Zach VorhiesUSsoftware-factoryrustbuild-speedarchitectural-lintopen-source-maintainerAgents run as a teamnixoskde-plasmaobswhistleblowerbuilder/builders/zack/avatar.jpg
The Google AI whistleblower who built his own software factory. He maintains FastLED, ships his own agent front ends and build caches, is porting OBS Studio to Rust with OBS Rust, and is pulling the best of Omarchy into NixOS with Omnix.
rigNixOS with KDE Plasma, building every platform from one linux box
found omarchyTried Omarchy, called its UI radical and exactly what we need, then started migrating the good parts to NixOS
modelsclaude, codex, deepseek
ai workflowClaude Code and Codex bridged by clud, his own agent front end, with post edit hooks and architectural lints keeping the agents honest
// shorts
Shorts
6 videos
0:44From Google AI whistleblower to a one man software factoryZach, builder #15: the Google AI whistleblower from 2019, now running a software factory he built himself and pointing it at OBS Studio, porting it to Rust (github.com/obs-rust).0:52Everyone loves Git worktrees. He hates them, so he fixed themGit worktrees blow your build cache. Zach's fix: reflink on BTRFS/XFS so each worktree is a copy-on-write clone of the cache, torn down when the PR closes.0:511X coder at Google. 4X the day Aider droppedAt Google he was a 1X coder. ChatGPT copy-paste, then Copilot, then Aider dropped and he hit 4X. Then he wired in test hooks so the agent iterated on each function until the unit test passed.0:56Claude Code refactored 200 files. So he wrote an architectural lintClaude Code put an include after a namespace and refactored 200 files while he was away. So every mistake became a post-edit lint, then an architectural lint.0:48Anything that can block gets a timeout: kernel APIkernel API: blessed functions only. Anything that can block gets a timeout, every app can dump its stacks, and the agent finds the deadlock in one step.0:504 MB hello world vs 1K of business logic in WebAssemblyA Rust hello world is 4 MB. The business logic is 1K. Compile it to WebAssembly against one API and you get kilobyte micro-apps and fast Rust builds.
// life timeline
2019Blows the whistle on Google's machine learning fairness
Early on Aider, builds his own front end and a test driven loop around it
Takes over maintaining FastLED and sets out to automate himself out of the job
Claude Code lands, he builds clud on top of it
Moves his stack to Rust and builds zccache
2026Launches OBS Rust, a fork of OBS Studio moving to Rust
2026Builds Omnix, Omarchy ideas on NixOS
// 00
The rundown
Zach Vorhies is a rock star. The world met him as the Google AI whistleblower. Builders know him as the guy with a software factory, a fleet of agents that writes, lints, caches and ships code faster than whole companies. He maintains FastLED, he is porting OBS Studio to Rust, and he is grafting the best of Omarchy onto NixOS. His own description of the workflow:
"I've got this Sisyphean boulder that I'm carrying on the back, and a Claude code exoskeleton that basically allows me to sprint up the mountain."
“I've got this Sisyphean boulder that I'm carrying on the back, and a Claude code exoskeleton.”Zach Vorhies · L293
// 01
Famous for a week
In 2019 he blew the whistle on Google's machine learning fairness. He jokes he was super famous for one week and has kept it going ever since. Today he runs a weekly show and turns up as a featured guest across the news, so yes, the camera setup is tight.
// 02
Aider to clud
He was in on the AI wave from day one. When Aider dropped he built his own front end for it, then wired in a loop that ran the unit tests and fed failures back until the code passed. Zach says that alone took him to about 4X. Then Claude Code arrived. He found it broken, so he fixed it with clud, his own front end with "God mode only, uh, always, and safety third." It bridges Codex and DeepSeek into Claude Code, and Zach says companies now run it to keep their chats off other people's servers.
// 03
The software factory
It started with FastLED, the LED library for Arduino. He took over as maintainer and asked the only sensible question: "how do I automate myself out of this position as maintainer?" The answer became a factory with his own tools up and down the stack. There is bosn for fast local CI and zccache, his replacement for ccache and sccache that caches precompiled headers and uses reflink so a git worktree stops nuking your build cache. Then a tweet convinced him the future was Rust, so he moved the whole stack there.
// 04
Architectural lint
Agents make the same mistakes over and over. Zach makes them pay once. Every failure becomes a lint in a post edit hook, and the agent fails on the spot and fixes itself. "I'm basically doing an architectural lint." His kernel API gives agents blessed functions only, where anything that can block has a timeout and every app can dump its own stacks. Now when something deadlocks he just tells the agent to figure it out. Zach says he is going about a hundred times faster than he did at Google. Modesty was never the assignment.
// 05
OBS Rust
Next target, OBS Studio. He forked it as OBS Rust and pointed the factory at it. Zach says the port is about 60% done with a team of about 10, and launch time dropped from about 4 seconds to under one. "It's 0.95 seconds to launch OBS Rust." He wants it to be the OBS for Omarchy.
// 06
Omarchy meets Nix
Zach loves what Omarchy did to the desktop. Its UI is "radical, and it's exactly what we need." Hyprland and vim as defaults are not his style though, so he went to NixOS with KDE Plasma and started pulling the good parts of Omarchy across. That turned into Omnix. Zach says it keeps the NixOS binary cache as the base, adds a second artifact cache spread across GitHub releases, and lets you pick Hyprland or KDE Plasma. His pitch to DHH is simple. "And it's like, hey, I've already done it. I've proved it out."
So here is Zach Vorhies, builder #15. Whistleblower. Maintainer. Factory owner. Rust evangelist with a boulder and an exoskeleton. Zach, you are a rock star.
L6Zack: Yeah, I'm a podcaster. I mean, I'm, I'm, uh, I do a weekly show, um, and I'm a featured guest on a lot of different, um, news organizations, so I have to have a tight setup, uh, in order to do that.
L28Nix: So I just want to make sure you're good with that and you will have a chance at the end if you go you know what I said this thing and I'm not cool with that so just don't put that in there.
L29Nix: There's a little surprise at the end, so be ready for the closing thing you say will build your own.
L30Nix: your own adventure, and you'll see what I mean when you get there. So, like, I'm so interested in this, all this, like, you're on podcast, you're on… you have your own, like, what in the… who are you, and how did I find you?
L32Zack: Um, yeah, so, uh, my name is Zach. Uh, I'm known as the Google AI whistleblower. Uh, in 2019, I, um…
L33Zack: Yeah, I blew a whistle on Google's, uh, machine learning fairness, which they were lying to Congress about and to the media. Um, they were using it to secretly censor what you could see online.
L34[redacted: private, lines 34 to 38]
L39Zack: I was like super famous for like one week. And then I've been sort of keeping that going.
L41Zack: Um, and I've been very into AI for a very long time, uh, and now that this AI wave has been hitting, uh, I've been in it since, you know, day one, especially with the, um, what was it called? It was called Aider. I made my own front end for that. That made it super easy to…
L42Zack: start using it, and then I was sitting there pretty much ahead of everybody else with my custom AI Aider, and I was, uh… I was like, oh my gosh, this is amazing.
L43Zack: Um, I got my, uh, the team that was reporting to me to all use it, and we started flying, and then Claude Code came out. Um, and then I made my own front end for that, called clud, two less letters than Claude.
L45Zack: and it's got a bunch of safety features in it. It's always got God mode enabled. It bridges other things like DeepSeek into Claude code automatically. So and then I've got corporations now using that because.
L47Zack: Uh, they want to use Claude code, but they don't want the security implications of sending all of their chats to, you know, Anthropic servers, so they can plug in.
L48Zack: uh, you know, DeepSeek, uh, into that. Um, they've been very happy with it. Uh, and then what I've discovered is pretty much every single tool that I want to use is broken in its each individual way. And so what I've been creating, uh, you know, for the last…
L50Zack: 3 years is, uh, slowly at first, but now really fast, which is a software factory.
L51Zack: And this software factory has, um, it's, it's basically proprietary tool sets up and down the entire stack, you know, like, uh, Docker tests. Uh, I've got my own little front end called bosn, which, uh, basically does garbage collection.
L52Zack: Uh, and allows you to run local CI really quickly on your own desktop. Um, I've got policy documentations that I roll out.
L53Zack: to my repositories, um, and this pretty much all started because I took over a…
L54Zack: um, a open source project, which is the second largest library on Arduino, and that, uh, library is called FastLED.
L55Zack: And that's basically… I was like, how do I… how do I automate myself out of this position as maintainer? Uh, and… and basically, the… the desire is to replace me as, uh, the, uh, the engineer, and that's been going very successfully.
L56Zack: And, um, now that I've been, you know, that's how I grew my software factory and developed it. Um, and then as the models keep on getting better, I'm able to replace more and more of my tool sets. Uh, that includes build tool sets in, uh, C++.
L57Zack: Um, one of the things that I did was, uh, I replaced CC cache and S CC cache. You may be familiar with those. They're caching, yeah, build systems. Uh, mine's called zccache and it does really interesting things like, uh, extends the feature set of which things, of which artifacts could be, uh, cacheable, including.
L58Zack: uh, pre-compiled headers, which are a big deal. Um, it has support for reflink, so that, uh, a Git worktree can be created, and then the entire thing can be essentially copied on write. So, so here, here's the thing. Everyone's like, oh my god, Git worktrees are so amazing. I hate.
L59Zack: get work trees, because they constantly blow the build cache of your build system, right? And so you sit there, and it's like.
L61Zack: Yeah, yeah, because the thing is, is that I'm sitting there, I'm like, why are all my CPUs pinned at 100%, right? And I go and I look, and it's like, oh, well, there's a Git work tree, it can't use any of the cache.
L62Nix: Because that's how we've always done it and we're trying to work backwards compatible crap.
L63Zack: Right, yeah, well, the thing is, if you have BTRFS or XFS, then basically you've got this reflink concept within the operating system itself, and you can just, like, basically copy the entire Git worktree verbatim with the cache, and then when the worktree needs to, like, update, because, you know, you've got, you're doing a changing of the code.
L64Zack: the specific build artifacts that are needed are then copied on right, right? Or they're modified and then replaced for that, and then the whole thing gets torn down at the end of, like, the PR.
L65Zack: Um, and so, like, this… this has been really great for, uh, removing huge bottlenecks. Um, and… and about a year ago, I switched over to… I started experimenting with Rust.
L67Zack: because of a tweet by the name of, um, Brock or something like that from OpenAI. He basically said that the future of all software is going to be done through Rust, and I went.
L152Zack: And so, all those optimizations together combine to make super, super fast CI. And essentially, at this point, I've got the fastest Rust CI local build software factory in existence.
L154Zack: I mean, right now, I'm basically… so, you know, I did this consensual takeover of FastLED, right? That's… that was kind of the first…
L155[redacted: private, lines 155 to 156]
L157Zack: Yeah. I've created a repo called OBS Rust. It's a new organization on GitHub. I've got obs-rust.com.
L158[redacted: private, lines 158 to 161]
L162Zack: The Rust migration is about 60% done at this point. We've moved to a full Rust build with C++ legacy features. And there's about a team about, I guess about like 10.
L163Zack: right now, uh, with about 4 heavily active, and we're just, you know, with a coordination system, you know, hammering these one by one as we're moving it all to Rust. And, you know, it's been amazing, like, Rust, or, I'm sorry, OBS Stream takes about 4 seconds to launch up.
L164Zack: With the changes that we've done right now, it's sub-second launch. It's 0.95 seconds to launch OBS Rust.
L165[redacted: private, lines 165 to 194]
L195Zack: you know, I guess this kind of leads into the next part of what I want to talk about, which is that, um, look, I tried to do Omarchy, I… Hyprland is not for me, um, it's just too hard to set up. I love KDE Plasma, and what I did is I switched to Nix. Right now, I heard that you guys are fixing it, that's great, I think that KDE Plasma is…
L201Zack: Right? And so what I've what I've what I investigated is, you know, I've got this software factory. Like, what can I do to, like, fix this whole problem with Nix and Omarchy? And I started getting in.
L215Zack: Right, but here's the thing. I mean, his avant-garde, you know, what was it? The way that he approached the user interface.
L216Zack: with Omarchy is radical, and it's exactly what we need. The problem is, is that I want to create the world of the year of linux, and I don't think that, uh, we should have.
L217Zack: them as the default editor, and I don't think we should have Hyprland as the GUI. No one's gonna…
L219Zack: For the lower users. I mean, I don't even like them. I hate them. Like, are you kidding me? Like, like colon exclamation point quit is the way that you like. That's not going to happen. Like the year of the linux is not going to happen. A bunch of people are bouncing out because of that. And so what I said is I said.
L222Zack: And so what I did is I went on a path where I said, I'm going to start pulling the good parts out of Omarchy and start migrating them to Nix. And I started creating my own metaphor within Nix with KDE Plasma, and it's working great. It works so good.
L223Zack: that I was like, hey, I'm just gonna go ahead and fix the whole Nix problem, okay?
L225Zack: And I started studying, and I used, you know, Claude Code to assist me in this, and what I found out is that there's a bunch of weaknesses within Nix that allows it to be effectively.
L226Zack: subverted for, uh… so that it can be used by everybody. And let me talk about what that is. So, the…
L227Zack: The whole, like, you've heard about the whole conundrum with them going really radical and…
L228Zack: essentially kicking out everybody. So there's a steering committee that controls write access to the Nix packages, and they want to basically have a lock on that so that, like, fascist software can't make it into the Nix repository. But they don't control read access.
L229Zack: And the read access is located at cache.nixos.org. Any operating system that in existence can read from that cache, can pull the dependencies from that. And that's really good because those dependencies are pre-compiled.
L230Zack: And it gets around this whole glibc problem that exists in Unix, which is like everything's got to be kind of compiled for the same glibc. I mean, this is really important for NVIDIA drivers. Like you can't pull an NVIDIA driver for glibc mini linux 2.17 and then run it on glibc 2.38.
L231Zack: It just doesn't work. You got to bind to certain shared objects that are part of the operating system. And so, you know, what linux has always been is they're like, hey, let's just recompile your entire kernel, right, with the drivers built in.
L232Zack: And that fixes the whole glibc problem, right? And Nix is like, hey, we can give you the ability to reconfigure your entire system. And the way that we're going to make that fast is that we've got Nix has an entire build system.
L233Zack: that's, I don't know, petabytes of information, I'm not entirely sure how big it is, but it's… it's got a lot of artifacts. They've got this entire graph system, so you can figure out what you're gonna need from your dependency tree, and then just downloads them, and then puts them into a local cache, and so when you reconfigure your system.
L235Zack: All your stuff's already pre-cached. It rebuilds the entire operating system from that local cache dependency. It's amazing. And I want that with Omarchy because, you know, I tried to do this. I tried to like rip out Hyprland and the whole operating system wedged itself so that I had to do a full reinstall, you know, two Nix.
L236Zack: And so, what I've been working on is a new operating system called Omnix, which fixes the whole problem with Nix. Like, we still use cache.nix.org.
L237Zack: for its base dependencies. And then what happens is that you have this distributed system where the dependencies that are supplemental to that are cached throughout the ecosystem on github.com.
L238Zack: Right? And you don't even need to use your AP…
L239Nix: I'm following you. I'm totally a sentence and a half behind you, like keeping up.
L240Zack: Yeah. So, you know, GitHub allows you to fetch the release binaries without charge as long as you know where the URL is. And the hard part is when you try to query.
L241Zack: query what the releases are, you eat up API credits, and you only get, like, 60 credits for an anonymous user, you get 5,000 if you're logged in, um, and so… and it's this stupid bottleneck that they have, but if you have… and I've hit this with my software factory on the.
L242Zack: A bunch of repositories where they time out because.
L243Zack: you know, it's like 422, you're making too many requests, and so what I've designed is a system that caches all those links, you know, does a scroll, uh, kind of a scan periodically.
L244Zack: And then, um, and then basically the distributed cache is distributed out through, uh, GitHub artifacts itself.
L245Zack: and then what you… so now you've got two layers. You've got the base layer, which is cache.nix.org, and then you've got GitHub, and the reason why this is so great is because anybody, if they want to get their…
L247Zack: their code into Omnix, they just have to have their Git repository as a submitted link into the repository finder, and their artifacts become, you know, possible. And that's, like, with Phase, like, 1, for example. Like, Phase 1 is that, like.
L248Zack: You know, Omnix doesn't host anything. Theoretically, if this were ever to get backing, you know, um, then, you know, Omnix would have its own supplemental artifact cache.
L249Zack: That could then basically you get both of them, right? Like if either Omnix is going or either Nix OS has it in their cache or cache.omnix.org has its own thing. And then those two working together basically gives you the entire cache.
L250Zack: a way to cheaply bootstrapping and being able to give a fully functional operating system. And so, once you have all this stuff, like, what can you do? Well, you can offer Hyprland, and you can offer KDE Plasma. And you just have to have a testing system that makes sure that there's no, you know, problems with this. And I know that.
L251Zack: What's his face? ThePrimeagen. He's doing this right now. He's doing a bunch of testing of different things so that it all works. And, you know, that's exactly the right thing. In fact, I'm doing something very similar, right?
L252Zack: like, what works with a bunch of different combinations of operating systems and architectures, um, and then if it works, I get, like, a little testing badge. You can see that at omnixlinux.com.
L253Zack: Um, and… and so on and so forth. And so, one of the things…
L254Zack: that I want to do is just basically kind of give, like, an option to DHH, so that he can actually have a future version of Omarchy, I'm sorry, Omarchy, on Nix OS.
L255Zack: And it's like, hey, I've already done it. I've proved it out.
L257Nix: Talk to me a little bit about Zach in 2022.
L258Nix: let's talk about before you really got on. I know you said you were early in the 19s, but let's… okay, let's go back to 19. We're going to call that Zach 1.0 in terms of power.
L259Nix: An ability to get s*** done. Before it… before, like, 2018 is when you were running without all this stuff.
L260Nix: And then maybe jump to 22, like, are you 3X better? Are you 2X better? And then kind of jump to where you are now. Are you 15X better? Which I don't think is unreasonable. And then talk about… and then leave…
L261Zack: Oh, that's way, way, way more than that.
L262Nix: And so then lead into… and you can brag, because it's fact.
L265Nix: how one person yourself, like, can make an impact now, and, like, you can outrun and do better coding than Microsoft can with 10,000 people looking at it. Like, give me that whole thing.
L266Zack: Right. So, you know, I'm working at Google, which is the I can't even do a git clone when I'm working at Google because I'm not even like they've got their own sort of proprietary system that doesn.
L267Zack: really, uh, work in the real world. Uh, and so I'm, like, basically a 1X coder, right? And then ChatGPT drops, and I'm, like, doing the copy and paste back and forth to, like, get stuff. I'm using…
L268Zack: you know, um, that code completion tool, Copilot, you know, and then, um, it was either 2022 or 2023, um.
L269Zack: you know, Aider drops. And I would say that all of a sudden, I was able to do, like, 4X the speed that I was doing. And that proto tool that.
L270Zack: existed back then allowed you to basically… you add a file to the chat, you describe what it is that you want to do, like, you know, this function exists, it just doesn't do anything.
L271Zack: fill in the body. And from that, you know, one function at a time, all of a sudden I can, I can get it to run.
L272Zack: Um, and then you can get it to also, like, um, and so that basically, like, 3 or 4x'd my productivity. And then what I did is I started adding in these hooks so that after the thing, after Aider ran.
L274Zack: Um, then what it would do is it would run a test, and if the test mess, uh, basically failed, then it would use that feedback back into Aider, and then it would autonomously iterate on the function until it worked in coordination with the unit test.
L275Zack: Um, and so I basically did that and that was amazing. And I rolled that out to, cause I had, you know, I'm a director of engineering for, uh, for us from a nonprofit. So I had a small team that was reporting to me. I got them all to use it. They were like, this is amazing, blah, blah, blah.
L276Zack: Then Claude Code came out. Uh, I considered Claude Code at the beginning very, very broken. Um, and I started to, um, essentially like fix it, um, with my own sort of, um, it's my own agent that I was using on top of that.
L277Zack: Uh, you know, I mean, basically, the first thing I did was I don't want to put in dash dash, you know, dangerously skip permissions. Like, I want God mode only, uh, always, and safety third. And so, you know, I started to do that, and uh, you know, I was basically getting, you know, let's say 6X.
L278Zack: Right. And then 10x. Um, and then, um, I started, and then once I started getting really fast, I started noticing that, um, my credits, uh, were draining really quickly. Right. Like, I remember back in, you know, the sonnet days.
L279Zack: with, uh, where it was a paid program only, I would use [redacted], and all of a sudden, I was, like, using [redacted], and then it went up to the, you know, to the 5X plan, and then all the way to the max 20, right? And so at the max… and then, at that point, um.
L280Zack: And because it's reading a lot more context in order to, like, do the tasks that I've assigned it to do, um, at some point, I'm like, okay, well, um.
L281Zack: I'm running out of credits within 2 days, I need something else, I got the Codex system working, and then I realized that there's this gap in how I use.
L282Zack: Claude Code with Codex, and so I bridged them together with clud. clud's my own agent, and it basically kind of normalizes the interface between them, and it actually allows you to use.
L283Zack: Codex as a plugin to Claude Code, right? Sets up a bridge and then it communicates back and forth. OpenAI allows you to do this. And so then, you know, now I'm at like, you know, 15X.
L284Zack: But I have to keep on going faster, right? Like, faster, faster, faster. How do I set up these, like, these hooks? Like, with FastLED, for example, um, it's a very… it's like a 250,000 line codebase.
L285Zack: um, there are very specific rules I've got to follow in order to, uh, support 50 different platforms, um, and so what I realized is that the AI just constantly was screwing up. Um, this was, like, around, like, Sonnet 4 times, right? Just constantly screwing up, wasn't doing the right thing, always having to, like, watch it like a hawk, and so I started.
L286Zack: Basically being like, well, I've got to create all these rules and these post edit hooks and custom C++ linting.
L287Zack: Right? And, you know, if you ever do a C++, you know, codebase, and…
L288Zack: You know, Claude Code introduces, uh, um, an include after a namespace opening. Uh, basically what happens is that all of your symbols start having the wrong namespace and it doesn't figure out what's going on. And you walk away for two hours, you come back and it's refactored 200 files.
L289Zack: Right? And so it's like, well, you know, I've got to create my own linting system so that as soon as a post-edit hook comes, the file gets scanned, it checks to see if an include was done after a namespace, and if it does, it.
L290Zack: Fails it right there. And then the agent goes, oh, I can't do an include after a namespace opening. I've got to close the namespace first. And then it, it basically started solving these entire classes of problems, one after the other. And as I started to introduce all of this custom linting, which by the way.
L291Zack: like, these agents are so good at being able to, like, write custom linting, and I started realizing I'm not… I'm going beyond just normal, like, program linting. I'm basically doing an architectural lint.
L292Zack: And that's when I got really, really excited. And now with Rust and like all these bottlenecks I've eliminated, um, with all these tools that I've created, I've, I, I, I think I'm going essentially like a hundred times faster than anybody else. I mean, it's, it's hard to imagine because.
L293Zack: I don't think that I could have ever done any of the things that I'm doing right now, um, with Rust. It's like, is it 100 times faster? Is it 200 times faster? All I know is I've got this Sisyphean boulder that I'm carrying on the back, and a Claude code exoskeleton that basically allows me to.
L294Zack: sprint up the mountain. And every single time there's a failure, what I do is I'm like, okay, well, this is going to be a new architectural lint. And from that point on, I'm going to be able to eliminate entire classes of bugs.
L295Zack: And so the, the final boss, I'm just going to tell you what the final boss is of all of this, right? Like, because as you go higher and higher abstraction levels and you, you give yourself more and more of your tasks and delegate it to Claude, you start hitting higher and higher, like, you know.
L296Zack: levels of bugs. So, for example, with my caching system, you know, I… I used to have it all synced with locks. That sucks. I moved it to async, and then every now and then, I would have, like, Claude Code would put… would insert a bug, or it would hold open a lock between an async await, which is a big…
L297Zack: freaking no-no, right? Because you can deadlock very easily if you don't release your locks before you go into an async await.
L298Zack: And so, you know, what I've realized is that the only way to get super lightning-fast Rust builds that are correct the first time is to create a subset of the language.
L299Zack: That the agents are allowed to operate in and if they want to go beyond that there's a facade barrier that contains like for example.
L300Zack: the facade barrier would be to a Windows platform, right? All the platform intricacies exist below that level.
L301Zack: All your business logic is gonna be contained with a, um, with a white-listed, like, level of, um.
L302Zack: things that they're able to call. No locks, async only, super fast.
L303Zack: And you create that barrier. And then once I realized that, like, okay, we've got this barrier, uh, this playpen, this sandbox that the business logic can live in, what I realized at some point is that.
L304Zack: um, oh my god, I've got this internal API, and I can, instead of building my apps for, um, you know, machines.
L305Zack: machine type, I can target WASM, WebAssembly.
L306Zack: And, and that's, that's the final boss. I haven't done it yet. I'm working on it. And the cool thing about, and by the way, the, the, the code base that I'm doing, people can help me if they want, it's called kernel API. Basically what it does, it gives you the, you know, if you want to.
L307Zack: operating system socket or pipe, you're going to, you can use this kernel API. It gives blessed functions. Like for example, you can't call a function without a timeout where it's potentially blocking. You always, always needs to be able to escape.
L308Zack: And let me tell you the amount of bugs that I've had where something deadlocks because there was an optional timeout that you weren't able to give. No, no, no. You got to get rid of that. You got to make the language extremely quick. Anything that can block has to have a timeout so it can, you know, escape.
L309Zack: You got to be able to dump the stack of the program. You got to be able to take a memory snapshot. All these different things that you need exist in kernel API so that when something actually does deadlock, I can just tell my agent.
L310Zack: figure out why this thing's deadlocked, dump the stack, the off-CPU stack, the on-CPU stack, which is involved in async await.
L311Zack: and then tell me exactly what's going on. At this point, it's flawless. Like, something deadlocks, I say, figure out what it is. All of my fleet of applications all have the same entry points for dumping the stack, seeing what the async await stacks are, the on-CPU stacks, you know, for tasks that are, like, calling into the CPU to do something expensive.
L312Zack: And then the operating, the agent is like, okay, it's deadlocked in this exact same space and there's no debugging hardly except for this like one step and then immediately issues a fix. And when that API is converted into WebAssembly.
L313Zack: then what happens is that I can create business logic applications. And so instead of shipping a application with all of its binary, like, you know, hooks into the operating system, which, by the way, if you just do a hello world, it's like 4 megabytes, right?
L314[redacted: private]
L315Zack: But in reality, the business logic, it could be expressed in like 1K.
L317Zack: And the thing is, is that if you compile that for a WebAssembly API.
L318Zack: You can now have a micro-app that reuses all these harnesses. The amount of memory that it takes is ridiculously small. You know, yeah, it's kilobytes for a full-featured application, right?
L319Zack: Um, and all of a sudden your Rust compiles will go really, really, really, really, really quickly. And here's the best part. You don't have to recompile them for every single, uh, architecture out there, right? Like, you know, for, for my apps, I'm doing it for.
L320Zack: Windows times linux times Mac times the architecture times the link type, right? So, with linux, it's two types of architectures, um…
L321Zack: ARM versus x64 times musl, you know, versus glibc, right? And then if you want to have like, you know, an old glibc so it can run anywhere, that's one thing. If you want to do something that is more modern, you know, all you could have.
L322Zack: 8 to 16 different linux builds just to support a variety of operating systems. With this micro… with this microapp architecture, what you have is you've got one WebAssembly. It runs on ARM, it runs on x64, it runs on musl, it runs on glibc, because the harness itself.
L323Zack: that it exists in is a micro low-level virtual machine.
L324Nix: Is this what Java wanted to do way back when?
L325Zack: Yeah, but Java runs at slow speed and has memory garbage collection. This one is you get the full power of the Rust borrow checker.
L373Nix: Like, I'm going to take this, and it's going to become your site. So is there anything that you want it to focus on? Like, make sure that you highlight that.
L382Zack: Um, alright, here we go. Ready? Um, alright.
L383Zack: My name is Zach Vorhies. I'm known as the Google AI whistleblower. I have a software factory that I've been building myself. It's incredible. I've turned that software factory right now on Obs.
L384Zack: Studio, I've created a fork that is porting it to Rust.