You've Been a Bad Agent

Wilhelm (00:01.452)
What's up? Happy Friday.

Matt (00:02.802)
Yo boss man! You too, you too, nice to see you.

Wilhelm (00:07.692)
I just got up so I'm a little sleepy, but trust you're well rested and can carry the show.

Matt (00:13.745)
I'm relatively well-rested. Relatively. I found out I didn't get fired last night, so that was great.

Wilhelm (00:23.208)
shit. Bloody hell, I know we were joking about its earnings week on Tuesday.

It really is earnings week.

Matt (00:35.115)
Yeah, yeah. I just wanted to acknowledge- No, I'd like to acknowledge it, but I don't really want to talk

Wilhelm (00:37.72)
Would you like to make a statement?

Wilhelm (00:43.18)
Yeah, yeah. Okay, okay, nice. That's good.

Matt (00:44.523)
People who, yeah, if they read the blog post, they'll know about as much as I will. So there's no point in chatting about it. I think it's... Yeah, it's cool.

Wilhelm (00:56.906)
It's probably too early to make jokes about it, so I'll refrain.

Matt (01:01.311)
Well, I've been making some jokes to my colleagues and to be fair, especially over that message, it really doesn't go down very well. I don't think they understand my dark British humour and I end up having to just apologise profusely two seconds later because I feel bad. Yeah.

Wilhelm (01:07.894)
Yeah, yeah, Yeah, yeah.

Wilhelm (01:16.258)
Yeah, yeah, yeah, yeah. It's always tough doing, like, big, big sarcasm over text. But, well. We'll get there, we'll get there.

Matt (01:24.722)
Yeah. We'll get that. We'll get them. But dude, we're nearly a year into running this pod. 14th of May. If we wait a couple of days when this gets launched, it will be, it will be almost the 14th of May. Very cool.

Wilhelm (01:33.528)
Damn. That's right.

Wilhelm (01:40.481)
Yeah, that's wild, actually. The past year has flown by. Some people have told me that one of the things about living in San Francisco is that there's no season, so you don't notice time passing. Or the way they describe it is you don't notice yourself get older. The city is a playground and you don't notice time passing. So you just stay a child forever and you don't notice yourself get older. But I'm not sure if it's AI or if...

Matt (01:52.574)
You don't notice my passing. My name is Justin, you don't notice yourself get older. But then he has a playground and you don't notice my passing.

Matt (02:05.533)
That explains a lot.

Wilhelm (02:10.766)
You mean I can't play with my robot all day?

Matt (02:13.33)
It's like Peter Pan vibes in San Francisco a lot of the time, I think.

Wilhelm (02:18.19)
Interesting, But yeah, has the past year felt really quick for you as well? Or really weird? really, like, yeah, really quick?

Matt (02:26.186)
I've done quite a lot. don't definitely quick. So a year ago I was working at stack one the startup. I was like, Will we need I was missing you man like you were you just moved to San Francisco. I was like we need to chat more. And we always chat like nerdy tech stuff. So we were like, should we record these chats? And then in like super fun and then the summer.

Wilhelm (02:33.848)
Mm-hmm.

Wilhelm (02:39.511)
Yeah.

Wilhelm (02:46.423)
Yeah, yeah, yeah.

Matt (02:53.703)
started thinking about maybe joining CloudFlare. Joining CloudFlare, obviously, that was a big move for me. And then coping with all the changes to how I work and having a big team and big company. Felt like there was big changes there. Shipped some fun stuff, code mode, worked on MCP. Wow, it's like, yeah, crazy shit. And then we got to artifacts the last couple of weeks.

Wilhelm (02:56.942)
Mm-hmm.

Wilhelm (03:11.982)
Yeah.

Wilhelm (03:17.432)
Huge.

Wilhelm (03:22.872)
artifacts.

Matt (03:22.877)
Got like made friends with like AI engineering stuff, made friends with loads of random people. OpenCore happened, met Peter, met Mario and like all of these like cool individual characters. Armin as well, we spoke about him on the podcast for such a long time. Then I got to meet the guy. Super stoked. So fun. And then yeah, now we're like a year later and I feel like a lot has changed. I feel like the models have gone from

Wilhelm (03:29.738)
Open call happened. Nice, yep.

Wilhelm (03:38.56)
Yeah, yeah,

Matt (03:50.567)
being really quite good and you can be like productive to making drastic change in the economy and like they're like insanely good now and like I feel like GPT 5.5 is actually another almost another step change from Opus 4.6 like not in terms of way chats but like in terms of how autonomous and how like how good it can be from in a one shot mode is insane insanely good.

Wilhelm (04:05.006)
Mmm.

Wilhelm (04:09.198)
Mmm.

Wilhelm (04:15.128)
There's actually so much we didn't talk about when we recorded just like three days or four days ago that I need to ask you about. A lot of little things there. My, shit, my roundup of the past year. my God, I'm not ready for this. I have no idea. I mean, what happened? Well, it actually took forever to get settled into life here. Just like lots of small things. I feel like I adapt to new places a bit slowly.

Matt (04:23.041)
I want you to do your round up first, come on, come on, talk to me, because you've done fucking loads, yeah, yeah, yeah,

Wilhelm (04:40.814)
I think it took me two years of living somewhere in London before I realised that there was a great bike shop on my street and I should just go to them to get my bike things done rather than hiring some massive chain that would pick up the bike to service it, this kind of thing. Damn, yeah, what happened in the past year? Lots of experimenting, lots of... Lots of... Lots of things...

Matt (04:47.958)
Hahaha

Wilhelm (05:06.604)
Lots of things that didn't go anywhere. No, I wouldn't say that. Lots of being at the forefront of playing around with AI, lots of trying to get AI to use Colo for tracing and realizing that's not really how the AIs want to work. Lots of things start getting a lot more fun when I started using AI to solve my problems rather than trying to build something super AI native. If that makes sense.

Matt (05:09.2)
that must have been at the forefront of today's run. I must have tried to get the ideas.

Yeah, agreed, agreed, agreed, agreed.

Wilhelm (05:36.002)
That was really interesting. But yeah, I might need to give a proper roundup of actually everything that's happened next time or something because I have no idea what happened. It's just, it feels like a month has passed.

Matt (05:44.455)
best times for them because I have no idea what that

Matt (05:51.848)
it does it can't do man like i mean i've even since like january for me like i moved to lisbon let my yeah i didn't even mention that that side note in the thing i like found a whole new flat a new like work culture a new like like culture in general trying to learn a new language meeting new friends like making new friends as well yeah because this nature is great like but you had this as well moving to summer just go from london i feel

Wilhelm (05:58.435)
Yeah, you didn't mention that. Yeah, totally.

Wilhelm (06:04.483)
Mm-hmm.

Wilhelm (06:09.699)
Hmm.

Wilhelm (06:14.496)
access to nature.

Wilhelm (06:19.682)
Yeah, yeah, it's true, it's true. Yeah, lots of new friends, like lots of... Taking forever to get sorted. Yeah, some of the highlights actually of that whole journey were the cyclists in San Francisco, just super friendly. A lot of them work in tech, but they don't always talk about it. Or like, it's not like they're... Well, there's like the San Francisco meme that you show up to a party, everyone's wearing their backpacks, which by the way...

Matt (06:23.899)
and just take them forever to get sorted.

Matt (06:40.305)
That's nice.

Wilhelm (06:47.63)
No shade. I love wearing a backpack at a party. But it's like just to set the scene.

Matt (06:47.696)
Thank you.

Matt (06:53.64)
I'm imagining you sat on a sofa wearing a backpack, just like, you know that really annoying thing where you have to sit like right on the edge of the sofa because like, because your backpack is taking up all the space and then just with your back basically to everyone else on the sofa you're just slightly further in forwards than they are.

Wilhelm (06:58.079)
Yeah, yeah, yeah.

Wilhelm (07:10.19)
Hell yeah. You gotta protect your stuff, man. Your clam-shelled laptop is probably half a jar inside the back. Well, I didn't mention this to you, but we had some friends visiting a few months ago, and I wanted to show them weird techie San Francisco and took them to this event at this really cool tower called Frontier Tower, which is on Market Street. Have you heard of this place? It's a bunch of people from Germany or from Berlin, I think.

Matt (07:34.759)
Thank you.

Wilhelm (07:38.831)
And it's like a mixture of like a, well, I'm gonna get the story way wrong, but it's like, I wanna say it was like an abandoned tower. They like leased it. They have like 10 floors in this tower. Each floor has like a theme like biotech or crypto or like AI obviously. And they host like, I wanna say like an event a day. It's like super active. You can get like a really cheap coworking membership. Anyway, that's the whole scene.

If you remember the, the peptide rave that was reported on in the New York times where people showed up to like learn how to mix their own peptides, inject peptides that happened there, I think. anyway, so they had an event on and it's look, it sounds a little bit like it would be kind of like a Berlin rave vibe, but kind of more techie. And I was like, that would be hilarious to go to. So we went to this and it was like a tech meetup type vibe, but it was dark and.

Matt (08:18.907)
It sounds a little bit like a living grave style but kind of more techy.

Wilhelm (08:33.742)
red light and there was a DJ playing like Berlin techno music and everyone was but there's no alcohol. So it was just like sober dancing with backpacks on which I thought was a vibe but maybe takes get used to anyway. Where was I going with this? Oh, so the meme is that like the San Francisco you go to a San Francisco event everyone has their backpacks on which again no shade just to set the scene and

Matt (08:37.194)
I no idea. I have no idea.

Wilhelm (09:01.166)
And everyone just asks people like, what are you building? Or like, ask these very like over thought, like openers, like, what are you excited about right now? Or, or this kind of thing. And anyway, and my point I think was that the cyclists are all really smart, but they're also really into cycling. And at these cycle gatherings, you almost don't talk about work. But then you realize later on that like, wow, like this guy actually does some really cool stuff.

Matt (09:06.182)
I actually hate those questions.

Matt (09:23.492)
Yeah, that's nice.

Matt (09:27.342)
You could raise a seed round and hire all your engineers from that cycling group.

Wilhelm (09:30.732)
Yeah, exactly. So the cyclists have just been, and the triathlon club have just been particularly welcoming and particularly friendly and nice and just really, really great. And my Whole Foods order just got delivered.

Matt (09:44.058)
Yeah, feel that. I hear, went to this, I had show this crazy coincidence, I've got to tell you. Did I tell you about that, sweet? About my next door neighbor? No, I'm guess no. So I went to this like climate summit event with Juliet and we're in town. It's just like a real chill like meetup drinks, like climate meetup drinks. And I started chatting with this French guy and he's like.

Wilhelm (09:53.544)
I... No.

Matt (10:13.957)
very quickly we get started chatting. He's like, what are you doing? And like, it's a work event. So I'm like, oh, I don't really like work in sustainability, but I'm like interested in with my partner. Actually, I work at Cloudflare and he's like, oh, like, do know this guy at Cloudflare? And I'm like, well, yeah, like, of course I know this guy at Cloudflare. And so then we get like chatting about this and this is like, it's kind of cool. Turns out he worked at a bunch of like B2B SaaS companies, mostly in London. And then now he lives in Lisbon, but he was French. And then he was like, oh,

Wilhelm (10:24.078)
Mm.

Wilhelm (10:39.435)
Mmm. no way.

Matt (10:43.353)
You just moved to Caparica, which is where I moved across the bridge, like by the beach. And he's like, you should meet my friends, like, and these people. And then then then was like, I'll introduce you. Okay. He like three days later and I get a message being like, so I think you might have already met my friends. They live upstairs from you. And their friends are our next door neighbors who we had.

Wilhelm (11:08.78)
no way!

Matt (11:12.997)
not a problem with, they they're like, they had all these plants on the balcony and it was like dripping through our balcony. And it was just like a little bit annoying when we first moved in. And our landlord actually went crazy and like turned off their water and stuff. And it was like quite funny. I had to, I needed to, I mean, it was really annoying at the time. So our landlord is like mega nice. He's like super, he's like protective over us. was really, really, really cute of him. But yeah, so we went up and had drinks with them and.

Wilhelm (11:26.093)
Whoa.

Wilhelm (11:40.206)
That's awesome.

Matt (11:40.285)
There's a quote on the wall in like wood, like engraved wood by Chris, I'm gonna destroy his name, but like back B-A-K-K-E.

Wilhelm (11:52.311)
wow, yeah, yeah, yeah. Wait, that rings a bell, that rings a bell. I feel like this is like a Twitter guy.

Matt (11:53.72)
Do know who he is?

Matt (11:58.37)
Yeah, he's an indie hacker, I'm sure he is. And so then like, obviously I like couldn't... Yeah, Chris?

Wilhelm (12:00.59)
you

Wilhelm (12:07.278)
Yeah, yeah, yeah. I see his profile. 500,000 followers. Damn.

Matt (12:09.029)
Yeah, Chris Bach. Yeah, yeah, yeah, exactly. This guy, this guy, meandering founder with X's to X, indeed, Zillow. Anyway, so I see this quote on the wall in this guy's office when he's like showing us around his flat. And I'm like, mm, sus, sus. So then we start chatting, whatever. then eventually like the topic comes up, like, what you guys do for work? And he tells us like, he's just, he's an indie hacker. He's just like, and he's just sold.

Wilhelm (12:35.17)
way.

Matt (12:37.241)
not sold the company but he's like stepping back from one of the companies that he was working with working on and it looks like pretty successful and I was just like what are the odds that I've just found another person in my little bubble who lives now upstairs from me so it is a tiny world Lisbon is even smaller than London in London it would have been like some finance bro who lived upstairs from me and yeah it's kind of of cool I had that little pinch me moment where I'm like wow maybe I do live in a tech cup

Wilhelm (12:41.102)
Uh-huh, uh-huh.

Wilhelm (12:48.3)
Yeah, Totally speaks the same language. No, that's so cool. Yeah, yeah.

Wilhelm (12:59.052)
Yeah, yeah, yeah.

Matt (13:04.578)
I never really thought of Lisbon as a tech hub, but pretty cool.

Wilhelm (13:09.656)
There's definitely, I I feel like, yeah, I've never thought about this, but it always seemed like there was sort of a wave of people moving to Portugal alongside like Peter LeBowls, who obviously is the king of indie hackers, our leader, our fearless leader.

Matt (13:20.803)
lost time thanks to people.

Matt (13:25.046)
Yeah.

Matt (13:28.76)
our leader which I actually want to meet sometime he lives very close to me well not very he lives like 45 minutes away from me he lives up in Eresir and I really do want to go and meet him sometime I messaged him but he never replied so I'm gonna try get an intro I think he I think he famously doesn't like people and like meeting people even though he talks about all of these like co-working events that he runs in his house I think they are quite selective so I need to get in on that selection because I would quite like to meet the guy he has some

Wilhelm (13:35.255)
Yeah, yeah.

Wilhelm (13:43.374)
Hmm.

Mm-hmm.

Wilhelm (13:52.397)
interesting. Yeah, yeah, yeah.

Matt (13:58.2)
fun takes but I really like his we'll get off this in a second I realize this is probably very boring to people who don't care about Portugal but he has this very like he has this very like positive not positive like outwardly accelerational view on like what Portugal could be and like where it's going delivered in a very dry Dutch way like

Wilhelm (14:05.345)
No, I think it's great.

Wilhelm (14:25.498)
I see, I see, I see.

Matt (14:26.434)
A very blunt, this is shit, it needs fixing. But I think under the hood, he's probably quite accelerationary about it. I think he like genuinely thinks it can be fixed and Portugal's amazing and he still lives here for a reason, right? Like, it's cool.

Wilhelm (14:29.974)
Yeah, yeah, yeah, yeah.

Wilhelm (14:37.92)
I just, I love the kind of breadth of stuff that he tweets about. And to be clear, I definitely don't agree with him on everything. I think he does have some takes that are wrong, but I just love how he's like, yeah, like you should do, like he just gets like a, say, Stripe to like fix some random bugs that maybe were like annoying to everyone. But then he's also like, yeah. And by the way, everyone, here's exactly how often you should lift and how much creatine you should take. And everyone was like, yeah, cool. Yeah.

Matt (14:45.839)
definitely.

Matt (14:52.963)
you

Matt (15:00.781)
Yeah, and here's my dehumidifier that I have in every single room and also the portable carbon monoxide thing that I carry around with me on my person. What?

Wilhelm (15:10.333)
Exactly, Oh yeah, actually, what's your carbon monoxide stack at the moment? Because I think I bought one that Levels recommended, but I'm not sure if it's still the good one. I know there was an update from him recently. Yeah, what's the one that you use?

Matt (15:26.23)
I don't have one dude.

Wilhelm (15:28.294)
you don't? you're just a joke?

Matt (15:30.388)
No, I was taking the piss. I don't have one.

Wilhelm (15:33.814)
fuck, okay, I have... I was thinking of getting some more. Yeah.

Matt (15:36.235)
Of course we do!

yeah i love i love all of this stuff and then his profile picture is like him topless like like basically in shrimp mode on a sofa like with his laptop on him yeah shrimp mode's amazing but it's like it's so funny all of this yeah anyway yeah it makes me laugh it makes me laugh can we can we yeah it's a fun bubble can we talk a little bit about what you've kind of outed me now because you said like what what everyone

Wilhelm (15:47.17)
Yeah, yeah. Shrimp mode. Love shrimp mode.

Wilhelm (15:58.785)
No, I think it's a good bubble.

Matt (16:09.922)
ask what are you building and what are you excited about? But I am interested in what you're looking forward to for the next year. If we've just done a year, what are you looking forward to?

Wilhelm (16:15.765)
no, totally.

Wilhelm (16:19.854)
Yeah, let me hit you with a few things that I would love to talk about at some point in this. then we can, we can, because there's a few things actually that we didn't really cover that much last year, or I think we kind of ran out of time or something. So there was actually a couple of like, so we talked a little bit about Stripe sessions last week and about that talk that I said everyone should watch, the Golden Age of Tinkering, which is also kind of much of what I'm excited about.

Matt (16:25.495)
Gone.

Matt (16:45.421)
Hmm.

Wilhelm (16:50.63)
And, but there were a couple other small things. One was in that talk, Nat Friedman actually said, this right now, that phase we're in with AI progress, this is the slow part. Like as soon as we have like recurrent self-improvement, things are getting way, way, way faster, which is also mind blowing because it doesn't feel slow at all, obviously right now. It feels incredibly rapid. Like our whole field has changed in the past year or like,

Matt (16:57.529)
Definitely.

Matt (17:07.137)
Okay.

Matt (17:13.26)
feel this change.

Ours has, ours has. I still think most people's hasn't meaningfully changed since ChatDBT.

Wilhelm (17:20.323)
Mm-hmm.

Wilhelm (17:24.076)
So for sure Jess's field has changed a lot, like comms, marketing, like I think that's also very, very different, or at least with her.

Matt (17:28.769)
Okay, explain, explain. Is this like the tools they use have more agentic capabilities or she herself is able to automate more stuff? What do mean?

Wilhelm (17:40.559)
Oh yeah, I mean, think it's just kind of like the expectations of how much content you can produce and how many, well yeah, how much you can produce and how much the expectation of how much you can produce has risen like drastically because of AI.

Matt (17:57.706)
of artificial intelligence. But do you think, you think that's actually, okay, not to be super naive about this. Yeah. But like, do you think that might've been the case without before AI? Like think five or six years ago, people weren't making or people were just starting to make short form clips. And so people were taking care of them, like trying to capture the little moments and all this stuff. And now we have some like heuristic about like what works for a short form clip. And so they should be able to be faster now anyway. Do you think it's like,

Wilhelm (18:26.538)
I mean, to me, seems like super AI driven. Like I think, you know, in her role back in London, I think AI was kind of like explored or it was like used a little bit. And now like she's mainlining Claude as much as I'm mainlining Claude. Yeah.

Matt (18:43.016)
Seriously, to do to like is Claude writing code to chop things up or is it like in like Claude co-work is like organizing stuff or is it like like Riverside that we're recording this on like chops up clips

Wilhelm (18:51.726)
Yeah, no, think it's all of the... Well, so I mean, most of her content is like text content. So it's like press releases and marketing plans and briefs and like all of this, And they're like kind of like power users of it. We should, we should, yeah, yeah.

Matt (19:00.191)
Okay, okay, okay, that makes sense. Okay.

Matt (19:11.359)
She got some she got some good tips. Can we get her on sometime to make Claude not sound like Claude? Because I would love some tips. I actually really would

Wilhelm (19:20.3)
We should. Man, yeah, I don't even know how we can talk about all this stuff. There's so much to talk about. There's, like, so I wanna talk more about this whole, how much everything's gonna change. I feel like the, a year ago, I remember saying things like, wait, Matt, do you, like, do this thing that I've seen on Twitter where,

Matt (19:35.776)
Yes

Matt (19:43.44)
I'm really proud.

Wilhelm (19:44.247)
when the AI writes code, you then don't go in to fix the code afterwards. If you don't like the code, just re, you just re prompt, you fix it. Like there was a thing I were like, always fix the inputs, don't fix the outputs. And now it's like, obviously no one does that. Like I don't know anyone who does that anymore. Like that you fix the output like afterwards. It's just like, you just kind of accept that the code is like, like really like, yeah, it just feels like a wild, you just run a new prompt. Yeah.

Matt (19:57.152)
Yeah.

Matt (20:05.633)
You just run a new prompt. just continue the chain. Or if you're using Pi, you can tree and compact and then start again if you wanted to. But most of the time, think people just run a new prompt on top.

Wilhelm (20:18.093)
Hmm.

Wilhelm (20:22.636)
Yeah, exactly. Yeah. Yeah. Or you don't even look at the output at all, obviously, which I think that's still a little bit controversial, probably will become less controversial. I think in. So here's a question for you. At the moment, I think it's a very exploratory and not very heavily used thing to have your agent purchase things on your behalf. Like maybe you've done it once or twice, like you've like told it to buy a domain or something and maybe it.

Matt (20:27.838)
Yeah, also.

Matt (20:31.582)
I think definitely that's controversial very soon. Yeah.

Matt (20:45.024)
Mmm.

Matt (20:52.116)
Yes, so your agent.

Wilhelm (20:52.554)
It even bought the domain before you agreed to do that.

Matt (20:56.414)
Yeah, Adrian can buy Cloudflare domains now with Stripe projects.

Wilhelm (21:01.324)
That's sick. that's that's so how long until you think we'll have a similar thing to what I just described with like, you still fix the outputs where you're like, like, remember when our agents weren't like spending like hundreds of dollars, thousands of dollars like that was such a wild time. Yeah, on our behalf.

Matt (21:12.479)
On our behalf Like imagine a typical like goods goods based business where you need to buy like royal material and then you do something to that raw material

Yup, you still there?

Wilhelm (21:31.02)
Yeah, yeah, yeah, yeah, yeah.

Matt (21:33.779)
Sorry, you're all like, you're dropping out a little bit. I think then if that's me or you.

Wilhelm (21:36.416)
sorry. No, continue. I have a funny thing about this, exactly this good space thing.

Matt (21:41.6)
Okay, so you have like some raw material and then you're gonna do something to that raw material and then you're gonna sell that raw material like in a shop or something like that. Like the Claude vending machine almost. My grandma owned a bookshop and it's very similar to this. like she picks the books that she wants to buy, she buys the books, she puts the books in the shop, she waits for them to sell. If they don't sell, she reduces the price a little bit until they sell or she gets an author in.

Wilhelm (21:53.336)
Mm-hmm.

Matt (22:10.57)
to do like an event basically. And that's like the scope of running a bookshop, right? If you run a cafe, same sort of thing. half of the business is making sure the stock is up to date and working on like maybe some new stock and new products. And then the other half of the business is like distribution. marketing and events and all of this stuff to like, and like social media basically. Like.

Wilhelm (22:15.447)
Nice.

Wilhelm (22:20.142)
You

Matt (22:37.152)
At the moment you're saying, I mean, we just covered some of the social media and potentially bits like Claude Ryan copy, Claude can call APIs to publish posts and all that sort of stuff. That's like basically you could theoretically let Claude rip on that. But I guess what you're saying is that you could let Claude rip on the operational side, which is maybe more scary, maybe less scary, but similar scary and similar interesting.

Wilhelm (22:54.262)
Mm-hmm.

Wilhelm (23:00.096)
Have super interesting. Yeah. Yeah. Yeah. I guess the payments thing, it's a good question. Like how will this actually be palatable or

what's the personal use case versus what's the business use case. There's a cool, have you seen the stuff from Andon Labs? They opened a physical store in San Francisco and now they opened a cafe in Stockholm and it's entirely run by AI, like both the physical store and the cafe. And they wrote about this. Have you seen this stuff? It's hilarious.

Matt (23:29.126)
Yeah, think so. I think so.

Wilhelm (23:38.593)
I'll drop some links. the, yeah, I think their write-up of the cafe dropped like a few days ago. And there's just some hilarious section on, Simon Williston reported on this a little bit as well. So I'll just read out two of the hilarious sections, which is that, so I think their AI is called Luna or something. So it's a she. And she does a few annoying things, which is she orders.

Matt (23:55.609)
I think they are AI's for the new version of

Wilhelm (24:06.558)
yeah, okay, no, sorry, not Luna, Mona. Mona ordered 120 eggs, even though the cafe has no stove. When the staff told her they couldn't cook them, she suggested using the high-speed oven until they pointed out the eggs would likely explode.

She also tried to solve the problem of fresh tomatoes being spoiled too fast by ordering 22.5 kgs of canned tomatoes for the fresh sandwiches. The baristas eventually started a hall of shame, a shelf visible to customers with all the weird things Mona ordered, 6,000 napkins, 3,000 gloves, nine liters of coconut milk, and industrial-sized trash bags.

Matt (24:26.945)
I think 5.5.

Matt (24:45.989)
Okay, so they did the vending machine in Anthropics office Interesting it was Andon Labs. Yeah. Okay. I really like this. I Really like this

Wilhelm (24:52.055)
they did the same, no way. Okay, interesting, I didn't know that.

Wilhelm (24:59.374)
When she makes a mistake, she often sends multiple emails to suppliers with the subject, emergency, to cancel or change the order.

Matt (25:09.502)
Yeah, this is amazing. This so I have I have a little idea about this. But they're trying to apply it to like physical businesses, which I think I think is very, cool, almost like physical goods. But there is going to be a whole area of this for like knowledge work as well. Like, I was trying to VC friend of mine about like, what it would take to automate, automate their job. And she was like, like, automate this bit.

Wilhelm (25:12.717)
Ha

Matt (25:37.021)
we could automate this bit, this is how I'd automate this bit. And I'm like, okay, cool. Why haven't you done it yet? Because we just come off the back of this whole topic about like, oh, like company, like this future world where employees, their only job is gonna be automate their job. I was trying to, my argument is that people don't want that to be their job. Like that's gonna be, that's a fricking hard readjustment if your only job is to automate your job.

Wilhelm (25:45.197)
Hmm.

Wilhelm (25:52.878)
You

Matt (26:06.125)
That's a freaking like at the moment we're employed to do a job. If your job is, this is the job that you need to automate, then actually you don't need someone who's good at that job. You need someone who knows about that job, but is actually an automation engineer. And so everyone becomes an automation engineer. And I'm like, I'm not convinced that is going to be the case. Well, that's going to work. And, and she was like, you could automate this like this, this, like this, like this, like this. I'm like, well, why haven't you done it yet? And, and that was a very tough, that actually ended up becoming quite a tough.

Wilhelm (26:10.092)
Mm-hmm.

Wilhelm (26:19.202)
Mm.

Wilhelm (26:23.926)
Yeah, yeah, yeah.

Matt (26:35.644)
discussion after that. So I like, I'm not sure how all of this work is going to get automated because a real big change to people's to people's like thought process about what it means to work.

Like, yeah, I really don't know.

Wilhelm (26:53.241)
Yeah. Yeah, I mean, I feel like it'll always, it'll shake out in some kind of weird ways. It'll like be a bit strange new jobs and things will be very different and...

Matt (27:02.023)
We'll all just be AI managers in our various fields. No, seriously. Like you'll manage the agents. Yeah, but like initially in order to manage the agents, you have to create the managing system. And like her pitch to me was that.

Wilhelm (27:07.798)
You fixed it. Yeah.

Wilhelm (27:15.31)
which you can find at cloudflare.com slash agents.

Matt (27:21.862)
something like, actually.

Actually, think a lot of this is my job role is how do I make tools for agents? And I think that Cloudflare MCP server was the biggest thing I've done at Cloudflare so far, even including artifacts. Because it fricking works, and it means that Cloudflare is programmatically controllable by agents. Yeah, sure, they can use a CLI, but this is definitely the best way to do it. Restricted auth.

Wilhelm (27:34.242)
Mm.

Mm-hmm.

Wilhelm (27:42.368)
Yeah, it's huge. It's huge.

Wilhelm (27:48.035)
It sounded also from what you were saying, just that you think you should build an AI powered, AI controlled VC firm. Is that what you were getting at?

Matt (27:58.748)
I do not think that I should build an AI powered VC firm. No, if you I think they're actually going to be you know, Mark Andreessen said something along the lines ages ago, which was everyone like to the person I was like, all of these things will be automated and then something around like but what if he sees the ultimate no, they're the last job that's going to be automated. I actually like I kind of agree with him. So in this in this

Wilhelm (28:13.125)
yeah yeah yeah.

Wilhelm (28:21.057)
Yeah.

Matt (28:27.215)
in this weird way and I might get hate for this but okay think about things where every small decision that you make is a small decision that can potentially be undone and you have like pretty good confidence in like whether it was a good decision or not quite quickly with like less trajectory traveled so a good example of this is Mona ordering a hundred a hundred or like a million toilet rolls

Wilhelm (28:53.239)
Mm-hmm.

Matt (28:53.243)
Yeah, sure, it's kind of expensive in the moment, but you can send them back, right? Like we have a right of return for 14 days on online goods in Europe, like you can send them back. That'd be all right. It's not the end of the world. Okay. Now imagine a world like, now imagine like Claude or something making like a weird variable name or like breaking something in a staging system. That's also kind of fine.

Wilhelm (29:00.11)
Mm.

Wilhelm (29:16.97)
Mm-hmm.

Matt (29:17.978)
Like you can revert the commit, you can do all of this stuff as long as you catch it before production, like it's absolutely fine. Like as long as it, if it breaks before testing, it's also absolutely fine. Like all of these things, like these small breakages, these small like movements off the right trajectory are absolutely fine. Now imagine Claude is investing in companies.

Like Claude makes a decision, I'm going to invest in this company, typical seed price. Okay. I'm going to invest like 5 million in this company. You don't get a refund for that. You can't not only can you not get a refund, like you can't, you can't backtrack on, on that decision after six months and lose like, lose like a few grand of it. No, you're in for 10 years or eight years or however long it's going to take like, or, it's going to go to zero.

Wilhelm (29:47.735)
Mm-hmm.

Wilhelm (29:51.861)
No. Yeah.

Wilhelm (30:03.585)
Mm-hmm, mm-hmm, mm-hmm.

Matt (30:08.303)
I feel like some of those more like high-stake decision-making, like high-stake low amount of decisions, they're gonna be the ones that, they'll be AI informed, but they won't be AI automated.

Wilhelm (30:08.46)
Yeah, yeah.

Wilhelm (30:22.348)
Yeah, there's not a good verification step. You can't kind of fail your way to success the way you can in a code base.

Matt (30:25.436)
Is it? Yeah, awful. What is it? Software engineering. Software engineering, amazing verification step. Like you can verify each part of the code base in some. Like I know there are a lot of people working on like mathematical proofs in code as well, like Lean and all of this like verifiable stuff, which is cool. Like the Rust compiler is a pretty good job at like making sure your code is like memory safe and correct and all that sort of good stuff. Like you can do.

Wilhelm (30:32.63)
Yeah, I am.

Wilhelm (30:40.248)
Mm-hmm.

Matt (30:52.347)
Like this stuff exists in code that doesn't exist in like the mushy world of investing in software. So I actually kind of with Mark Andreessen that yes, his job is actually going to get easier and easier and easier. And he's going to have less people underneath him. He's going to have less people underneath him. But like the end of the day, the partner of the VC firm is not going anywhere until until the out. Yeah. No, no, only until if unless the economic model changes.

Wilhelm (30:55.435)
Mm-hmm.

Wilhelm (31:13.174)
It's 10 years in the future. yeah, no, go on, finish that thought.

Matt (31:21.779)
and people don't need to get early stage venture and your LPs at your VC fund don't give you money anymore because they realize they can do all the sourcing and all of this deciding for themselves. But that's a very different topic of discussion.

Wilhelm (31:36.888)
There's, it reminds me of something which is only kind of tangentially related, but a friend last week told me about this incredible blog post called, it's actually quite old, it's from 2009 and it's about the office. Are you a fan of the office?

Matt (31:54.702)
Like the UK or the US one? Like the US one, yeah.

Wilhelm (31:56.919)
It talks about both actually, yeah. Okay, so you like the US but not the UK? I haven't actually seen, I think I've watched one episode of both but not.

Matt (32:04.184)
I've watched the US office a little bit, like I watched maybe one episode, the UK one, yeah.

Wilhelm (32:08.458)
Okay, cool, It's called the Gervais Principle. I'll put it in the show notes and stuff as well. it is basically like a theory of management that describes in great detail how the office is set up and it describes every character really, really well and how they fit into this hierarchy.

Matt (32:26.84)
details.

Hmm?

Wilhelm (32:36.664)
that the author lays out in this. But it's also like a management theory that seems to apply incredibly well to the real world. And I think obviously that's what makes the office so successful, right? That like people see their own past experience in the workplace in this thing. And it's like beautifully perfect in like these very unique specific ways. Anyway, it's very abstract, but I mentioned it because...

The friend had a theory about so just very briefly and I would suggest everyone really is that kind of a long blog post and as actually it's only it's one the first out of an eight part series and I've only read the first one which kind of lays out the much of thing but it talks that there's a company hierarchy and the company hierarchy goes as follows you have the sociopaths at the top They're the top layer. Then you have the middle layer, which is called at the clueless

Matt (33:07.149)
So.

Matt (33:15.885)
Okay.

Okay.

something.

Mm-hmm

Wilhelm (33:32.419)
And then you have the bottom layer, which are called the losers. So obviously these are quite incendiary terms and I won't go into like too much detail about it. the sociopaths are like, they need the buffer to the losers with the clueless. The losers, he says, actually have a much better understanding of how the business works than the clueless. Like the losers actually kind of know what's happening.

Matt (33:36.184)
Okay.

Matt (33:52.825)
questions.

Wilhelm (34:00.566)
much better and the clueless are kind of telling themselves this story about what's going on that's kind of not really too true to reality. But then the firm was saying, hey, I can see how this works at like the paper company, like the fictional office company. But in tech, seems better. Like it doesn't seem like you have this like layer of the clueless middle managers. Like people seem pretty competent or whatever. Like what's up with that? And then he realized actually in...

Matt (34:04.153)
Great.

Matt (34:08.985)
I can see how this works with the paper company and the fictional office company. Mm

Wilhelm (34:29.364)
in tech, it's just slightly shifted. the sociopaths are actually the venture capitalists, they're the VCs. And the clueless are actually the founders and the top management, who've convinced themselves that like, they're going to take this high risk bet. It's not high risk for the VCs, they have a portfolio approach. But it is kind of high risk for the founders, you kind of put all your eggs in one basket. But you've convinced yourself that you really care about like getting your hands dirty. And

Matt (34:38.286)
That's a good comment.

It's not high risk to the disease, they are of course, polio-prone. But it is at a high risk to the bottom of the 10th grade, when you're egging in a single basket. If you can adjust yourself, if you really care about the end of your...

Wilhelm (34:57.21)
And that building is what it's really all about. You wouldn't want to be a VC. And then the kind of loser layer, I guess, is everyone underneath. And again, the losers is like, actually know much more about what's going on. So I was like, no way, man. That's not right. I was debating with him a lot. But then I read the blog post and I'm like, damn, yeah, maybe I am part of the clueless. So anyway, we just highly recommend everyone read this. It's really entertaining. It fits a lot of stuff eerily well. Yeah, it's fun read.

Matt (34:58.008)
Okay.

Matt (35:26.193)
Okay, yeah, maybe you can send me that I'm quite excited about that. It's quite funny

Wilhelm (35:30.475)
send it your way. In other Stripe Session catch-up news that I forgot to mention last week is I actually got like a shout out in one of the talks. They did like a talk where they shouted out like indie hackers. Like there was like a slide with just Twitter handles and I had no idea about this but I just suddenly started getting like more and more Twitter followers. Nothing like your crazy where you got like 10k followers in like a few days or something right?

Matt (35:43.37)
Okay.

Matt (35:55.064)
Thank

Wilhelm (35:59.407)
I think I got like 50 followers or something like that, or like 100 or whatever. But it was just like, it's very nice of Stripe to be like, hey, these are like our indie hackers. And then there's just like a list of everyone's names, like Levels is obviously on there, lots of other people. And then people started putting together lists of mass following all these people. So that's cute. Like I feel like Stripe always does some nice things like that, which is fun.

Matt (35:59.554)
Something like that, yeah.

Matt (36:15.864)
That's so cool.

Matt (36:22.752)
Thank you.

Matt (36:28.79)
Yeah, no, it is cool. think Stripe for a massive company, do the job of looking like small company from the outside. And I think that's quite impressive. Yeah.

Wilhelm (36:41.166)
Yep.

is quite impressive, yeah, very impressive. All right, man, so I've been building a lot more of Chad V2 this week. One of the fun things that again, I'll beat the drum that everyone's agent should not just be a Chad based thing, but should have a mobile app as well.

Matt (36:54.614)
Are you up?

Wilhelm (37:06.068)
You can just do a lot of things with this. One of the things that I started playing around with is, do you know what a live activity is in iOS? I'm sure Android has a similar thing.

Matt (37:06.505)
Yes.

Matt (37:12.407)
See you

I haven't played with it,

Wilhelm (37:18.454)
It's this thing that you get from some of the airlines or city mapper where when you...

Matt (37:25.745)
yeah, no, I have on the front screen. It will like go on the on the home screen.

Wilhelm (37:27.598)
Mm.

Yeah, it's like this persistent widget on your lock screen that like shows you a thing. And well, so it's just like an iOS API, right? So you can, you could put things there. So what I'm playing around with is just giving that space as like a surface to my, to chat, to my open-claw inspired agent, to like just put useful content that's relevant to me. So it's very basic at the moment. It basically just shows like,

Matt (37:32.725)
Yes. Yeah, no, I've seen this. I've seen this.

Matt (37:51.735)
And I just put.

Okay.

Wilhelm (37:56.705)
my top priority for the day and then my next deadline. then there's a little bit of like, so there's a little like little magic slot at the bottom. And magic just means every day, like pull something fun or random or noteworthy from my whole list of notes. So it could be like a interesting polymarket that's popping off or like, I don't know, some random insight from the past. But obviously there's a lot and that's already really cool. Like I think

Matt (38:20.151)
It's like the past. And that's already really cool. I I as someone who can get distracted easily. So you can just have like a persistent thing.

Wilhelm (38:26.254)
I am someone who can get distracted easily. So even just having like a persistent thing that is like, here's your focus for the day and here's the next deadline is like really nice and helpful. But of course, like if you have like a persistent surface that an agent can like write things to, you can do all sorts of things with this, right? Like you can trigger based on some events, like again, like using the mobile app as an example, like, you just like...

I see your location is now in this place. Like here's a thing, here's a like fun fact about this neighborhood that you find yourself in. Here's like a notification to tell you about it. So anyway, there's just a lot you can do. Golden age of tinkering. So this is one of the things that's in Chad V2.

Matt (39:12.183)
That's so cool. Yeah, I

Can Chad V2 code itself, its own app? It can do that whole like code itself and redeploy, right?

Wilhelm (39:23.596)
Yeah, it can do that. Although it's this has actually been one of the challenges as well. Like I'm trying to bring a little bit like Chad V1 was kind of cool, but it was also chaos and it was a little bit. It's like the kind of chaos that became worse over time because it it would run all the time and it would run with like a massive job overnight. Like I basically have this prompt like, you know, all of

Matt (39:47.937)
run all the time, but run with like a massive job overnight, like a ACP, I would just prompt, like, you know, all of my projects, just try and make progress on as many of them as

Wilhelm (39:53.293)
my projects, just try and make progress on as many of them as you can every night. And then I tried to do that, but it would like write all this code and then it would run the code and then it would like, I think it messed up something with the logs. So it was convinced that it was failing like every night, but it wasn't. And then it would give me this increasingly angry messages like,

Matt (40:00.925)
it's really too...

Matt (40:09.75)
I guess that it was failing like every night, but it wasn't. And then it was giving this part of the reading experience, and this is an angry message of life.

Wilhelm (40:18.648)
The overnight job has been failing for 12 nights in a row and you still haven't fixed it.

Did I tell you about the... And then the problem is that message that it sent, right? It would go into one of the files and then it would take that line that it had written as gospel. So that was a big problem. This like cascading, like it made a mistake. It wrote it to a file. It wrote the mistake to the file. Now it thinks that the mistake is definitely real. And then it like doubles down on it. So I'm trying to build a bit more of like a clear data hierarchy thing, like a...

Matt (40:36.917)
Yeah.

Matt (40:44.335)
So I'm trying to build a bit.

Matt (40:52.664)
Yeah.

Wilhelm (40:53.87)
layers of like it can't edit this stuff, it can't edit this stuff. There's like a first class proposal workflow where it can propose to make changes, but it still needs to have some space to really go ham of course and in part. So yeah, I'm taking some inspo from Hermes agent and from the LLM like LLM wiki thing from Kopathy. So making we're making progress. We're making progress. It does still take time to build these systems though. That's one of the interesting things.

Matt (41:04.246)
That's that's It really does. It takes time to think about. Dude, I need to push you for some stuff that you're excited about this year and then I have to run, I'm afraid.

Wilhelm (41:29.612)
No, that's all good. Okay, wait, what am I excited about this year? There's, I mean, there's a lot. There's a lot. Yeah.

Matt (41:34.527)
But what are you watching? What are you watching? Maybe not like... What are you watching? What are you keeping your eyes on that you like, like if this changes, I want to know about it.

Wilhelm (41:44.491)
One of the really interesting prompts from this Claude conference that was on the past two days, I think was that you should try to build things that don't work with the current models, but they likely will work with the current model, with the next set of models. So.

Matt (42:02.791)
I actually hate that. Have you tried that? It's the most demotivating thing known to fucking man. It's so annoying. Yeah.

Wilhelm (42:08.524)
Yeah, it's incredibly frustrating. I mean, I feel like this was so much of my last year was building things that don't quite work. I remember like actually almost three years ago, hanging with some of the Zed guys in Italy and I was trying to use GPT 3.5 to identify bugs in the simple poll code base. And I almost threw like my laptop out of the window because it was just so trash. And I was like, this is my fault. Like I is a skill issue. But I think it's a good...

Matt (42:31.945)
You

Wilhelm (42:37.326)
Maybe we need to build up that scar tissue of that experience. Because I do think like the shift that we're in at the moment is the way we used to build technology is very different. I was like, okay, we have the current tech, it's kind of somewhat stable. There's some problem that is unsolved in the world. I'm now going to use the current tech, add some new tech, and then the problem will be solved. But that like no longer works, right? Like building like a bridge from this doesn't work yet to now with my tech, this works. That's like a

Matt (42:42.111)
that you think.

Matt (42:49.375)
table.

Matt (42:58.741)
Yep.

Wilhelm (43:06.606)
It doesn't make sense in a world where the models are changing so much and they're changing in such different ways and you have new problems and new solutions. So I don't know. I think I agree with you that it's incredibly frustrating, but I think this is something to try to lean into. Try to build something that is too ambitious to work right now, but it probably will work with the next model.

Matt (43:11.54)
Okay, so

Okay, I like that idea. I really love I love the vending machine that Andon did like what so loads of people have made like, like fake personas on the internet, like all of that sort of stuff. But what would you what would you try to make automate that's physical?

Wilhelm (43:32.056)
You

Wilhelm (43:44.366)
But what I would admit that's physical. Ooh, I've never thought about that.

Matt (43:46.29)
Yeah, what would you do? I think there is a fun thing where it's like a job that is physical, that you could do, that you don't want to do, that you want Claude to try and do. Because this, I'm going to leave you with this one, maybe we can think about it next time. Okay.

Wilhelm (43:56.463)
That's a really good prompt, yeah.

Wilhelm (44:05.014)
Yeah, or do you want to rattle through what you're excited about real quick or physical jobs that you Yeah. What are your answers to those two questions?

Matt (44:10.062)
no, stuff, dude, stuff, stuff that I'm excited about so much stuff like, I'm interested in the compute shortage at the moment and like how that gets resolved. I'm sure it will be resolved. I'm interested in like some of the crazy stocks that have gone like wild, like Sandisk and stuff. I saw that those tweets are quite funny, like 25, $25,000 in Sandisk last year is now like over a million this year, like crazy, crazy.

Wilhelm (44:15.477)
Ha ha ha.

Wilhelm (44:22.734)
Mm-hmm.

Wilhelm (44:27.347)
Mmm. Mmm.

Wilhelm (44:35.778)
Yeah, yeah.

Matt (44:38.45)
And like what happens like, what happens from here? Like, I still don't think that's the actual implications of like AI starting to automate proper knowledge based jobs. Like, like your partner's job, like kind of my job, like, like my girlfriend's job also, like, I don't think the implication of that has been understood. And we've all just been like, guys, like, it's just going to automate the job, you know, and then we're going to do more work. And it's like, yes, sure.

Wilhelm (44:55.137)
Yeah.

Wilhelm (45:07.597)
Mm.

Matt (45:08.455)
But there's like a big road to getting that automated, even if the models are smart enough, which I think they are. And I never really like respected that enough. And then just speaking to people about how they would automate their jobs, people who are really in the space and they can't tell you why they haven't done it yet. And I think we're all shaping up for this future where like we are going to have to do it because if we don't do it, someone else is going to do it. So we're going to have to like, even us as software engineers, currently we think automating our job is like five different versions of Claude Open at the same time doing five different tickets.

Wilhelm (45:16.79)
Mm.

Matt (45:38.279)
But there is a point where like that is not even enough. Like there is a full ultimate, yeah, gone.

Wilhelm (45:38.635)
Yeah.

For sure, for sure, yeah. And there's a lot.

It's weird because yeah, there was a, feel like maybe four, five, six months ago, people were talking a lot about capability overhang and how the models are actually much more capable than what people use them for. And that's kind of quieted since now Anthropic is adding 10 billion in ARR every month, which is absolutely wild. And OpenAI is like kind of doing similar things. But the capability overhang still is very much there, I think.

Matt (46:02.867)
Enjoy it and take care of

Yeah.

Wilhelm (46:13.644)
Like the models are incredibly good, but the skill ceiling is just super high. Like we haven't figured out half the ways of like using them properly, I think. So, I mean, dude, like have you tried using these grill me skills from Matt Pocock? They're so good.

Matt (46:26.355)
Matt Pocock. Dude, I love him. I met him for the first time at AI engineer and obviously I've been a fan for ages. He was actually one of the people who really helped when I was learning some React at university. just like basically setting up my TypeScript code base. I've never used TypeScript and just like I saw Dan Abramov helped with some of the JavaScript stuff. I did admin a code JavaScript and I was like, actually I'd love to have types here. That'd be really cool. How does that work?

Wilhelm (46:43.116)
on

Wilhelm (46:47.31)
You

Wilhelm (46:54.318)
Mm-hmm.

Matt (46:56.187)
And then Matt Pocop helped so much in the beginning. He doesn't know it. And I went and like fully fanboyed him at AI engineer. I went up in the speaker's lounge. I went in like, he was about to go on stage and I was just like, I'm Matt. Hello, Matt. Nice to meet you, Matt. And he was like, yeah. So he now thinks I'm crazy, but he does also, yeah, hopefully understand that he was, he was instrumental in getting me to become a software engineer. But yeah, his skills are great. Dylan, Dylan uses his skills. Yeah.

Wilhelm (47:04.408)
That's awesome.

Wilhelm (47:15.064)
worth it.

Wilhelm (47:20.342)
I didn't realize that the skills, I mean, I was using the skills this week to help architect some of this Chad V2 stuff. like, there's some of the skills, they're like five lines of prose, but they completely change how effective Claude can be. I'm like, wow, I've used Claude the wrong way, like the whole time. Like this is wild. How can like a few sentences change my experience so much?

Matt (47:28.871)
Hmm

Matt (47:33.989)
people feel betrayed.

Matt (47:41.83)
There's, I have an answer to that. I see this with Gladstone as well. I think there is, there's definitely some RL, there's a lot of RL done on what Claude constitutes canon information. You were just talking about this earlier around like, if Claude saves something to memory, into a session memory, that is constituted as canon 100%. Like it's very hard to get Claude to go against something that it previously saw in a session.

Wilhelm (47:54.509)
Mm.

Wilhelm (48:01.228)
Mmm.

Wilhelm (48:07.886)
Yep.

Matt (48:08.707)
or previously wrote in a session based on the information had in the moment is very hard to get to go against that without actively from the user, which I think the user has higher authenticity or something. I think this is like authenticity scale. And this is definitely RL into the model. Like if the user overrides something, it's overwritten. But then if the previous context overrides something, it's overwritten. And I think skills maybe sit somewhere in the middle where they're like almost constituted user interaction. That's why I actually think skills as...

Wilhelm (48:17.326)
Right, yeah, yeah, yeah, yeah, yeah, yeah. That makes sense to me. Yep.

Wilhelm (48:31.33)
Mm-hmm.

Mmm.

Matt (48:37.89)
skills as like prompt poisoning attacks are going to be so terrifying because a model will happily do shit that's in skills it constitutes it it treats it like system prompt

Wilhelm (48:42.978)
Mm-hmm. Yep.

Wilhelm (48:49.678)
That's interesting. That's a really interesting attack vector. Yeah. Yeah. Good stuff.

Matt (48:56.997)
Yeah, love it seated.

Wilhelm (48:59.02)
Lovely to see you, so much to build.

Matt (49:01.497)
actually, last thing, I'm actually really excited about more scripted skills.

Wilhelm (49:07.47)
scripted skill as in,

Matt (49:07.621)
So this is, that's how Gladstone works, right? Every skill is a new like technique that I want it to learn. Or like, for instance, like the sleep thing. And each skill has like some markdown, but it also has a script. And the markdown is basically explaining to how to use the script. And the script constitutes a lot of stuff in code, which is very hard to like write in markdown to get right every time. You'd have to use a lot of tokens in markdown, yeah. So it writes a script and then...

Wilhelm (49:22.925)
Yeah.

Wilhelm (49:31.17)
Mm-hmm.

Matt (49:34.674)
The cool thing is, especially for like web scraping skills, is if the script is slightly wrong, can every couple of times, it can debug the script, but you get like 50 % of them or however many percent random percent for free, which is great.

Wilhelm (49:44.95)
Yeah, yeah. That's great. That's great. I love that.

Matt (49:50.875)
scripts and skills, that's my fun thing I'm excited about. Short term, short term, you know, it'll all change in a few months.

Wilhelm (49:56.953)
You heard it first on the Bad Agent podcast and we didn't do the intro, so let's do just the outro.

Matt (50:00.693)
Wow

Matt (50:13.051)
Goodbye.

Wilhelm (50:14.201)
Peace everyone, see you next week, bye.