The Pi Updates are insane

What's up, everyone? It's another Syntax Weekly. We're coming at you with a whole ton of stuff. Yes. Last week was a wild week for updates. Tons of stuff got updated. We're gonna be talking about both software libraries as well as AI stuff. What got updated? What got released? What all the new hotnesses? Stuff with pi, SvelteKit, OpenAI release stuff, effect, all kinds of crazy stuff. My name is Scott Tolinski. I'm a developer from Denver. And with me, as always, is Wes Bos and CJ Reynolds. What's up, guys?

Man, hell of a week. Stoked to talk about everything that's going on because some some super exciting stuff.

Yes.

Super, super duper exciting.

And, man, guys, I... I'm so excited to get into all of it. And I know, CJ, you had something up first, which was example.com.

I I need to get schooled on what the heck this thing is because I saw this in the in the show notes here. I got no idea what you're talking about. Yeah. Let's get into it. So if you're not familiar,

example.com is actually a live website.

And I believe it's owned by the IANA, the same people that handle domain name registry, like the top tier one, not individual companies, but charge of all that.

But they host example.com, and it's supposed to be used in things like documentation. And you've probably seen before, like, if you're looking at dev docs or, I don't know, something, it has a link to example.com when it's talking about websites or URLs or whatever else. Or you might see a fake email like userexample dot com. Example.com is a real domain, and they keep it live for those very purposes, for use in testing and documentation and monitoring and stuff like that. But it finally got a facelift.

I would show you what it looked like before, but the web archive is currently down, so I can't.

But it was pretty, pretty simple before. Like, now it has multiple languages listed. It's nothing crazy.

I will say that I first learned about this from Oliver Dunk.

So they actually reached out to the IANA asking what happened, why is there new content? And they replied to his email that it's a placeholder that really no one should be accessing, but people still do. So they actually updated what it looks like to help with that and clear up confusion if somebody actually goes to it. It's really funny. Wes is the top comment here. How many CIs did this break? Because people probably use example.com,

maybe even as, like, a health checkpoint or, like... Yeah. Oh, I'm sure people, like like, curl it and then check the, like, byte output to, like, as, like, a sanity check that they are getting the real Internet.

Yeah. That's funny. I I wonder how many it broke for them. And is it... I can't get to it right now if I visit it. If I visit example.com, it just says, like, DNS issues, which maybe they broke something. They broke it in Canada.

No w w w, just example.com?

Yeah. Straight up example.com.

DNS probe error.

They don't want you to have it in Canada. Must be a Canada issue. Yeah. Yes.

Unfortunately.

So I thought this was fun because...

I I know I've been to example.com a lot because they create a lot of tutorials over the years. And

so... Yeah. But... I think it would be hilarious if they just start publishing all the emails that get sent to example.com.

That's, you know, that's what I would say. Fun blog. Because, like, everybody's...

Everybody is sending all of their test data to example.com.

That's exactly... That's... They say, if you need to don't put in something assuming it's not going to go somewhere because we've all been there where you accidentally switch your prod keys or something, and then you're sending real emails.

And so what you do in that case is you use example.com knowing that that email will never go anywhere. But they should just... As a as a joke or as, like, a data collection endpoint, maybe they'll... They can go full agentic.

Just start publishing everything that people send to example.com.

Yeah. That's it. Exactly. So I went to DefCon this year. I'll have to find the paper because I don't remember the exact domains, but there there was a hacker that did a presentation.

They bought a couple of innocuous domains, very similar to example.com, but they actually listened to all of the incoming emails. And they reported, like, the number of API keys, the number of, like, just random stuff people were sending to these domains that they thought nobody was actually listening at. Yeah. So, yeah, be very careful of the domains you use and tests and and stuff like that for your internal secret data. Yes. Alright. Alright. Let's talk about the death of pull requests. This is something that is

kind of unfortunate, but I'm seeing it. I thought probably, what, four or five times last week alone.

And what is happening is is prolific open source maintainers are simply just coming out and saying, like, is... Sanjay Saurus is due to AI. I have disabled external per request on all of my repo. So Sanjay is the, like, creator of... He probably has, if I were to estimate, probably five, 600 npm packages that he maintains. And he's he's, like, probably one of the four pillars holding up the entire npm ecosystem.

And he's just simply closing external pull requests. He's open source as we know it was fun, well lost. It'd been fifteen years for me. I will still maintain projects and handle issues. Essentially, everybody is getting this SWCRS due to AI generated pull request. We are no longer accepting external PR. Sorry for the change. Appreciate your understanding. Right? Andreas Clings, works on ladybird browser.

Every open source project that matters will soon do this. If you've always wanted to get a PR somewhere, now is the time. It's just everybody is getting slopped. Hanno is closing their their pull request as well because it used to be that a pull request was golden because somebody would put some effort into solving that issue, figuring out the issue, and then sending you the code. And now that the actual implementation is essentially free, it's it's not the actual writing the code that matters. And you'll you'll notice a lot of these people are not saying we're we're not closing issues because it's more important that people submit really good detailed issues of what the actual problem is. And then the the open source maintainer can stick an agent on it themselves, knowing full well that they're... They have a little bit more confidence in the code rather than these drive by pull requests that people are getting slopped with.

Yeah. I like the change ultimately because it... I mean, you could say that going only to issues is just gonna be agents talking to each other. But I've had this issue with some of my open source repos where people make a PR and they didn't talk to me first.

And then I feel bad because I close them like this. I don't need this. This doesn't align with the project. And I can only imagine how much worse it's gotten with people thinking that they're solving an issue because the AI is writing code. So issue first makes sense because then you can actually talk about it, decide what needs to change, what should be added. And then you could just throw an agent at the thread and you could figure out how to implement it. But with the discussion first, then you're more likely to get more quality submissions.

Yeah. You know what's another really frustrating thing is that people aren't contributing back to projects. They're simply just forking them and making their own version of it. So here I've got this nimblue, which is I have this little, nimbot label printer. Right? It prints out these little thermal labels and I absolutely love it. And the the iOS app is is kind of awful for it. Right? So I figured out there is a whole open source. You can use it in in Chrome, and there's a whole open source implementation of this printer called nimblue. Right? And I thought, this is awesome, but it's not exactly what I want. So my gut reaction was, I'm just going to make my own version of it. My... So I said, let's stop.

Let's see if we can make this version better because quite honestly, I'm really excited about something right now. I'm going to fork it. I'm gonna make my own version about it. And then I'm gonna I'm gonna lose interest in it. There's gonna be a thousand other. And if it seems like the guy behind this, this repo is, like, pretty invested in it. So, like, it...

It's so hard to get people that actually care about these projects and will continue to develop them. So I'm like, let's let's open an issue. You know? Let's talk about it. So I was like, okay. Well, there's no label preview. So I have these little, like, pig labels that we print on. And I... There's a couple, like, gotchas here where it's like, can't print over the pig's face. You have to keep it in the bounds. You should be able to preview it. And then also there there is an NFC tag in the labels themselves that the printer will read.

So it knows which label you currently have. And there is an API for pulling all of the information about these labels, you know, size, printable area, etcetera. So I was just like, let's let's add this. This is a great idea. And then, like, he didn't want to. He's like, we shouldn't... I don't wanna do this. You know? Like, I I... It's serverless. Okay. You can you can you can do it entirely from the browser. They could start blocking requests at any any moment. I'm just like, damn it. Like, I really Yeah.

Yeah. You tried. I really wanted to do this. So, like, stuck an agent on it, made a a version that's 10 times better. But I tried.

Yeah.

Yeah. I I think that's the issue, though, is, like, I, psych myself out of, Well, I don't want to wait for a response or they're not going to like my idea. It's sad to say, but especially in the agentic era where I can literally just fork it, add the thing that I want, It works for me. I mean, it really is the era of, like, more personal software, but that also means there's less collaboration, there's less sharing.

So it's unfortunate. Honestly, though, like, in my personal software is decent, but it's not gonna be as good as if Right. 80 people have input on, like, little tweaks and whatever, like like, this type of thing. And then I think that's gonna be the rarest thing is the people that simply just care enough to continue on with a project. Because every... Right now, every time I search for something, I find eight GitHub repos that have been touched four months ago and and nothing since. Yeah. Because everybody vibes up a version of it. They posted on Reddit announcing blah blah blah studio, open source version of x y and z. And then a month later, it's crickets. It doesn't work anymore. And nobody's nobody's interested because they're moved on to to making motion graphics in Claude or something like that. Yeah. You know what? Paulo Paulo just says as a maintainer,

Paulo works on SvelteKit and a number of things. The fact that that now every issue comes with a PR is nice, but the fact that every issue comes with a PR is a nightmare, which is hilarious because anybody who's been in this industry for any amount time, like, yeah, you would have been so delighted if every issue came with a PR before.

It's just such a weird swap that now they're just inundated with so much sloppy hours or sloppy quests that you have to act as such like a gatekeeper of this repo.

Otherwise, who knows what's gonna get let in? And and therefore, the the most, you know, optimal solution there is just to turn off PRs. Yeah. I made a Chrome extension couple months ago called JSON Alexander,

which is essentially just like a JSON highlighter for Chrome.

We talked about it on the show previously.

But as soon as I released that thing, like, the feature cannon got shot at me. And, like, everybody wanted to add every feature under the sun.

And I I ended up just starting to ignore everything because I was just like, I am happy with the current status of it. I don't want to have to be implementing all... Taking all your slot pull requests. And, like, this is just a stupid little Chrome extension that has, like, 15 features in it, you know, let alone, like, felt.

Yeah. Something that's actually real. Yeah. You should have, like, a bot that auto replied, like, an AI Jason Alexander just been like, it's slop. Very Believe it or not, this is some good slop.

Yes. Please leave a message at the beep. Yes. Exactly.

Oh, what a weird world we are right now. And even, like, look at this. There's, like, a thousand. At the time, everybody was using it because there was no good alternative JSON formatter after the, like, one that everybody used went, like, paid.

And now there's bazillions of them. Yes.

Billions. Yes. Bazillions.

Alright. Got anything else there, Wes? PRs. Let's do the next one. Yes. Pull requests are dead. Rest in peace.

Pull requests are dead, and pi is alive and well. Pi has hit version one point o, and there's been a number of wild new features in addition to a new harness entirely.

And we're gonna be getting into all of that here. So PIE, if you don't know, is the minimal extendable harness. It's the harness that you can build and make it like you want. And because of that, it does not include things like MCP or many other things that could just be added on. Well, in a reverse, of course, pi has added MCP because today, the MCP world has changed. Right? MCP is not what it was a year ago in terms of, like... Remember back when we were first getting into TypeScript West, there was TypeScript and Flow and and and not that MCP had all these competitors, but it wasn't so ubiquitous. And now it is. No, it was like it wasn't stateless.

And like, MCP has been changed. What? Like, probably six or seven times over the last year or so. And it's it's finally at a spot now where oh, six, seven. I can't stop saying it. It's finally at a spot now where it's like, it's actually really good. Yes. Well, pi in m c... Well, pi always had MCP through extensions, but now it's it's built into pi. So, wonderful.

There another cool thing is is code mode. I'll talk about code mode in just a second.

Extension support for virtual models, deferred tool loading, so it doesn't load all, like, tool, descriptions and titles, I believe, at at at the jump. Correct me if I'm wrong about that. Cash warming for anthropic models, mid conversation system messages, a new two e theme, and full screen mode by default. The big, big billboard features here are obviously MCP and code mode. Now I I I feel like code mode is one of those things that, like, you hear about, you don't necessarily know the benefit of. And so I wanna just break it down real quick, what the heck code mode is and why you should care about it being added to PI. It's just a simple configuration turn on. But the way that your agent or PI worked before is that the agent, you know, searches, reads a file, it reads another file, it compares results, it does this or that it writes a script, right? These are the things that your agents doing with code mode. What it's doing is it's writing a small program, and it's doing so inside of a small sandbox. So it goes off, it writes a program, it does its stuff. So instead of the agent kind of doing each thing, what it's doing is is it's inside of a sandbox, it's able to do things like perform a search, read the files in parallel, filter and combine results, and then get the useful results only and then send that back to the agent. So ultimately, what you're doing is a massive amount of context saving by the amount of unnecessary things that are taking place essentially behind the scenes of a general call.

So that sandboxing is basically giving you a more efficient way of working with tools, MCPs, even things like JEV. Right? You could run inside of code mode. You could have it go off and classify a 100 things or whatever, but that's not going to bloat the context. The only thing that matters is what comes back from that. Yeah. Cloudflare was, like, I think the first to like popularize this whole idea of code mode, which was their their MCP was absolutely massive. And and what you had to do is

create a tool for absolutely everything that you could do on Cloudflare. And then Cloudflare was there, okay, these are my tools. I will call this one. And then the response came back. Okay, now I will check your dashboard. And then the response comes back and like, it's like, okay, well, these these elements are pretty good at, like, calling tools, but you know what they're they're really good at? They're really good at writing and running code. So if you give them essentially, you give them the whole TypeScript definition of here is all of the functions you can call in all of the arguments, and then here is somewhere you can run it. And then all you also have all of JavaScript available to you. Right. You know, the entire entirety of the language to run and then just go nuts and and run it all. And then like Scott says, you just simply just get the answer back. And then that goes...

It's it's much faster. It's much more flexible, and it's not gonna blow out your context. So it's great to see code mode really catching on.

Yes.

People who don't know too much about context bloating in general, the more you fill your context with, obviously, the more money you're spending, but also specifically, the worse your responses are gonna get because everything that's in context is being passed through the LLM. So therefore, you might have a bunch of unnecessary stuff in the thing you're trying to do.

So basically, you don't want your context to be too big at any given point. You want just enough so that the agent can solve the task with enough information and knows how to access that information.

And this helps that. So I'm pretty stoked about that because pi's already great, And it just got greater. Another thing that they just announced is pi durable, which is a durable agent harness.

So pi durable is a completely, I don't wanna say completely different project, but it's not like pi one point o feature.

Basically, long running durable agents that can run anywhere.

They have storage. That storage contains the conversation log. So all user messages, model responses, tool execution responses, It stores tasks.

These things are excellent because they can be marked as replay safe. They can be resumable.

So that way, if something happens and the agent work stops, there's a crash or anything like that, it can just pick up exactly where it left off on resume.

And if something is marked as replay safe, it will just replay it. If not, it won't.

You could run many things in parallel. So this opens up, like, better forking. It opens up better sub agents.

It's pretty cool. It's kind of like PI's version. Like I said, it's a durable agent harness, and it can run anywhere. You can run it on durable objects. You you know, you you can have this to build your entire agent system if you want. In fact, the Sentry has an agent called junior, which is a junior's like a Slack bot agent has access to a bunch of stuff. And I think they're rebuilding it right now in py durable because this solves so many of their issues.

So

really cool stuff. Man, I'm I'm trying to think, like, like, what what is this useful for? Like, I I understand, like, you could, like, pause and resume. Yeah.

Yeah. So I'm I'm... I have a personal assistant I'm building that's basically, like, a more locked down, less give it all your data version of OpenClaw, and I'm building it on top of PIE. So I have a web interface where I can have different chat threads.

And anytime I'm chatting in that thread, that actually fires up a PIE agent, passes it the context.

But I had to manage everything myself. So I have my own database of message history.

I have, like, try catches that are able to, like, launch an agent, and if it fails, relaunch it and, like, reinject the context. So, like, all of this stuff I basically had to build from scratch on top of Py, because Py is really good as kind of, like, the core agent because it can connect to various AIs. It has a few good built in tools. But with I can imagine replacing basically half of my code base with PyDurable. So instead of having to rearchitect all of the storage of messages and launching of agents and making sure that when they crash, they reload and that kind of thing, just use PyDurable instead, it basically handles all the extra details for me. Oh.