Wil and Matt discuss tech, startups, and building really cool things with AI. Sometimes joined by (actual expert) friends.
Matt Carey (00:00.142)
have no idea how you play the song. Hey dude!
Wilhelm (00:02.301)
Hello, we're back.
It's the end of February, kind of. It's March on Sunday. we have, it's been a while since we recorded actually. I was away two weeks, New Zealand. You've been doing all sorts. my God, yeah, we need to get the full update from you because I feel like your life has changed since we last spoke.
Matt Carey (00:24.354)
How's it changed, Will?
Wilhelm (00:26.304)
you'll find out after the jingle.
Matt Carey (00:27.512)
Hahaha
Matt Carey (00:34.414)
Are you playing it?
Wilhelm (00:35.715)
shit.
Matt Carey (00:46.722)
Yo, so I need a haircut. We can talk about my hair first. Look how long it is.
Wilhelm (00:51.473)
That's you giving crazy genius.
Matt Carey (00:55.884)
Yeah. It's like, it's also, it's a bit salty. I went for a surf this morning. Well, I tried. I couldn't, I couldn't even get out, dude. The waves were so big. I just got absolutely annihilated for like 20 minutes and was like, I'm done.
Wilhelm (01:01.607)
Isn't that?
no way!
Wilhelm (01:09.863)
Seriously.
Wilhelm (01:14.013)
When you go surfing, feel like, so in New Zealand there was some surfing going on. I didn't go out, but I didn't realize there's very different sizes of surfboards that exist. What's the one you use? Is it like one of the massive ones?
Matt Carey (01:23.171)
Mmm.
Matt Carey (01:26.727)
I... Man, just left.
Wilhelm (01:30.174)
wow, okay, that's pretty big. I love that you just have a surfboard in your living room.
Matt Carey (01:32.111)
It's not that big. It's like... Yeah, it's like 6'2", I think. It's a bit small, actually, for me, maybe. I probably need something a little bit bigger. Because the big one gives you more stability, bit more... Yeah, like it's easier to catch waves with bigger boards. And because I've done loads of other water sports, I'm always like, oh, surfing can't be that hard, it? It's actually really bloody hard.
Wilhelm (01:52.955)
No.
Wilhelm (01:59.143)
Right.
Matt Carey (02:00.142)
So normally you'd start on like a long board, or like a longish board, like eight foot plus. And I did a bit of time on that and then I bought this one and probably stunted my progression, but we'll see.
Wilhelm (02:07.133)
Mm-hmm.
Wilhelm (02:15.645)
Okay, interesting. is a smaller board harder? Okay.
Matt Carey (02:21.698)
Yeah, yeah, a smaller board's much harder to like catch the wave, to pop up, you have much less time to do everything. But you maybe need a smaller board on some bigger waves. Or it's like, not necessarily need, like if you're really good at longboarding you can do it, but it's tough.
Wilhelm (02:35.517)
you.
Wilhelm (02:40.861)
Is the smaller one, wait, I would expect the long one to be a bit harder to maneuver because it's so massive, but that's not how it works.
Matt Carey (02:49.294)
Yes, no, they are harder to turn. But like most people's problem with surfing is like the first like two seconds rather than the rest of it. If you get that far, I think you're doing pretty well.
Wilhelm (03:02.225)
I see, I see. And then one thing I saw as well, which I had never seen before is like someone was applying like, like sometimes for the triathlons beforehand, you will apply like body glide to yourself, like under your armpits. yeah, it's kind of like Vaseline. and you applied like anywhere there might be any rubbing. So when you're swimming, like, there might be chafing like around your neck or under your arms, but someone was applying that to their surfboard. It looked like body glide, but to their surfboard.
Matt Carey (03:14.542)
Like Vaseline.
Matt Carey (03:30.935)
No, it's wax.
Wilhelm (03:32.898)
okay.
Matt Carey (03:34.444)
So it's to do the opposite thing, to stop yourself sticking. No, to make yourself stick. I had no idea where you were going with that. I was actually kind of worried. what's he about to say? Yeah, look. It's like, you see how the top of the board is like a little bit mottled and like white. The board is brown, but the top of the board has got these like white mottled bits on it. That's wax. Like almost like candle wax.
Wilhelm (03:45.78)
Okay, so you apply it to the surface that you stand on. It's not to the bottom.
Yes. Yeah.
Wilhelm (04:02.727)
Gotcha. Okay, that's really good to know. I thought they were applying it to the bottom so that it gives you more float or something, I don't know.
Matt Carey (04:04.621)
It's kind of similar, a bit softer.
Matt Carey (04:11.795)
Nah, to the top. That would be like really overpowered. I don't think anyone's doing that.
Wilhelm (04:17.073)
Well, because you know how people who swim across the channel, the English channel, that what it's called? The British channel? The Euro channel? Of course. Excuse me, I have swum the Euro channel. Okay, but people who do that, douse themselves in like duck fat or something, To have more...
Matt Carey (04:24.705)
The English channel. English channel. Swim. I swim the Euro tunnel, Yeah, no.
Matt Carey (04:37.453)
Yeah, isn't that because of cold? Isn't that because of cold?
Wilhelm (04:40.855)
Okay, yeah. Could be. Is that something you'd ever do?
Matt Carey (04:44.397)
Not anytime soon.
Wilhelm (04:49.213)
Apparently there's a long waiting list for it. And then still it might get, you might wait like a year or two years and then it still gets canceled because of the weather.
Matt Carey (04:56.493)
That's crazy. We were going to kitesurf the channel at one point. We were going to do it as a record attempt. And then some guys that we know in France just did it for giggles one time. And then it kind of like, I was all hyped up to do it as a record attempt, like a speed record attempt. And we were going to get, we had boats ready, we were going to get some funding, we were going to make it a legit thing. And then like...
Wilhelm (05:17.01)
Yeah.
Matt Carey (05:25.323)
the French, like half of the French team just did it on their lunch break from, not even from the, not even the shortest part. They went from, they went from like Brittany to Eastbourne and they just turned up on the beach in Eastbourne and had an ice cream and then went back again. And it was like, yeah.
Wilhelm (05:32.029)
unbelievable.
Wilhelm (05:42.253)
Eastbourne? Eastbourne. Like, in the UK, like, right. Jesus Christ. And then they did...
Matt Carey (05:47.084)
Yeah. And they just did it for a giggle. Yeah. And because they were on these hydrofoiling boards that are super fast and they were like a French race team and they just did it for fun. And I was like, kind of of kind of knocked me a little bit there because it doesn't seem that big now if they just do it for breakfast, does it?
Wilhelm (06:05.821)
That's wild. Do you know how long it might have taken them?
Matt Carey (06:10.447)
did it? It took them like two and a half hours.
Wilhelm (06:12.397)
Okay.
Matt Carey (06:14.157)
It was absolutely rapid and it's not even the shortest part. It's actually quite a large part. It's got to be at least double the channel. No. Yeah.
Wilhelm (06:20.157)
Yeah, it's not a dover at all. That's wild. I hope they brought their passports.
Matt Carey (06:28.563)
they definitely didn't. Mostly they scarpered pretty quick.
Wilhelm (06:34.649)
That's amazing, that's incredible. By the way, was catching up with Jess yesterday. I was chatting with Jess yesterday and she was saying...
Matt Carey (06:41.431)
Dude, no. I've told you that to stop saying catching up is like a friendship thing. You can't say that. We're just circling back to our chores for the week.
Wilhelm (06:51.029)
Yeah, exactly. I was like, oh yeah, Matt and I, had like a really nice catch up on, oh sorry, we had a really nice conversation on Friday and we talked about all the stuff, all these crazy things happening in Matt's life. And she was like, oh yeah, how are they doing in Lisbon? I was like, oh yeah, we didn't really talk about that. And she's like, how can, this is such a male thing. You talk for like two hours.
Matt Carey (07:01.164)
Yeah, nice.
Wilhelm (07:19.297)
Like there's a friend of mine, I think two years ago, he had a baby and I was catching up with him for like a while after not talking for ages. And she was like, yeah, so how's the baby? Like, we didn't talk about that at all. So let me get that out of the way. How are you guys doing in Lisbon? How's life? How's settling in?
Matt Carey (07:36.526)
Yeah, we're good. We're good. good. We got flat, got like stuff arriving that left the UK ages ago, but is arriving like next week or sometime this week, I think, actually. And then, yeah, we're in our last few days of Airbnb, which is good. Although the Airbnb we're in is, yeah, it's pretty swanky, actually. It's pretty nice.
Wilhelm (07:48.518)
Nice.
Wilhelm (07:57.612)
right, that's where you are now.
Wilhelm (08:01.853)
It looks cool.
Matt Carey (08:03.051)
but you can't see the other side. The other side is literally just like double door windows out to the sea. And so I woke up this morning and I just see the ocean and people surfing and I was like, I want to be there. And then I got annihilated. Absolutely washing machined. But yeah, it's pretty sick. Like just some admin to sort out then we're pretty done. Like be here for the next month and then we go to the US for MCB Dev Summit.
Wilhelm (08:14.095)
Yeah, that's amazing. That's so cool though.
Matt Carey (08:32.587)
and then I'll be back for a bit. And then I got a wedding back in the UK.
Wilhelm (08:32.797)
Hell yeah.
Wilhelm (08:36.381)
Are you making it out to SF on that trip in the end?
Matt Carey (08:40.609)
Well, if I say yes now, this will be the third time I've promised to come to SF and not come. So I think I do want to say yes. I do want to say yes. Nothing's been approved yet, but I would like to.
Wilhelm (08:45.575)
then you have a lot. I see.
Wilhelm (08:53.947)
I see, see. OK, OK, so you're set. It's out of your hands at this point.
Matt Carey (08:59.389)
as in I haven't even asked yet. So if any of my bosses are listening to this, like, you are going to get an email at some point. Can I come to San Francisco? But yeah, I want to do an event there. It'd be really fun. I'm going to MCP Dev Summit. I'm speaking about some cool stuff I've done recently there. And that's in New York. And it'd be really nice to come back via. It's not really back via, but it'd be nice to come to CSF because I've been in the year. It's quite fun.
Wilhelm (09:04.337)
Hahaha
Wilhelm (09:09.095)
lost it.
Wilhelm (09:12.572)
Yes.
Wilhelm (09:22.193)
Bye, guys.
Wilhelm (09:26.705)
I mean, to be honest, it does feel like kind of vibe because you can just, you can fly direct from SF back to Europe, right? So it is kind of like, what's one long flight, you know? You can, you can probably even sleep better on the flight home from San Francisco than from New York. it might be.
Matt Carey (09:37.559)
Yeah.
Matt Carey (09:41.12)
yeah, the flight home from New York sucks because it's a bit too short.
Wilhelm (09:44.367)
Yeah, exactly. And they don't really, yeah, that's the whole thing, right? Like, I feel like it's like a six hour flight or something. And then it takes them like two hours to get done serving dinner. So then you have like four hours left to sleep and then they wake you up obviously.
Matt Carey (09:54.349)
Yeah, yeah, yeah, it's, it's, yeah, it's ridiculous. You can't do anything. And we need to stop talking about admin, man. We can't talk about admin. We need to talk about fun stuff. There's so much fun stuff,
Wilhelm (10:00.925)
fruit and sonic travel when?
Wilhelm (10:06.941)
There's so much fun. I do believe that there are a lot of listeners to the pod who cherish the first bit where we just talk about our lives and tune out as soon as it technical.
Matt Carey (10:14.7)
This is the reality. The reality TV side of our pod.
Wilhelm (10:20.305)
Okay, wait, last question before we move on from Admin. Weird request, but could you send me your floor plan of the flat that you got? I just love floor plans.
Matt Carey (10:24.075)
Yeah.
Matt Carey (10:32.236)
Yes, yeah, yeah, I can do that. I'll send you the Idarista, the right move page. Yeah, just all for you, all for you. no, before we go on to the other stuff, how was New Zealand?
Wilhelm (10:41.949)
Please, yes, great. That sounds great. Thank you. I appreciate that.
Wilhelm (10:52.893)
Great. Yeah, really good. Did we talk much about it? am so we do that for two weeks
Matt Carey (10:56.394)
No, we didn't. You just said you went and it was great fun and then we moved on to other stuff. But I never asked properly about what you did.
Wilhelm (11:01.18)
No, I said it.
Man, was awesome. We were there for two weeks. It's summer there, right? So it's it's great. It's nice to spend some time of the Northern Hemisphere winter in the Southern Hemisphere. And we did a bunch of It was a very family-focused trip. So it was not like exploring the mountains. It was more like get as much face time with like Jess's parents and wider family. And that was really nice.
Matt Carey (11:12.076)
Mmm.
Wilhelm (11:34.334)
Her mother is a counselor in an all-boys school, and she's just so good at her job. It's incredible. I just love talking to her about all the therapisty things that she does. It's just fascinating. So that's really, really cool. So lots of good family time. But then we did see a couple things. There is...
There's this beach town I really love called Fongamata, which is like a couple hour drive away. We went there like a couple times, including on my birthday. And they have this really, really cool store called Sonny's. And it's like any little boy's dream. Like they sell like really weird sweets, like sour, like it's like chewed tape. It's like tape you can eat and like a weird little stick you can lick.
Matt Carey (12:27.628)
I've always dreamed of like munching on sellotape.
Wilhelm (12:31.737)
Exactly. Or they sell like little toy guns or like Minecraft figurines and like weird trinkets. So we just spent like an hour in there. I became a loyalty member, got a free t-shirt. And then we spent time at the beach as well. There's like an island you can swim to. Yeah, it's like a really cool beach town. Like New Zealand beach towns I think are my like favorite everywhere because it's so, walkable, everyone's so friendly. It's super cheap. You can do also, you can surf. It's like good weather.
The UV is pretty intense. then, yeah, it's just great. Yeah, so a of New Zealand stuff. But yeah, I think in general, just life in New Zealand is pretty chill, pretty nice. Everyone's very friendly. Yeah. There's lots to love, even beyond the incredible nature in New Zealand, there's lots to love about just day-to-day life. I'll just give you one other example.
a gas station called, I mean, they have the gas stations, but the main gas station is the BP, like British Petrol. Every single BP has like the same layout and they just get the little things really right. So for example, they have like a proper cafe in every single BP with like espresso based drinks. It's so comfortable. Like one day I just, I just worked out of the gas station.
Matt Carey (13:37.504)
Yeah, yeah.
Matt Carey (13:50.572)
it makes you feel comfortable, it? You walk in, you're like, I've been here before. So nice.
Wilhelm (13:59.964)
like for like half a day because they have like great air conditioning, really clean toilets, like obviously lots of snacks and drinks because it's a gas station. And then they have like a full proper cafe with like really nice like espresso based drinks. So it's like they have free wifi power. So it's just like, it feels kind of foreign to like a European gas station vibe because the coffee there is like never good, right? But yeah, so that's just one example, like lots of beautiful little things in New Zealand, like the gas station.
Matt Carey (14:28.652)
don't know. The coffee is never good in central European gas stations. Like France, Germany, Spain. I'm just offending people left, and center, but like it's dreadful. But if you go to Italian, like auto grill, the coffee is better. Yeah, auto grill is great.
Wilhelm (14:34.533)
Okay.
Wilhelm (14:47.49)
yeah, Ultra Grill is an institution.
Matt Carey (14:51.83)
They'll have this massive espresso machine that you know has never been cleaned and you can kind of taste every single espresso that's ever come out of it. But it's kind of good in a really, really bad kind of way.
Wilhelm (15:02.917)
Yeah, yeah, yeah, yeah. That's interesting. and then also in New Zealand, I made, I told you, I'm building this, my own personal AI. I understand this is a bit of a trope now that everyone's building their own like open cloth thing. And like, it burns like hundreds of dollars of tokens just to deliver you a personalized morning report. But I'm building that. I'm building that for myself as I think should everyone. If you, if you...
Matt Carey (15:04.278)
running.
Matt Carey (15:13.355)
Yeah, yeah.
Matt Carey (15:28.288)
Yeah, I think so too.
Wilhelm (15:29.743)
Yeah, it's just like so much stuff. mean, I don't know. It's going to be very niche. It's going to be like, I feel like all of us just have really like niche needs. And I think it's hard to pitch them to like a wider audience because it's like there's something I really might care about and that it's done for me automatically or like on the reg. But that just doesn't generalize well. And I think people building these solutions that generalize, it's just, it's just
Yeah, I don't know, getting your favorite sports teams results, it's just not very, it's just not that compelling to most people.
Matt Carey (16:06.953)
On the reg. Sorry, can we just go back to on the reg? People do this on the reg. Is that a Kiwiism or is that a SFism?
Wilhelm (16:16.849)
that's a Britishism, right? Doing something on the reg?
Matt Carey (16:21.493)
Is it?
Wilhelm (16:23.549)
I think it is, yeah. But everyone should build their own personal AI. I was spending a ton of time doing that in New Zealand. It kicks off at 3 a.m. every night, does all the stuff.
Matt Carey (16:24.723)
Okay, we'll roll with you on that.
Matt Carey (16:31.359)
They should.
Matt Carey (16:36.191)
Have you tried, is this OpenClaw or have you tried Pi? Like, because OpenClaw is a bit on Pi, right?
Wilhelm (16:40.453)
Okay. So yeah, yeah. it is. Yeah. I have spun up an open claw thing, but I wanted it to be all in Python so that I can use my own proprietary Python tracing technology, COLA on top of it. so like it's yeah, PIE and open claw are like TypeScript, right? So like, that doesn't really work. I don't know. Does it matter?
Matt Carey (16:54.567)
Ugh.
I'm
Matt Carey (17:04.971)
The right language to do this in, yeah.
Matt Carey (17:12.809)
Okay. So you're like Pydantic AI or something?
Wilhelm (17:16.209)
No, no, it's all from scratch. I mean, it's built on top of Cloud Code SDK.
Matt Carey (17:21.742)
okay, so it's not fully from scratch.
Wilhelm (17:22.779)
Yeah. Correct, yeah. I think there's some, yeah, I this is actually one of the topics I wanted to talk about with you, which is that AMP blog post. mean, AMP is just doing always such incredible things. I feel like we need to get Torsten back on the pot, because the world has changed since August or whatever. But I think it's.
Matt Carey (17:26.26)
Okay.
Matt Carey (17:36.107)
Mmmmm
Matt Carey (17:41.899)
Yeah dude, was the only August we had him on.
Wilhelm (17:44.837)
Yeah, I mean, which is like an eternity, right? But I think like...
Matt Carey (17:46.251)
Whoa. Yeah, I remember we were talking about like, Sonnet 4 or something. Sonnet 4, yeah, Sonnet 4 had just come out when Thorsten was on the pod.
Wilhelm (17:52.189)
Probably.
It was a different... So I think my main focus with my personal agent right now is actually not building it in some kind of newly, beautifully abstracted way, but just getting it to do stuff for me constantly. So the focus was more on actual output and doing real stuff and basically project managing my life a little bit, telling me what to prioritize and actually executing some things I don't want to have to do myself, which is everything from like...
Matt Carey (18:08.235)
Mmm.
Wilhelm (18:24.647)
help figure out tax returns with the accountant, which is coming up in the US, to building actual simple poll features, which are very self-contained and easily testable, which are all already tracked in places. Now it just needs to be done. The GitHub issue exists. Someone just needs to write the poll request and test it sufficiently, and then maybe do a Canary deploy, whatever. So there's lots of...
Matt Carey (18:40.959)
Mmm.
Wilhelm (18:51.695)
Anyway, long point, but the focus for me with my agent so far has just been on actually doing things. But there is this interesting trend emerging with Py, right? Which is that actually you don't need something as fancy as Claude Code SDK now, maybe. You don't need all these different tools. Like the beauty about Py, right, is it has like four tools, right? Like read file, write file, bash, and one more.
Matt Carey (19:16.883)
You don't need, you really don't need very much. I think like pretty much just bash. And you're like, you're kind of sorted. It can do everything with bash, but then read and write file. think there is especially write file. It's like the patch tool is like pretty, pretty good.
Wilhelm (19:19.664)
Yeah.
Wilhelm (19:25.137)
Right. Yeah.
Wilhelm (19:33.598)
Yeah, and that means, and the reason for that is that the models have become so good at it. Like I don't think you could do this like six months ago. Like you need more tool or more harness infrastructure, which I think actually is something you've said for ages as well, right? Like be careful how big your harness is and how much scaffolding you give to the model, because at some point the model will just become good enough and...
you won't need as much scaffolding. I get in the way actually.
Matt Carey (20:04.778)
Yeah, I think for stuff where like...
Yeah, no, I fully agree with that statement. I don't have anything more to add there. I've got some other stuff, but it doesn't really fit.
Wilhelm (20:14.545)
Hahaha
Yeah, and the one interesting thing, okay, so if we want to get into this AMP blog post, it's, yeah.
Matt Carey (20:24.542)
Yeah, let's talk about that because I think it's awesome that they are able to just do this, not just a pivot, but full sunset a product that was probably doing very well in a direction that is seemingly the direction that people are going. And they're like, no, no, no, no, no, no, no, no. What we need to do, so the story is that they've sunsetted or they are about to sunset their VS Code extension.
And instead they're focusing full grind on agents, like autonomous agents that you don't need to have an IDE. You maybe don't even need to have a terminal running on your laptop. It might be like cloud agents. might be basically all types of agents is what they're focusing on. And the CLI is like going to be their next frontier. And they've had the AMP CLI for a while, but that's what they're focusing on.
Wilhelm (20:57.991)
Yeah.
Matt Carey (21:22.3)
And, but even at the bottom of the blog post, I'm pretty sure if I remember it correctly, it's like, this is what we're doing next. We don't know what the future is going to hold. It's kind of like, Star Trek quote of like venturing out into the unknown is seemingly what all of their posts sound like. And I don't know if that's Torsten's writing or they also have like quite like a Star Trek inspired theme to that blog, which I think is kind of cool, but it always makes me think of that. Yeah. Go forth and prosper. When I read their blog posts.
Wilhelm (21:49.702)
It's so impressive how fast they're moving and how committed they are to staying at the frontier. I think that's really admirable. I think the biggest shift to my thinking that this blog post did is just that we've seen, or I feel like the state of the art so far has been like opus and sonnet in that they're so trigger happy with their tool calls, exploring lots of stuff, and...
like executing a lot. But I think in there, in here, there's like hints that like actually with codex 5.3 in particular, it's.
it's a bit less trigger happy than Opus and it's able to think for longer, think through more like edge cases, explore more, and then actually give you a more fully finished solution. whereas Opus might yield back control to you more frequently, to be like more of a collaborator, I guess, it's like a collaborative approach. With the latest codex model, you...
the better you can spec up what you want upfront, the more of a complete, well thought through, and even, maybe it even made decisions about edge cases where Opus would have given you control back, but Codex just does a good version that considers the trade-offs for you, and then gives you a much more complete finished product than Opus would have done. And...
That feels kind of new and kind of great and a bit of a new direction. And it's a bit, yeah, I'm still processing this. I don't really know exactly what it means yet or whatever, but like, it's interesting.
Matt Carey (23:35.893)
Yeah, I've been experimenting with like handing off to just in my own coding. have a skill which I just, because I do most of my coding with Claude, right? I'm trying to Claude and I really like the like how Claude speaks back and I like all of that. I the vibe of Claude and I feel like I can kind of trust how it's going to act, which I think is really good. And that is like, that is the vibe they're trying to create, Anthropic, right? That I can trust.
Wilhelm (23:58.918)
Yeah, yeah, totally.
Matt Carey (24:05.886)
how I know the model is going to react based on some input. It's not going to go and brick my file system or anything like that, which is kind of cool.
Wilhelm (24:12.285)
Which by the way, Claude did for me while I was in New Zealand. It did break my file system and almost
Matt Carey (24:16.458)
did you see the meta? Okay, wait, wait, one sec. So there's a skill that I have where I'm just like, like, once you've write your plan and whatever, and then just get codecs to review it and bounce off some ideas. And then it inserts the skill and it runs codecs headlessly with open code. And then Claudio can see the response and then they bounce off ideas from each other. It's kind of mad seeing two bash processes bounce ideas off each other.
Wilhelm (24:30.672)
Yes.
Matt Carey (24:45.746)
And then you get back like something that is definitely greater than some of its parts. I'm not saying this is a world changer or anything, but it is very, very cool. I think this is maybe me building too much harness, but I think for the moment, while they definitely have different strengths, those two models, I'm quite enjoying that.
Wilhelm (25:03.803)
Yeah, no, I think that's an elite workflow. That's interesting though. So you actually get them to bounce ideas between each other?
Matt Carey (25:11.88)
Yeah, just for the planning phase, just for the planning. Like, I'm just like, build a plan, or maybe I've got a GitHub issue that I'm working off, and then I have some like, pre-conceived idea of how the thing should be done, or sometimes I don't, and I'm like, maybe you could just like, I load all of the relevant stuff into context, and then I'm like, Claude, right, make your plan, and then I have a look at the plan, but I'm also just, while I'm having a look at it, I'm like, like, get Codex to review the plan and see what it says and incorporate the changes.
Wilhelm (25:20.22)
Yeah.
Matt Carey (25:41.502)
And then I look at it, I read it again, maybe a couple of times, maybe I do something else while that's happening. And then normally the changes are all really good and they're always edge cases that Claude has like, not understood properly.
Wilhelm (25:42.076)
Yeah.
Wilhelm (25:52.733)
Totally. Yeah.
That's what, yeah, it's funny. I have a very similar workflow. It's called like second opinion. And yeah, it's like plan and code. And then the second opinion skill like shells out to the Gemini CLI and Codec CLI. then, but then it just feeds back into the plan once the way in my workflow. And then Opus consolidates like the three opinions, I guess. But then it doesn't go back to, yeah, yeah, that sounds better. Like.
Matt Carey (26:02.238)
Hmm
Matt Carey (26:07.369)
It's like literally the same.
Matt Carey (26:20.005)
Okay, mine they can just keep trying. So my skill, my skill just explains to Claude how to call another model, like using the open code CLI and just says, this is the model you should use, call it, give it this type of input, and then see what it says. And then you can reply to it. And so they bounced off ideas here. And they create like a session. Yeah, they create like a session ID. Yeah, they create their own session to chat to each other.
Wilhelm (26:29.532)
Right.
Wilhelm (26:40.709)
Okay, yeah. That's great. Yeah, it sounds like I should add the, can reply to it. Yeah. they create their own session.
Matt Carey (26:49.961)
It's really weird. I'll send you a gist or a guest.
Wilhelm (26:50.181)
Man, new things I love more than a session ID.
Wilhelm (26:56.829)
It's definitely a gist. Okay, let's not start this.
Matt Carey (27:00.751)
A gist. It might be a gist. Is it a scone or is it a scone? my most unpopular opinion on Twitter was that I like my scones the Cornish way with cream on top. Yeah, the cream on top of the jam. baby. I was going to say the meta super alignment lead that deleted all of our emails. Do see that?
Wilhelm (27:05.957)
It's not a guess.
Wilhelm (27:14.493)
the cream on over the... Yeah, that's disgusting. That's disgusting.
Wilhelm (27:27.847)
yeah. No, I didn't see it. So, okay, this is also hilarious at the moment because we should talk about your reaching a new state of Twitter fame. At the same time as I am not on Twitter anymore at all, I was like, I should give up something for Lent. Okay, what's like my most unhealthy thing that I do? Okay, it's like fucking going on Twitter all the time. So no Twitter for me until Easter for me.
But at the same time, you're blowing up. So wait, how many followers you at now?
Matt Carey (27:59.73)
like 21?
Wilhelm (28:01.745)
Holy crap, 21K.
Matt Carey (28:04.169)
Yeah, well, okay, so I did a blog post on Friday that's been like the culmination of a few weeks work. It's something that I'm really excited about. It's about like code mode, so code execution from code that the client has created. So like your agent writes some code, your MCP server executes that code. And so I'm calling it like server-side code mode. Previously, code mode was like a client-side abstraction where the client executed the code.
and I kind of didn't, we didn't see enough adoption. So I was like, why don't we just move to the server and then we could have our club for our API be served by MCP, like the whole API, all 2,500 endpoints or whatever. And I was like, that would be sick. And it turns out it was sick. I did a little demo. This took me like a day to build. I did a little demo. I built on the Tuesday, I built on the Monday, got very excited on the Tuesday, got even more excited on the Wednesday.
had loads of other work to do on the Thursday, but then I demoed it at Friday demos on the Friday and it went down really well. And so immediately afterwards people were like, right, when's the blog coming out? All of this stuff. And it took me another two or so weeks, three weeks to do the blog, but it finally came out. MCP server is released. mcp.cloudflare.com forward slash MCP. And I think the blog post you'll be able to find it on the Cloudflare blog. But yeah, I released it on Friday and then.
for like the whole day and my phone was like super hot and I had to like close the Twitter app because it's just getting annihilated by followers, which is kind of a cool thing. went from, yeah, I started the week on, I started Monday at like 2000 followers and then Dylan posted about me and then I ended up like by the end of Tuesday on like 5,000 followers. And then on Friday I was on like 5,000 K Dylan Mulroy, he made like.
Wilhelm (29:56.285)
It's our goose, Dylan.
Matt Carey (29:59.786)
He was part of the team that made Vessel domains. That's an awesome app, that. And yeah, he joined CloudFlare, I don't know, a couple of months back now. But he's doing some really cool stuff in town, actually, that I can't talk about. yeah, Dylan's a legend. Anyway, he posted about me and a bunch of other people about cool people to follow in AI, right?
Wilhelm (30:03.121)
Okay.
Wilhelm (30:09.777)
Nice, nice, nice.
Matt Carey (30:27.977)
It was like mostly, I think it was me and Sunil and some other people. And I was like, I was very grateful. Because, well, I don't know, it was just cool because I ended up with like 5,000 followers and I was chatting to I was like, oh mate, being on 5,000 followers is insane. When I think like I did like five years of like professional sport and so much effort to try to get 5,000 followers on Instagram because it was important for that sponsorship and stuff. And then in Nerdland, all you need is like one of your mates to repost you and you get 5,000 followers.
Wilhelm (30:52.283)
and
Matt Carey (30:57.801)
And then on Friday, posted a blog post and within like four hours I was on 15k. I was like, what the fuck just happened? Yeah. So now I'm on, yeah, 21. 21, 21, 21, which is pretty cool. I might have to start being careful about what I say or just not chatting so much rubbish, but I don't know. I'll try not to do that. Try to still chat rubbish.
Wilhelm (30:58.183)
Hell yeah.
Wilhelm (31:09.101)
That's amazing. And now I'm 22. Or 21.
Wilhelm (31:19.741)
Mmm.
Wilhelm (31:25.511)
So what is life like? Give us a little taste for the rest of us.
Matt Carey (31:30.105)
you know, wake up in the morning, gold spoon in my mouth, paid for by an insert sponsor here. Yeah, no, it's, yeah, I actually have no idea. I don't know what this means for me on Twitter. I have no clue. I guess like, I was gonna actually ask Sunil today because he's on like 45, right? And I'd say his tweets are still as he is as a person. He's pretty good at it, but yeah, I don't know. Hopefully I can keep it still the same.
Wilhelm (31:41.733)
Hahaha
Wilhelm (31:49.03)
Yeah.
Wilhelm (31:54.086)
Yeah.
I feel like back in like, I don't know, like 10 years ago or whatever, people with high follow accounts would do this like lazy web thing. Like they would just tweet like lazy web colon and then like some question that they can't be bothered to Google themselves.
And then people would just reply to them with like the full answers like laser web like what kind of I don't know what kind of mouse pad should I buy and then you would get like personalized recommendations Now you can now everyone has this yeah, exactly That's that's wild
Matt Carey (32:27.952)
Now you can just ask Gemini on Google search.
Matt Carey (32:35.528)
Google have democratized the high follower count.
Wilhelm (32:38.555)
Well, all that I ask is that you remember those of us that followed you back in ye olde days. The London DevTools community.
Matt Carey (32:46.652)
Oh, dude. Dude. No, but what is kind of mad, right, is like that blog post is actually, like the blog is pretty badly written and quite rubbish. That's my fault. But the actual idea is fucking sick. It means, yeah, gone.
Wilhelm (33:01.702)
It's really cool.
No, no, sorry. You say this first. I have some questions. But yes, it is really, cool.
Matt Carey (33:07.674)
No, no, no, go on, no, hit me, hit me. I'm just turning the light on because I can't see anything. there we go. That's bright.
Wilhelm (33:12.517)
I think code mode is really, cool. I don't think I quite clocked that it's server side. Does the client still generate the code and the server just executes it, or is there some prompt that happens on the client and then there's infrastructure server?
Matt Carey (33:26.344)
Yeah, exactly. No, no, So previously in the summer when Kenton and Sunil released the first version of code mode, the idea was like, if you were building agents, you would implement this like code act pattern, like code execution pattern where the user would tell you something and then based on what the user told you, it would write some code that could call like some SDK or something.
and then you would execute that code safely, but it was all to do with how you build agents. Like, MCP's just been shouted at as like a context problem for ages. And I was just like, well, this kind of solves the context problem on the client, but it is actually, it's so hard to implement because if you think most MCP clients are like CLIs, if you think of like how, like,
Wilhelm (34:18.97)
Yeah,
Matt Carey (34:23.884)
Claude code, for instance, if you wanted to do code execution in Claude code, you're there are some there are some options now like, Pynantic just built one that runs in process, but there wasn't really this like code interpreter option, like REPL basically, that they could make one but you end up increasing the size of the Claude code binary a lot it becomes much less portable they often use like
Yeah, just like the portability of it all goes out the window. And so what CloudFlare did was they were like, we have this primitive that is called dynamic worker isolates and you can run, you can basically run a CloudFlare worker from a string. So a string that a model generates that is code, you can run in a fully sandboxed way. But like, no one is gonna add that to all of that to their clients unless they're building a client on CloudFlare.
Wilhelm (35:18.173)
Yeah. Right.
Matt Carey (35:18.556)
And so we're like, yeah, sure, we're going in the direction we want to go. We want to get more usage and all that sort of cool stuff. But like, it's not going to see mass adoption yet. Like, I think it would, the blog post was a little bit esoteric and people then maybe didn't quite understand. we were literally just saying there's a better way for you to do code act, which is like a blog post that out ages ago or like hugging faces.
Or was it that Schmalt code or something was another one that came out? Basically, there were these blog posts that came out like two years ago when models started writing code kind of well. That was like, just let the model write code. It will do everything and you can make agents, like code agents basically. And then Sunil and Kenz took that and were like, we could make MCP better by generating an SDK from the MCP server and then giving that to the model. And then the model can do code act.
Wilhelm (35:46.695)
Mmm.
Matt Carey (36:10.405)
And I was like, yeah, that is also, that is really cool, but it requires like architecture in the agent. Like what happens to everyone who's just running Claude code that's not going to implement this anytime soon because they have to call out to an external API if they do that. Like, how do we do that? Well, why don't we just, why don't we just put the code execution on the server? And so that means the Cloudflare API is like the Cloudflare MCP server just works so well and anyone can use it. Like the most simple agent loop ever.
that generates code and you can use the Cloud for MCP server. You don't need like to work out how dynamic loaders work. You don't need to work out how any of stuff. And then the server developers, the people who like, the people who genuinely want to make it very easy for people to use their product, they can invest the time to work out how all of that works. They can invest the time in their, in their infra to make like a good sandbox or use dynamic workers and...
Wilhelm (36:45.275)
Yep.
Matt Carey (37:05.499)
Yeah, like they can invest that time. Whereas they're like making every... You sort of have three people when you work with MCP. You have like three customers. You have like the server developer. They're like super invested in giving this out to as many people as possible. Then you have the server user, like the agent user basically. And they can be anyone. Like such a broad section of society, like from...
Well, for tech, right, there's some very non-technical people playing with it. There are some hugely technical people playing with it. There's some very, like, non-AI native people playing with it. Maybe it's preloaded at their work, you know, like all of this sort of stuff. And then you have the client developer, who is a completely different person, and you kind of have to have all of these three people working together in a way, and they don't know, all these three people don't know each other. And so that's like how,
Wilhelm (37:53.565)
Yeah.
Matt Carey (37:59.036)
That's the beauty of a unified protocol. And that's why I'm really bullish on MCP, because at some point, they're going to have to decide the lines of communication that they're going to chat around. But what we did was took some complexity from the client's developers, which clients are already stupidly hard to build, and we just moved it to the server.
Wilhelm (38:16.509)
That makes a ton of sense, well, I mean, there's a lot of cool things about this. Like guess on the one hand, you're also like advertising this pattern by having a sick Cloudflare MCP that just works for anyone who wants to use Cloudflare, right? So you advertise code mode in a way. You advertise that people can build this into their clients if they really want to.
Matt Carey (38:39.847)
Yeah, and don't get me wrong, I think the end state of this is that everyone's clients will support code execution. Like, don't get me wrong, I'm 100 % sure Sunil and Kenton are right. I just think they're probably a year and a half too right, or like a year and half down, maybe two years too right at the time.
Wilhelm (38:55.965)
A year and a half too right is amazing. It's an amazing statement.
Matt Carey (39:01.703)
I think that, no, I think there probably are because there's quite a lot of, you need to have a lot of faith in your sandbox to just be like, to the model, to be as a client developer, to be like, we're going to go all in on our models generating code that we run. Like, there's this whole new world we're entering of that, which didn't exist before. And it didn't exist before because if you ask any, like, like,
Wilhelm (39:08.987)
huh.
Matt Carey (39:29.507)
I don't know, developer, would you let your customers generate, would you let your customers write code that you just execute without looking at it? They would tell you, obviously not. Like there's whole companies that have been built, and I kind of know because we did some of this at Stack One, previously, around building like DSLs, like domain specific, like almost languages, like some markup scripts, like all of this stuff was built. Terraform, like, you know, like all of this stuff was built because...
these declarative things because we don't trust people to execute code. And Terraform is maybe not a great example, but there's so much of this stuff because code was expensive and people built wrappers so they could control the output of code. But if you can steer the code generation via a prompt or something like that and you...
Wilhelm (40:09.789)
Totally.
Matt Carey (40:26.043)
and you can execute it very well. And the thing that's generating it has got to like a level of control where you're pretty confident with the steering and you can put enough like permissions and guardrails in place from both to protect both the user and to also protect your system. So dynamic worker loaders. Then why can't you get all of your customers to generate code? And why can't you run code on demand whenever? Why can't, why can't.
Different customers to your website see different websites to suit them. Like previously you'd have like you do things like with a CMS and you like a D test or whatever and you like inject different different different static assets. But like why can't you just have different bits of code running? Well, that gets tough.
Wilhelm (40:57.539)
Yeah, totally.
Wilhelm (41:04.317)
Yeah, yeah, yeah.
Wilhelm (41:11.579)
Yeah, it's It's interesting because one, yeah, no, I think the vision you make is very, the vision you kind of sketch out is very compelling. It's, and you can see how, and I think in a way this is part of the promise of like AI, right? That like, even if you have like, even if you are a customer of some company and your use case is like really, really, really niche, like it's not how most customers of that company use the service or the product.
you're kind of always a bit fucked, right? Because you're just not the priority. Like any feature request you have, it doesn't help most customers. So like it'll always be deprioritized. But now, like if AI makes the stuff really, really easy to build, maybe we can actually solve the really, long tail of edge cases and niche use cases really, really well, which I think is kind of glorious. The one thing it does make me think of, which I'm sure we'll figure this out, but it's kind of interesting how like it loops back.
I saw a little bit at GitHub, what it was like to run a public graph QL API from the inside and graph QL had some of these aspects, right? It's not like single get endpoints. It's like, you can craft the query, but turns out like basically every week, someone had managed to craft a query that was so like complex that it like broke the database in like a new way.
Matt Carey (42:32.676)
Yeah, so when I published the Komo blog post, it's really funny you mentioned this, I got a fair few comments being like, wow, you just created, you've recreated GraphQL for MCP. Wow, look at you go. And I was just like, yes, maybe I did. But also I 100 % didn't because under the hood are like predefined rest endpoints with permission sets with like,
literally with a token that only works for a certain set of endpoints with like everything is like role based. don't know. Like there is, there is a different, there is a, it's, it's, it's tractable how different they are to what I did. I just meant the interface on top of the rest API is now very fuzzy, but it's still a rest API.
Wilhelm (43:06.801)
Yes.
Wilhelm (43:18.683)
Yeah, yeah, I'm with you. I'm with you. I think that is such a good distinction because the problem with GraphQL, right? And actually some of the GraphQL queries and mutations at GitHub were powered by rest endpoints underneath. But the problem was that they weren't actually hitting the rest endpoints, right? Like GitHub took on that complexity in the backend of translating GraphQL to rest, which meant also the rate limits and all this stuff had to translate in like awkward ways, right? So like...
Matt Carey (43:32.122)
Yeah.
Matt Carey (43:44.826)
Yeah, that's crazy. That's crazy. Yeah. Yeah.
Wilhelm (43:46.278)
That's not trivial. That's so every week it was just like, I think it was this guy, Robert, who was like writing all this. He was spending so much time just like, building new data fetching patterns into the back end just so that you could fetch that without taking down the database.
Matt Carey (44:00.231)
and
I would say that GraphQL in that way is such an anti-pattern and is actually reminiscent of the DSLs that we're trying to get rid of. We're literally trying to get rid of some DSL as much as possible. I don't care about your markup language. I literally don't care. I just want the model to write TypeScript, Python, whatever, you know? And then I want to...
Wilhelm (44:13.574)
Yes, I am.
Matt Carey (44:30.49)
have a very clearly defined interface under the hood of what the model can and can't do. And then the model can just like, the agent can just arrange those requests however it wants to do it. if the fact that it can arrange the requests like that means that we get much better context or much lower context utilization because the TypeScript acts like a mini plan for it. Like a very compact plan, Like you can do...
Wilhelm (44:55.911)
Yeah, yeah.
Matt Carey (44:58.968)
five or six requests in a map and then like drill down into them and get stuff out and return it. And you can only do that because the model has the types. It knows what it's going to get back and then it can do that. But it's still working with fucking rest endpoints. Like it's you're not taking it out of distribution. You're not trying to make it. It can't. It can't. So the giggle is I even did this for the graph QL API. So the cover API contains a graph QL API and
Wilhelm (45:02.823)
Yes.
Wilhelm (45:11.537)
That's really cool,
Yeah.
Yeah, yeah.
Wilhelm (45:23.631)
Mmm, whoa.
Know that.
Matt Carey (45:28.856)
Yeah, it's the analytics API for like HTTP requests. And so the model can write GraphQL inside TypeScript to query the GraphQL API. It can write like a query or mutation or whatever to query the GraphQL API. But like, it's such a different feeling to what we're doing. And everyone was like, you're just like, your search endpoint is just like, how is that different from introspection? And I was like, my.
Wilhelm (45:32.884)
fair,
Matt Carey (45:56.187)
God, the fact you used the word introspection in a sensible conversation.
Anyway.
Wilhelm (46:04.431)
Interspection, what a throwback. Yeah, no, this is really, cool. And congrats on the ship. Like, how cool. You deserve all those followers and more.
Matt Carey (46:07.387)
now.
Matt Carey (46:13.318)
Thank you. Jesus. I wonder what percentage it bought. I reckon it's so high because it got a lot of, like the blog post itself got a lot of traction. I think like over a million views just from my Twitter post, which is kind of cool. And then lots of big accounts reposting it. mean, was on my news section of Twitter for like two days.
Wilhelm (46:23.665)
Nice.
Matt Carey (46:42.296)
Up until yesterday afternoon, I think, it was on my Twitter news page.
Wilhelm (46:46.609)
Yeah, damn, nice. That's very good. I'm just looking at what other stuff I want to go through. Can we do a quick, what's a sparking joy in your life and what's not sparking joy?
Matt Carey (46:50.34)
It went well. It went well.
Matt Carey (47:04.31)
Sparking joy. dude, there's so much fun stuff happening. Like for me, moving country is a massive deal. I know you did the same last year and it takes like quite a long time to get back on your feet and like understand that I still haven't moved into a house with a bed yet that is mine. Like I'm still in Airbnb, know? Like all of this stuff takes a lot of time. Yeah.
Wilhelm (47:05.981)
I
Wilhelm (47:14.461)
huge.
Wilhelm (47:26.109)
Took me ages to be honest, much longer than I expected. It took like, I think six months. I thought it would take like two weeks.
Matt Carey (47:32.57)
And like, well, like the last two weeks I haven't been as productive as maybe, well, maybe the last week I was pretty productive because I was just in London, like grinding. But the week before I was definitely like much less productive than normal. Like the week before that definitely as well. Just sort myself out and like getting ready. But the payoffs are to be so good. Like it was so beautiful today here. The weather is amazing. I have an open core meetup tomorrow in Lisbon, which I'm very excited about. I'm going to...
Wilhelm (48:01.115)
open.
Matt Carey (48:02.921)
OpenClaw is a meetup happening in Lisbon tomorrow. And so I get to meet some of the, yeah, get to meet some of the deaf community. There was a guy from SuperBass who was at my meetup that I did in London last week. So we're going to organize something. I don't know, there's some like, I'm really excited to meet more people around here and have friends, you know? And the office is awesome. We've been climbing a bunch together. And I think that there's...
Wilhelm (48:03.869)
yeah, no way! The lobster shall reign.
Wilhelm (48:19.613)
I'm gonna sick.
Wilhelm (48:25.031)
Totally. Yeah, yeah, It's huge.
Matt Carey (48:32.131)
That sort of thing is sparking joy, like a new place. There's a bunch of AI stuff, but like, that's like the really fun thing.
Wilhelm (48:40.221)
That's awesome. Can I just ask, can you give us a very quick recap of demo days last week? I wasn't done.
Matt Carey (48:40.997)
What about you, dude?
Matt Carey (48:48.167)
yeah. Mate, it was awesome. Okay, so who do we have? We had Fred from Alpik demoing SkyBridge, which is an open source thing for deploying MCP apps. We had Max Drake from TL Draw. I'm just going to run through them. Demoing fairies. I don't know if you've seen a demo of fairies on...
Wilhelm (49:10.333)
Is it still live? I haven't. I saw it when it was released in December, but then did they shut it down or is it still around?
Matt Carey (49:15.397)
Dude, I don't know. They said they were going to shut it down. I don't know if they did. It's like agents acting on the canvas and every agent has a terminal and you can SSH into their terminal. They could chat to each other. It was wild. Like all of the stuff that they've been doing. Yeah, it was crazy. like Steve's running like project trackers and stuff. He can build any app that he wants on the canvas with his primitives. It's kind of, it's just wild. like under...
Wilhelm (49:22.054)
Yeah.
Wilhelm (49:26.773)
my god, that's amazing.
Wilhelm (49:31.613)
Incredible.
Matt Carey (49:43.034)
The reason, I'm just going to talk a little bit about more of this. The reason why I love it so much is I know under the hood, TLDRAW is a durable object. And so I know what Steve is doing is just a better version of what we're doing with code. And that just makes me like one, it just makes me very like happy and impressed and like, there's a warm fuzzy feeling inside to know that he has like mastered what we're trying to do with code and developer SDKs.
Wilhelm (49:50.737)
and
Wilhelm (50:03.568)
Yeah.
I didn't realize this was the case, that it's built on top of durable objects.
Matt Carey (50:10.629)
Yeah, well, I think TailDraw is Cloudflare all the way down. Like Dribbble Objects for the canvas. Ferris has a bunch of sandboxes that they use. Man, it's super cool. Like, it's just, every time I see their demos, I'm just like this, the, yeah, magnificence of what they do is crazy. Yeah, it's beautiful, it's beautiful. And it's all just like, it's got such style and like...
Wilhelm (50:15.238)
Right.
Matt Carey (50:38.981)
I know, I really want to say like, it's just very cute. It's very cute and cuddly. And I love that for them. Anyway, sorry, that was massive sideshow. But then just after that, we had Nourish actually from my team demoed sandboxes. So was like a really good segue. And he's been doing some cool stuff with sandboxes. he is Mr. Sandbox. And like he definitely doesn't get enough credit for the crazy stuff that they do. They shipped...
Wilhelm (51:00.743)
Nice.
Matt Carey (51:05.565)
PTY, which I don't know what it stands for, but it's like terminal, like a hosted terminal that you can just open on a website. Yeah, exactly. And you can just like open a website that your sandbox is running on and you just like have a terminal, which is fucking cool. And you can also watch like what an agent does on the terminal. You can watch it. Like it's just like, it's mad, like some of the orchestration you can do. then we have Fixer, which is like a...
Wilhelm (51:11.333)
like a pseudo teletype.
Wilhelm (51:19.805)
Yeah, that's cool.
Matt Carey (51:34.593)
superhuman productivity tool, demoed some draft stuff, like how they generate drafts, that was kind of interesting. And then to wrap it off, we had Samuel from Pydantic, like another London DevTools, and it was another sandbox, but like a runtime sandbox that he made, something called Monty, which is like built in Rust, Python interpreter, ID, like it's not built to run code that you write, it's built to run code that you generate.
Wilhelm (51:45.574)
Mmm.
Wilhelm (51:54.187)
I saw this. Yep.
Matt Carey (52:04.202)
I think it's a good angle. He should definitely, I'm sure he is going to keep going with it, but it's a great angle. There's going to be so many local process, local stuff where you really want, wait one second, wait one second, wait a second.
Wilhelm (52:04.53)
Nice, yep.
Wilhelm (52:17.093)
Yeah, yeah, yeah.
Wilhelm (52:26.161)
these hold color.
Matt Carey (52:44.164)
I'm back, I'm back. no, yeah, Monty was really cool. The cool thing about building like a sandbox runtime for code that you generate is that he can do lots of really fun things with the, it's not an AST, but it's kind of an AST in Python because they run Pydantic, so.
Wilhelm (52:45.095)
We back.
Matt Carey (53:09.026)
He can do all of these, I don't know, you know more about this stuff, but he can like see like if they don't return something, he can add a return. And so he can just fix a lot of the issues that models make when writing code. I don't know if he should invest too much time into that because that feels like building a harness for a model that's going to get better. like, it's still cool because he can really get to a quite a high level of reliability. Well, that's the idea anyway. And it follows, yeah. But it's brand new.
Wilhelm (53:15.387)
Mm-hmm.
Mmm.
Wilhelm (53:33.881)
Yeah, that's interesting. I didn't realize they do that stuff. That's cool. Yeah.
Matt Carey (53:38.401)
And it's a similar idea. I'm so excited about sandboxes. Like I'm turning into Kenton. like, mean, not as smart as Kenton, but...
Wilhelm (53:43.776)
yeah, go on.
That's interesting because I'm not very sandbox-pilled actually. So what's the pitch?
Matt Carey (53:50.597)
There's lots of different types of sandbox. And I have a blog post that maybe I'll write about this sometime. You have the VM, the micro VM, which is basically a computer that you can do whatever you want to do in the sandbox. that's like, if you really need file system rights, if you need bash, that's what you should go with. And then you have something slightly in the middle where you have some state to it.
Like basically you have a durable object that sits somewhere in middle and it's kind of a bit of a strange sandbox, but it's, it is still a sandbox, right? Like, like it's running, it's running its own code inside V8. It doesn't have access to a file system. It has access to a SQL lite API that you can give it. Like it has some sort of the stuff of a full VM, but it doesn't have everything. And then like moving a little bit further towards the bare bones, you have, you have like a, like a worker.
Wilhelm (54:43.783)
Mm-hmm.
Matt Carey (54:49.848)
like a Cloudflow worker or a Lambda function or even for cells like fluid compute thing is like kind of a sandbox in its own way as well. Like, especially a Cloudflow worker, you have access to nothing on the underlying host, like absolutely nothing. they, you have no file system access, you have literally nothing. So there's that. But then if you go that one step even more wild, it's just like.
Instead of having a script that you deploy, you can just run a script in a sandbox. that script has no access to a file system. Actually, you don't even save it necessarily in a deployment thing. It's just some code that gets executed. And that's the dynamic worker. And that's the furthest to the lightweight side.
Wilhelm (55:23.983)
Yeah, yeah.
Wilhelm (55:39.236)
Okay, I understand.
Wilhelm (55:44.933)
Yeah, Okay, I see.
Matt Carey (55:45.283)
and that's what Monty is as well, but it's local.
Wilhelm (55:49.32)
Got it, got it. Okay, no, I think what you're saying makes sense. There's lots of cool things to do with sandboxes in general, but I think my gripe is not that the sandboxes are too heavy, but that they are like, like I feel like the biggest cool thing at the product layer we've seen with AI in the past, whatever, year, is OpenClaw, right? And the, yeah.
Matt Carey (55:49.614)
So Monty is similar.
Matt Carey (56:08.91)
Do really think?
Matt Carey (56:13.784)
Seriously.
Wilhelm (56:14.683)
Wait, why is that controversial?
Matt Carey (56:17.796)
The coolest thing we've seen in the last two months is OpenClaw. Yeah, 100%. It's not the coolest thing we've seen in AI.
Wilhelm (56:26.225)
mean, at the product layer, I think we've seen cool models. We've seen cool output for models. But in terms of...
Matt Carey (56:31.874)
Okay, come on, come on. So, Claw Code was like probably the fastest-grossing product to however many billions of ARR. And you're gonna still say that OpenClaw is the coolest thing you've seen?
Wilhelm (56:45.693)
To me, you make a good point, but I think you can say yes, because mean, OpenClaw is just Claude code, but with more power, right? Like OpenClaw can just do, it is Claude code.
Matt Carey (56:50.574)
You
Wilhelm (57:03.035)
Basically, like it can write code, can write its own skills, but it can also restart itself so that the skills are immediately loaded, something that code code doesn't do. And it has a rich ecosystem of tools where you can talk to it via WhatsApp, something code code can't do. It's a superset of code code. Yeah.
Matt Carey (57:16.468)
Yeah, it's Claude code. All of these things are something plus an orchestration thing. Like it's Claude code plus some orchestration. But like as Claude code is a superset of like the AISDK or something like or Pi or something like that. Yeah. Yeah. But each one was interesting.
Wilhelm (57:28.891)
Yeah, I think it's a virtue of being the coolest thing because it is newer. No, I agree. mean, it doesn't, it's not really my main point that it's the coolest thing. My main point is that it is a really, really cool thing. But fundamentally it's going the other way of sandbox. I feel like the big unlock where all of the labs, think,
Matt Carey (57:38.052)
Okay. I was just a bit shocked by that. It is cool. It is cool, but like, it's not...
Matt Carey (57:48.878)
Wait, how is it the other way? How is it going the other way? It's literally going the way of sandbox. How is it going the other way?
Wilhelm (57:55.706)
No, it's more access to more stuff and unlocking new youth cases without a sample.
Matt Carey (57:59.32)
Yeah, which is sandbox. But how do you do that? No, it's with a sandbox. How are you running cloud code on a Mac mini? That is a fucking sandbox.
Wilhelm (58:07.965)
But how is OpenClaw, but a sandbox implies restrictions on a Mac mini.
Matt Carey (58:13.887)
Sorry, how are you running OpenClaw? No, how are you running? No, it doesn't. A sandbox requires like a separate compute requires a separate compute requirement. It doesn't necessarily require like restriction. Like it just means that you're defining like what you can do in and what you can do out. Like if you run your your OpenClaw on a Mac mini, why are you running on a Mac mini and not on your laptop?
Wilhelm (58:39.867)
so that it's always live, always available.
Matt Carey (58:43.648)
Yeah, okay. So you're running it on a machine. You're running it on a VM. Some people call VM sandboxes.
Wilhelm (58:46.045)
But to me, like sandbox. Nah. Okay. I mean, maybe we have different definitions for this, but to me, a sandbox is a VM plus restrictions or more niche types of things like a worker or like whatever.
Matt Carey (59:00.675)
It's a VM that you don't care about if it gets bricked. Well, you actually do because they have sandbox snapshots.
Wilhelm (59:04.111)
Okay, but you do. But you do care if it gets bricked on the Mac Mini, right? It's not a throwaway thing.
Matt Carey (59:11.171)
Why? You just reinstall Mac Mini. You just reinstall like Mac OS.
Wilhelm (59:18.173)
I don't think this is true for like most people or 99 % of people. Like you have files on your laptop that can get lost. Like it's not just a server, it's not cattle, right? Like, but my point, right, is that like...
Matt Carey (59:23.117)
Okay, so, okay, so.
Yeah.
Matt Carey (59:32.737)
No, no, but it is just a server. The reason why people are using a Mac Mini is so you can log in to all of your stuff like WhatsApp that is kind of hard to do on a headless server. But that's like the furthest. OK, so if we were talking about VM versus dynamic Isola or like like an interpreter like a Mac Mini is like over here somewhere, like on the VM side, like past the VM.
Wilhelm (59:39.345)
Correct. Yeah.
Wilhelm (59:56.062)
OK, so the scale that I'm talking about is like your agent has shed tons of access to shed tons of stuff. And the other end of the scale I'm talking about is you.
Matt Carey (01:00:04.567)
Yes, agreed.
Wilhelm (01:00:10.269)
run untrusted code safely, lots of restrictions, no network access, like kind of like a Dino, like lock the runtime down, lock the network down. And I think the open claw, the most interesting thing in the past year, as I've claimed, is like goes all in on full access, internet access, like no worry about the lethal trifecta, the model decides.
Matt Carey (01:00:29.545)
Nah dude, nah, nah, nah, nah, nah, nah, because, okay, they're literally the same thing because like what I call a sandbox, like dynamic worker isolate, they're all, all of these things are sandboxes all the way down because dynamic worker isolate, if you want to give access to the internet, like that is fine.
Wilhelm (01:00:44.071)
But wait, so you're saying everything is a sandbox? Like my laptop's a sandbox, like everything's a sandbox. If a sandbox is just a computer, the word doesn't really mean much.
Matt Carey (01:00:48.195)
Your laptop, okay, your-
Yeah, but that's a sandbox is a computer with varying levels of access control. Something that you that's at least that's the way I'm thinking about it. I mean, it to probably has a description, but like it's a computer that if it got wiped and restarted, I wouldn't care about that much. That's like if you have VirtualBox or whatever, the people use that VMware on your laptop. You can run an OS inside and you could run that daily and you might care if it gets bricked, but
Wilhelm (01:00:58.597)
Okay, cool.
Matt Carey (01:01:22.751)
It's not going to brick the underlying host. I think that's like.
Wilhelm (01:01:27.229)
Okay, but what made OpenClaw powerful is that you ran it in dangerously pipe-bass permissions and you give it access to your email, all this cool shit. It didn't become powerful from adding network restrictions or putting it in VMware or...
Matt Carey (01:01:32.664)
Yeah.
Matt Carey (01:01:39.779)
I'm not saying to add network restriction, I'm saying like you... It's a, yeah, okay. From a sandbox you have like a server on one side. Because at every level something is a sandbox, right? Like you're not like playing with, it literally is like how low do you want to go? Dude, like...
Wilhelm (01:01:50.718)
But, but okay. The way I grew up, like sandbox is like literally it's kids playing in a sandbox. It's like a safe environment where you can mess shit up. That's what a sandbox is. So it's like a separate thing from your like re from the real world, from the real environment. It's like a lockdown thing.
Matt Carey (01:02:11.425)
If you're, okay, your Mac Mini, your Mac Mini, if you don't give it your banking details, it doesn't have your banking details. Your laptop, if you run it on your laptop, it does have your banking details. That is why your Mac Mini is a sandbox.
Wilhelm (01:02:21.977)
it does have my banking details. And that's why my...
OK, but no, no, but I think people are using a Mac Mini as opposed to a VM in the cloud because it has access to their iMessage and because you can sign into Chrome.
Matt Carey (01:02:35.831)
Yes, 100%, 100%, but there's still a signing things that it has access to. It's still a sandbox.
Wilhelm (01:02:43.675)
Okay, I would argue that people choose a Mac mini because you can easily give access to more stuff.
Matt Carey (01:02:50.893)
Yes, it's just a user experience thing, but it's still a sandbox. Functionally, it's still the same.
Wilhelm (01:02:57.309)
I mean, I'm signed into everything on my Mac Mini. that's like... Okay. I understand what you're saying, but I don't know. I feel like most people use the definition of sandbox as like a... You use a sandbox when you want to lock things down. Like when you want to restrict.
Matt Carey (01:03:00.771)
And that is entirely your choice. If it wasn't your choice, it wouldn't be a sandbox.
Matt Carey (01:03:22.359)
No, I think you want control. I think you need the control,
Wilhelm (01:03:26.929)
But you can get control on your laptop as well, right? Like you can...
Matt Carey (01:03:31.565)
Yeah, how would you do that?
put it in Docker, in a sandbox. Yeah, if you want to. If you run something on unfettered access on my laptop and then it bricks my machine, that's gonna be very annoying.
Wilhelm (01:03:35.502)
You... Yeah?
Wilhelm (01:03:41.196)
okay.
Matt Carey (01:03:50.401)
But like it's, but, I also don't give it express permission. If I run it as root on my laptop, I don't get it express permissions of like stuff that it can access. It could access all my passwords, whatever, like anything.
Wilhelm (01:03:57.435)
Okay. Let me just say.
I got the Mac Mini so that the thing can always run. And I gave it the Mac Mini so that it can easily access all my stuff. I didn't get the Mac Mini so that I could be intentional about what I give it access to.
Matt Carey (01:04:05.334)
Yeah, that's because you got it as a server.
Matt Carey (01:04:16.898)
Okay, okay, so are you signed into every... Okay, that might not be why you didn't... Okay, they're the same thing, dude. It's literally the same thing. You're always intentional about what you give it access to, even if that intention is to dangerously skip permissions.
Wilhelm (01:04:23.813)
Hahaha!
Wilhelm (01:04:34.339)
Okay, sure. Yeah.
Matt Carey (01:04:36.138)
We're too early in the curve to be like, mate, like, I don't know what it has access to. Yes, of course you do, you set it up. If in the future there'll be like something that runs externally to you that you don't know how it works and like that thing has an entire knowledge of your life and universe and everything, then yeah, that's not sandboxed, whatever. But I'm not entirely sure even.
Wilhelm (01:04:54.077)
I'm
Wilhelm (01:05:00.263)
This is great because we now have a title for this episode. It's the one where we fight about sandboxes. I think these days you have to make rage bait titles to get clicks. So that's a great one.
Matt Carey (01:05:12.576)
Yeah, don't know if fighting about sandbox is the most interesting thing in world, but like, you don't agree with me that all of these different types of machine are a sandbox in one way or another. It just depends on what you want to, what you could be bothered to restore when it breaks.
Wilhelm (01:05:27.869)
Yeah, let me think about this. I just, like, I guess, you know, I agree with... Yeah, yeah, go on.
Matt Carey (01:05:30.819)
Like, like, OK, for instance, for instance, I've got like a Raspberry Pi. I have a Raspberry Pi. Like my Raspberry Pi is is a sandbox as far as I'm concerned. Like I can shut off Internet access. I can only load. I can only load the email accounts that I want to load. I can only install the CLIs I want to install. Yeah, sure. The model could install more. But like I think.
The point is, if you can escalate your permissions, then maybe you're not necessarily sandboxed that well. What is it, if like there's, yeah, I don't know. Like that permission, the permission side of thing is like different. I'm genuinely more concerned about like, if it all gets bricked, is it going to stop me doing my day to day? And the answer is no. whatever it is, um, code code or like, what is it, um, open core running.
Wilhelm (01:06:11.933)
And.
Wilhelm (01:06:16.187)
I say I I say I
Like, yeah.
Matt Carey (01:06:25.796)
I have something running in WhatsApp, like just a WhatsApp agent that I've had for ages. it's actually in Telegram. And I've had it, it's on a Lambda, right? Ages ago. And I put it on a Lambda, not on an app on my machine because I didn't want to care about when it ran or like it killing battery from my phone. It wasn't because like I cared about like the permissions it was going to access. Like Apple apps are inherently sandboxed.
Like I put it somewhere else so it could run all the time. But it happens that like, like also Lambda is also sandboxed, you know, like it's, it can try and break the execution environment and the next request will reset, but it can't because Lambda's like every, this is the, this is the cool world that we're entering where you can send a service code that solves your problem, which you couldn't before. That's why all these people have made all these sandboxes because
Wilhelm (01:06:58.141)
Mm.
Wilhelm (01:07:06.3)
Yeah.
Wilhelm (01:07:11.004)
Right.
Matt Carey (01:07:24.194)
As soon as you give a customer the ability to send you code, you need to have a pretty good sandbox. So any machine, anything. No, no, but isn't that like a Mac mini?
Wilhelm (01:07:31.633)
Totally. I'm totally with you on that. I'm totally with you on that.
Well, I would say this, you wouldn't want to run.
Matt Carey (01:07:39.242)
I can write code on a Mac Mini. Apple gave me a machine that I can write code on. It's a sandbox, dude.
Wilhelm (01:07:43.774)
But I feel like when you're talking about running untrusted code from your customer on your own computer, you're not going to run it on a Mac Mini, right? You want a real sandbox with restrictions. Like, you want to run it in workers, probably, or something like it, so that it's really isolated and you're really intentional about what it has access to.
Matt Carey (01:08:07.071)
Or you're just like getting rid of the things that it doesn't need to have access to. okay, if I had lots of Mac minis and I didn't care about them getting destroyed, then yeah, sure. I know there are people on Twitter who are like renting out Mac minis for OpenClaw. then you are, like, it's just a machine at the end of the day. Like, whether you want to call it a sandbox or not, I think they're all varying gradations of sandbox.
Wilhelm (01:08:19.719)
Yeah.
Wilhelm (01:08:23.748)
Yep.
Wilhelm (01:08:35.685)
Okay, but then everything is a sandbox, right?
Matt Carey (01:08:39.839)
Well, I would say your personal laptop is not a sandbox because if it breaks, there's no sandcastles. It's quite sad.
Wilhelm (01:08:47.823)
And to be clear, I would be upset if my Mac mini broke. Like there is like state on there, there is like stuff on there. Some I'm starting to back up.
Matt Carey (01:08:54.945)
Yeah, and I think that's probably full for you. I think you probably should do some backups.
Wilhelm (01:09:01.277)
So there's actually quite an elaborate backup system now, but still like...
Matt Carey (01:09:05.295)
didn't you break it? What happened to your backup system? You said something. I think we should stop talking about sandboxes. There was something funny that happened about your backup system. Yeah.
Wilhelm (01:09:09.659)
Yeah, do you wanna, do you wanna hear the whole story? Yeah, it was kind of bad. like, I...
So I work on my laptop, but then the code runs on my Mac Mini for my personal agent.
And so the code gets kind of written on one machine and executed on the other machine. And they're like on the same tail scale. But basically, it's this one folder, basically, which has all the code in it. And I wanted that to be synced, like the code to be synced from the laptop to the Mac Mini and vice versa. Because the Mac Mini might generate files as well that I want to see on the laptop. So I want a two-way sync.
Matt Carey (01:09:57.505)
Mm-hmm.
Wilhelm (01:09:58.78)
And I thought, OK, this is basically just like iCloud, right? Like, I can just use iCloud. Like, that's literally what it's made for. And I already pay for it. So just put this whole folder in iCloud. But then as soon as I started doing iCloud, and iCloud, think, syncs all the time, like syncs constantly. Like, it's just instant and kind of behind the scenes in the background. But as soon as the Mac Mini held a lock over a file or something because it was editing it,
Matt Carey (01:10:25.588)
yeah, this was this crazy Apple thingy. Sorry, current.
Wilhelm (01:10:28.871)
the sync just failed or it duplicated the file. So instantly it was just tons of duplicated files. It's like code files just duplicated with the same name and a two at the end. So maybe that's OK for your photos or your documents or whatever, but it's pretty annoying for this use case because you don't want all these duplicated files constantly that Apple just happily duplicates for you in the background.
So I was like, screw this, I don't need iCloud. Move it back into a non-iCloud folder. I'll just have the agent sync the code, right? And I was thinking back to, do you remember that Hacker News comment on the Dropbox announcement thread where someone was like, this is cool, but no one will use this. You can just wire up some rsync scripts and...
And then you have free Dropbox for yourself. And obviously no one in the real world wires up R-Sync scripts. They just want to use Dropbox. But I was like, hey, AI is here. Let me just tell the agent to do what that Hacker News comment said and just use R-Sync to send the files around, right? Like how hard could it be? And now I'm fully in control. The problem was that the way the agent wrote R-Sync scripts is it passed a flag to it where like it deletes files if they're not...
If they're, yeah, it deletes, I know, there's like a, I think R-Sync is like one way as opposed to two ways, so you have to decide if you pull or you push, and then if the file isn't present on one machine and you're pulling, I guess, and you pass a flag to R-Sync, it will actually delete the file if it's not, because it's like the correct version is the file's not present, therefore the file's Anyway, so my whole folder with like all my AI exploration stuff was just completely deleted.
And obviously called, apologize profusely, but it was just all gone. like, and luckily there was some random Apple backup process running, which does like snapshots, it persists them only for like 24 hours. And we were like 22 hours into it. then luckily we could, we could remount that old snapshot and all the files were in there. So yeah. So.
Matt Carey (01:12:12.657)
Dude
Matt Carey (01:12:27.153)
Awesome. That's good.
Matt Carey (01:12:34.133)
Are you saying we as in you and Claude?
Matt Carey (01:12:39.937)
I saw an amazing tweet from Ricky which was like, I, well, mostly Claude.
Wilhelm (01:12:46.589)
Totally. This is how it feels. Man, I mean, I feel like I've been deep in like Claude psychosis the past few weeks. I'm like staying up late. I stayed up until 3am, like Friday night. And then my whole weekend was just ruined because of that. Like I don't think I can do like all nighters anymore, to be honest. Like it's just, it's kind of rough. But it's fun. That's the thing. And it's so flow state inducing because it's just, even now that the agent tell the agents tell you like what to do next.
Matt Carey (01:12:49.153)
Yeah, sick.
Matt Carey (01:13:06.635)
Don't do that, dude. Don't do that.
Wilhelm (01:13:15.953)
You know how both Claude and Codex at the end, they're like, want me to do this next thing? And it's like, yes, you can just type in yes. And it builds the next thing. So speaking of, I now have a better sync system based on Unison, which is something Claude recommended, which is some really old OCaml syncing thing. But it's working well for me. You can define rules and all this stuff. And I have a separate backup. I think I have three layers of backups now, so this doesn't happen again.
Matt Carey (01:13:26.559)
Yeah, crazy.
Matt Carey (01:13:37.392)
yeah, you said that, yeah.
Wilhelm (01:13:46.285)
and the syncing system works reasonably well. And also, I'm using a new service which is in my Sparks Joy category called healthchecks.io. Have you heard of this?
Matt Carey (01:13:59.721)
No, I have to run quite soon by the way, it's getting quite late for me.
Wilhelm (01:14:01.692)
Because... Yeah, and...
Let's wrap up in a sec. Yeah, this health checks thing is really, cool because I'm now running all these, like, because on the Mac mini and on the MacBook, you're using like launch D to run all these background services, right? That could fail or could not fail. And like, who actually looks at whether they work or not? Like, sure, you can like ask Claude to look into whether something was working or not. But with this health health checks thing, like the way you set it up is you say, I expect this script to run every 15 minutes.
Matt Carey (01:14:20.479)
Yeah.
Wilhelm (01:14:36.165)
And then as part of the script run, it pings this health check service. And if it doesn't run every 15 minutes, you get an email from the health check service saying it's down, like it hasn't checked in, basically. So it's pretty cool. And then I can just tell Claude, hey, one of the health checks is down. Can you investigate? And then it investigates, fixes the issue. Obviously, the next step is to wire up the health check service to send an email or a webhook to my.
Matt Carey (01:14:49.024)
I mean, that's pretty cool.
Wilhelm (01:15:04.091)
Mac Mini agent so that it can then automatically look into what the issue is and resolve it.
Matt Carey (01:15:09.152)
and then send you an email with, then send you a WhatsApp with like all of the, like an issue report that happened in the past and is now fixed. Yeah, literally. This customer had this problem and it's now fixed with this PR. This PR is merged.
Wilhelm (01:15:17.401)
Exactly, Dream. Total dream.
Wilhelm (01:15:24.539)
Yeah, here come the software factories.
Matt Carey (01:15:28.404)
Yeah, it's happening, it's happening, it's happening. Okay, right dude, I have to go. It was a pleasure, as always.
Wilhelm (01:15:37.393)
Big pleasure. Good fight about sandboxes.
Matt Carey (01:15:41.417)
Yeah, maybe I misguided. I just see them as there's gradations of sandbox. Gradations are there's gradations of compute really from not sandboxed to very sandboxed. Would we agree on that? Do you want to agree on that? Okay. Okay. Okay. We can leave with leave with some unity.
Wilhelm (01:15:46.567)
Yeah, no, I can't...
Wilhelm (01:15:54.684)
Right. Yeah, yeah. I can agree on that. I can. No, for sure. I mean, I think we're mostly saying the same thing. I think we're mostly saying the same thing. It's just.
Yeah, maybe I need to think of like a good analogy.
Matt Carey (01:16:15.056)
Yeah, I think it really depends where you draw the line like I'm sure Kenton would be like well a Mac Mini is obviously not a sandbox Because it's like a machine with a file system and like other stuff and things like that I actually I don't want to put words into his mouth, but I
Wilhelm (01:16:28.347)
Right. Yeah, and it can.
He seems super cool by the way, Kenton. If anyone hasn't seen the Linus Tech Tips video of Kenton's house, it's so cool. It's wild.
Matt Carey (01:16:43.024)
yeah, yeah he has the LAN party.
Wilhelm (01:16:46.415)
It's so cool. Like I want to be Kenton when I grow up.
Matt Carey (01:16:48.096)
Yeah, he has hasn't he has an amazing talk from I think from last year. I don't know if you just search his name on YouTube, you'll find it. I'm talking about dynamic workers and sandboxes and just like his whole journey with sandboxes. So I don't want to like, I think I think if
Wilhelm (01:17:06.405)
Ask him if one of the LAN party computers, if he would consider it a sandbox.
Because I think they'll like, they probably don't have any...
Matt Carey (01:17:17.108)
Just tweet him. I want to see the answer, but I don't know. Yeah. It'd be nice to that in a public forum.
Wilhelm (01:17:24.197)
Yeah, that's true. Sweet. Any closing?
Matt Carey (01:17:28.32)
I guess if they're okay, okay, actually, here's another is another thing. If they're fungible, if the machines are fungible, as in like you can swap one out for another. Is it a sandbox? Because like, if they're fungible, and you don't care if one gets bricked, because you have another one that you can swap in, or you have the potential for another one you could swap in. Then to me, you have some level of sandboxing because like,
Wilhelm (01:17:40.925)
Thanks for explaining the fungible means.
Matt Carey (01:17:56.148)
The only reason why you sandbox stuff is for that pain in the ass moment if something, stuff gets broken.
Wilhelm (01:18:02.696)
Yeah, you're trying to cover your breast.
Matt Carey (01:18:04.832)
But it's not just broken, there's also a security issue to sandboxing as well, like from external actors rather than internal. Like we have this, the security thing's changed, right? Because we had all this thing about external actors previously. And now it's like whether your own machine is going to destroy itself.
Wilhelm (01:18:11.804)
Mm-hmm.
Wilhelm (01:18:22.383)
Yeah, yeah, yeah, yeah, right, right, right. It's an internal actor that you want to lock down. I mean, me, the defining characteristic of a sandbox is just like, it's like lockdown in some way. So that bad things don't happen.
Matt Carey (01:18:24.032)
Which is crazy.
Matt Carey (01:18:34.144)
Yeah, I think, dude, your Mac mini is locked down in the way that like, if you don't put a password in passwords.txt on your homepage, it doesn't have access to your passwords.
Wilhelm (01:18:39.579)
in
Wilhelm (01:18:47.737)
I agree, it's locked down in minimal ways, but I'm working hard to lock it down less so that my agent is more powerful. All right, man.
Matt Carey (01:18:53.108)
Yes! Agent be free, agent be free. Give agent hands, claws. Bye bye. Big love.
Wilhelm (01:19:01.97)
Big love, big love, bye bye. See ya.