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, and welcome back to dot net rocks. I'm Carl Franklin, an amateur camp This is episode number
two thousand and five. Here we are, so, I guess we should start with what happened in two thousand and five?
You want to start? Where do you want to go? I got a bunch of weird things.
Yeah, I'll save all the good stuff for you. But there's some, you know, not so good things that happen in two thousand and five, like Hurricane Katrina. Oh right, yeah, I was down in New Orleans after Katrina for jazz Fest and Robert Plant and what's her name? Alison Krause came out and did when the levee.
Breaks a little on the nose but okay.
And also I remember going to a bar where they said it was a makeshift hospital during Katrina and on the you know, people would lay on the bar and doctors would come in and work on them. Sure, all right, Iraq war, bad war? In Iraqi elections. John Paul the Second died in two thousand and five. Pope Benedict became Pope Well. German Cardinal Joseph Ratzinger became Pope Benedict, and he was a sixteen.
He was a close confident of John Paul's.
I recall, yeah, the London July seven terrorist bombings.
MS.
Were you around when that happened.
I was in London when that happened. I was Yeah, that hit very close to home. Well glad you're still here. Yeah, me too.
Sodom Hussein's trial begin, Well, we know how that'll go. Yeah, we know. The spoiler alert didn't go well for him. Terry Schiavo the right to Die case. Remember the legal and ethical battle over removing the feeding tubes from Terry's chabo. Yeah, yeah, national debate about end of life decisions, government intervention, family rights. Record breaking Atlantic hurricane season, not just Katrina, but Rita Wilma, just one of the worst hurricane seasons. UH Israel withdrew
from Gaza. Things were looking up.
There, we're gonna be elections only one.
Yeah, and Angela Merkle became Chancellor of Germany, first female chamer beginning a leadership era that would shape European politics for more than fifteen years. Some honorable mentions. Michael Jackson was acquitted, Prince Charles married Camilla Parker Bowels. French and Dutch rejection of the EU Constitution. Let's talk bad about the French and.
The Dutch, right Arguably the moment the E that's the moment the EU broke.
Yeah, in Pakistan there was an earthquake that killed tens of thousands of people. In high oil and gasoline prices worldwide, I'm not so sure that they were as high as they are now. I don't know, probably not. Yeah.
No.
Also, I got married in two thousand and five. Yeah, ah wow, which nearly made it to twenty years. Yeah, but you know, most of them were good and it was it's amicably dissolved, good, good, good.
Some lovely progeny. As I recall, I do have two.
I have two of the most amazing kids anybody could hope for. Yeah, that's just apps. I mean, they are special, they are and yeah, I'll never say that, but my kids are awesome, all right.
So Top ten grossing movies from two thousand and five. Number ten Hitch, number nine, Batman Begins, so many Batman's, I'm Batman, Charlie and the Chocolate Factory, Mister and Missus Smith, Madagascar, King Kong War, The World's The Chronicles of Narnia, Lying the Witch and the Wardrobe, Number two, Star Wars Episode three, Revenge of the Sith, and number one Harry Potter and the Goblet of Fire, grossing eight hundred and ninety five point nine million dollars.
I'll see. If I could only keep one of those, I'd go with Madagascar like that one was a great Winds just killed.
Yeah, yeah. Richard, Space and Tech, I guess this was a pretty good year for space. Huh.
Well, uh, this is the year that Airbus reveals the A three eighty find the building an airplane bigger than seven forty seven, and they kind of botched it because, actually, music,
they'll take far longer to build than planned. Deep down, if you're into aeronautics, they had they knew, they expected that they would build stretch versions of the A three eighty, right, they only ever built the one model the eight, but they were planning on building stretch models, and so in anticipation of that, they overbuilt the wing so that it would be able to handle the stretch versions, and that there are for by making the main version, the initial
version less efficient than it could be. Okay, that, combined with the general failure of spoken hub flying systems, this method of using big airports that everybody flies into and everybody flies out of, meant the A three eighty was the wrong airplane for modern times. And so build a less than two hundred of them, and then they'll shut on the line and it'll be a major financial loser.
For air US.
I don't remember any of that. I just Airbus built other planes since then, right, Oh yeah, No.
Arguably the A three twenty is the most successful airplane out there. You know, he's neck connect with the seven thirty seven. But I can go nuts on aerospace, as you well know. Yeah, let's keep going on, I would argue. One of the most interesting things happened this year. The
year was the DARPA Grand Challenge. This was the beginning of automated driving, and when Stanford's stan Lee vehicle completed the challenge successfully, they had it they'd been running for a couple of years at that point.
That was the best show I ever watched on the Discovery Channel.
Yeah, it was great, And here we are twenty years later, still trying to figure it out. All right. On the space side, January is a crazy month. It starts with a group of astronomers identifying a Cooper object they called Aras at the time, when they identified it, they thought it was larger than Pluto and kicked off the cascade that results in Pluto being demoted as a planet. Right, So the issue being here, now that they found Arison,
they're pretty sure there's other objects of that size. They'll laterally find them, things like said and so forth. Initially they say, okay, well there's ten plants in Solar system. Then there might be eleven or twelve or thirteen, but they're different from the main nine or the main eight.
But they do or at the Sun.
Right all or at the Sun. So they had to The reality is there was no definition for a planet at the time.
Right right. I remember Neil de grass Tyson was one of the guys who put the nail on the coffin of Pluto as a planet.
That's actually how Neil Degrass Tyson became famous.
Yeah.
Yeah, it was support of this argument. The problem is that the definition is weak orbiting the Sun, fine is round or has hydrostatic equilibrium aka enough gravity pull itself into shape. Fine, here's the one they added to make it planet has quote cleared its orbit.
What does that mean?
So that's exactly so it sucked up everything which is in its orbital path itself.
Yeah, to be clear, no planet has actually done that. Every planet has both Trojan and Apollo asteroids following in its orbit. But one would argue that is clear that they are on the same orbital path, just in gravitational equilibrium, so they never get to the planet. You know, Pluto isn't on the plane of the elliptic, it doesn't actually orbit flat with the other eight planets. It also crosses
Neptune's orbit like it's a different object. Now. At the same time, they also had Serres, which is the probably the remains of a planet that existed between Mars and Jupiter and was torn apart, and it's a planetary core, so they considered they called that a dwarf planet as well.
Was originally called an asteroid, even though it's huge, and now they had ris as well, also not on the planet ecliptic also, you know, doing some crossing and asymmetry, and so the dwarf planet category sort of picks all those Cooper objects up.
I remember being blown away by the images of Pluto. That was a voyager.
Oh when new Horizons gets there, new Horizons.
Yeah.
Yeah. My favorite thing about that whole sort of dwarf planet drama is that Eris is named after the Greek goddess of mischief and discord and chaos, which is very appropriate. Completely through the whole astronomical establishment into disorder for years, you.
Got people wound up and still wound up. There's a movement again to make Pluto a planet.
Yes, so Hailarius. That's right, it's not over yet. I recently heard that.
Yeah, all right, well let's move on. That's a fun one. Also in January, the Huygens probe off of Cassini lands on Titan and gives us our first images of the surface of Titan because it has a very thick hydro chemical atmosphere methane and methane primarily, and we can see as it's descending and sends back these photos erosion patterns, just not erosion caused by water. Water is frozen solid there,
it's like rock there. This is done by hydrochemicals, by things like ethane raining out of the sky and eroding the terrain. So amazing pictures. Great accomplishment there, and also in January. We're still in January the Deep and Impact experiments. This launched off a Delta two to impact on the Temple one comet as it went by, so they get there by July. It was a very quick intercept, but
mission will go a little bit wrong. They will have the impactor, but the sensors will be overwhelmed by the debris. It'll kick up far more than expected.
I remember that.
But the Stardust mission that's already up there, that's been doing collections off of other comets, will be able to do a pass on Temple one's debris cloud as well and get material that way, so they'll recover from.
You want to land on the head of the comet, not the tail.
Well this they only sent an impactor in it. The idea was to kick up debris to try and measure the contents and actually collect some of that de breach. It didn't work out. Finally for space There was only one flight of a shuttle after the Columbia disaster in two thousand and three. There were no flights in two thousand and four. There's one in two thousand and five.
It was supposed to be the return of flight. It was the Discovery Shuttle bringing supplies to the space station that brought up the Rafello, the MPLM European capsule that they would take out of the cargo bay of the shuttle and dock to the station to leave it there that they could refill it with garbage and or experiments and bring it back. There will be enough problems with Discovery, including damages to its tiles, that they will not fly
again that entire year. They'll have to do further corrections. This is when they add the additional arm to be able to inspect the hull, and that'll be later. This one they did of they because they went to the space station. They did the inspection by pulling Discovery off of the space station and flipping the shuttle over so they could photograph it and saw some problems. So the issues of the tile system are becoming more and more apparent, and it will ultimately result in the end of the
Shuttle program. On the computing side, in the Microsoft lands, we get XP National sixty four bit, the first sixty four bit version of Windows.
I ran it, yeah, I remember that show. Yeah, we did on sixty four bit, and that also prompted the sixty four bit question yeah, which we end our conference shows with.
It also goes alongside dot net two and Studio two thousand and five, which includes sixty four bit support, except that they had law offered the option as compile any, which really meant running thirty two bit and crash in sixty four bit.
Yep, I remember.
Any wrong? We did that show. Don't do it any If you're not sure you're going to run in sixty four bit, set it to thirty two bid. Oh you will be sad. And really the only thing you were getting was do you get more than two gigs of address space?
Right?
And I think I said this on the show at the time. If you need more than two gigs of address space, what do you do? Do you doing, especially for a web page? I want to run. I want to run is sixty four bit moode. It's like more than two gigs for a web page?
What do you doing?
Like a tudy stories.
Coming out of the dot com bust last year. In two thousand and four, we had YouTube going public because the SEC required they had too many shareholders, so they start making some our Google goes public and this is when YouTube is founded and the first YouTube video called Me at the Zoo is published. This is before you know, Google hasn't bought them yet, right, but they will. But
my Google does launch Google Maps in two thousand and five. Yeah, and a little bit on the hardware side, Microsoft launches the Xbox three sixty, the weirdo version of Xbox that all the game developers hated. The first r duinos are sold, and this really begins this wave even before Raspberry Pie and things of the little micro controller you can program
and do cool computer things. Yeah, and Lenovo buys IBM think padline and becomes you know, today, I think the coolest laptop maker on the planet, not that I want one, but they do cool stuff.
The think pads where coveted. Yeah.
The coolest maker on the planet that is an Apple. Let's put it that way.
The Apple doesn't make a lot of machines. They make the highest quality machines, like by far. But when I say cool, I mean like you never know what logos Leno was gonna do. Is like this one has two screens, this one has three screens, this one full. Yes, yes, it becomes right like they're nuts. Yes, all right, standing down. They make it. They make a dedicated knock for teams because everybody hates themselves, Like I don't know it all right, that's all I got, thanks.
For Okay, before we talk to MS. Let's do better. No framework, roll.
It awesome man, you got.
So there's this project on GitHub called Sophia Script for Windows.
Oh what is that?
And basically it's a PowerShell module for fine tuning Windows on GitHub. It's got one hundred and fifty unique functions to configure Windows using Microsoft's officially documented ways without without harming your system. Okay, And there is now or there was recently a Windows version of this. So this is Sofia app in the Sofia community. So they say, it's the next chapter of the Sofia script project. And this is a free, open source app for fine tuning Windows
ten and Windows eleven. Offers a modern UI UX with more than one hundred and thirty unique tweaks and shows how Windows can be configured without making any harm to Windows. So I haven't used it, but there's you can install it with Chocolate Scoop and beta versions you can just you know, expand them and run them and yeah, it's this is good stuf.
Yeah, sure, you know. Over my other life on Windows Weekly, Paul the Rod has been working steadily on how do you make Windows eleven not suck? Yeah, right, like just turning removing stuff and getting the base configuration down, you know. And if you don't want to deal with one drive, here's how to actually make it go away without nagging you. If you don't want a Microsoft account, here's how you
actually can figure it. But this looks like a tool just make your life a lot easier for that sort of thing.
Yep. There's one feature that I'm going to check to see if it has, and that is this persistent the first persistent feature quote unquote that anytime you're in file Explorer and you've got wave files and MP three files and stuff, it insists on using the music format for the fields that it shows rather than regular documents. And I don't care about artists and album and contributing artists and you know, all this stuff. I want to see the file sized, because I'm you know, I deal with files.
Yeah, I don't care about how big the file is.
And it's and you turn it off and you turn it and it comes back, and it's just really really aggravating. I just want that feature out of Windows. So I'm gonna look for that. Stop fighting me, right, That's that's what I'm talking about. Who's talking to us?
Richard grabbing camming off the show nineteen sixteen. You know, back in the World War One days done with one Rendel called how simple is as simple as possible? Because we get good show names, generated a ton of comments. Of course, this is when we were talking primarily about things like is our micro services essential? Like what are better ways to build software? How do you keep this just as complicated as it needs to be? In no more than that? And just got a ton of comments
off of this show. Yeah, and I'm gonna grab this one from Robin Osborne who says, thanks so much. When I'm really doubting myself during the decade long tsunami of JavaScript all the things, and I'm thinking, surely this isn't the best approach. Am I wrong? Do I just not understand something that everybody else gets. It's a load off my chest to hear people I respect in the industry talking about lovely things like HTMX and referencing Alex Russell's
great articles. I love Polly, I love twelve factor back in the day. I love making things as simple as possible, simple to build, simple to maintain, and still for the end user. Gagny and kiss for the win. Shit, I'm not going to argue with that. How about kiss? You mean, keep it simple?
Stupid?
Yeah, I don't know where you want to go, but you know what ramin where on board. You obviously heard it on that show and probably hear more of that today too. So thank you so much for your comment. A copy of music Coby. It's on its way to you. And if you'd like a copy of music Cobe, I write a comment on the website at dot at Rocks dot com or on the facebooks we publish every show there and even comment there when I read it on the show, we'll send you copy of music Cobayah.
Music to Koba, of course, is a project I started a long time ago to create twenty five minute music tracks that can slip into the background and help you focus and it's been wildly popular. Now we have twenty two tracks. Twenty three is coming, I promise, and music too by dot net if you want to get it in MP three, wave and FLACK versions. Okay, let's bring back ms Rendel. Ms has been writing code since childhood, getting paid for it since her teens, and inflicting opinions
on conference audiences for the last fifteen years. The former CTO with deep roots in dot net, c sharp APIs performance engineering and system architecture. She's now working in financial technology by day, then spending her spare time building ambitious, weird, and occasionally beautiful things at the intersection of ar AI and human connection. Welcome back with MS.
Hey, Hey, thank you for having me.
Thanks for participating in the intro. Yeah, fun, Yeah, I'm sure since the ast time we talked, you've got a lot of things on your mind.
It's been a wild couple of years. Yeah, yeah, I just I've continued trying to keep things as simple as possible. There's a bit of a movement has kind of popped up in the last couple of years. Actually, Chris Woodruff, who I know, you guys know, has been doing a lot of writing sort of around the simplicity first approach.
She's a big advocate of that. There's the use what works initiative, which is sort of the whole buy don't build kind of trying to suggest to people that just because an open source project is asking for commercial support money, that's not a good reason to go off and build your own messaging stack or warth authentication stack or whatever it.
We wrote this code not because it was easy, but because we thought it was easy.
Yes, we wrote this code because we thought it would save us five thousand dollars.
And I think it's a trap of modern AI development, which we're all finding ourselves doing some somewhat, is that you think, you know, oh, I'll just I need this, I'll just have my AI write it. Where the first question should be does it exist? You know?
Yes?
And that's literally in my prompt that I use for exploring new ideas, there's a sort of thing very near the top saying if something exists, let's just use it, tell me about it.
It kind of reminds me of the days when you know, everybody wanted to build an iPhone app, an Android app, and they come to you with ideas. You know, hey, I got this great idea for an app, I'll tell you what you build it it will split the yes, okay question number First of all, no, in question for you? Does it exist? Did you love to see if it exists? In the apisode? Oh no, I didn't do that.
All right, Well, yeah, where do you exists? Bet? It does?
So anyway, I'm building something that definitely exists.
Nice, but I wanted to talk.
To you guys about it anyway because I am Actually it's been a while since I built something kind of big, and in dot net there's been a bunch of sort of interesting projects, but they weren't really they were either kind of closed source and I wasn't allowed to talk about them, or they were outside the dot net ecosystem. But yeah, I've actually I've been working on a thing and it's in dot net, and I thought you guys would be interested.
All right, tell us about it.
Well, you don't need to tell us about it. We can just talk about something else if you like.
No, So, I'm building my own coding agent harness. Oh thing, So like clawed code or codex or open code or pie or any of the millions of different ones that are out there.
I got to tell you before you go any further. I tried that and it I was tearing my hair out so much that I just gave up.
It's going quite well.
It's not easy.
It's not easy, and there's a lot of things that sort of pop up along the way where you kind of go, oh, I didn't think of that. Oh that's interesting. Oh yeah, I suppose that is, but yeah, no. Doing it in dot net and c sharp is provide some interesting opportunities that I was kind of like, well, if I was using dot Net, then I could use new
get to distribute extensions. I would have kind of very firm interface boundaries, and I could have an abstractions and an extensions package with different kinds of extension, and people could extend around those. And there's a phenomenal first party ecosystem around building AI tools and agents with dot net, because you've got like Microsoft dot Extensions, dot Ai, You've
got Semantic Kernel, you've got the Multi Agent framework. These are all things that are being built and supported by somebody else, which takes a lot of the sort of more complicated bits of effort.
You know.
Just talking about providers, you've got sort of three anthropic open AI and Gemini and then you've got like local providers like Olama, Lama, cpp LM Studio, OMLX and all the other things that are out there that are like open Ai compatible, like deep seek and Kimmi and GLM and all these different things, and there are either pre existing rappers for them. You know, there's a Anthropic ship, an implementation of I Chat client which is in the
Microsoft Extensions project. There's an Alarma sharp out there which interacts with Alarm. So yeah, there's all these sorts of things, and so it makes a lot of the building of this thing more like lego and sticking sort of pre built bricks together, which really really helps.
So, yeah, do we need another one?
Yes, yes we do. I do. I definitely do.
You find all these other ones are lacking.
So my sort of daily driver and the thing I've been using at work for the last year is Claude Coat and claud Coke is amazing. But it works with Claude and you can do sort of interesting hacky setting anthropic environment variables and get it to work with other engines, especially when they have an anthropic compatible API. But we are approaching the rug pull. Yeah, the point where in the same way that Uber went right, We've destroyed the competition.
Now we can rack the prices up. The vcs are kind of like going, it's probably time that we start charging what it actually costs and make it profitable. And there are multiple outcomes that could happen from that.
Yeah.
I mean, I'm going to wrestle with this being a rug pull so much as a reality check.
Yes, yeah, okay, yeah, No, definitely.
It's is like, hey, this stuff costs more money than you were charging, so how much is it really.
Well, it depends on what side you're looking at. If you're looking at the LM providers, it's a reality check. If you're looking at users, it's a rugbull. Yeah, no doubt about it.
I mean that's certainly for a lot of the less technical users who don't understand the constraints and limits and energy use and everything.
I just want to tell you what I'm doing. I'm using GitHub copilot Cli against a local machine that's running Olama, and I found a couple of models that work really really well. Which ones so that so Quinn yep, Quinn three point six latest is the best? Yes, and even two point five thirty three thirty two billion yea Quin coder. Yeah, works works the best, that's the second best. And then
we did this on code with Ai dot com. The last two episodes were local language model roundups, and I did and I tested a bunch of them with against my Olama machine, and those two are the only ones that I would use going.
For I've been using Gemma four. I tested at the thirty one billion parameter and it's good. Quinn is better for coding, yeah, but Gemma four I'm using for a project that I'm not ready to tell you about it but because I don't want to. It's it's in stealth and at some point somebody's going to give me sacks of money to actually productize it. Still, this is kind
of at the core of that. But yes, I think I don't know whether sort of the big funded cloud LM providers and Thropic and open AI, it's it's kind of questionable whether they even survive because they're a real questions especially, I mean, you Knowthos. Was Mythos held back because it was exposing sort of vulnerabilities in ten thousand two power fundamental internet projects? Or was it held back because the compute capacity did not exist to run it?
Yeah? Yeah, it's pretty big, you know, and.
Thropic of basically least Colossus one and two off SpaceX now to try and run this stuff. But you know, I keep seeing analysis that these we're expecting to build hundreds of gigawatts of capacity per year, but we're actually building tens of gigawatts if we're lucky.
So well he did. Yeah, it turns out you can't just build power plants over night. It's not that quick. Yeah, you know, as the guy who does a bunch of talks about power, it's like, hey, pretty much the fastest you can build one power plants about five years. Yeah, and this thing hasn't been going on for five years. So how many power plants do you think we've so far forward?
One of my favorite things about the power plants was, I think it was Hank Green on YouTube did a video. There's lots of old shutdown coal power stations which have got all the everything that they need to be turned into small nuclear power stations. You could just stick a reactor core in there. The reason they can't do it, certainly in the United States, is that they're already radioactive to a level that exceeds EPA regulations. For how radioactive
and nuclear power station can be just from burning coal. Wow, which is it's lovely. Yeah, Trump plants to refire these coal plants. Yeah, I mean that's also crazy. Yeah, but yes, I kind of see a potential future where you've got to exist in between the cloud llms and the on device things by providing like custom training. So a customer comes to you and goes, I want coding agents that I can run locally that all I care about is
dot Net and typescript and you know, React. And you either get like a completely ground up trained thing, or you get a fine tuned trained thing training and fine tuning models.
I think so too.
A it's compute expensive, and B it's it's not easy, it's not sort of straightforward. It's a it's a skin.
Well you know. The other the other option is to take something that does know, you know, c sharp, typescript, all of those, but also knows a lot of other stuff and just pare it down to what you need it for. The distillations kind of like stubbing out when you're compiling, you know, stub out the stuff that you don't want to link in.
Yeah, and so yeah, I want something that I can make work with Claude or Gemini, or I can make it work with OMLX running a local thing, or I can make it work with Olama running on a DGX Spark sitting in my house. And yeah, so, and I wanted something that I had more control over. I looked at the source code for the Pie Coding Agent, which is the closest thing, and it's written in typescript, which, apart from anything else, means you have to have Node installed.
And I have a long history of railing against the need to install mode and so yeah, and I thought I would I would sort of build my own Pie inspired thing in C sharp. And it's so nice to be back in c sharp and to have the agent generating code that I fully understand and can refactor with kind of sharper and those sorts of things. And so yeah, it's uh, it's it's out well, it's not quite out
there yet. I haven't done a proper launch, but it's public on GitHub at demonic ai with an A and then the demon d m o N dash core.
And will add that to the list.
Yeah, I'll send you the I think I sent you the link in the pre show thing. And yeah, the core of it runs as a separate process, and it's it's a Jason RPC server, so you I'm currently building a terminal client for it, like a net console app, but there's going to be an Avalonia app, and it is also going to be there will be another thing, which is a public gRPC server that exposes it over gRPC so you can talk to it from your phone. And yeah, but the things I'm kind of excited about.
So dot net means I can compile it as a standalone thing, so people don't have to necessarily have the dot Net SDK installed, although if you have the dot Net SDK installed, you'll also just be able to do dot netool installed demon and it will be distributed that way as well. And I've borrowed so the process isolation thing is is cool. It's using Microsoft Extensions AI, which means there's a bunch of stuff out there that I
can already import and wrap quite lightly. I've borrowed the session system from Pie, so you have a tree of sessions and when you spawn a subagent, it kind of creates a fork from the current session, and so you end up with a tree. I don't know if you've looked at pie at all. But it has this concept of sessions as trees with sub sessions and sub sub sessions and all that sort of thing. It's really quite clever. And they're all just files in the directory.
Course, so you're basically persisting the history in all the context around that is in MD files, yeah, yeh.
MD files and JSONL files. And then and this is the bit that I'm sort of I think is my my value proposition is the extension model. So you can create an extension. You have a new Get package that you can get, and you build your extension based on that, and then you publish it to new get with a tag of DMA nash extension and with the source code available.
And then Nougat is already one of the more secure package platforms out there, whereas you know, NPM has been having some disasters lately, and PI has its share of issues and everything else. New get is pretty secure. And if you publish it as source available and you get the new package that goes alongside the new package, then
you install it using the agent. So you go, hey, I'm going to install an extension, and it goes off and it gets it and it gets the source code and it scans the source code for vulnerabilities before it allows you to load the package. So if someone tries to distribute a package that's got a Bitcoin minor or something that's going to extract all your bank details and credit cards, it's got a built in secure scam before it does anything.
How far along are you with this project?
It's working, I can use it to build itself. And yeah, I'm expecting it's as we're recording this. It's NDC Copenhagen next week, and I'm expecting to be making a lot of noise about it at NDC Copenhagen and trying to persuade people to give it a look. But yes, and then the sort of the much much bigger thing, the startup that's being built on top of it, once the core of this is done, should progress fairly quickly as well. And that's not only kind of personal AI and playing
in the kind of open claw and Hermes space. It's also adorable, okay, And I'm going to bring adorability and beauty to.
The manifest itself.
Actually at the moment, as a bird, oh.
Bird bird, is it an avatar you interact with or is it just I will.
Go I will come back and tell you more about it at another stage. But it is, it is. It's leveraging a lot of the things that I've been working with over the last couple of years and bringing them together.
I can smell the smoke and see the wheels turning in the brain of yours.
Oh but no, I've been I've got kind of like a proof of concept app installed on my phone that is really incredibly simple that is already blowing people away. They're kind of go, oh, that's really that's amazing. And then yeah, well, hooking it up to aiye is just going to do very interesting things.
I think this is a good time to take a break, so we'll be right back after these very important messages. Don't go away, and we're back. It's darting at rocks. I'm Carl Frank, and that's my friend Richard Campbell, and that's our friend M's Randall. Hello, talking about some new stuff, a new way to interact with your local language model or your large language model.
Yep, yep, I am circling back on this whole. Aren't we going to get to a place where you have an LM associated with an application that can modify the application.
Yes, I mean that's that's one of the things with this is especially the LLLM can briefly extend itself by writing a CSX script, which you can you know, distributing the Rosslyn things. We can compile that, load it into an assembly context. That becomes a temporary extension, and then there's actually a built in promote command, right that then converts that into an actual sort of dot Net project with a cs pro file and ne get packaging and everything else.
In the work Vanilla was doing was extension. He was using squad dash to to modify a squad dash.
Yes, and there are there's there are platforms out there where it is literally this, you know, create a new application and it just starts and it's an empty screen on your iPhone or your desktop browser or whatever, and you literally start telling, hey, application, this is what I want you to be, this is what I want you
to do. And yeah, it kind of builds itself from the inside out, and if you want a new feature at any point, you can just go, hey, come back and I need you to add a new feature, right, And I think that is there's that level of things where it's kind of like build an application so that
I can track my to do list or whatever. And then there's the point at which, and this is kind of one of the things that's very heavily on my mind is we have ways of interacting with agents that the idea of having an actual application with a specific user interface will just be silly. Why would you do that?
Well, yeah, text boxes and buttons and things seem silly. The screen will be used to present information.
Yeah, sort of. ALI is generating infographics and charts and pictorial representations so we can all forget to learn how to read and all these sorts of things.
Look, the line goes up to the right, this is good. You should get a raise. Or the line is going down, you are going to be down.
So yes, the Onion did a good story a couple of weeks ago about high schoolers lacking objectssistance skills, which
is something that toddlers developed about eighteen months. So yes, But yeah, know that the bigger idea I think is, and and this is the kind of core of having this as a standalone thing with an STDO RPC interface is does it need an avaloniar application, does it need a command line application, or can it sort of just run as a server somewhere and I can talk to it using my phone and my AirPods and that sort
of thing just persistently, and it can run locally. I can run it in my house, I can access it over tail scale or a cloud flare tunnel or something, and it can keep all my data locally. To me, yeah, that's important on my own sort of you know, And that's I mean to me, that is a huge thing. That is, it's very much. I don't want to send every email email I get, and every Apple message I get and and all these different things through a hosted
cloud model. It would be much better if I could run them locally.
And so I'm with you, and I think that the future is, you know, big machines running locally that are running some you know OLAM or some kind of local model host and because that's the big thing, right for the expense can be absorbed by big companies. I mean, but you know, guys like us don't want to be paying a couple of thousand dollars a month for large language model access. But but you know, privacy is a big problem.
Privacy is a really big thing. And you know, I can see a business model where actually, what we're selling is like an appliance in the same way that people would buy an Alexa or whatever they called those things, an echo, or they would buy a Google Home device.
And you could sell, like I mean right now, a DGX Spark clone, the thing with one hundred and twenty eight gigs of unified memory and a petaflop of COUDA and tensor cause and you can pick one of those up for three thousand, three and a half thousand of your local dollarish currencies. And that's the first generation of that thing. And so in five years time, that spec
is going to be available. As you know, it's not going to be quite Raspberry Pie, but it'll it'll have come down under sort of close to the Mac Mini level of pricing. Mac minnis, you know, with sixty four gigs or one hundred and twenty eight gigs would be plenty big enough.
My machine is a MD Rising nine nine ninety x twelve core ninety six gigs RAM, the Nvidia g tour g fors rtx fifty ninety with thirty two gigs of v RAM. And yeah, like I said, quenn three six runs in there just fine.
Yep, And yeah, I do. I think that's that's where we're going to end up. One of the things, particularly with with Demon, one of the sort of features that it has because I couldn't figure out how to get other things to do this is every extension can specify a different model, and so it's like you could have read, write, Bash edit all these different things. The extension that talks to your Gmail account and those run on your locally
hosted model. But web search, which is a pain to do locally because you need an API endpoint that you can hit and you have to pay for all of those.
There are MCP servers for that. Yeah, so you could just plug those in as you want to do.
But or you could, you know, just say, use the Gemini model whenever you want to do a web search and you can just drop out quickly go to Gemini, or maybe use sort of a slightly bigger, more complicated model on Olarma cloud because you're doing a web search, so.
It's a different process. You can use it different mechanism.
Yes, I suppose technically what you searched for or is useful information, but somebody's recording that anyway. If you search at Google, it's not like they're not keeping track of that sure, and so you know, so it's like, yeah, the ability to be pragmatic about privacy and security and where things are hosted and all this sort of stuff and to be able yeah, like I say, to use a completely different model for certain extension types and be able to configure that is a useful thing.
We have a new customer at appinax. In reading the contract, there's a clause in there that no date, none of their data will ever be entered into an AI large language model for privacy reasons and you know, none of their code and all that stuff. So then we get working and we've got to get up repo with all the code that has access to other data, and you know,
kind of had to clarify with the customer. You know, what is the reason, what is the thing that you're most concerned about that you have this clause in here, and it was privacy. You know, it's basically they've heard the stories of people just taking requirements dumping them into chat GPT. Chat GPT now knows all about them, and you know, they lose control of their intellectual property and
all this stuff. So it's taking some convincing, but we're convincing them that with getthub copilot, there's a way to opt out of them using your data to train their models. And they have your data anyway in GitHub, so it's not like it's going outside of that. The only thing that I need to be sure of is that if we're using Claude sonnet for example, does Claude know does Anthropic now know about your data?
I mean, this is what happens when you mix a match, right is now you have to think of each of the layers.
Yeah, so that's the stuff we're working through right now.
So yeah, it's it's the thing I'd love about all of this stuff. And you know, the work I've been doing with various kinds of AI and everything over the last two years, it hasn't settled yet.
It's like we're far from it. We're nowhere, But I say, what do we know?
Here comes the rug.
I've done a lot of kind of historical talk looking at the history of programming languages and that sort of thing, and Fort Tram and Cobold and Algol and whatever, and yeah, there was a time when people were inventing the paradigms and that sort of thing. But then at some point we had C and then we had C plus plus and it's kind of like everything since then has just been variations on a theme to one extent or another. There was the dot com thing and the internet thing,
and that was it was cool. But yeah, this is still at that kind of pioneering gold rush stage where there's a lot of stupidity going on, but there's also a lot of like really creative work still to be done, and it's fun to be a part of that.
Well. Part of this is while there's too much money around, you don't make good choices, like we do need to be more efficient and more disapinted. And so as you see the money starting to come out of the system and they you know, your chances are becoming a billionaire now and are disappearing. So those folks leave and good riddance to them. Yes, we get to the engineering side of can we make this actually practical and useful? You know?
Comparing this to the dark com boom, it's like it's two thousands, the end is coming, but the heck's not going to go away, just the dumb. The Dumb's going to go.
Yes, Yeah, And one thing I'm hoping with this is that you know, back in the day with not a mollusk I did. I got a lot of engagement with that project. People wrote kind of a Mongo dB adapter, and I think there was actually simple data for Twitter.
Somebody wrote a thing where you could write simple data queries against the Twitter fire and I think, you know, I know there was still a ton of net shark programmers out there, and I know they're using sort of generative AI, but this is something where I'm kind of like going, hey, look, here's the actual harness is written in dot net, and it's kind of moved it a little bit further along from just the Microsoft extensions, which give you fine grained building blocks, and it's kind of
here's an actual agent and you can hack on it, and you can share your hacks and you can contribute back and so forth. And I'm hoping it's going to open up this kind of the joy of experimentation to a bunch more people who, for whatever reason, you know.
I'm talking to teams now who are no longer arguing over futures, are just building all the versions they can think of and then comparing them.
Ah.
Yes, as the code generation goes so quickly, Yeah, I could just think about code differently now.
Yeah, and yeah, it's like which of these is I've actually so I had Demon dot Console, which was just a standard console application, and then briefly there was Demon dot TUI, which was using the terminal dot guy newgat package, but that was very broken for a variety of reasons, and so then it reverted back, and now it's Demon dot terminal, which is using Spector Console. But yeah, and it's like going down the TWI dead end would have wasted a week of development time a couple of years ago,
whereas this time it was just one evening. We got halfway through the evening and I was kind of like, this isn't working, and we're going to revert back the last few changes and pick up from the console application again. And yeah, it's it doesn't cost anywhere near the same, so you can actually spike you can. Actually I've always had this thing about you should build it once and then throw that away and then build it again properly with all the lessons you learned from that first build.
And that's actually a reasonable proposition. Now it's kind of do a really dirty, exploratory vibe coded build, but get the agent to keep a record of all the things and the decision record architecture decisions and so forth, and then you can literally use that to generate an insanely detailed spec and effectively do a clean room implementation of your prototype. The prototype is.
A sort of off the back of the experiment.
Yeah, and then you know, you get people going all vibe coded stuff. The code's an absolute mess. But this is a solution to that. Use the vibe coding to prototype and then with all the constraints and clean code and don't repeat yourself and yagny and kiss and whatever else to a ground up reimplementation.
I find myself doing a lot of reading when I'm working with an AI agent. In other words, I sort of give it an overview. Is if it's a greenfield project, I'll give it an overview, and then I'll ask it to come up with a plan, you know, in multiple steps, and then write a customer facing document that even I can understand, well that even my customer can understand and I can understand. And then I go through the plan and figure out is this step necessary, why you know?
And what about this step? And it'll come up with things that I hadn't thought of. But you know, sometimes it might be a little over ambitious and but but I like to work that way first on the architecture, on the plan, before I even do anything coding wise, and that is the most important part of the project, if you ask me. But it's a lot of you know, talking and a lot of reading.
Absolutely. Yeah.
Yeah, So we have the link to it on GitHub. I think you already provided that the show notes.
Yeah, so yeah, and yeah, by the time it's this episode goes out, I will have been making noise about it and hopefully that'll be.
The force goes out during the conference.
Nice, Oh that'll look brilliantly then, yeah, almost like we had a plan. So yeah, I've actually I've turned up with my list of things to cover and so forth. So yes, but I think anything covered, No, it's I think we've covered everything. It does local models, it's the permission models and all this sort of stuff. It is. It is properly open source. It's Mozilla Public License two point zero. And yeah, I hope people will at least
take a look at it. If nobody, I'm definitely going to tell if nobody engages, I do actually need it for my bigger more elaborate plans, but it would be nice if people did like it and use it and play with it and stuff.
So, because the way all good things start, you build a thing you need and you wonder if other people need it.
Yeah, Frankly, well, I'm going to check it out, and I'm really looking forward to your more adorable project that we will learn more about sometime in the future, and you will come back and tell us about that.
Right, you will be the first people I tell fantastic.
Yeah, Endell, thank you very much. It's always a pleasure to talk to you, and h this was no different, mind blowing as always doing the awesome.
Thank you.
I will We'll talk to you next time on dot rocks.
Dot net Rocks is brought to you by Franklin's Net and produced by Pop Studios, a full service audio, video and post production facility 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 for RSS feeds, downloads, mobile apps, comments, and access to the full archives going back the 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 by the
Home the Texas