You've Been a Bad Agent

Wilhelm (00:01.004)
Alright boss man, we're on.

Matt (00:02.924)
Yo, we're back!

Wilhelm (00:04.942)
I wanted to beat you to the boss man this time on the recording.

Matt (00:09.169)
Nice. Okay. It is, it is. Okay, why are you saying that?

Wilhelm (00:11.202)
This is very exciting day.

because we have a shared agenda for the first time.

Matt (00:19.93)
Dude, it's only taken us a year to work out what we want to say. Do remember the first one? We like had like WhatsApp and we were just like sending each other WhatsApps during the podcast being like, yeah, we should talk about this. And then we kind of went off that.

Wilhelm (00:30.309)
yeah, yeah, yeah.

Yeah, yeah, yeah, yeah. That was a good one. That was a good one. Yeah, man, the echo is still quite bad. the... I guess you can't hear it. Actually, there's not really that much echo. It's just, it sounds like you're sitting in a room that's not massively far.

Matt (00:44.677)
The end.

Matt (00:50.732)
sitting in a room that's not got many things in it. yeah I need to get some soft furnishings to this place. It does echo quite badly actually. Wait I'm gonna be back in one second. Yeah yeah I'm gonna grab I have I have a massive bean bag I'm gonna go grab that give me one second.

Wilhelm (00:59.278)
Can you hear it? Okay, okay.

Wilhelm (01:04.856)
Hell yeah, yeah. Go for it.

and I will keep the audience entertained with a story. No, I don't know. Probably Matt can still hear me, actually. Can you still hear me? In which case I could talk about the first thing, which is actually a bit of a monologue anyway.

Wilhelm (01:28.974)
Okay, don't think you can hear me still. Well, dear audience, this podcast is unedited and raw and very authentic. That's what you came here for. That's what you're paying us the big bucks for. Okay, nice, you can't hear me.

Matt (01:42.288)
I can't hear you. I can't hear you, dude. I can't, I can't. I have this massive bean bag, which you can't see, but it's huge.

Wilhelm (01:50.518)
That is massive. Wow. Okay, I think it's a bit better. I think it's maybe a bit better. Yeah. Okay, I feel like we actually have too much to talk about. I mean, maybe, maybe we, don't know. I don't know, I have no idea. We've never done this before. But is there anything you really want to start with? Or actually, wait, should we talk about things that aren't on the agenda to start?

Matt (01:57.712)
Yeah, I think so. Okay.

Matt (02:05.613)
Okay, I'm gonna...

whoa.

Okay, first of all, I want to give us a little bit of breakdown of what we're going to talk about because anyone who's been listening now has been listening for like five minutes and has had no like dopamine rush. So here is the dopamine rush. are, yeah, we're going to talk about security. VS code extensions are a security risk. GitHub has been hacked. NPM is broken basically. Capacity joins Anthropic. The Pope also is basically an Anthropic.

Wilhelm (02:23.383)
Shit, yeah, yeah, yeah, go on.

Wilhelm (02:32.923)
my God, yeah.

Wilhelm (02:41.677)
Ha

Matt (02:42.224)
What is going on right now? Capacity constraints, we had Google I.O., Slack is back, AMP Labs, really interesting this actually. You had PyCon, there's been some banger, banger tweets, some by Rune. There was also one about the SF bubble and people making too much money, which I found personally quite interesting. That's not on our agenda, I'm just making this up right now.

Wilhelm (02:46.231)
I hope that's chocolate.

Wilhelm (02:51.069)
god, yeah.

Wilhelm (03:07.354)
yeah, yeah,

I'm adding it.

Matt (03:11.651)
But yeah, absolutely loads. Where do you want to start? I'll show you on the intro.

Wilhelm (03:15.725)
yeah, that's from the intro.

Wilhelm (03:31.307)
Woo! Okay, yeah, I feel like we should start with the security stuff. That feels kind of important and serious and whatever.

Matt (03:32.751)
Woo! So now we can stop.

Matt (03:41.837)
Yes, and as I work for a security company now.

Wilhelm (03:45.784)
is that right?

Matt (03:47.151)
We have thoughts. we in a, no, we're are we in a, we're a security company, aren't we?

Wilhelm (03:52.131)
yeah, I always thought of you as a sort of infrastructure cloud provider. But I guess, yeah, I'm a security company, sure.

Matt (04:00.281)
think before that we're a security company. I think it was first a security company, then a cloud, now an AI.

Wilhelm (04:02.445)
Yeah.

Wilhelm (04:06.317)
An AI cloud. Yeah, okay. So to give some context on the security stuff, it feels like everyone's getting hacked all the time. There's all these worms, the shy Hulud, all these packages are being compromised constantly. GitHub seems to be pwned, or at least in the sense that attackers access their internal repositories, which I think means like the private repositories that GitHub has, such as GitHub itself.

Matt (04:32.622)
Bad things, bad things.

Wilhelm (04:34.827)
I don't think it means everyone's private repositories were compromised. So it sounds like a, I don't know, maybe like a staff token leaked or something like that. Or like an employee was compromised or something like that. It doesn't sound like all of GitHub is pwned. But yeah, there's a noticeable increase in various security attacks. We've talked about a little bit on the pod before. I mean, obviously this very much fits the whole narrative of Mythos and Anthropic that like...

AI tools make it much easier for attackers to execute these attacks. They can do them more quickly. They can find zero days more quickly. I think we talked a little bit last time or the week before about how like the old mental model of like, if you're not a valuable enough target, you probably won't get popped by a zero day. That probably has to change a little bit. And I've spent a bunch of time thinking about like how I can improve my like own security on the back of this. But before I go into like what I'm doing,

Yeah. Any, any more thoughts from you on, on, on all of this? mean, yeah, I guess you're closer to the industry on security.

Matt (05:40.366)
just think it's crazy, even... Dude, everyone's getting done. people I see on Twitter with moderately successful NPM packages are getting like, oh, by the way, don't release. Someone just like, or someone's literally outing them being like, guys, don't install this one because this one's got a shy delude. It's crazy, like mad stuff. I'm not gonna name names because some people are friends of mine, but it's just bonkers. Everyone is a target.

Wilhelm (06:01.772)
Right.

Matt (06:09.997)
They seem to be getting in mostly through GitHub CI, seems to be like the most common attack vector. Like, it's wild. I think people should definitely think about putting minimum ages on their global NPM config. Super easy to do on your computer. You should just do it. Like minimum age three days. Like definitely do it. It's the simplest thing to do.

Wilhelm (06:21.101)
Oh, I just saying, yeah.

Wilhelm (06:36.577)
Yeah, Python has this as well. And I think there's even a way, this was discussed a little bit at PyCon where I was for last week and over the weekend. There's even a way, like I think my first thought with this cool down thing was like, like, okay, but I'm making packages. If I can't install the package I just published because it has a three day cool down. That sounds really annoying. But I think there's a way to exclude.

Matt (06:52.269)
Hmm.

Matt (06:57.099)
No, you can have workspace. You can exclude anything in your workspace. So if you have your package locally, then you can use that. can also like, which you should be doing anyway, if you're building packages, you can use that NPM link. I don't know how it exists in the Python world, but you can install via a path rather than installing via NPM. Yeah.

Wilhelm (07:03.405)
Yeah, yeah.

Wilhelm (07:10.7)
Yeah, and

Yeah, yeah.

Wilhelm (07:20.139)
Yeah, I think people have figured out this problem that I'm imagining. Or there's even a thing like if you pin to like a specific package version that's like newer than your cooldown, it will just install, or at least on NPM.

Matt (07:31.565)
Actually, that's crazy

Wilhelm (07:35.82)
That's what I heard from someone. take it as the truth. No, no. Yeah. Worth double checking. Yeah. And it just seems, yeah, it is seem a bit, I think it's worth thinking about security and updating our prior beliefs about like security, like a little bit. It is wild. And so, so the GitHub thing you mentioned, the GitHub actions, compromise things. So this is like someone pwns your CI and the CI has like trusted publishing set up or something like that. Right. Which means that

Matt (07:40.684)
Okay, we need to look that up.

Wilhelm (08:05.901)
NPM or the Python Package Registry or whoever will just accept like a new package version if it comes from a certain repositories like CI from GitHub CI, yeah. Which was supposed to be a massive security improvement over like maintainers publishing from locally with like a token or with like username and password. Like this was supposed to be a step up, right? But it sounds like people are pwning GitHub actions or...

Matt (08:15.5)
Get out, CI. Yeah, Yeah, so.

Matt (08:26.252)
Hmm.

Matt (08:31.381)
Yeah, OIDC publishing is now no longer gold standard. It's just an additive thing, which it probably should have been already, but people consider it gold standard. This is a verified package. And actually, if your contributors are allowed to run some parts of CI or some parts of CI get run on your contributor's code, then actually your CI can be poisoned. And so what they were doing, I think, was they were storing some...

Wilhelm (08:57.015)
Yep.

Matt (09:00.972)
executables in your CI. They were actually storing it in the machine, in the runner, think, actually, that runs the CI. And then when it was run again on main, then you got pwned.

Wilhelm (09:14.647)
Damn. Yeah, that's wild. I'm pretty sure there was also one attack that, so there's like different GitHub action triggers for when a, when, when just like a random person on the internet makes a pull request to your repo so that it doesn't trigger CI with all the secrets loaded. Like it's a different event name. I think it's called like pull request target as opposed to like pull request. that tells GitHub, okay, this is like someone who doesn't have access to the repo. So run different workflows, run like different GitHub action runs.

Matt (09:30.476)
Mm.

Wilhelm (09:43.83)
But I think someone figured out a way. So, but you might still want something to run, right? When someone like a safe workflow or like a non-publishing workflow. But I think I saw a thing where like someone found like a like shell escaping thing by putting like a whole command into like the GitHub pull request title. And then that managed to like escalate permissions to the more serious workflows. then boom, now you can publish the package.

Matt (09:44.235)
Yeah.

Matt (10:11.82)
Yeah. I think actions are, I mean, it's pretty scary. Like in a lot of cases you're letting your contributors run some code in the same environment that then you're releasing your stuff in. It's like quite wild. Yeah. I don't know. It doesn't feel that bad until you think about it. Go on.

Wilhelm (10:25.377)
Mm-hmm.

Wilhelm (10:28.735)
One thing I saw was that you just, yeah, right. One thing I saw, and again, I'm just like regurgitating hacker news comments here, but there was something about like how even with the like trusted publishing workflow thing where the CI publishes it, it still should send like some kind of notification to the maintainers who then like have to approve, like hit an approve button basically. So there's just like a manual approval step, like a one click email or whatever.

And that would mean compromising CI isn't enough. also have to compromise. I don't know, someone's email or something. That seems like it would be better.

Matt (11:05.578)
Yeah, yeah, that seems better. I think a lot of this stuff is like about depth, isn't it? Like you need to have more than one way, you need to have more than one point of interaction with a human. Yeah.

Wilhelm (11:12.845)
Mm.

Wilhelm (11:18.965)
Yeah, yeah, OK, so one thing I was keen to talk about is like VS code extensions as being a security risk. And I've not seen this discussed sufficiently. It was actually mentioned by Mitchell Hashimoto in his recent tweet about this whole security situation where he talks about like, I mean, he has quite an extreme stance, I think, where it's like you shouldn't update your dependencies ever unless

there are specific features that you want from the new, from the upgrade, and then you should review every line of code that the new dependency has, which I mean, no one in the JavaScript ecosystem does this. You can kind of follow the reasoning a little bit, but yeah, anyway, I mean, it's just not really what most people do, but I guess people can judge for themselves if they want to do that. But he briefly mentioned VS code extensions or like editor extensions in general. And I think the approach he suggested was like,

Yeah, just like don't have them or just fork them and implement like, I guess, slop fork them and have the minimal feature set that you're actually interested in. And the reason I think that VSCOs are particularly, a particularly risky surface is because having made one of these, right? Like you install this thing and it can just shell out and has full file system access. So you install like one little, you know, theme.

or like, like a little rainbow brackets for my CSV or whatever, or like nice colorization for my CSV, that like little JavaScript package can fully shell out, execute whatever it wants, like spin up a whole server, has full write access. And then also, VS Code updates these extensions for you in the background, right? So like you restart your VS Code, you have the new extension version. Who knows, it could have this complete malware.

Matt (12:52.617)
Mm.

Matt (12:57.992)
Mmm.

Wilhelm (13:13.205)
And there's a lot of really, really popular VS code extensions. Like I'm talking like 50 million installs that are run by like one person. And obviously I'm an indie hacker type person. want to support people who are like have, you know, who are independent developers making like useful software. But I think the security calculus has just like changed a little bit and we don't want to be getting pwned by like our, I don't know, dark mode extension that suddenly like takes over your computer.

Matt (13:13.545)
It's crazy.

Matt (13:43.412)
Yeah, it's crazy. I think we'll see much, I mean, Nix was so before its time, but I think we'll see much more like this with like deterministic builds where like, where like everything on a developer's machine will be, it'll probably be the same in the cloud as well, right? Like you'll have a machine, maybe it's locally, maybe it's in the cloud, but you just take it with you with your Nix config and your config just like the same way that we have it with, with,

with like applications, it will be the same with your dev machine. And I think the reason why, I've been trying to work out why this feels so appealing to me. I think, cause it is just annoying as well. Like people have this a little bit with their dot files if they put them on GitHub or whatever, and then you have this stuff. But I think the reason why this is so appealing to me is Claude code, like these agents, they cross the bridge between running deployed and running locally.

Like a lot of the time you're running them against your deployed application, even if it's deployed in like some sort of staging environment. And so this thing, this thing is kind of interesting to me how like, there was always this dream that everyone would like work on a deployed environment all the time. I think AI might be have it making this, making this feel possible. And then with the security stuff, I think it's like making it feel almost inevitable because

Wilhelm (15:14.285)
Mm.

Matt (15:14.621)
because like you have your application that you deploy, you pin everything, like everything is pinned by a lock file and it's done. It's done like that because of these supply chain attacks and because of all of this stuff. So. Like, feel like someone needs to make a really nice version of brew that runs on Nix.

Wilhelm (15:24.842)
Yeah.

Wilhelm (15:31.105)
Yeah, yeah, yeah.

Matt (15:40.644)
and then just has like lock file for your entire computer, that would be really nice. Someone should do that.

Wilhelm (15:40.918)
Hmm.

Wilhelm (15:48.225)
That's interesting. Yeah, the whole like local versus cloud development thing is interesting. It's almost, yeah.

Matt (15:54.456)
You can see it just with Brew. if you ever install anything with Brew, it updates EVERYTHING.

Wilhelm (16:01.853)
It does, yeah.

Matt (16:03.561)
It's insane. Like you install, I don't know, what you want to install with Brew, like a poplar to work on PDFs. And now you're updating like your Safari and like loads of like random shit is just being updated. But before that, you're updating your Rails version, your Ruby version. You're updating like your Node version. You're updating your Postgres version because you wanted to install a PDF tool via Brew. It's like.

Wilhelm (16:12.727)
Mm-hmm.

Wilhelm (16:18.305)
Mm-hmm, mm-hmm, mm-hmm.

Wilhelm (16:23.129)
Mm-hmm.

Mm-hmm, mm-hmm, mm-hmm, mm-hmm.

Wilhelm (16:32.139)
Mm-hmm.

Matt (16:32.818)
Frickin wild and like I get home brew is the best thing at that time because Mac didn't have a package manager but like Jesus Christ that has to change

Wilhelm (16:41.973)
Yeah, it's a lot. It's a lot. Yeah, I think it's a good time to re-evaluate this stuff. I feel like also, so I did a big, just told Claude to audit all my VS Code extensions recently and...

I think it's a good time also because I don't actually need all these different themes or like various little extensions anymore because I mean, how much time do you spend in the editor or like, what do you do in there? Right? You're like reviewing diffs or like you're looking at some mock-down files or something. And otherwise you're just kind of like in the terminal. So.

Matt (17:05.352)
You don't need any of it, dude.

Matt (17:13.126)
Your diffs, you can run like CLI tool, can run like the Pierre Diffy, like just in CLI. Like I have a Pi extension to look at diffs. these guys, we were right when a year ago we were like, is Codecode the new IDE? It's like, yes, 100 % it's the new IDE. like we're building the, the Pi is actually the new IDE because you can build extensions to actually make it an IDE rather than Codecode feels a bit more like an app.

Wilhelm (17:19.265)
Yeah, yeah,

Wilhelm (17:30.815)
Mm-hmm.

Wilhelm (17:35.489)
Yeah.

Matt (17:42.885)
A quote catch was a bit more canary.

Wilhelm (17:43.309)
Totally, Yeah. No, it's very true. Also...

Matt (17:51.25)
But the brew thing, someone should actually do a new brew. We need a new brew. A new brew based on Nix. Like, NixOS exists, but... Yeah, NixOS exists, but Jesus Christ, like, it's a set-up pain. yeah. And Nix has all of these, like, weird nomenclature, like, to do things. Someone just needs to make Nix for dummies. That, like, does the thing. Yeah. I don't know.

Wilhelm (17:54.614)
Someone showed, yeah. Revised in Rust.

Wilhelm (18:07.212)
Mm.

Wilhelm (18:13.473)
Mm-hmm.

Wilhelm (18:21.099)
Yeah, I must admit I've never properly played around with Nix. Maybe now is a good time.

Matt (18:25.734)
It's great. It's just a bit too overcomplicate. And it does so much stuff. Like you can version your server, you can version your laptop, you can version your OS for Linux, you can version everything. If you want everything to be versioned, then it lives on this like crazy idea that like everything should be locked in space. And I kind of get that like as, okay, my thesis on this is like, thesis, horrific word. My thoughts on this is that as LLMs make software more plastic,

Wilhelm (18:33.686)
Hmm.

Wilhelm (18:38.733)
You

Wilhelm (18:49.602)
Hehehe.

Matt (18:53.379)
as a more like, as in like you can bend them and then they pop back and you can like play with them and you can like get a heat gun and like push things into different shapes and prod it, you know? And it stays where it is. Like you deform it beyond its like plasticness. Was there before you used to like patch packages. Now you can just like go into node modules and just like fuck around with everything like, and just like completely remake shit.

Wilhelm (18:56.844)
Mm-hmm.

Wilhelm (19:04.341)
Yeah, yeah.

Wilhelm (19:20.619)
you

Matt (19:21.467)
I think people are going to like, vendor more packages themselves. I think the whole of people's build environments is going to become more like in-house because you want to change something in a package. You don't want to like, submit a PR to a maintainer for them to review and they have like millions of things to review. You're just going to change it yourself and then run your package locally. And then there needs to be better ways of sharing that like new package with your team rather than...

rather than patching or vendoring. think like they're still quite clunky. Like vendoring, you now have to copy the whole of the package contents in and maybe they have a different build system. Patching, you take their built-up artifact and you add some horrible diff on top, also kind of grim because now you can't really update stuff. We use patching a lot, by the way, in agents.stk. That's someone who was a big fan of patching. We patch loads of our dependencies for like stuff that we want to change on them.

Wilhelm (20:09.632)
Interesting.

Mm-hmm.

Matt (20:15.514)
And I think there's something in between that doesn't exist yet that I think will exist where we basically just take all of our dependencies when we want them at the moment that we want them. And then we have them all. And then if you want to upgrade to a newer version, you're just like, LLM upgrade this version. And it goes and it has a look at the commits and it just applies your commits on top. It applies the commits. Yeah. Yeah.

Wilhelm (20:20.417)
Yeah.

Wilhelm (20:31.885)
Yeah, yeah, yeah.

Wilhelm (20:37.121)
like a more like AI native package management workflow. Are you saying you're here to announce today a new package manager? Call it Matt PM.

Matt (20:46.842)
God no.

Matt PM, Jesus. No, no, no, no. I'm not announcing anything. I just think those two things should exist and maybe they're actually the same thing. I don't think they are, but maybe they are.

Wilhelm (21:02.381)
One more question for you on the security stuff, is so far a bunch of these attacks that we've seen, especially some of the worms and stuff, I feel like they have executed deterministic code locally, right? So it's like, check all these directories for credentials and then exfiltrate them to this random place. One thing we haven't seen yet, which feels kind of scary is one of these things just shelling. If everyone has Py or Claude or Codex installed, what's going to stop one of these?

like malwares to just like shell out to like codex in YOLO mode and say, yeah, why don't you find all the credentials and then send them somewhere and maybe dress it up in a nice story as in like.

Matt (21:42.128)
Didn't... Didn't we, didn't we have that?

Wilhelm (21:46.822)
we already had it?

Matt (21:48.365)
I feel like that existed. might be like going crazy, but I'm sure that existed already. like a few months ago, like December time or something like that, we had this like crazy malware that was like targeting your local Claude code instance.

Wilhelm (21:59.18)
Mm-hmm.

Wilhelm (22:04.477)
wild yeah it makes sense and it feels I mean that feels like incredibly hard to like prompt inject protect against right because this is just like yes this is like just a user request starting up

Matt (22:18.147)
Yeah, it's fricking weird. Yeah, the security stuff scares the living shit out of me. I'm not gonna lie. I think we're gonna see more and more and more. Everyone is a target. Like when it's just a prompt away, you're gonna have people hitting everything. Like, doesn't matter what you're making, you're gonna have people attacking you. And not in a way of like, where is my about.php? Or like, look for your .env. It's gonna be like sophisticated shit.

Wilhelm (22:42.953)
Yeah, yeah, yeah, yeah.

Yeah.

Matt (22:48.365)
looking for like CRS, CRCF headers or whatever it is and like trying to see if they can steal other users credentials and using your site for phishing links and loads of crazy shit.

Wilhelm (22:57.463)
So what can we do about it? What else can we do about it? Like what should people be doing now?

Matt (23:05.261)
just the trusted publishing one is like the easiest step ever. Just, just do that. It's a no brainer, like absolutely no brainer. And get all of your teams to do that. This is, this is another thing I think like, I don't know about you, when you worked at GitHub, was your laptop like entirely locked down? Like normally developer laptops, you can install whatever you want, right?

Wilhelm (23:09.665)
Mm-hmm.

Wilhelm (23:23.775)
Yeah, it wasn't really locked down at all. It was a little bit,

Matt (23:26.114)
Yeah, I heard like Google, I heard like Google is like pretty locked now. Like you have to install from their registry, from their NPM registries and all this stuff.

Wilhelm (23:33.966)
They have a really cool system. Have you heard of Santa? I think that was a core tenet of... Yeah, that's actually a really cool story. I might get the details about this wrong, and I think this is not the entire story of their endpoint security. It's a Mac OS system called Santa, where basically you can never... And I think other tech companies adopted this as well, I think it's open source. Anytime you want to execute a binary on your...

Matt (23:39.287)
No.

Wilhelm (24:03.415)
Google employee MacBook, it has to get checked by Santa. And there's like a central store of certain approved binaries. And then if you want to like run some new binary, right, like you're a developer, you know, there might be some legitimate reason.

I think there's different levels of approvals, but the lowest level approval is that just like some other employee has to approve it. So you just need to get someone else to say, yeah, that seems like a legit binary or there is like some reason to use it. So it's very like peer approval. And then it's allowed. then yeah, Santa basically checks everything. I actually have Santa running on the Mac mini because it's like a, yeah, I think it has some like logging capabilities as well. I think actually it's evolved a little bit as a product.

I think the people who built it at Google left Google and are now working on it as like its own company. And it can monitor and block USB connections and all this stuff. think Santa is a cool tool for this. it's watching and checking things a little bit on my Mac Mini.

Matt (25:12.128)
I really like that. Yeah, I think we'll see more and more and more of this. It makes sense, but I think like before you had to have like tech, big tech company infrastructure in order to do that. Like we have this with network traffic a bunch, like I have warp, like after warp and like there's like a bunch of security stuff there, but it's quite web based. So we have our, we have internal NPM registries and stuff, but I work a lot in open source. So I need like,

Wilhelm (25:25.377)
Yeah, yeah,

Wilhelm (25:33.975)
Yeah.

Matt (25:40.675)
public NPM registries. So I'm kind of a gaping open hole in the system. Like everyone who works just on internal stuff, they don't need public NPM registries.

Wilhelm (25:40.78)
Mm-hmm.

Wilhelm (25:47.553)
Yeah

Wilhelm (25:52.43)
Yeah, yeah, yeah. You heard it here first, everyone. If you want to attack Cloudflare, go for Matt. He is the gaping wound. No, I think the networking stuff is interesting. Yeah, I think this is one thing that I...

Matt (26:02.531)
Please don't.

Wilhelm (26:11.309)
used to just not really think about this much, but I think it's really worth adopting now. Just like put everything into like a private network, like use tail scale or cloud flares equivalence. And then even if an attacker steals your token to like some internal service, if your actual server isn't public, it's just behind like a VPN, it's in your private network and you need to be on the network to actually hit the...

the service with the token and then attacker exfiltrates your token and they can't really do anything with it. So it's like a, it's a really nice like update and it's, and it's really not that hard now. think pre-Talescale like running all this WireGuard stuff yourself was hard or even like using like AWS's own like VPN infrastructure. I set that up ages ago for us and it was a giant pain and it's incredibly straightforward with Talescale. So I think everyone should be Talescale maxing or

Matt (27:01.207)
Yep.

Wilhelm (27:05.203)
or use Cloudflare's thing, I think there's a Cloudflare thing, but I actually don't really know how it works.

Matt (27:06.819)
Yeah, we have we have tunnels Which is how you give your internal servers external IP addresses and then we have mesh which used to be called Something else that was really hard to remember. No, used to know that's tails tails thing. We have mesh our mesh used to be called warp client, I think something that had nothing to do with mesh, but mesh is sick because mesh is like

Wilhelm (27:21.431)
funnel.

Wilhelm (27:25.962)
Okay.

Wilhelm (27:30.228)
Okay, yeah, right.

Matt (27:35.466)
All of your devices can have mesh installed, they have the mesh client, and then they have like VPC peering basically, like you'll see them in your local network. They'll all be together. It doesn't matter if they're separated by time or space, but as long as they're behind, they're inside the car fly network, we will like put them in the same VPC, which is sick.

Wilhelm (27:44.343)
Mm. Mm-hmm.

Wilhelm (27:57.08)
they rejected the coldness of the open internet and embraced the warmth of VPC peering.

Matt (28:04.916)
Warmth of the orange cloud, exactly.

Wilhelm (28:06.815)
the warmth of the orange cloud. This is me trying to turn that Mamdani quote about collectivism into a networking thing. Nice, nice, nice. Okay.

Matt (28:14.531)
Yeah, I get that. Right, let's move on from security, because it actually stresses me out a little bit. It's like the world is going crazy with mythos and all this stuff. Yeah.

Wilhelm (28:24.589)
All right. I want to tell you about this experience I had with ATC.com. Are you ready for this? This is going to be a bit of a monologue. So one idea I wrote down ages and ages ago in my long list of ideas of things to build is, you know flight radar? We can see planes and stuff and you can see a plane in the sky and you're like, oh, I know where that's going. The idea was just that that was already public data. Flight radar exists. There's a bunch of APIs for it. Another thing that's public.

Matt (28:30.218)
Okay, tell me. Yes.

Wilhelm (28:52.721)
is like ATC audio. So it's actually interesting, like in the US it's fully public, so like you can listen in, there's various websites where you can just listen into like the San Francisco tower frequency and just hear what pilots are saying, hear what the controllers are saying. In the UK, funnily enough, there's actually some ancient laws, like really ancient laws that prohibit this kind of thing. So that's actually not, there's no web listening into UK ATC.

which is a story for another time. But the idea was just like, okay, have this flight, like actual routing and flight radar is public, the ATC is public, wouldn't it be cool to build like a flight radar that just like overlays the like ATC audio onto the real flight thing? Anyway, wrote this idea down ages ago, didn't really think about it again much. Then I come across this thing called ATC.com, which is a mobile app. I think it's an iOS app only.

And they've basically taken some version of this idea and made it an absolutely incredible experience. Just like, absolutely. It's so well built. I mean, it's a three letter dot com domain. They're executing it incredibly well. Like they're doing influencer marketing with all of the like ATC replay YouTube channels. That's where I found out about it. So just like really, really well done. And I was testing it out coming back from LA where PyCon was. And I just thought, okay, like, you know, this is a great time. Like let me fire up.

ATC.com and you can literally follow the plane that you're in. Like you can click on the plane, you can see it taxiing on the app and then you can hear all of the ATC transmissions that your pilots are making and that the plane is receiving like in real time, like as it's happening. So you're not like clicking around a bunch. So it's really, really cool. And then actually I was really delighted because I hear the controller say,

Matt (30:36.821)
That's crazy.

Wilhelm (30:46.727)
United 1195 or whatever I was just heads up there's going to be a one hour delay for you. And I'm like my god are you serious like we're in the plane we just left the gate and it's going to be a one hour delay but I was really curious like how that was going to be communicated to like the passengers when it was going to be communicated like the pilots just seemed like pretty cool about it on the radio and they didn't actually let the passengers know until

35 minutes later. So we just taxied around LAX for 35 minutes to some really remote part of the airport where we could just like chill. And then they announced it as like, oh yeah, it's a 25 minute delay. And I'm like, that's genius messaging. It's not an hour delay, it's just a 25 minutes. But I could follow along live with this whole thing. Anyway, so I'm just like enamored with this app. think it's absolutely incredible. It's built really well. It's clearly built with AI. Like you can see a little bit of like AI in how the app was made, but it has so many features. You're almost like,

Matt (31:25.473)
I'm

Wilhelm (31:41.666)
The only way they built this so quickly is with AI. And then I start digging into like, okay, who are the people behind this? Like, this is so impressive. I imagine this is some kind of indie developer who just went hard after something and is executing super well. Turns out it's a YC company behind it called Enhanced Radar. They just raised $7 million, actually. But they're actually not building this app as their main thing. The app is a side project for them.

that is basically just like a consumer showcase of the like ATC AI models that they're building and selling as like a B2B service to airports and airlines and like the Federal Aviation Authority, I guess. So they want to like make like all this ATC stuff like way more safe and like help avoid like overload. I don't mean there was tons of stories in the recent past about like

ATC misses like causing all sorts of accidents like with the, was it an Air Canada flight in LaGuardia or whatever that caused like a pretty severe crash because like there was a fire truck going over the runway at the same time as like a plane was landing. Because in aviation, right, so much of this is voice, right? It's like you have pilots responding to these voice commands, like you have.

Matt (32:50.304)
Mmm.

Wilhelm (33:05.375)
ATC issuing this stuff, the controllers have to take breaks constantly because there's so much going on at the same time. And yeah, I mean, I don't know exactly what this company is actually selling, but they're developing their own like ATC models that like can transcribe audio into text. And I just love that they have this complete consumer app play just to like show off how great their tech is.

Matt (33:30.642)
Wild.

Cool. I love, yeah, I guess I really love when people do like really high quality side projects. I'm such a fan. I'm such a fan. I guess like, is this what AMP is doing right now?

Wilhelm (33:33.623)
So yeah, applied AI.

Wilhelm (33:44.205)
Mm-hmm.

Wilhelm (33:51.084)
Yeah, I actually don't really, I read their thing and I like agree with it, but I also don't know exactly what they're doing. Do you want to give people some context on like, yeah, what this is or what?

Matt (33:59.197)
Yeah. It's like... No, you can do You're the former employee.

Wilhelm (34:07.221)
No, well, no. I very briefly worked with Thorsten. can't, I don't think I can claim I'm a former AMP employee. I mean, I don't really know. mean, so I feel like the AMP people have been a little more quiet recently. So you could kind of tell that like something was in the works and they published like a manifesto. They published a new site called Amplabs.com. And I mean, it's a very short blog post, but let me just read the first line.

Matt (34:08.969)
You

Matt (34:27.847)
Hmm.

Wilhelm (34:34.337)
We are assembling small teams of exceptional software builders to bring the full power of artificial intelligence to a single company per industry and region with CEO mandates. So it sounds like AI plus some smart people, hyper-focus, one company, one industry, one region. Go crazy.

Wilhelm (35:00.385)
but I don't know what that looks like. Probably an example would help if they published one.

Matt (35:00.456)
Yeah, I think-

Matt (35:04.402)
Yeah. I think it's cool because it probably opens up a whole new like GTM for them where they can be like, yeah, we've been working on this thing in the abstract, but actually we've got to something that is way like the original AMP was super sick for individual developers and also great for teams, right? But it was super sick for individuals. Like, and they had like leaderboards and like, it made you feel a bit like a community, but it was like,

I felt like it was great for individuals. And then they started adding like more collaborative features and it felt like it was more teams. And I kind of went off it a little bit because it wasn't really for me. Like a lot of people were not using it. Like no one I knew really knew was using it. And I'm sure it was hugely productive for the people who were using it. And it always felt like a premium tier, like coding agent. And then they remade it to be just the CLI. And that was like

Wilhelm (35:41.741)
You

Wilhelm (35:56.897)
Yeah, yeah, yeah.

Matt (36:00.392)
probably a good move because now they're not holding legacy stuff and I really rate them for this like, this like, sort of like, iterate and cut type mentality. Not only do they iterate and change stuff, they also like, really ruthlessly get rid of the old shit. And I love them for that and I think that's why they have so many like, good, like, I think that's why they have so many fun takes because they're like super, they're super-

Wilhelm (36:11.277)
Totally.

Wilhelm (36:16.823)
You

Matt (36:28.828)
They're very good at leaning into those takes and so everything feels quite comes across quite strong and I like it. It makes everything exciting. Like when they announced AMP free, it was like no one had really ever done free LLM tokens before. And now they're like, we're going to do free ones supported by sponsors. Let's just roll with it. This is great. And it's like a whole new way of a whole new monetization model. I think they've done that a few times. And now with the new version of AMP, it's gone even more, more enterprise-y.

Wilhelm (36:48.365)
Yep.

Matt (36:58.363)
Like it's like we're going to run AMP, but we're going to run it on our infrastructure. We're going to run it.

Wilhelm (37:04.311)
This is the AMP Neo.

Matt (37:06.713)
AMP Neo, yeah, it's like we're gonna run it remotely. It's remote AMP. Like this is the next version of AMP. This is how software engineering is gonna get done. You're gonna be working async with agents, not with them, but on top of them. You're gonna hand off, everything is gonna be a hand off experiment to agents. And so for that, agents need to run when your laptop closes. I think that's like the main thing.

Wilhelm (37:12.089)
I see, that what, okay, interesting.

Wilhelm (37:18.061)
Mm-hmm.

Wilhelm (37:30.499)
I see.

Wilhelm (37:34.978)
Mm-hmm.

Matt (37:35.707)
And so that's what AMP Neo is. But when you think about that as a feature, it's very much an enterprise feature. Like, it's very much like a feature of teams that are making money. It's not like an individual, even like me, like Cloudflare, unless Cloudflare dives into it, it's not that useful for me. What would be better for me is just to stick Cloud code in a VPS or in a Cloudflare container or something that I have access to.

Wilhelm (37:43.019)
Yeah, yeah, interesting.

Matt (38:04.807)
something that I can hack with. A friend of mine, cloufler, Luis, actually made this amazing Pi extension where you essentially push everything, you push Pi to a container. So you can run Pi in a container and then you can pull it back again to your local. So you can work on it locally and then you can hand it off to a container and set it running again and then you can pull it back. And it's a really cool thing. It's like Git plus Pi. It's like Git.

Wilhelm (38:23.405)
Mmm, cool, yeah, that's great.

Matt (38:34.458)
with the coding agent included with the dev environment. So you can pull the dev environment, run Pi, and as Pi is running, you can push it to a remote. And I thought that was super inventive and super cool. And I think we'll see a lot more of these different types of UXs, but leaning into the remote stuff feels very enterprise to me. It feels very like Gitpod. We had Matt on the podcast, didn't we? Yeah, owner.

Wilhelm (38:50.827)
That's really cool, yeah.

Wilhelm (39:01.633)
That's right, yeah, yeah, yeah.

Matt (39:04.144)
Gitpod, like that dream of the remote dev environment, which really only matters for like, well, at least I think it matters for bigger enterprise. think like the sort of the more individual you get, the more you're happy with having your own machine and having been responsible for your own equipment, I guess. But me, yeah.

Wilhelm (39:24.203)
Yeah, It's interesting. feel like there was some cursor guys at a meetup a few months ago, they were basically saying, I think they were going really deep on cloud agents or background agents. And they were saying, they were pitching it as a bit like, if you want to do a lot of work in parallel, then you kind of need cloud agents. If you want to really scale up.

Matt (39:35.12)
Yeah, they love their background agents, don't they? Yeah.

Wilhelm (39:50.912)
and have not just two or three, but like, I don't know, five, 10 agents working on the same code base, like at the same time, all shipping things, then you need this kind of thing. And I kind of buy that a little bit, but my own adoption of the CloudAgent stuff has been like...

slower than I expected it would have been like I don't know six six months ago yeah

Matt (40:11.462)
Definitely. Yeah. No, no, no, no, me too, me too, me too. And I think like jump, that's a big hurdle to jump. Like to be like running it locally to them being like, I'm actually going to kick off card agents from Slack or from Chicha or from, from discord. like, it's like in order to do that, you really have to have this collaborative environment with colleagues because otherwise you would just run a single, you would just run multiple instances of

Wilhelm (40:18.337)
Yeah.

Wilhelm (40:25.901)
Yeah, yeah,

Wilhelm (40:36.481)
Right.

Matt (40:40.924)
Codecode in a VPS and you would manage that yourself. The only other reason to have all of this extra info is if you have other people that might look at the results. At least that's my take on it. Like when I work independently at Cloudflare, it's like I'm very happy handing off my stuff into a container and then like even using like Codecode's handoff version is pretty cool. I can put it on my phone and I can like check back in on it. Like all of that stuff is great.

Wilhelm (40:50.912)
Yeah, right, right,

Wilhelm (40:54.957)
Mm-hmm.

Matt (41:10.556)
Or I can have like five terminals and just have them next to each other and like just have stuff running. Like all that's fine because I'm the one reviewing the output all the time. But if someone else was to review the output, then it's really nice for them to see the prompt that set it off, to see the trace, to see the PR and to have all of that in one place. And I think you get that really well with some of the cloud agent experiences.

Wilhelm (41:10.903)
Yeah.

Wilhelm (41:19.809)
Mm-hmm.

Wilhelm (41:26.953)
Mm-hmm.

Wilhelm (41:33.291)
Yeah, totally. The other thing is when you actually have to pick up your laptop or close your laptop and move about the world and then your agent stops running.

Matt (41:44.241)
yeah, definitely. Yeah, definitely. Then implementing something like this is also good. It's also good. like, it's, but that, in that case, really, who doesn't have a machine that's running all the time anyway, somewhere, like if it's not your laptop. I think most people, well, not most people, but a lot of people have a dev server running somewhere that you could run. Even a Mac mini, you have it running at home that you could SSH into, start a prompt.

Wilhelm (41:50.829)
I, but even for this, yeah.

Wilhelm (42:02.4)
Yeah.

Wilhelm (42:11.821)
Mm-hmm.

Matt (42:14.243)
let it run, like just let it go. Like you could have, used to use GitHub actions for this. You make an issue in GitHub tag core. That was like the OG background agents as far as I'm concerned.

Wilhelm (42:15.095)
Totally, totally, yeah.

Wilhelm (42:26.765)
Totally. Yeah. It's funny actually. I feel like this is surely an outdated path and surely I'll just figure out my cloud agent set up soon or whatever. But I mean yesterday I had like three different cloud things running. One of them was really deep, like locally as well. Right. So they're actively running. I haven't got a workflow yet for, for handing off. I don't know if there is a thing for this, like say your cloud is like actively running. It's like deep in some work.

like handing it off to like the Mac mini feels like kind of hard or I don't know how to do that. I...

Matt (43:00.869)
Yeah, that's, you need an extension. You can have an extension to that. So that's what, that's what Luisa's extension does. It like sends the current chat to a container and says, continue.

Wilhelm (43:07.722)
I see, see.

Wilhelm (43:12.543)
Okay. But then you also need the whole state to be across, right? So if you have lots of like dirty files or whatever they need. So it just feels like a little bit messy. but so I was in the office late yesterday. It was like 9 PM or something. And I just really wanted to go home, but I also really didn't want to stop the, the Claude, which is how I got to being at the office at 9 PM in the first place. Just one more prompt, bro. So I was like, okay, let me like, are you familiar with caffeinate? It's.

Matt (43:16.76)
Yeah, yeah, yeah.

Matt (43:40.59)
Yeah.

Wilhelm (43:41.515)
It's like built into Mac natively actually, which I thought it was like a third party thing, anyway, but I think that's not quite good enough. So I installed, recently, I see.

Matt (43:47.266)
Recently built into Mac. Recently. It's like the last version of Mac OS.

Wilhelm (43:54.452)
Interesting, okay. And I was shopping around a little bit and I landed on amphetamine, which is another alternative to caffeinate. And you can turn it on and then even when your laptop lid is closed, everything keeps running. So what I do is before I leave the office, I connect the MacBook to my like personal hotspot, turn on caffeinate, close the laptop, get back home, back at it.

Matt (44:10.873)
Yeah.

Matt (44:21.55)
You've got to be careful doing that because it can overheat in your bag.

Wilhelm (44:28.213)
Apparently it's less of an issue with the series laptops.

Matt (44:32.226)
Okay, I've heard some horror stories.

Wilhelm (44:33.217)
but it was warm when I took it out the backpack. Okay, what are the horror stories?

Matt (44:38.158)
Yeah, like, we're not allowed to use amphetamine in California. There's like, it's like a bunch of security problems with it as a general thing.

Wilhelm (44:46.391)
That's just the company drug policy. No amphetamines during the workday.

Matt (44:49.857)
Yeah, but no, there's like a security posture of like, if you're preventing your Mac from turning off, like why, what are you trying to, there's a security posture thing there. Can you still hear me?

Wilhelm (45:04.951)
I can still hear you, I'm just grabbing a drink.

Matt (45:07.737)
Okay. But, but yeah, I've, I've, I've heard some, I've heard, I've heard some smoky stories. You gotta be a bit careful. Don't put, don't put it in, don't put it in your bag when it's like overheating. Right, should we?

Wilhelm (45:21.805)
Just tell them, hey, boss man, I'm creating shareholder value. Mile up while I'm commuting.

Matt (45:25.785)
UGH No, no, I see at that point I just prefer to work on my phone I think push everything to a container

Wilhelm (45:32.587)
Yeah, well, that's the thing. You can work on your phone. So your cloud code remote control is what you look at on your phone while the laptop's in your backpack. So I was on the public, like on the train yesterday, looking at like the code in the cloud app where I could see what it's doing on the laptop.

Matt (45:53.429)
Okay.

Wilhelm (45:54.529)
That was it my backpack.

Wilhelm (45:59.682)
Let's do it. What do you want to talk about next? Let's do it.

Matt (46:01.251)
Can we talk about capathi?

Matt (46:05.879)
Well, he's joined Anthropic, guys.

Wilhelm (46:09.238)
You

Matt (46:10.585)
It's kind of hectic. It's kind of hectic. mean, on Jacob Pathy, mean, probably everyone listening knows who he is. But if you don't know who he is, he is, he was a co-founder of OpenAI, which I actually didn't realize that much, but then he worked, he was like head of computer vision or something at Tesla for a while. And then most recently he's been known for

making these insanely detailed and well, like the last four years, insanely detailed and well written YouTube videos about like everything and anything and very ML related. Like some of his best ones are, or some of my favorite ones from him are like building GPT from scratch, when he talks about basically building a GPT-2.

like model from scratch. Then there's nano GPT where he gets to GPT-3 like performance, I think, but on much less code and on much less tokens using more modern techniques. And he goes through basically the last 15 years of deep learning. It's kind of crazy in some of those videos. He goes through attention. He goes through the whole transformer architecture. And then what some of the other ones that he did, I don't know, just like...

I've definitely slogged through a lot of Kapati's videos. They've been, they're amazing. The depth, like, yeah, just like the way he speaks about, about machine learning is very intelligible, very legible. People like to say that word at the moment. Very legible. And I really, like, I mean, I've learned a lot from him. I think I was, it's definitely started. Yeah, no, no, no. Okay. So.

Wilhelm (47:39.595)
He's a total legend. Yeah.

Matt (48:02.265)
So I did, I did a, like a bachelor's in reinforcement learning for mechatronics systems. And then basically stopped doing any type of deep learning after that. And that was in 20, 2020. So looking a lot at like some of the stuff that came out, the RL stuff that came out in 2017. And that was like really, really fun. And like, yeah, we've got like a bit of an appreciation from that. And then when I started doing software engineering as a, as a thing, I like stopped doing all of that and like learn about AWS and all this like cloud stuff.

and servers and then when Kapathis videos came out it was just like my god like deep learning machine learning it's just it's so cool and so like fun and kind of ethereal and like I really think he captured the imagination of a lot of people not just me you know from some of from some of his videos and you can see that he has mad Twitter fame millions of followers just like an absolute legend and then he left then he started well

Wilhelm (48:34.509)
Mm-hmm.

Wilhelm (48:56.182)
Yeah.

Matt (49:00.448)
Supposedly started an education company based off the back of some of his videos And the fact he's an absolute genius and was like was gonna go do something like that and then now in a very big twist of events Twist and turn he has joined Anthropic

Wilhelm (49:18.579)
I saw an interesting article recently. I forgot to send this to you actually. I think you'd like it. But it talks a little bit about how everyone is figuring out all this stuff at the same time at the moment. And...

And sometimes it feels like you stumble across a new concept, right? Or a new thing that seems to work. Like, I'm sure you've had this where you're like, damn, I haven't seen anyone like talk about this. And this article is basically making the point that for anything that you discover where you feel like you haven't seen this discussed, probably 10,000 people on the same day also discovered the same thing. So it, it, it feels like everyone is kind of on the same journey together. And that's why it can feel a little bit like echo chambery or while everyone, everyone is

shouting in the same direction I think is the article that put it but but it singled out Kapathi as being like one of the few people that consistently wins the like rights to like name the thing that everyone discovered so like he is consistently the one who gives the thing a name and because of his like reach and

Matt (50:08.022)
Mmm.

Wilhelm (50:28.715)
history and whatever. He actually like gives a name to the thing that everyone has kind of come across. So obviously I think most famously Vibe Coding in like February last year. I think then the LLM...

LLM Wiki was like another one from like a few months ago. A lot of people were building similar things at the same time but he actually like gave it a name and like gave it a simple description and kind of like popularized it a little bit. So yeah, he's a special cookie.

Matt (51:01.131)
Definitely. Some... Yeah, special cookie. Interesting dude. I would love to have been in the room when Dario convinced him to join Anthropic. Or maybe not convinced him. So there is two schools of thought about his thing in Anthropic. So he's actually not super senior. As in like, he's not going into management or anything like he's going to become an IC.

Wilhelm (51:13.741)
Mm.

Matt (51:29.632)
He's like three or four layers down from the CEO. He's going to work in a team that is using Claude specifically, like very much like using AI to create better algorithms for pre-training. And this kind of tracks for me as like...

Wilhelm (51:52.396)
And the recurrent self-improvement, right? I thought that's what he was working on. The like RSI. Right, right, right.

Matt (51:55.478)
Yeah, well like the auto research thing. Yeah, yeah, like how can you use Claude to take the algorithms that we have, some of the most costly algorithms in the world really, like they're consuming huge amounts of compute, huge amounts of capital to run that compute, and can we make them faster, more efficient, yeah, better in all of these ways. I think it's...

Wilhelm (52:22.711)
Mm-hmm.

Matt (52:25.376)
Like.

Tesla, he was definitely a director head of handpicked by Elon type vibe before that he was a co founder of open AI like this is not a step down, but it is a

It's a step in a different direction, definitely going back to being an IC. Now what is a very large company.

Wilhelm (52:51.115)
Yeah, I thought there was a thing as to like the why he's doing this. I I feel like there was some podcast interview or something where he said that he...

because he hasn't worked or hadn't worked at a Frontier lab in however many months, he felt that he was a little bit distant from the Frontier at that point. And you could see that a little bit, right? I think in October, November, he was saying, like, agenda coding doesn't really work yet. He just prefers autocomplete. And then by December, he's completely changed his mind, which I think is quite different from what people inside the labs who could really see where the journey was going and are more at the Frontier would have said.

think he even said in that podcast, like, yeah, if there is any frontier lab that like would be interested in having him, like he would be interested in going back to the frontier and like continuing, like, yeah, doing that kind of like frontier research again. So sounds like he's doing exactly that.

Matt (53:48.917)
think I remember that one actually. Everything sounds great in retrospect, doesn't it?

Wilhelm (53:53.966)
And then should we talk about the Pope who has not joined anthropic contrary to popular belief?

Matt (54:00.861)
Dude, I have no idea about what the Pope... I just saw, I just saw like an amazing meme generated picture of Dario dressed as the Pope and that was all I know about what happened. have no clue. Yeah, please.

Wilhelm (54:10.836)
I see,

Wait, can I tell you about this? Because this is incredible. And as backstory for this, right? this is, I might get the details wrong, but the story is funny. So this Pope, right, is an American Pope. He's from like Ohio or something. He's not actually from Ohio. But I think when he became the Pope, he said something like, you know, the last time we had an American Pope was during the Industrial Revolution.

And he believes that similarly to that guy, he really sees the intelligence revolution and the AI revolution as a massive change to humanity. And he views his role as an American pope to help humanity through that, help it deal with that, or embrace that. Just like his American predecessor did. Which is just hilarious in general, but also pretty cool.

Matt (55:00.532)
Help America.

Wilhelm (55:12.981)
And I think the news recently was that one of the anthropic co-founders was helping the Pope write an encyclical letter. And I didn't know what this was. And so just going to read out the definition from the Catholic Church in England website. So a papal encyclical is a formal pastoral letter written by the Pope. It is primarily addressed to Catholic.

to Catholic bishops, but frequently extended to all the faithful and people of goodwill to provide teaching guidance or analysis on matters of doctrine, morality, or contemporary social issues. So it sounds like

The Pope, with the help of one of the anthropoconfounders, is writing a letter about AI. And it's coming out in three days, and it's called Magnifico Humanitas, or Magnificent Humanity, and it will focus on AI and human dignity. So, isn't that incredible? This is history in the making.

Matt (55:53.768)
Mm-hmm.

Matt (56:17.492)
Wow.

Matt (56:24.68)
Yeah, I...

Matt (56:30.215)
I do not want to get myself in trouble by speaking right now. Yeah, I just, I just like.

Wilhelm (56:36.299)
Really? I'm curious what you would say.

Matt (56:46.95)
struggle with a lot of this.

Wilhelm (56:47.447)
Do you think the Pope has mythos access?

Matt (56:54.007)
I... probably not. Not the Pope, but like, wouldn't be surprised if the Vatican has some levels of mythos access, I wouldn't be surprised. Dude, okay, speaking of like weird access, did you see that? Okay, can we move on?

Wilhelm (56:55.287)
Like, do you think Anthropic would give the Pope mythos access?

Wilhelm (57:12.224)
Yeah, yeah, go on, go on.

Matt (57:14.283)
There was that tweet about, I don't know what all you guys are doing, and I don't know why you're paying so much money, but in China, we pay like 10 cents on the dollar to, or even less, to resellers, and that's how we get all of our tokens. I'm here like token maxing, I don't know what it was, like a thousand pounds of tokens a day, and like, you know, I'm paying like nothing, and it's great. Like this is like, this is how we should.

Wilhelm (57:36.013)
Mm.

Matt (57:43.699)
I don't know what you guys in the West are doing, over in China, this is amazing. And we're using all the frontier models. Not many people are using the Chinese models. We're all using Opus 4.7, GPT-55, but you know, we don't pay money for them. We just use our resellers. Did you see that tweet?

Wilhelm (58:02.569)
No, no, no, I didn't see that and I don't really get it or like, yeah, tell me more.

Matt (58:05.308)
Dude, it's basically a tweet about how I think he was a university student in China was making, was like, they're spending not, I'm gonna find it. They're not spending very much money on models and they are.

Wilhelm (58:22.551)
They're buying tokens from resellers. that the... I say yeah.

Matt (58:23.154)
Yeah, from resellers. Yeah. So Chinese students are buying GPT 4.5, 5.5 and Claude API access from, I'm going to butcher this name. I'm so sorry to anyone speaking. Zhan Yu and Tiao Bao, proxy sellers for almost 96 to 97 % cheaper. People are apparently burning a hundred million plus tokens a day for like a dollar and vibe coding nonstop.

And there's a lot of idea about like what this means. Like they use these things called proxy stations. Basically someone buys a huge amount of API credits, accounts and routes, everything through their own service and then resells it for dirt cheap. The only downside is your prompts and chats are probably getting logged somewhere too. 100 % they're getting logged. I mean, this is how they're doing distillation, right? Surely.

Wilhelm (59:19.063)
interesting. I see. mean, there was a... Right, like, okay, I was going to say something else. I mean, the distillation thing kind of makes sense, but there is a pretty big discount.

And there was a thing at the Stripe sessions thing where they were talking about how there's unbelievable amounts of fraud with AI. And basically you have various companies that get free tokens for whatever reason, right? It could be like a startup grant or something else. And obviously you can see how it's pretty appealing. If you, if you're a startup and you have no revenue, but you have 200 K in tokens and you can just sell them for some haircut. But I mean, that's like, obviously it's not allowed by the providers. It's like not the idea at all.

Matt (59:36.646)
Hmm.

Matt (59:43.056)
Yeah, and they just sell them, yeah.

Wilhelm (59:58.176)
It is like fraudulent. So I thought that's what was going on. mean, the distillation thing makes sense as well, but yeah. I don't know. It sounds weird.

Matt (01:00:05.073)
I mean, they definitely, that definitely is happening. Like anyone with AWS credits now can use them on whatever the AWS model provider is called, but they can use them on called Bedrock. Yeah. Anyone with that can use them on Bedrock. So I have a bunch of AWS credits, like theoretically I could resell them illegally, you know, and like make some money, make 10 set, make what like?

Wilhelm (01:00:16.172)
Bad rock.

Wilhelm (01:00:27.979)
Mm.

Matt (01:00:32.273)
What is it like 10, three or four cents to the dollar I could make, which doesn't feel very worth it. But the other way of looking at it is like they're getting all of this stuff. The resellers are maybe, maybe I'm selling my tokens to a reseller and I'm selling them like 20 cents on the dollar. And then the reseller is selling them to like a lab at like maybe 50 cents on the dollar. So the resellers make some money.

Wilhelm (01:00:37.249)
Yeah, that's wild.

Wilhelm (01:01:00.279)
Mm-hmm.

Matt (01:01:01.285)
And then the labs are getting them at 50 cents on the dollar. And then I like taking like a 90 % cut on those to get the reasoning traces. And I saw like two tweets in line and the other tweet below, which was kind of fun was like, just take, I'm going to read it out.

Wilhelm (01:01:08.641)
Mm-hmm.

Matt (01:01:20.594)
The other tweet was talking all about RL traces. And RLs, I'm really a big fan of RL. I'm so happy that it worked because it means that stuff that did when I was 18 and 19 actually means something now, which is very, very cool. I feel for all the people working on GANs. But whatever your frontier LLM users think their taste in managing agents gives their labor a comparative advantage.

Wilhelm (01:01:32.524)
Mmm.

Hell yeah.

Matt (01:01:47.163)
follow these steps. whenever your frontier LLM users, basically whenever your users think their taste in managing agents gives their labor a comparative advantage, basically saying like, LLMs can't do this thing that I can do because I'm so smart. Follow these steps. Record their traces, replace that tasteful user messages with agent chain of thought. Basically turn them from user messages and just inline them into the assistant message.

And then make the trace a single long horizon trajectory with a single high level goal inferred by an LLM critic. So users are actually very bad at citing what their goal for a session is. So you can actually just generate that goal and then inline all of the other user messages for the whole session. And then do SFT on the traces or like RL on the traces or whatever to teach the LLM taste. Repeat until the benchmarks break basically. And...

Wilhelm (01:02:27.063)
Mm.

Matt (01:02:43.537)
I think those two tweets have something to go with together. They're definitely talking about the same thing. And I think it's very cool.

Wilhelm (01:02:49.533)
Interesting. Right, right, right. I see what saying. Yeah, yeah, yeah.

Matt (01:02:55.162)
So I love it when you get like a look behind the curtain of what's happening in some of these places. Like I have no concept, but I'm sure this is happening.

Wilhelm (01:03:00.011)
Yeah. Why is it phrased, that's true, it's phrased in such a hostile way, like, your users have taste, actually they don't, let's replace them with AI.

Matt (01:03:11.907)
Yeah, which I think is maybe why it didn't get as much exposure as people thought because you have to actually actually kind of know something about ML and like the first of all, you have to know something about like reasoning choices and then you have to have some like knowledge to even understand the tweet. then secondly, you have to have some like knowledge about the backwards and forwards between like seeing like a lot of engineers at the moment still maintaining that they have like something to offer about LLMs.

because they have taste, I guess.

Wilhelm (01:03:42.903)
Mm.

Wilhelm (01:03:47.02)
Actually, this is a good, yeah, I should say I need to run in like five minutes or something. So we can talk maybe like two more things on the list. There's genuinely so much happening, but I want to briefly mention just the experience I had at PyCon, which I think we already talked about this a little bit when we chatted off camera earlier in the week, but I was just surprised how many people, so I went to PyCon, big massive huge Python conference.

Matt (01:03:47.993)
Tough.

Wilhelm (01:04:15.841)
biggest one in the world, all the legends of Python are there, like the creator and some of the core devs and so on. And everyone I met, I basically asked them how AI-pilled they are. And I think people seemed actually medium AI-pilled, which was maybe a little better than I was expecting, but a lot of people were also still writing code by hand. And...

That kind of surprised me because, I don't think I've really done that this year, especially Python. Yeah. The first language that AI could write. Well, maybe, but yet.

Matt (01:04:45.455)
Especially Python,

Matt (01:04:54.889)
Yeah, if you're making cred applications in Flask and Django, you can probably generate those things.

Wilhelm (01:04:59.435)
Yeah. But yeah, I think it was just a bit of like a moment of like, wow, yeah, I am in a bit of a bubble. Like, there are a lot of people or like, yeah, this kind of thing changes slowly. And people still think about this like very differently. I think also a lot of people were saying things along the lines of like, yeah, AI is a tool. Sometimes the output is good. Sometimes it's not. And

I don't know how to feel about that kind of take to be honest, because I feel like if you say... Right, right, right. Yeah, yeah.

Matt (01:05:32.943)
This is how we spoke about like a year ago, dude. This is like how we talked about a year ago. It's like, sometimes it's good, sometimes it's not. Do you go back and change the inputs or do you fuck with the outputs? And like, I was kind of, well, we need to pull up those, what we talked about then, because I think one of us was very fucking with the outputs and one of us was fucking with the inputs. And I actually think fucking with the outputs is the forward thing to do, right? Because you have more trust in the other one.

Wilhelm (01:05:41.835)
Yeah, yeah,

Matt (01:06:01.763)
Fucking with the inputs is like changing your, yeah, changing your grounding.

Wilhelm (01:06:08.245)
Maybe it's just like a timeline thing. like you say, and this is just like kind of what we were doing like a year ago. But yeah, I just feel like if you have a take like, yeah, AI is like a tool. It's kind of like underplaying it a little bit. Like our industry is changing completely, right? it's, I don't, yeah, I don't know if I would consider that take slightly.

Slightly harmful or or or just a bit dangerous like you

Matt (01:06:32.566)
Reductive. I just think it's, I think it's dangerous to these people's job prospects, first and foremost. I don't think, you never get paid for doing like artisan stuff. That can be like a hobby and maybe you'll get paid enough for your hobby that you don't, that's fine. like work does not reward artisan behavior in like handwritten organic stuff.

Wilhelm (01:06:50.743)
Mm-hmm.

Matt (01:07:02.318)
in so much, in so many ways. You look at the rise of supermarkets, you look at the complete decimation of anything to do with the farmers market in the UK that exists outside of central London where people are earning 100k year. It's just like complete decimation, the rise of Lidl, know? Supply chains and all of this sort of stuff. People, like this...

Wilhelm (01:07:11.949)
Mm.

Wilhelm (01:07:17.249)
Hahaha.

Right,

Matt (01:07:27.424)
No one, no, I'm using as a food example, which is kind of scary anyway, because like this is also like the rise of fast food and the rise of like really unhealthy food and this sort of stuff. I don't think they're necessarily, I don't think causation equals, I don't think correlation is the same as causation, but I think there's different things that work here, but I think you're gonna struggle. you previously made floorboards by hand, for instance,

Wilhelm (01:07:30.753)
Yeah, yeah,

Wilhelm (01:07:36.703)
Yeah.

Wilhelm (01:07:42.86)
Yeah, yeah.

Wilhelm (01:07:47.305)
Yeah.

Matt (01:07:54.646)
And then now a machine comes that can cut floorboards. Yeah, sure. You might find, you might be able to use better materials than the machine that cuts floorboards because the machine that cuts floorboards, they're using worse materials because they've realized that they can like bulk sell it and they're optimizing every part of the floorboard cutting process. But, and maybe you have these like solid wood, like beautiful floorboards and maybe the machine that cuts floorboards is using some engineered wood or whatever. But like, like that machine that cuts floorboards is now not only

Wilhelm (01:08:02.487)
Mm-hmm.

Wilhelm (01:08:12.107)
Yeah, yeah,

Matt (01:08:22.987)
100x better than you, it's now 1000x better than you and you are now relegated to a hobby.

Wilhelm (01:08:28.203)
Yeah, it's a bit like if you're standing in Manchester in the mid 1800s and you're like, you know, mass manufacturing, it's a tool. You win some, you lose some. Whereas the Industrial Revolution is popping off around you. Yeah, exactly. It's like, you know the printing press, it's a tool, but I kind of prefer, yeah.

Matt (01:08:41.677)
Yeah, like a printing press.

Matt (01:08:50.209)
The printing press is a tool to get ideas out, know? Like sometimes it works, sometimes it doesn't. You I can probably just still write my paper by hand and like the people who get my paper, they'll love my paper, you know? Like if I go to a printing press, the quality of the paper might be worse, you know? Like it's just, might be tough at the beginning, you know?

Wilhelm (01:08:54.413)
Yeah, yeah.

Wilhelm (01:09:02.951)
Or when the internet... Imagine you're in like... In 1995, you're like, yeah, computer networking, you know, it's a tool. Like, you don't have to send everything over the network. Like, that's silly.

Matt (01:09:20.397)
Yeah, that would be, that would be weird. know, like floppy disks, know, like floppy disks, can still, we can still, like sometimes the packets might not get there. And so I'll just send a floppy disk as a backup every time, you know, we'll just send it, we'll just send it as a backup, you know, it's like postal service. You postal service, super reliable, genuinely amazing postal service. You know, the thing that we're all super used to, I that's actually a really great, great analogy because, because like, as we know, the postal service loses shit all the time.

Wilhelm (01:09:34.965)
Yeah. Yeah.

Wilhelm (01:09:43.817)
Yeah.

Matt (01:09:49.752)
and it's just like, because it's human error, like all the time it's human error but packets get lost quite deterministically.

Wilhelm (01:09:51.638)
Right.

Wilhelm (01:09:55.338)
It's UDP.

Wilhelm (01:09:59.602)
yeah, interesting. I guess we're all trying to figure this out together or something. Okay. Yeah. I should run. There is, man. I, I, so I went to the Slack event two days ago. I need to tell, yeah, I just think Slack is back open for business. Like it is wild. They, I think now might be the best time in a long time to build something new on the Slack platform. Like.

Matt (01:10:12.958)
You wanna talk about something? Tell me.

Wilhelm (01:10:29.293)
I haven't had that belief in a while. mean, so, you know, Simbubble was built over 10 years ago on the Slack platform. Of course, like worked very well for me, but they're making like, it's, it's funny because I think Slack for like a year just spent like all their engineering resources on like integrating with Salesforce. But now it sounds like the platform and the like app ecosystem is like a top company priority again, or the company priority.

And the way they think about this is that like Slack is the best and maybe only place where you can have multiplayer AI agents. So we're like, you can collaborate with others instructing AI agents and they kind of want Slack to be like the front end of like people interacting with AI agents. Like they'll be running on various like servers in the backgrounds and like various cloud providers will power them and you'll have various specialized agents for like

all these different business activities that you do, but they'll be driven through Slack. Like they really care about that. They really want to make that happen. And the people building this are like really, really competent and smart. like, anyway, I just had a really good time at this event. I didn't expect it at all.

Matt (01:11:36.044)
Yeah. I hate this. I hate this, but I believe I hate this, but I believe it like 100 % believe it. Like I remember chatting with someone that stack one. So that must've been like over a year and a half ago now. And we were like, dude, Slack is the perfect UI for agents. Unfortunately, it's so annoying, but it's perfect. Like there are so many annoying issues with that. People get really into like messaging apps, but like discord is okay. Discord.

Wilhelm (01:11:44.845)
Yeah.

Wilhelm (01:11:53.836)
Right.

Matt (01:12:05.942)
Discord is okay. Slack, unfortunately, is the place where people live. GChat's an absolute clusterfuck. We use that at Cloudflare and there's just so many problems with it. any messaging app is a place for agents, because it's a place for people. And you need a place for people and agents to chat. we tried all of these custom dashboards and things, but actually, agent is not a dash... It's very hard to distill what it's doing into a dashboard.

Wilhelm (01:12:07.904)
You

Wilhelm (01:12:13.483)
Yep.

Wilhelm (01:12:23.169)
Right.

Wilhelm (01:12:35.624)
Right. Yep, yep, Mm-hmm. Mm-hmm. Mm-hmm, mm-hmm, mm-hmm.

Matt (01:12:35.872)
Distilling it into threads and channels makes way more sense because it's like time sensitive. Dashboard is gonna show you something from years ago at the same place as something from yesterday. like the whole idea of messaging is actually agent first. think I'm fully with you on this and I hate it so much, but yeah, I'm with you.

Wilhelm (01:12:55.553)
That's a very good description. Yeah, I like that.

What, yeah, I was just very surprised that like Slack is got like the foresight and the kind of internal alignment and commitment and whatever to actually invest in this now. There was this meme going around that like Mark Benioff, right, the CEO president of Salesforce, that he didn't really get the point of Slack. And that obviously is a bit of a problem when you've acquired a company for like 30 billion or however much it was.

but it sounds like he really gets it now and he's really bought into this stuff as well. I know. So if you're listening to this and you're building stuff, I mean, I feel excited about building stuff on the Slack platform again.

Matt (01:13:33.15)
I don't think, I've seen a few like Slack competitors and unfortunately it's such an easy roll up into a larger business organization. really need like, like with Teams for instance, like Teams is rolled up into every Microsoft subscription. GChat is rolled up into every Google subscription. Like it's so that the competitors are free and like, and like, so you really have to have that like.

Wilhelm (01:13:52.279)
Right.

Matt (01:13:59.742)
deep integration with company culture and like Slack kind of has that. All of our customer channels are in Slack, you know? Like you can't get away from it. You can't get away from it. And there would have to be a big agent of disruption for Slack not to be still be here for the next generation of agents. really do think so. Can we talk, I know you have to go, but can we talk very briefly about the money thing, about the SF bubble?

Wilhelm (01:14:09.227)
Right, right, right, right, right, yeah, yeah, yeah.

Wilhelm (01:14:25.165)
yeah, yeah, please. Yeah, yeah, go for it.

Matt (01:14:28.735)
And how people, people just, and I see this on Twitter as well. I'm definitely guilty of feeling this way. I think I genuinely think guilty is the right word of like, you see, you see friends getting acquired by open AI and Anthropic and one half you feel for them like, shit, they lost their agency. And the other half you're like, my God, these companies are going to print trillions of dollars in the next few years. An example I heard was the Anthropic's revenue.

Wilhelm (01:14:43.533)
Mmm.

Wilhelm (01:14:54.145)
Yeah, I am.

Matt (01:14:58.62)
in the first three months of this year was larger than Databricks and Snowflakes is in the full company valuations. Like just like, just like, just like, just like think of that. there, it just doesn't make, that just doesn't make any sense to me. Yeah.

Wilhelm (01:15:13.173)
Yeah, I think if the growth continues, if the growth continues, think Anthropic will have more annual revenue than Apple next year.

Matt (01:15:23.434)
Yeah, yeah, yeah. So, and then when you're thinking about these numbers, it's of brought to home to me a little bit when there was an amazing article from someone called Nan, who we can link below about the philanthropy that might come off this. And Americans are really big on philanthropy, especially when they make billions. And this article kind of brought some of the money to home as well because like,

Wilhelm (01:15:25.077)
Like what?

Wilhelm (01:15:47.911)
yeah, yeah, I saw this, I saw this, yeah, yeah, yeah.

Matt (01:15:53.322)
on average people donate sort of like billionaires specifically donate like something like 10 % of their money to a foundation that foundation invests in like charitable projects and like sometimes they get a return sometimes they don't but they try and like make impact on like social impact on the world and the whole article was about how like

Wilhelm (01:16:15.179)
Mm-hmm.

Matt (01:16:18.516)
the amount of money that the Anthropic founders and first employees and OpenAI Foundation are gonna create. It's gonna consume this like entire industry. It's just gonna like feed into this entire industry. Even at like conservative values, there's gonna be these insane amounts of money flowing into social impact, which is a very like positive, potentially positive outcome of all of this. But just like this, it just brings to home like the scale of this like value creation.

Wilhelm (01:16:43.693)
The scale is wild, yep.

Matt (01:16:46.441)
And I think a lot of people are talking about the scale of money because we've just never seen value creation in this way. And that flowing down to an individual, especially individuals who are sitting kind of like me and you actually, sit like, we sit looking at a lot of this stuff, very, very front and center. Like you're in San Francisco and I work a lot with like these companies and you see people like this who are making generational wealth and you're a little bit like, oh my God, like why is this not me? Why am I not? I mean, I'm doing very well for myself.

Wilhelm (01:16:54.605)
Mm-hmm.

Yeah.

Wilhelm (01:17:03.964)
Yeah.

Wilhelm (01:17:14.069)
Yeah, man. Yeah, yeah, yeah.

Matt (01:17:16.381)
But like, why am I not earning billions, know, like doing this? There was a tweet, which we'll link below, that brought this into a very high, like, front, mind and centre because like the SF brain rot of money is insane. They like worship this like deity called dollars and it feels very unhealthy and yeah.

Wilhelm (01:17:26.509)
Yeah.

Wilhelm (01:17:32.749)
You

Wilhelm (01:17:44.077)
Clearly Matt, we need to pivot into non-profits, given that there are going to be these hundreds of billions that flow into the charity space. Why don't we start a charity?

Matt (01:17:52.455)
I'm very excited if that does happen because this is big up. I think if I ever started a company, like again, like a proper one, would consider, would definitely start a public impact company. Like what do they call it? PPM something or other PMC like a public benefit corporation because dude, the worst, there was another amazing podcast by Eric Reese, the lean startup guy. He was on Lenny's.

Wilhelm (01:18:08.823)
Public benefit cooperation.

Wilhelm (01:18:18.765)
Mm.

Matt (01:18:19.688)
podcast and he's talking about his new book, which I really can't wait to come out called incorruptible. They use Cloudflare as a great example of a company that is supposedly quite incorruptible because Matthew Prince and Michelle Zatlin own, they have hectic founder control, very similar to Gazette, where investors can't force them out because they have this crazy founder control. So for instance, if like the tobacco company, Philip Morris offered them 1.01.

Wilhelm (01:18:26.445)
Mmm.

Wilhelm (01:18:35.373)
Hmm.

Matt (01:18:47.91)
dollars to the dollar of share, they wouldn't have to sell. that's what most companies have for duty duty to sell that stuff. And Anthropic also has the same thing. And they have this like public benefit trust that has like an oversight trust about whether they're keeping safe and proper AI and they have controls and all this sort of stuff. And if an acquisition offer came that went against those ideas, they wouldn't have an obligation to sell, which I quite like. I don't know how we got to this bit.

Wilhelm (01:18:52.903)
Hmm, no I didn't realize that.

Wilhelm (01:19:07.212)
Mm-hmm.

Matt (01:19:17.052)
But yeah, these are like very big roundup of the last things that I've been interested in.

Wilhelm (01:19:18.197)
Yeah, yeah, yeah, No, I think there's definitely a vibe, the whole permanent underclass thing, the whole like, yeah, I mean, it's just, I think it's, in the 2010s, we had exits and you would have millionaires or like, deca millionaires, as in you would have like tens of millions from your colleagues overnight or whatever, and that already was a bit like...

I know, maybe an unusual experience to have your colleague just suddenly be super wealthy in that way. But now it's going to be bit crazier because you will have like, think actual ICs like becoming billionaires, When these, if the growth continues and open-air anthropic, et cetera, go public. So...

Matt (01:20:07.429)
Yeah, friends of mine in Anthropic, like a lot of them will get something along the in the region of 10 to 100 million like ballpark, like current valuations even potentially. Or maybe it's slightly projected valuations, but like that's wild. And like you could even see if we went 10x again, that then that suddenly starts becoming a bit ridiculous. Well, it's already ridiculous, but becoming insane.

Wilhelm (01:20:15.87)
Yeah.

Yeah. So.

Wilhelm (01:20:30.277)
I do think it's worth touching grass, which I think you've been an advocate for. And you're a grass advocate.

Matt (01:20:42.98)
I think I'm a sand advocate now. Like, I think you really have to like lie in a sand pit and just realize that there is more to this world than tokens.

Wilhelm (01:20:50.027)
But also besides touching grass, think people do get obsessed with the numbers or whatever. It's really not about the money. And I think this is very easy to forget. It's not about token maxing. It's not about building as much as possible. It's about like creating, using these tools to create what you care about, right? And build what you want to see in the world and craft that in a way that feels true to you. I think that's like, that is easy to forget and to get lost in all of the...

various maxing of various numbers.

Matt (01:21:22.135)
I love that one. Did you see? Okay, last thing and I promise and just like kind of like echoing what you just said. There was an amazing article, article, Twitter article by Will Mandesis. I think we'll try to link all of these below, but this was an article like against what he's calling grind slop, which is like.

Wilhelm (01:21:26.582)
Yeah.

Wilhelm (01:21:42.017)
Mmm.

Matt (01:21:42.982)
You always have that friend who everyone says just works really hard and then some people are like oh man He works really hard like you know, it's amazing works with how what does this company do? I don't know but he works really hard, know, like he's he's in that place He built a cap and I think he started that he started the start of the article is taking the piss out of a tweet that Harry Stebbins made recently where he met a founder that works like seven days a week and Built a cafe inside his office and has a bed underneath in the corner off it in the corner of the room

Wilhelm (01:21:54.817)
Yeah

Matt (01:22:13.096)
so he doesn't have to leave and all of this like crazy ass shit that really like people need to like judge themselves under a microscope and be like my god there are other things to do in the world but and you only have one life and you should really live it because like this is probably probably you'll look back and think this isn't living it and there are some very like common regrets that people have before they die and like one of them is not seeing family one of them is not seeing friends one of them is not traveling and like you're not doing any of this stuff if you're

Wilhelm (01:22:26.925)
Mmm.

Matt (01:22:41.314)
if you're hiding in the corner of a room in San Francisco grinding on something that maybe doesn't hugely matter in the world. I really like, I kind of hate, there's two things that I kind of hate at the moment, and this is going to be outing me a lot, is I kind of hate people's obsession with building stuff that...

Wilhelm (01:22:49.389)
Mm.

Matt (01:23:06.703)
probably doesn't matter. Like there is already a way to do that on the internet. Like you're building a new thing because you think your thing will be able to make a little bit of money, maybe like 20 % or 30 % of the market value of the old thing. And you know, it's probably never going to succeed the market value of the old thing, but you think there's space for another thing. You can maybe cut the margins and have cheaper infrastructure or whatever. And you think this will make you a bunch of money. And it's like, no one needs the nth piece of accounting software for dogs. know, like it's,

Wilhelm (01:23:34.686)
You

Matt (01:23:35.173)
You can find other things to do. Your life is way more worthy than that. that's, very, the other thing about this is like the YC thing of getting you to pivot, pivot and pivot until you find this very small niche of people that love your idea. And I think that's great when you say it like that, but some of the weird outcomes of that is that people just end up converging on making the same stuff. And that same stuff is really like.

Wilhelm (01:23:48.119)
Mm.

Wilhelm (01:23:58.862)
Mmm.

Matt (01:24:01.605)
you're trying to be an opponent, none of you are making that much meaningful impact in the world. You're all doing just what you can see right in front of you because you're just very like short-termist about these things. I think it's one of the really exciting things about Nan's post about potentially this huge influx of money coming to social impact is maybe we'll get a whole new generation of founders are really trying to like solve things in the world. Like I did some, I just back like just crazy little bits of clawing. Apparently 300,000 people die from drowning a year. Just like.

Wilhelm (01:24:30.679)
Mmm. Mm-hmm.

Matt (01:24:30.906)
just drowning. Most of them are kids, like very young kids. This is like one of the easiest solved problems in the world. Like you teach them to swim and you put a gate around a pool if you had a pool. Like they're not like, it's not, they're not hard problems. And we don't do it like comparatively around 500,000 to 600,000 people die from malaria. And that has been, that is coming down rapidly from similar, very small interventions, like getting nets.

Wilhelm (01:24:43.33)
Hmm.

Wilhelm (01:24:53.325)
Hmm.

Matt (01:24:59.523)
like getting mosquito nets, having fans, like this, these types of stuff. 2.5 million people die. I mean, I Googled this stuff after the article. 2.5 million people die a year from burning solid fuel inside homes and they die from things like lung cancers and stuff. Like that's horrific. Like we need a whole new generation of founders. And I'm very excited if there's the money to do this where like people can build smarter solutions to actually like elevate their fellow man. And I'm really stoked about that.

Wilhelm (01:25:14.743)
Hmm.

Wilhelm (01:25:29.197)
Go forth and build, everyone.

Matt (01:25:33.519)
Yes boss. Dude this was the most philosophical one we've had in a while.

Wilhelm (01:25:37.581)
Yeah, I feel like we need to have some kind of papal meme in the title.

Matt (01:25:45.157)
We do. Or we need to have us dressed as popes. Or is that, is that, is that not allowed? I don't know. Could I be Jesus?

Wilhelm (01:25:46.625)
Hahaha

Wilhelm (01:25:50.701)
Is that too much? I don't know. Cool. Come back next time to find out. All right. Peace, everyone.

Matt (01:25:57.701)
dear.

Thanks guys, see you soon.