Two. You know, the original form of that our craft was quite narrow, is a two x two seating, and then they widened it to a two to one two seating, and then they widened it again into a three three seating. Programming wise, the remarkable extraordinary found. John McCarthy at MIT developed the LISP programming language in two
fifty eight. It wouldn't get deployed for another couple of years because we barely had computers at that point. LISPS short for the list processing. But yeah, he had defined the language already. At that point, you barely had a computer functional Interestingly.
The functional language that was used a lot in was it used in AI at the time or what they called AI.
Yeah, the term AI had already been coined at that point. McCarthy or Minsky coined the term.
But I don't know what they meant by that back then. Was it machine learning kind of thing or.
No, it was just that was more decision tree modeling. What what Minsky got financing from the military for was to do resource allocation and management for the US military, which by the way, worked perfectly.
So expert systems that kind of thing.
No, just resource management, okay, like decision treat mesia management. How do you how do you move a battalion? What do you got to move first? Like what those pieces look like based on the aircraft you got to the ships? You're using that kind of thing. Okay, One more computer one the RC five oh one, which was one of the very first computers only nineteen fifty eight to use
transistors instead of vacuum two hm. No, discrete, it's all discrete electronics at this point, right, we haven't got ICs yet, but we are beginning down the transistorization of computing. Wow.
You know, it just occurs to me when we have a show number that's the same year as the year we recorded in the world's gonna explode.
I know, I'm with you. It's I think we're doing two thousand and two is going to be a problem.
Yeah, all right, Well, let's introduce Mark Miller. He's been on the show many times. He is fondly remembered in dot Netrock's history, of course through Monday's hilarious bits on there, and also just in early dot net Rocks. He's been an innovator in many, many ways. So he's a multi year c Sharp Microsoft MVP and a leading expert on
user interface design. He's a chief architect of the ide Tools division at Developer Express, and he streams live c Sharp coding and design on Twitch, dot tv, slash code rushed. Mark has been creating tools for software developers for over four decades. He was also the He was the guest on the first show that I was co host on show one on one.
Wow.
Really that's cool.
Yeah, I did not know that. I feel like I must have. I feel like after the show, Carl MSSA got a call for Richard. Richard's like, there's no way I'm doing this never again.
I think I think I got off that show and said only fifty shows. Yeah, that's right. I mean, how hard could it be? Fifty show fifty and.
Here we are the longest running dot net podcast ever and probably one of the longest running podcasts ever that's still going.
Yeah. Probably, Yeah, it's hard to know.
That's interesting. Yeah, that's right.
We are the og podcast for better or worse.
Congratulations, guys, I forget Why don't we switch it up. I'll interview both of you and we'll talk about that.
Hey, all right, I don't think dang.
Okay, So this morning, Mark Martin my time. You gave me a demo of your new code Rush tools that use AI and speech recognition, and yeah, wow, I was just blown away. So maybe you can tell everybody about this give us the elevator pitch.
Okay, sure. It's essentially it's kind of like as if AI was kind of like a suit of armor for you as a developer. It is unlike the agent mode of Copilot, where it kind of goes off in space. You kind of give it some boundaries and say here, this is the issue I want you to go after. The tools that we have in code Rush are really much more much closer to you as a developer. It's kind of like, you know, leveling up in a game or something along those lines. Your power and speed increase.
The essence of the tooling is making changes to code, creating new code, and also doing essentially semantic searches through code.
Those are the areas so that you can you can use the power of AI to essentially offload a lot of cognitive effort, right, and so anytime that I've got something that is tedious would otherwise be tedious to do, or or if i have something that maybe I'm not sure how to do right, both of these are great, great places for AI to come step in and and and the what makes them feel like a suit of armor is, I think is the is the ease of
use the user interface. Right. By by going after this in my usual approach right to doing creating any kind of tooling, you know, looking for any kind of frictional force, right, any kind of bump and trying to smooth out every bump, we've gotten to this place where you can essentially do what they're calling, you know, vibe coding to some degree, right, but you can do it I as I need it, right when I need it, I can say okay, you know, as an example, I think if I showed Carl this morning,
I said something like, let's create a new class called customer with a first name, last name, an email address, a street address, and a birth date. And I can just do that. I don't even have to say the names of the types. I can just say here's what I want to do. And if I want to specify things, I can. I can go into more detail. But if I don't, and I often don't, and this is what really lowers the cognitive load. Right as I'm using the tooling, I say, hey can we fix this, or hey can
you add the method that's missing? Right can I? And I love talking in clues to AI because it often figures it out and puts in the thing that I need, whatever it is, and then I can just go from there. I can start writing the code or doing what I need to do.
You know, almost exactly one hundred shows ago eighteen fifty we were talking to you about open AI. This is in twenty twenty three, and you hinted at this whole idea of Hey, I talk about my code when I code anyway. Yeah, Like, why wouldn't I be saying it. Wouldn't it be nice if it was listening? Like this sounds like you just built what you were talking about.
Yeah, this is when we were upstairs at the guard you and me Mark sitting across from each other in the attic there where we had the booths, but in twenty twenty three. But I think the I wanted to say that. The one of the first impressions that I got was how out of the way the speech thing was like, I'm always nervous that if I hit the wrong key and then it starts listening and it's going
to create a bunch of garbage. But you've got it so that the right control key by default you tap it once and then you hold it down and speak right. It's not like you can just hold down control because you know what, sometimes you need it as a control key. You want to hold down control and you see or V and so I really appreciate the fact that you came up with like an action that's deliberate that you
would not do otherwise. Yeah, as opposed to like, you know, the agent that just when you type a few characters, says hey, how about this in a page of grade out code?
You know, yeah, no, I think we have solved a lot of problems that would otherwise be blocking forces to making this vision come true. Right, and you know the the you know what you just said about the keyboard. You know, I've been using this, this this feature has been in development. I've been using it actively for probably at least seven months. I want to say. Yeah, and so I've forgotten how kind of unique that solution is. Yeah,
And it wasn't until you just mentioned that. I was like, oh, right, yeah, we've made this. There's a couple of things that happen, right, But basically we said, look, there's on programmer keyboards there's no push to talk button, right, yeah, weird. So we have to with something and we can't make it be something where you have to reach the mouse, and we can't make it for something where you have to hold down the control and yet another key. We can't. It's it's got to be effortless to get in.
And so we I thought it was a brilliant solution.
Well thanks, yeah, no, And and so that's that's part of what we're doing is we're making things effortless, right, And there's actually about four different kind of cross section criterias for success that I think we've done really well in that we've kind of leap frog other existing technology, and one of them is ease of use. On the ease of use side of things, it is effortless and it is also fast. Yes, right when you when you say I want to do this, and then you release
the control key because you're done talking. On average, our our voice to text conversion times is less than a half a second, so within a half a second of that release, we know what you're asking for, and at that point we throw it up to open end. Right, and we also include with that a lot of context and a lot of detail that opening eye can use to figure out what you mean.
Let's talk about context, okay, sure, Yeah. So one of the things I like about chat GPT and I pay for the premium version, is that it has it remembers a lot about you know, it knows that I primarily work in Blazer. It knows, so therefore when I say I got this new application and I got this file here,
it knows it's a Blazer app. Nless I tell it otherwise. Yeah, so it has a history and a lot of these agents, and I know the Microsoft Copilot agent does not have this memory, and I think you said that Code Rush doesn't either. But it does understand the context of your application, right.
Yeah, yes, it does, so just briefly, no history yet, it's stateless right now, right, But we're leaning in a direction. We're considering a move towards a essentially a summarized history in case you want to reference something you've done before.
But that takes more tokens and costs more money and blah blah, so they have to be an option, right.
Well, yeah, well that's why I'm saying summarized. In other words, brief really a brief history, but so that we can get context. So let's just I just want to say that this version that's coming out effectively, I think tomorrow from our recording date, so should already be out by
the time you're there from the Code Rush. This version is what I call one point zero, but there is already a strongly visualized two point zero version of this that we are going to commence work on essentially tomorrow that is going to take things to a whole new level of what I think is you know, if you want to call it the mind blowing scale, we're going to go from this version I call this a five, and we're going to go to a ten. Yeah, you know, we're going to get to this killer level and we're
going to solve this. But so briefly, the history right now, it's stateless, yeah okay, but with regards to context, the context is rich. We send, for example, a summary of what's going on with the project, what your new get packages are, that sort of thing. We send the current file, we send code behind or the designer if you're in
one of those two files, right. You can also do things like say, if you mentioned the word clipboard, there's a something called triggered prompts, and it detects for that word click clipboard and then adds an additional prompt that sets out and says the contents so my clipboard are and shows what they are.
And I like that because it's contextual, like a system prompt. Even in GitHub Copilot is copilot, dash instructions, dot MD, and I think you put in a dot GitHub folder if my memory serves and it's general, it gets sent with every single you know, user yeah, but I like the way that you have broken these out into contextuals based on keywords and things.
Yeah, and no, we did have by the way, we worked on a feature that allowed me to like hold down the alt key and say the name of a class and then release it. But we are not documenting that. We're essentially not gonna We're not going forward with that feature. We're instead going forward with a you can talk about any class you want and we'll figure it out. That's
not now, that'll be in version two. Nice, So version in version two you'll be able to effortlessly, AI will be able to infer what you mean, what you're talking about, which will impact the context that we submit to AI. So so we're going to solve all this in the two point zero I guess I shouldn't be talking about this so much, but but what we do now is
a relatively rich context. So if I want to work with another file and I want to work with it in this file, like say, for example, I'm a Zambal designer and I want to create a data grid and I want to do data binding on a particular control class, I can copy that class to the clipboard and then say, hey, I want to create a data grid here, and I wanted to have instances of this customer class, and I want in the code behind for you to create fifty
sample records that have realistic looking data sample data so I can try this out and work on the UI. And then that's my problem.
I saw this. I saw this myself, that you did that. And I want to talk about the speed too, because in the demo that I did and blazer Train, I basically created these long prompts. I could create a one prompt to create a data manager for every model and the models folder, which there was like twelve or thirteen, right. I did it against north Wind, and I said, create a data manager, and I want creud methods for each one of these using ado dot net objects and blah
blah blah. Trycatch and here's some return class types. Right, success, failure the data and then error messages. That kind of simple thing. Yeah, and you it basically took like fifteen minutes for it to do that and come back and say, yeah, here's your you know, here's the here's the change set that I'm going to check in or that that you
can check out. But I like the way that code rush works because if I was going to do that, I would create the data manage your class skeleton, or I would tell it to and then I would say, hey, let's create cred methods for the album or the artist or the customer or whatever, and just inspect those because that would take like no time at all for code Rush to do. I have a feeling, yeah, and then and then from there I would tweak that and say, okay,
now let's duplicate this for all of the other models. Sure, which is something I wouldn't have had a chance to do if I because I used a big prompt like that.
Right. Yeah.
I actually saw an example today where I was in one class that had implemented I property notified change. I asked for a new class, just saying here's the properties I need, and it because I started in a class had this style it built it followed the same style in the second class, which is interesting.
Nice.
But with regards to speed, we have totally solved the speed problem, like really really solved it well. If you compare the code rush aigen feature against leading competing AI tools, it can be up to ten or one hundred times faster than the other tools, depending on what you're doing and what's going on. It is and at the very least it is a little bit faster, right, it's about five percent faster if anything.
Else, so five to one hundred times faster, five percent to one hundred times.
No, it's like five to ten thousand percent faster. It depends on what you're looking at and what you're doing. But in some cases, and basically what the I guess the secret the big reveal there is the leading tools that are out there have not addressed performance in terms of the integration of AI results with code. They've not. It seems pretty clear to me at this point that in the current versions that I'm looking at, they that
is not a concern. That's not something that looked at or not even something that they've conceived of that they could actually do faster. Wow, And that is that's you know, when I was talking about these different criteria, these different cross sections for how you evaluate something.
Speed is one of those Yes, yeah.
And from and why And I'll give everybody the big secret here. I'll tell you what the secret, sauce is. How we're able to do it is that we make it so that when AI comes back, it doesn't have to be flowery. It doesn't have to give me the whole class. It can just give me the changes. And because it's just giving me the changes, I'm sending back fewer tokens, which means it's cheaper, it's faster, and it has a lower environmental impact.
Because your code rush, you know exactly where the code should go when it comes back. Yeah, so you do that part of it.
Yeah, we do all that part.
So, yeah, you do everything from visual studio.
Yeah. So the agent technology that's inside code rush is essentially an expert at taking the response back from AI and immediately folding it back in the But the speed benefits are really all gained in conversation with AI. That's where they are. And nobody's doing this yet because you can tell because they're so slow.
Yeah, just what are you truly changing the workflow? You're just speeding up my existing workflow.
Yeah, I'm speeding up your existing workflow.
Yeah, which is what you've always done.
I'm changing it in a way and that now you can ask more effortlessly and get the answer and see the results more quickly. And so because that offer is now available, you lean into that offer more frequently. I think and so there's kind of this natural change where you're like, oh, they this tool is pretty cool. Let me try this. I'm going to slash, you know, try to slash away at this problem.
Right, But this also seems like not an all in like do everything through Agentic that you're still doing your own coding. There some hints from these tools.
Yes, yeah, it's very much in that space.
Yeah.
Like, often I'm like just writing code normally, and then i might be in a place where I'm like, okay, wait, I don't know how to do this. This involves some new framework or new new thing I've never touched before, and I just ask, right, and I just get it and I take a look at it. And often I'm delighted because I'm learning at the same time when I
see it come in. Sometimes I ask for something that's like a tedious thing, and I'm delighted because I see a new way of doing it that I didn't know before, for example, which is also kind of a cool learning experience.
Yeah. Yeah, but I also could see, like with the Agenic thing, there's sort of scut work coding that you could do, but it's easier just to hand it to the tool. Let it generate.
Yeah, hey, this sounds like a good place to take a break. So we'll be right back after these very important messages. And as a reminder, if you don't want to hear these messages, you can opt for an ad free feed by becoming a patron. Got a patreon, dot dot nerocks dot com. We'll be right back. You know, dot net six has officially reached the end of support, and now is the time to upgrade. Dot Net eight is well supported on aws. Learn more at aws dot
Amazon dot com, slash dot net. All right, we're back. It's dot net rocks. I'm Carl Franklin. That's Richard Campbell, hey, and that's Mark Mellers Mela, Mark Miller. All right, all right, so I have another question. Is you support what project types now and what project types will you be supporting in the future.
Okay, so that's a great question. It's our In terms of project types, it's basically, if you can hold onto c sharp, we support that project type. So we support c sharp, and we support right now is examal and so all the various flavors of that. Right so that if you're working or just WPF or whatever way you can work in that conver that that combination. If you're working in other UI frameworks that are not supported, but you're still working in c sharp, we've got to cover on the c sharp.
Side of it, right, But for now, you wouldn't be able to say, hey, build me a form in Blazer that calls into these methods.
Right, No, we cannot do that. Now. Now we're we are close again, and the two version of this, the two version, a couple of things are coming into A one is Blazer support will be coming great. And a second thing that we have there's a current limitation is the only way we can create new files right now is by creating a new class.
Okay.
So in other words, there's no way for us to have a conversation with AI where a AI comes back and says, here's a new dot Razor file and here's a new Dot Sammle file that doesn't that's currently not happening. It's not hard to get there. But we wanted to ship, and so we're shipping what we have now. Okay makes sense?
Yeah, I wrote an agent myself in WPF that used the open AI and I got the response back and parsed out all the comments. So that I just got the code and then I could tell it to you know, write code files. And it does, but I have to use I basically inject into the thing that does that, compile it to Roslin and inject in system io and things like that, so it can actually write files. But yeah, it takes the result that comes back, compiles it with Roslin after you inspect it, of course, and that can
also create files and things. Yeah, but it was a fun exercise, but it's nothing compared to what the actual real tools do.
Yeah.
So when we're talking about that, right, the inspection, right, and this idea of user interface, right, one of the things that happens. I think in a lot of places, a lot of tooling will give you that inspection phase in it beforehand in another window. And our approach instead is let's make the code changes, inject them everywhere, but give you a really high quality tool for navigating among
all those changes, navigating through all those changes. And I think that the latter approach is certainly better for.
This this feeling.
Of I'm going to I'm going to go after this, I'm speaking with intensing what I want, so why not just give it to you faster by putting it right there in the code. Right, If you think about it, if I have to go to another window and I'm looking at code, sometimes I'm looking at code in that other window and talking about other tools. I'm looking at code.
That is where a small change is among a giant reproduction of all the code I had, right, and I have to scroll up and down, visually scan it, then copy it, maybe copy a feel and go paste it in. Maybe it's in sections and I have to go copy one and paste it in, copy another, paste it in. Or maybe there's a button I can click and it puts it in. But still I got to click and wait and then go to the other one, click and wait and do that. And there's this whole thing where
you're not in the editor, right. And and what I'm trying to do here is we're you know, with all of the tooling that we're trying to, you know, create is make it so that look, we are not leaving the editor, right the editors where we stay. The editor is our world. That's the world we you know, we slash and build in you know, we create, and we go after the tasks and the challenges there in the editor.
So there's just no way, we're leaving that space. And as a result, it's more effortless, the cognitive load is lower, and as and the other things.
You have this.
Oh, I just I wanted something that was tedious and I and now I have it. Well, I'm ready to you know, I'm still refreshed. I'm ready to keep going.
Yeah right, yeah.
So one of the problems that I have with the open AI for code and even the co pilot, the getthub copilot coding agent, is that it's nondeterministic. So I give if I gave it the same prompt twice the second time, I may come back with you know, I
had this problem with it creating a data manager. It decided to break that out into three partial classes, which wouldn't be bad except that one of two of the partial classes had the same method in it, and you know, I had to specifically tell it, don't use partial classes, put in everything in one class, and don't have any duplicate methods. I had to tell it that.
Yeah, yeah, you know, yeah, yeah, I was just gonna say, well, when you first said it's non deterministic, I'm like, welcome to AI. And then as you were talking, I was thinking, well, you can kind of control it, corral it in a bit, right,
you could do that. I actually find that one of the things that's useful is if I see like this kind of variance and it goes against my style, I kind of put in like these what we call the system kind of system proplems add ons to the prompt, right, right, so that no matter what it'll say, hey, I always like to use the latest fluent assertions or something like that in my test case.
Yes, yes, right, I guess what I'm saying is.
I just those are those are ways that you can kind of get to there's you know, it's it's still not deterministic, but you're getting closer, right, You're getting.
Yeah, getting closer. Yeah, you're tweaking it. I found that the coding agent like to downgrade to from dot net nine to dot net eight constantly.
Yes, yes, I've seen that before too.
It's in my system prompt and it's in my user prompt, and then it still doesn't and I say, do not downgrade you know, oh okay, oh yeah, yeah.
So this is really interesting. So one of the things I've discovered when you're working with working with code that's being generated off of sources that have changed over time. Right, classic example is end unit test cases. Right. They used to be able to say a cert dot is true, and now they essentially broke that. And now you've got to say, now the code you often wanted to be a cert dot that. Right, So if you have essentially breaking changes or code, that is the way you do
it has changed over time. And AI is saying, well, I'm scanning everything for the answer, right, then you're gonna you're gonna get this kind of variance more often. And the the more either edge case it is, in other words, the fewer examples there are, or the more that they've broken it over time, right, the more this variance occurs. Right, And you really, you really, you really start to need
to corral things in with these very explicit pieces. But what's interesting is that, like one of the things that we had to do to build this right, Essentially, one of the things that our agent does is our agent allows for variance to come back from AI and it gets it back in the right place. So so we've got like a ton of code that and test cases
that are all based around this idea of defense against hallucinations. Right, sure, and so when a hallucination comes like it's we as soon as we spot it, it's like gold to us, right yeah, we're like you grab it and we have. One of the things that's really cool about our implementation is that for every interaction you do, you do there's a log file that's created.
Yes, I saw these amazing.
So a couple of things that are beneficial about this one is that log file is beneficial if you're ever going to exploit the the ind nification against copyright which is offered by AI tooling right writ shout GBT has something along the lines.
Because you have a log for the prompt, you have a log for the return, yes, and then you have other logs, but they're individual logs yes.
And therefore, so for each request you make, there are essentially four logs that are generated and and that gives you a fleet verifiable audit trail and audit auditible trail in terms of what how I have used ai UH in this application and how I've integrated the One of those four files that we create has a dot test extension, and that dot test extension is a test case that if you had problems or something didn't work the way you were expecting you contact support as it Support might
ask for the test case. You could send it, and that test case would allow us to reproduce that scenario. And that's how we've grabbed onto these hallucinations so so fast and quickly. Right when we when we see a problem, it's gold because we grab that test case and we go put it in and then we start looking at it. How do we you know, what's what? We look at it? We analyze what are the chances that hallucination happens again? Right?
How do we what defensive code can we add? And what's interesting is we have a very specific spec that we deliver open A but we have a very wide tolerance for what open AI can come back with. Right.
What we tolerate is would be in normal scenarios would because I think a lot of people would almost consider that's ridiculous in terms of how much variance or or you know from the spec you're allowing for, right, you know, but we were basically allowing for open eye to respond back with things that are almost completely wrong and then we fix them.
So there is a temperature setting right when you when you issue an API command.
Yes, we're not touching that or playing with that at all.
You're using the default. So what is the default?
Like a very low temperature, so low tolerance for hallucinations, like zero.
Yeah, I understanding is that I don't know if it's zero or not, but I'm understanding it's pretty low.
It's pretty low.
Yeah, yeah, yeah, and it might be even lower when you're doing code related requests.
I imagine it should be. I'm also thinking about what Carl talked about with the system MD, like shouldn't have be this be stuff that's basically a part of every prompt every time, like resist hallucination, make sure it's accurate. I keep reading about folks doing agenda GAI, where it's like, make sure you do every step. You know, you cannot proceed to the next thing without showing proof of this. Building more testing into each step so that it actually shows that the test passed.
What I think I hear Mark saying is that we're foregoing most of that in lieu of taking the stuff that comes back from it and using our own smarts to discern what is real and what's not.
Yeah. And also we're doing things a step at a time, and each step is directed as a developer directed action right, So the developer might say I need some test cases. And then the developer might say, okay, I want to All these test cases have got hard coded numbers inside of them, and I want to change that to be parameters to the test case and use the test case attribute. So convert all these test cases to me the more flexible, do that for me. And so that's the second step
or whatever. And so the developers are doing these steps one at a time in our model, and so the developer essentially approves after each one. Right now. One of the piece is the things you mentioned there that I really like that we do not have yet, but I think some form of this is likely to come in the two point zero and that is the ability to validate or verify by looking at, for example, compiler errors
after a build like before and after a build. Right, compiler always gets to say before and after a change, I mean not a build. Yeah, So I think that's interesting. That's a that's a really interesting piece. Giving giving that, including that as part of context, and make it triggerable. So if I use the word error, or I say the word fix, can you fix this something along those lines, it will say, well, what's going on.
In that list of triggered prompts that you showed me where you have like clipboard and then a prompt that has is there one? Can you set just a general system prompt in there that goes with every single request?
Yes, each one of those prompts has a regular expression that triggers it, and the regular expression essentially searching that text, that speech to text translation that came back.
So you get just to a wild card everything.
Yeah, you dude, dot star for the for the regular expression to match, and it'll match everything. But even with that, you can specify an additional code rush context, which could say something like that, make this happen only in see sharp files. Yes, we'll make this happen only when i'm uh, when there's a selection, or only when I'm debugging. Right, So you can set different contexts, rich control over that.
I love that. I love both, and that all.
Happens automatically, so you don't need to say any of those things. Right.
I like both.
I like having a general system prompt, and I also like having a context based triggered prompt. Yeah yeah, yeah, that's very very smart. Wow, I don't know where to go? Where else where else can we go besides Austin here.
I mean, I'm blown away, right. You know, there is one thing that became a parent when we were talking about this, is that the way that I did it in my Blazer train thing was to create a big prompt, a big giant prompt, right, and then you know, like creating a data manager for example, that's got you know,
fifty sixty methods. You know, we got cred methods for idio dot net for each of these models, let's say, and it takes you know, fifteen minutes, and then I can check it out and one at a time tweak things.
Right.
Well, that's a different way to approach. And if it doesn't work out, you can just blow the issue away and start again, right, Yeah, close the issue. But that's a different approach than what you're doing. What you're doing is, like you said, very close to the code, right, it's in the code. You're probably not doing big things like that. You might say, you know, do some data binding over this particular model or something like that in WPF and later in Blazer, but you would probably do them one
at a time and in a more controlled way. It's kind of what I'm thinking. But it's two different ways to approach you.
Well, well, I do like I I kind of approach by groups. So like for data binding, for example, I would say, well, like, I want to I want you I've got a class on the clipboard. I want to do data binding.
All.
I want you to find all the controls on the form already and bind them to the appropriate properties in this class. So I'll do that. Or I'm like say, hey, I want you to take all the buttons on this form that don't have a style, and I want you to add a new style to that, and I want you to set the default font for that new style to be twenty four. So I will kind of group things together.
And I'll digestible pieces, yes, and.
I'll do those all at once check the results. Sometimes it's a run to check those results and stepping through the code maybe, and sometimes it's visually scanning the code and seeing what's going on.
And so what happens if it does something that you don't like?
I mean, well, you've got so that's a great question. I love that if you ask, because this is this is actually.
Another Hey have we met? This is my freaking job, man, This is I don't know that you should have a podcast or something. This guy's good.
So so yeah, this is a This is actually great because again this is a great point of differentiation. If you look at other tooling, other tooling makes changes in multiple places and requires multiple undos right to do that, whereas the code Rush aigen does everything, wraps everything up in a single undue unit, so you can undo, right,
go back, or go forward again. And when you go forward, that navigator I talked about that really high level you know, ability to look around, to go from changed to change, and to see the differences. That comes back up on a redo. So if you undoing then redo, you can get that navigator back up and scan the differences inside if you need.
To do that. Yeah, that's cool, so so.
And it's fast, right. The main way time is waiting for AI to give you that composition that responds back. That's your main once you do that, for code Rush to make many changes through many files is generally less than a second. It doesn't, it doesn't take that long.
So I'm also thinking in terms of how much experience you need as a software developer to keep able to use these tools successfully.
Well, you cannot, I think you cannot be a brand new developer.
Yeah, I agree, Right, you can't, like you don't know when the tools misbehaving.
Well, you could, you can get I haven't seen scenarios where you ask for something and there's just not a lot of content out there in terms of what you're asking for, and so what you get is something that maybe doesn't work, or maybe it's on a you know, on a framework that's been changing over time, So you get old code that doesn't work or something like that. And if you don't know anything about that, you don't know how to solve it. You're stuck.
Here's a great example, and I brought this up in a previous show. Mark you'll like this. Yeah, tell the AI to create a collection class that is thread safe. Now, if you didn't know, there's already a thread safe collection in dot net, right, it's going off and it's not going to know that, and it's going off and creating right. Right, you didn't say, hey, is there a thread safe collection in dot net? You just said, create me a thread safe collection object, right, a junior divisse.
It's going to do what you were saying.
Yeah, it's going to do.
Exactly what you say. And now you're committed to that thing and you use that everywhere and oops, there's a problem with it, right because it's not debugged and tested and all that or whatever.
Yeah, yeah, I think that I think you have to be you know, I think you have to be just kind of if you're going to if the code you're going to generate and the things you're going to build are are kind of, you know, within the norm whatever that is of development in other words, the things of which we have a lot of content that AI has studied, then I think that your user level needs to be
maybe below you know, intermediah. You can be. You don't have to, you can be somewhere between beginner and intermedia and you can benefit And there are some ways where you can benefit really hugely from this as a beginner, when you say I want a new class that does this, this and this, and then you've got it and you're done.
There's it's so much easier. And and right now we're only supporting c sharp, but when we start supporting other languages, the ability to now code in another language becomes transferable because I'm still using the exact same spoken English right to interact with it.
But getting back to my getting back to my thing, my example of the thread safe collection. Yes, you probably is. It's a good idea, even for more advanced developers to first have a consultation with it about what the best way to go ahead implementing such a feature would be.
Right, yeah, no, I agree. And that's where this this feature is not really a consultation feature. Yeah, you can ask it questions and it'll come back then and the navigator and give you the answer. But it's not interactive.
But at chat GPT would be good for that, right.
Yes, I will switch over to chat GPT for that kind of a style of dialogue.
For that kind of consultation.
Yeah, yes, yeah, I'll do that in chat GPT. I use Copilot for other things as well. Copilot's great, great at like hey, take a look at this error or something along those lines. Sometimes Copilot's great for just these kinds of questions. Because it's with code rush, you can actually talk to copilot using voice. Yeah so, and I didn't tell you this this morning, but if you hold down the left control key, just hold it down and start your your prompt with hey co pilot, oh, and
then ask the question. It will open up the copilot window fill in the prompt with your question and all you have to do is just you know, hit enter, okay and send it up.
Wow.
So I think we're in agreement in that we're and I brought this up in our talk with Scott Hunter and also offline with Hunter, and.
That the.
As a developer, there's just as much value on your creativity and imagination now as there is on your ability to write code and the things that you remember how to do, maybe even more like you have to be able to imagine things and in order to create them. Whereas you know, for the last fifty years of my career, it was always like, oh, I need to do this, Well, how do I do that? Oh, st break it down into steps, go research those steps, figure out how to write the code. And now it's like, you.
Know, and no one to consult with on that. By the way, you're right.
No one to consult with.
Yeah, and now it's big picture think all right, so start with chat, GPT or whatever, and you know, consult on the architecture. What are the ways that I can go about doing this and open up your imagination? Like you you might as well just try it because you can always blow it away.
I almost wondering if you just need an architect. Now, that's just the architect can use the tools to get a lot of the code written.
Yeah, it's it's you know, for for this journey. One of the steps along this journey was I realized we needed a great way to show differences inside the code, right, And I checked out what tooling that we had already and and we didn't. We didn't have what we needed essentially that was already built in. So I started that process, and I and I think I started a discussion with job GPT and and and to just kind of understand how the technology works, and ultimately I ended up with
an engine that does essentially the same thing. But there's two variations of the engine. One variation works with lines and the other works with characters. And so we use the line engine at the high level to see what's changed, and then if a line has differences, then we feed that to the character engine right to see what is changed, right.
What's the actual difference in this line to that line?
Right? And I also use the tool to help build
the visualizer as well. And the visualizer is super interesting because there's no built in difference visualizer that's out there, and so we were like okay, or I was like, okay, let's try I think the rich textbox or whatever I think is called WPF, and I start to do that and look at that, and we start to make progress, and then we realize that, okay, well it's great, except for the ends of the lines don't look great because the lines end at different you know, you know is
and the background highlighting goes only to that end of the line.
Are you talking about the get difference visualizer.
I'm talking about the one that's built into Aigen's navigator, the Coat Rush navigator, the one that we built from scratch.
Yeah, the one that's in Visual Studio is pretty good.
I mean, if you make a change to a file and then you double click it, you see the original on the right and the new one on the left.
And yeah, okay, so anyway, there is uh that's I'm not saying. I have I have some problems with it. I don't like a number of things of it, and so I implemented something different. Yeah, it's basically what I ended up doing. And at any rate, we got the highlighting right, but the ends of the lines wasn't working. And I'm like, well, how do I get this to work?
And that was the question essentially asked, and the answer is, you can't do it with the rich textbox, but you can do these adorners where you throw them in there. So then I'm like, okay, wait, I can put adorners out on the ends here and line them up. Yeah, okay, let's do that. But now I scroll, they don't scroll. And then so now my new question is how do I get the adorers to scroll? Let's make them scroll
and work. And this whole thing is this inertive process where I am doing this a piece at a time, but finally get it so that it's essentially solid. Right that you've got this, I'm using the rich text I'm not doing my own custom control, using existing rich text box and getting and making it behave as if it's a difference view. And that's done largely through the help of AI assisting and making changes interactively as I go through.
It's very cool, yeah, so cool.
Yeah, But you know your question about what level do you have to be? I think at least intermediate in general, and then the further off the beaten path you go, the more of an expert you kind of need to be or the more of a debugger, you know, kind of let's figure this out kind of person that you need to be.
You need to know what questions to ask, right, That's that's what it all comes down to.
Yeah, yeah, what to ask for You're right, Yeah, you do. Even if you're working with something you've done. You know, if you're working with a new framework, right, if I'm working with a control set, what do they call their grid control or what do they call their their toolbar control? Right? Right, If you know what that's called, then you can ask the.
Question and hopeful you'll do research first. That's the whole idea. The less the less familiar you are with what you're going to end up with, the more research you have to do up front, however, to know how to ask the questions.
However, Carl, I'm today, I'm in this code and I'm like saying, hey, I'm seeing that a whole bunch of these methods are passing in the it's an on notified property change. Yeah, thing I hear, And I'm like, and my prompt goes like this, I say, hey, I'm noticing all of these calls to property change are passing in the name of the property, But isn't there some way to check from the call stack what that method calling
method is? And that's all I said. And so in other words, you could know kind of genly about it and speak about it. And then it goes in and it ends. It puts in I think member calling name attribute.
Member calling name yeah, on.
The parameter definition, and it does that, and it's great, and I'm like, oh, but wait, you left all of the calls with their arguments. They don't need those arguments anymore, that's right, yeah, And then it fixes that and so that instead of me doing any of those things manually, it's just you know, speak, wait a few seconds, and then speak and wait a few seconds. And that's a different flaw.
You also have to have some pain, right like if I've had pain in the past with making base classes for models, for example, which I do not do anymore because you get into all sorts of problems with that, right then you know, you could easily take that in a direction where it would say, well, let me just make a base class for you that does all the notified property change and blah blah blah, and now you can just create models that are in that space class
and right, Yeah, you know what I mean, And you wouldn't have the experience to say, no, I really don't want.
To do that, right right, No, I get that, I get it. Yeah, But you know, you can make things work a lot of different ways. I can have the same set of test cases and they can all pass regardless of which frameworks. And sometimes I don't understand the problem with the framework until it's time to go to version two, right, right, right, and then I realized, Okay,
that was a mistake. I got to really really make this architecture great, right, And that is like that's something that you don't see with what we've done here, but we have actually made this architecture really good. We stopped we were moving towards bringing Blazer support in this one oh release, and we stopped it to focus on the architecture.
And as a result, I'm like way more confident about our ability to go in and support Blazer in other languages because of the changes that we made to simplify and decouple the architecture behind this.
It's awesome.
So feel good about that.
So where can we get this?
So? Okay, go to def express dot com slash code Rush and you can get you can download a copy codeer is for free. Code Rush is free. There's no charge for this. However, you're gonna have to supply an open Ai API key, an open Ai API key, and you're gonna need to supply if you want in voice, you're gonna need to supply an Azure Cognitive Services speech key.
Which is free for.
It's essentially you can get a free tier that gives you I think up to five dollars a month, and I use maybe three dollars a month in all my months of my daily use of the product, so you're probably not gonna pay anything for that. And Azure Speech Cognitive Services is excellent. I just I cannot, I cannot tell you how good this service is.
It's really good.
It's it is way more accurate than anything I've seen built on a machine running off of a machine. And it's fast. It's like generally we get times of processing times of a about from the time you stopped talking, we get a response back that's about four hundred milliseconds.
That's amazing, Mark, congratulations, and I'm really blown away. And it's always good to talk to you. I always learn things and what can I say? Thanks again?
Thanks, guys, all right, and we'll talk to you later on dot net Bros. Dot Net Rocks is brought to you by Franklin's Net and produced by Pop Studios, a full service audio, video and post production facilities located physically in New London, Connecticut, and of course in the cloud online at pwop dot com.
Visit our website at d O T N E t R O c k S dot com, RSS feeds, downloads, mobile apps, comments, and access to the full archives going back to show number one, recorded in September two thousand and two. And make sure you check out our sponsors. They keep us in business. Now, go write some code, see you next time.
You got JAD Middle Vans And then I'm