You've Been a Bad Agent

Wilhelm (00:01.102)
Вот Happy Monday.

Matt (00:01.34)
Yo.

Yeah, this is is this the first for us recording on a Monday?

Wilhelm (00:08.642)
Back on the pod, fresh from the weekend. You're covered in salt, I'm covered in sunburn.

Matt (00:15.527)
You s did you get sunburnt this weekend?

Wilhelm (00:17.331)
man, yeah, I always get sunburnt. I always get sunburnt. And it's so tiring. I'm so dehydrated right now.

Matt (00:25.051)
I've been chugging water today. I've got an empty glass because I just absolutely drained it. yeah, it was how hot is it?

Wilhelm (00:31.672)
Hydration is just it's so t tiring, man. I feel like my throat hurts from all this water I'm drinking.

Why is there no AI for keeping me hydrated?

Matt (00:45.32)
You're gonna have a robot in like ten years, you're gonna have a robot walking around behind you just with like a straw.

Wilhelm (00:51.864)
Just holding my Seren like my what's it called when you have like a my my drop, yeah. I have a robot just to hold my drop to follow me around. That'll be two thousand thirty.

Matt (00:54.549)
Your IV drip.

dear. That's not long enough away for that to be the ki that to happen.

Wilhelm (01:04.802)
How's it going, sir?

Wilhelm (01:10.232)
CloudFlass stock all time high, you must be happy.

Matt (01:13.903)
Is it?

Wilhelm (01:15.618)
Brother, yeah, it's popping today. Markets are on fire.

Matt (01:16.893)
Nice. Nice. Good stuff.

Wilhelm (01:22.924)
Never been this high. all right, we'll leave that at that. wait, you had a you had a very active weekend as well, right?

Matt (01:32.503)
Yeah, I ran a trail race in a vineyard. twenty two twenty two Ks. yeah, it's really nice. Some good some good hills, like almost eight hundred meters of climbing, which was really good.

Wilhelm (01:37.816)
Yeah, vineyard.

Wilhelm (01:48.45)
Wow. That's I that's a lot of climbing. how wait, eight hundred meters over how

Matt (01:54.3)
It's not bad. It's like I think like 750 over 22Ks, which is not awful. Not awful.

Wilhelm (01:59.386)
through it I mean I did the cycling well the in the triathlon the cycle was I think six hundred meters of climbing over forty K and that was a lot of climbing for me. I'm not a very good climber, so hats off.

Matt (02:13.307)
Yeah. But this is this is running, like as in like there's a lot of walking involved up trail. But Juliette beat me by like she says five minutes, but it was more like it was more like eight or ten.

Wilhelm (02:24.878)
She's on fire. That's awesome. no way.

Matt (02:26.927)
Yeah, she she podiumed actually. th little little third place for Jiuji, yeah. It was really cool. They gave us wine, loads of random fruit, and some veggies, and like a massive thing of like desiccated cod, like salted cod. Quite weird. It's been it's on the side in the kitchen and I don't really want to touch it.

Wilhelm (02:44.67)
Ooh. That's that's not

Wait, was that the prize? The or what did you just describe?

Matt (02:54.897)
That everyone got that. That wasn't even a prize. We have two loads of that. I have that as well. That was like your finishing medal.

Wilhelm (03:01.346)
That's I mean to be there, if I was like a race director for one of these things, I would do some weird ass finish line stuff like, yeah, sure, let's get everyone some COD.

Matt (03:06.982)
So

The the aid stations in Portugal are amazing. at the beginning, the first one, I was like, this is a bit like

Wilhelm (03:14.243)
Mm.

Matt (03:18.278)
It's a bit low-key, isn't it? It's just like some fruit, some Coca-Cola, and some peanuts. And I'm like, and there's some banana as well. I'm like, okay. It it's got the basics, but like, you know, it's not good, it's not bad, it's it's okay. And then the second one, there was nothing. And I was like, that sucks. And they just stamp your thing. And then the last one, it was not very far from the end, maybe like three, four K from the end. So I basically didn't stop. but they had

Wilhelm (03:25.1)
Yeah.

Wilhelm (03:43.51)
Yeah. All right.

Matt (03:44.784)
They had barbecues going, they were giving out like bifanas, which are like this Portuguese meat in a wrap, in a in a in a bread thing. they had like the wine was flowing at that aid station. There was there were shots, it was insane, and then like a kilometer down the road, there's like these two kids in the middle of the in the middle of the path with with a table just giving you giving ginger shots.

Wilhelm (03:59.756)
No mine was flowing.

Matt (04:14.468)
So ginger is like a cherry cherry liqueur. it's it's alcohol, like really quite strong alcohol. Yeah, and they get

Wilhelm (04:15.67)
Mm.

Yeah, yeah, yeah. It's not a ginger shot you would buy in Pretta Manger. It's a different kind of ginger shot.

Matt (04:26.156)
No, it's not like yeah, it's a different type of ginger shot and ginger, yeah, it was it was intense. Anyway, I wasn't really feeling like a ginger shot at that moment. But there was a lot of wine involved in the last aid station and the people at it were very drunk by the time I arrived. It was it was it was good fun. And then we f got to the finish line. some of the Portuguese blokes had started by having a cigarette and a beer, and by the end everyone was having a cigarette and a beer, so

Wilhelm (04:30.659)
Mm-hmm.

Wilhelm (04:44.034)
That sounds like a great time.

Matt (04:55.493)
That was it wa it was a different type of race, that's all I'm saying.

Wilhelm (04:59.416)
That's awesome. Yeah, why don't we come visit you and do something like that together? That'd be sick.

Matt (05:03.471)
Yeah, i you're more than welcome dude. There's there's a bunch of these type of things going on. and it was pretty intense. Like the first the first people the first the first guy who finished, I think he did it in an hour forty two for twenty two Ks with decent elevation. It's pretty fast. Hour forty two. So like it's kind of half marathon on road, but normally you would say it takes about double the time for you to do your half marathon.

Wilhelm (05:10.273)
Mm-hmm. Yeah.

Wilhelm (05:20.353)
Mm-mm mm mm.

That's very fast, yeah.

Wilhelm (05:32.44)
Seriously. Wait, so this guy road marathon equivalent would be half of that.

Matt (05:33.808)
Yeah.

Matt (05:37.637)
So this this this guy's probably yeah, this guy's probably well he's probably not half, but he's because at the top end it goes down a little bit. But he's probably doing like just over an hour, like maybe maybe an hour five, an hour ten. Like it's f like he's properly pro. Yeah, yeah, yeah. Like properly pro. yeah. So he's doing he's doing some good stuff. And the fact he did like he did this in an hour forty four, it was well i it was wild fast. Like

Wilhelm (05:45.099)
Yeah, yeah, fair.

Yeah, wow, okay, yeah.

Let's wrap it.

Yeah.

Wilhelm (05:59.021)
That's awesome.

Wilhelm (06:04.47)
Yeah.

Matt (06:07.619)
It's insanely quick. yeah, and they're like falling down the hills. Yeah, no, it's cool. But yeah, dude, how was your tri Nah, I'm not very good. My my ankles hurt right now and I d f fair, Judith left this morning. She went back to London to see her parents and her brother's graduation and stuff. Dude, I don't know how she's walking. Like my legs are not working and my ankles hurt so much because you're basically just falling down this hill and trying to stop yourself.

Wilhelm (06:07.838)
Sounds like a great vibe.

The photos looked awesome. Yeah, yeah, yeah. Are you a good downhill runner? Are you a good downhill run? okay, okay.

Wilhelm (06:29.538)
Mm. Mm.

Wilhelm (06:35.511)
Yeah, yeah, yeah, yeah.

Matt (06:37.423)
And it's not like the cheese roll style fall where there's like plenty of space to go. These are like very narrow paths and they're like winding and you've got to just you've got to go for it. Yeah, no, she was very fast.

Wilhelm (06:42.988)
Yeah, yeah, yeah.

Wilhelm (06:46.872)
Does your back hurt in weird ways from the downhill?

Matt (06:50.927)
Not back, like my traps weirdly hurt. Like

Wilhelm (06:52.302)
Okay. Mm. Yeah, that's what I mean, I think. Yeah, yeah. When we did a we did a downhill race last year and it was just like, you didn't even know that I had like muscles there that could hurt in that kind of way.

Matt (07:04.79)
It it's very strange. Anyway, she was so fast. I think that's why she got all her time on me. Is like well she got a lot of time on the downhill. Like we ha there was one downhill that was it's a good like eight, nine minutes of going down. and I think she got I think she got three minutes on me. Like Yeah, seriously strong. Seriously strong. yeah.

Wilhelm (07:09.514)
Hm, on the downhill.

Wilhelm (07:24.15)
That's strong.

That's very, very strong, very, very strong. yeah, the try was was was great as well. It's it was awesome actually, yeah. It's it was on the like edge of Yosemite on a lake called Bass Lake.

Matt (07:34.244)
Where was it? Where was it?

Wilhelm (07:40.774)
And you swim in the lake and then you do a massive loop around the lake, which was really cool. And it's really quiet up there, like there's no one on the roads. I've never done a race where like it's just you and like three other guys on a bike or something. And you maybe see like three or four people the whole race. There's no cars, even though the roads weren't closed for the race, because it's like such a s it's a really tiny race, like I think two hundred people do it. and so they they I guess they can't really close the roads. Also, like there's not really enough

Other roads that cars could go on instead because it's like so close to like a national park. There's like usually one road. but despite all of that, it was just empty and beautiful and some really nice long downhills. It's the first time I rode the TT bike since September, and as just like so fun. I'm not a good climber, like I said, so the like I was tough going back up, but

Matt (08:18.285)
Yeah.

Matt (08:27.489)
Mm.

Wilhelm (08:36.958)
It's just really fun going like really fast. I had like my aero helmet, like the weird-looking sperm helmet. And like on the arrow, it was just really, really good. And then the run was like a little bit hilly and but it was a great temperature. We did it with like two friends. Yeah, it was a great time. Really, really good time. And Jess Podium does well, actually. Look at us. So yeah, good, really good time.

Matt (08:58.234)
That's awesome. That's so good.

Wilhelm (09:05.932)
A lot of driving to get there and like driving to get back home and like staying in like Airbnbs and stuff. I think it it was like a five hour drive there and like back Yeah. But I mean that's that's with like stopping and like yeah.

Matt (09:09.443)
How far away is it?

Matt (09:15.021)
wow. Wow. I think five hours and I can end up yeah, I think five hours I can end up in Madrid.

Wilhelm (09:22.798)
Yeah, I can believe that. Yeah.

Matt (09:25.77)
Yeah. I can end up in like a different country, like halfway into that country.

Wilhelm (09:31.384)
A hundred percent, yeah. Yeah, California is massive. And then you you do end up spending a lot of time on these like windy roads w as you get closer to the to the nature parts. But yeah, really cool, really small race, really cute. It's funny, I think they actually messed up all of all everyone's times a little bit. Like that's the kind of quality of the race that like they they really care about making it beautiful and it's like an amazing vibe. There's like free waffles on the finish line and all the stuff. But they were like mega stressed about

Matt (09:39.725)
Yeah.

Matt (09:51.65)
Nice.

Wilhelm (10:01.034)
yeah, and every I think everyone's times were like off by ten minutes or something. So it's not awful, yeah. But would highly recommend Bass Lake Olympic triathlon if anyone is inclined.

Matt (10:06.688)
Okay. Well that's not awful.

Matt (10:17.386)
So are you loving the Olympic distance?

Wilhelm (10:20.492)
Actually, I haven't done this one this distance in in three years. It's a cool distance. It's actually a cool distance. Yeah. The swim is really long compared to the like longer distance triathlons. Because like I guess if you break down my splits roughly, it's like half an hour of swimming, like an hour twenty on the bike, and then like another forty, fifty minutes of running. So the the swim is like a much bigger percentage. and I'm not a very good swimmer.

Matt (10:29.89)
Yeah.

Wilhelm (10:50.498)
But yeah, it's an interesting distance. It's like Yeah, yeah, yeah. Yeah, yeah. So even with a wetsuit, I'm not a good swimmer. yeah, how was your how was your kiting race? That was last weekend.

Matt (10:51.468)
Was it was it a wet suit swim? Were you in wetsuit? Yeah.

Matt (11:01.912)
Yeah, I don't winging race, which is basically the same thing now, but yeah, that was brutal. Jude was asking me on the way home from this one and she was like, like which was harder because they're like two like this trail race where there was some very good people but it was it was just a trail race and then the winging race last week where there were like the world champion wasn't there but like a bunch of the top ten were there.

Wilhelm (11:29.528)
Tim.

Matt (11:31.697)
and like like like maybe five of the top ten were there. So that that was really cool. and she was like, which is harder? And I think she thought I was gonna say the trail race because like it's like exercise and stuff. But the wing race was way harder because like the barrier for entry is ridiculous. Like the trail race, like I'm like a decent trail runner, like I'm not not bad by any stretch of imagination. And so like I knew I was like nowhere near gonna get timed out of the

This I think the timeout was like four hours. and I I knew I was like gonna be sort of like around two and a half hours, and I was like 2 31. and so, but like I knew that that was gonna be the case, and like I knew that even if I walked like all the uphills and just like jogged the downhills and stuff like that would be the case. but the winging race man, like, even to get to the start line, holy my

Wilhelm (12:00.589)
Mm-hmm.

Mm-hmm. Yeah, yeah, yeah, yeah.

Wilhelm (12:11.202)
Damn, nice.

Matt (12:30.27)
God, you need a lot of skill. And and like I guess just the barrier for entry is just sort of two, three, four years of competition in wind sports plus being really quite good at winging. It's it's this is basically it. It's like it's wild. So so I I I was playing around, so I was there and I was there with a bunch of friends who I used to windsurf with. One of them is the

Wilhelm (12:32.246)
Really?

Wilhelm (12:37.517)
Yeah.

Wilhelm (12:46.422)
Yeah. Yeah, yeah, yeah.

Matt (12:59.776)
Is the reigning win like wind Olympic windsurfing world champion. So he's like the reigning world champion. He's the currently still the best in the world at as of the last the last world championships. And me and him came last and second to last. Or second to last and third to last, I think, in this like winging event. And you'd think that it would translate. Like he is a wind four, he's like windsurf foiling. And winging, if you look at them next to each other, they kind of look the same. It's like a board and a sail.

Wilhelm (13:05.27)
Mm-hmm.

Wilhelm (13:22.572)
Yeah, yeah. Right.

Wilhelm (13:27.895)
Right, right.

Matt (13:29.164)
with a hydrofoil on water and he did two races and I did three races. Like like and there was probably I think there were like 12 races over the over the two days and we only we only completed two and three between us. So it's just like it is so hard. but we'll get there. We'll get there. It's it's just a completely different sport. Yeah, yeah. High skill floor I think.

Wilhelm (13:31.734)
Mm-hmm.

Wilhelm (13:36.685)
Yeah.

Wilhelm (13:44.802)
Wilhelm (13:50.978)
High skill ceiling.

Wilhelm (13:55.276)
High skill floor, yeah, yeah, right, right, right yep.

Matt (13:56.641)
High school full. Like I think you have to be really quite good in order to even be able to race and get around a course and all this stuff. So yeah, the the winging.

Wilhelm (14:03.512)
We need to put like a picture or something of what this looks like because I feel like it's hot. Maybe we make this this week's album art the like you doing the like some version of our normal one, but you doing the race. It are there a lot of crashes? Like it looks from the picture I saw that there like it's like it looks a bit like a one of those massive sailing racing things where like you could easily crash into another person's machine.

Matt (14:13.802)
Ha ha.

Matt (14:17.344)
Yeah. E

Matt (14:26.144)
Yeah.

Matt (14:30.785)
There's a fair amount of crashing. not so much how I'm involved, because I try and stay away from most people, but it like in general, there is some crashing. like I'm crashing into the water quite a lot, not necessarily that much into other people. In like the professional racing, they are crashing into each other quite a lot. And there is like some hectic accidents.

Wilhelm (14:58.478)
Yeah yeah.

Matt (14:59.539)
Le less so much in the winging because it's still newer. but in the kiting and in the i in the windsurfing, those two have it this is racing specifically. They have like some hectic accidents. Like I think one guy broke his leg pretty recently in like one of the most recent windsurfing accidents and like like equipment is regularly destroyed, like boards snapped in half and like stuff like that, yeah. Yeah. Yeah.

Wilhelm (15:09.976)
Yeah, yeah, yeah.

Wilhelm (15:21.462)
Okay, yeah, that's brutal. all right, should we talk about bad agents?

Matt (15:27.807)
Yeah, let's do it. I just like one thing that's just kind of fun. It's like like both of us have moved out of London and both moved somewhere where there's just like different stuff happening. I think it's really fun to see how we're just getting into doing more outdoors stuff. Like I used to do a lot more of this stuff when but before I moved to London and that wasn't like I was doing less techie, nerdy computery stuff. I was just living but somewhere where there was more access to it.

Wilhelm (15:30.749)
yeah, yeah.

Wilhelm (15:48.086)
Mm-hmm.

Wilhelm (15:57.185)
Yeah, yeah, yeah.

Matt (15:57.238)
And so it's really exciting for me to like be back in that in that scenario. And generally I think I feel I I think I have like a much better mental state during the week when when a few days on the weekends. Yeah, I can absolutely like annihilate myself up a hill in the north in the north of Portugal and then drink wine with like some seventy year olds in the evening. And that's like it is very good fun. Yeah.

Wilhelm (16:06.968)
When you do so in the weekend. Yeah, yeah. Yeah.

Wilhelm (16:15.394)
Mm-hmm.

Wilhelm (16:20.364)
Yeah, a hundred percent. No, I was feeling I think at the s like middle of last week, before this weekend, like, damn, yeah, it's really time for like a bit of a break. There's been like a lot of a lot of really fun work stuff and a lot of like a lot of a lot of weekends spent working out of choice and out of excitement. But

it is nice to have a break from it all and touch some grass or whatever. And yeah. And we're lucky to have lots of nature near us. both of us do.

Matt (16:47.573)
Definitely.

Matt (16:51.945)
Yeah, no, it's really really nice. And I guess that I wanted to bring that up because I do want to like segue us into the first topic I had for us, which is which is there was a tweet and I think it was a LinkedIn post as well, 'cause he's like platform maxing, from Harry Stebbings about the Corgi co-found about the Corgi Founder, which is some like financial SaaS insurance platform for AI.

Wilhelm (16:58.85)
Go on. Go on.

Matt (17:22.049)
and the tweet was about like I've never met a founder. Well no, if you want to be a founder, you have to work seven days a week. This founder doing so well put a cafe in his office because he couldn't find a cafe open at 3 a.m. outside his office, sleeps in his desk, sleeps under his desk, two and a two-thirds I if I'm paraphrasing correctly, two-thirds of his workforce have a corgi tattoo.

Wilhelm (17:37.054)
Yeah.

Matt (17:50.247)
Yadda yadda yadda yadda this guy lives for it seven-day grind. This is what you have to do. Let's all bite our computers and sniff off our keyboards or something like something along those lines. And and and I just thought some of the reactions to it were very funny. Because like first I'm like, this is just his brand. And yeah, I'm like, I mean respect to Harry Steppings for having a brand, for one.

Wilhelm (18:13.038)
Spawn some good memes.

Matt (18:20.188)
And like, I guess like living his brand. I showed this to Judyette and she was like, this is a really long tweet for someone to basically tell me they have no hobbies. And and I thought that was a very like French way of looking at it, which I really liked. but but in general, I I'm just like I saw some of the replies, and one of them that really liked was Carrie, the founder of Linear, replying,

Wilhelm (18:21.006)
Mm-hmm.

Wilhelm (18:37.335)
Yeah, yeah, yeah.

Matt (18:48.787)
And basically saying, yeah, sure, Linear, like we have like hard patches before releases and stuff, but in general, I like am home, I have dinner with my family, I go to bed at normal time, I don't work on the weekends, I have this like work-life balance that I think protects the the the the design and the craft that I have in my head and what I want something to become. Because you have to have some times of off-period.

Wilhelm (19:08.194)
Mm-hmm.

Matt (19:16.818)
And I mean we talked about grind slop a while ago, but holy

Wilhelm (19:19.456)
Yeah, yeah. You mentioned it last week. I I I hadn't read the post I think when we last spoke about it. I've I read it since and it's gone around so I I have some I have some thoughts for sure. But yeah, I think the Grantslaw post was referencing this corgi Harry Sebing stuff. Or like i maybe it actually s sparked it. Like the author of that Grant Slot post wrote it. Yeah, yeah.

Matt (19:23.987)
Okay.

Matt (19:37.17)
Yeah, I think it prefaced it. I think it prefaced it because Harry Stebbings posted a couple of weeks ago about he'd met this like crazy committed founder and like every founder must be like this person because this person like rocked his world basically. And it was just like wow.

Wilhelm (19:51.158)
Right. And turns out it was about Corgi insurance.

Matt (19:54.332)
Yeah, and then but the memes that have come off it were so funny, like the Corgi founder relating his financial AI SaaS platform to the Oppenheimer Phoenix project, I think. That was the sorry, the Manhattan Project, that was it. Gee, I'm gonna get hate from Americans now for not knowing what that project was. but yeah. Yeah, it it did. And I actually I watched the film, which I thought was

Wilhelm (20:07.426)
The Manhattan Project?

Mm.

Wilhelm (20:15.939)
Had a big impact on the world, that one. Yeah.

Matt (20:24.263)
Wonderful. just bring that out there just to say I'm not anti the project. But yeah.

Wilhelm (20:25.591)
Nice.

Yeah, it's interesting this thing. I th there was I read the Grindslot post and I think yes last weekend or something, and there's actually like I someone sent it in a group chat and was like, you know, is it like this? What's it like? Blah blah blah. And I wrote down this like very detailed reply, like various things I agree with, disagree with. I was like engaging with it seriously. And then I realized that like this is just the

Twitter debate that comes up every six months about should you work hard or should you not work hard. And then people present their incredibly like polarized views, because that's what fits in a tweet. And then we have this extremely stupid debate every six months. And like it just s like clockwork. And this is just the latest iteration of it.

And then I was just like, okay, screw this. I'm not engaging with this anymore. This is a complete waste of time. Like, I don't want to be having this debate. which I think is still mostly by position. But then I went on a bike ride with a with a friend, and this thing came up again. and he was saying he I think was reporting a little bit like from the front lines about some of the like

grind slop attitude that he sees a little bit day to day. And that made me realize, okay, maybe there are like more people, employees, founders than I realized, who really do care a lot about

Wilhelm (21:52.088)
presenting how much they work publicly, which I think that's clearly like such a stupid thing to be. Or like that that that's I completely agree that that's like a bad goal that you need to be constantly posting about how hard you grind or how much you work or this kind of stuff. And actually that matters more than like what you're building or why you're building it or like

creating some new beauty in the world. Like that's all secondary to like, no no, like I'm posting 10,000 times a day. I work seven days a week. Like that's obvious like I so maybe this is a bigger issue than I thought.

Matt (22:20.145)
Yeah, I think

Matt (22:25.501)
I think it's I think it's actually quite more yeah, I think it's more fundamental than maybe what you're thinking. And I think it it's exacerbated by a lot of mainly American influences, which is why I think it's quite very funny to have Harry Stemmings on it, because obviously he's not American. But I ha know a bunch of people who went to YC and there was a big it was around the time of like nine nine six where they were like talking about working nine to nine six days a week. I know, but they were

Wilhelm (22:41.806)
Mm.

Wilhelm (22:49.954)
Yeah, exactly. We had the nine nine six debate like six months ago. That was like six months ago.

Matt (22:55.142)
They were like glorifying 996. Like in all of their job ads, it was like and you see them all the time. Like you must be willing to do this, you must be willing to do this. And that that's the way they want to run their company, like no matter what. But then I look at some of I guess like personally I don't think it's the most wonderful way to run a company. Like I when you see you see a lot of people running and it's like whether are they running in a direction that even makes sense?

and I guess that was like Carrie's argument. But you are right. It is this fun th this movement of people bringing up that they work really hard, and then people bringing up that you don't have to work that hard, you have to just like think a bit more, and then we have this debate like every six months. But yeah, nothing ever gets done without working hard. But yeah, obviously you have to work hard. Like we're not like it's not the same. but I do think one people are losing the fact that you're you are trying to make

Wilhelm (23:37.24)
And then someone else says, actually, no, you do have to work hard.

Matt (23:53.925)
maybe they're not trying to make beauty in the world. But like it would be nice if people tried to like spend a little bit of time just to think about how they could make a little bit more beauty in the world. I I I think that.

Wilhelm (23:55.213)
Yeah.

Wilhelm (23:58.904)
Yeah.

What feels right. Well i i I I I think that's maybe the unique twist on this current moment in time, that like it like and we talked about this a bit last time, this like

You know, you all the people around you are like getting crazy rich, and the only way you can hop onto the bandwagon is by like grind slopping and grind slop forking someone else's company or whatever. so you like you end up building exactly the same thing as another company, and you're grind slopping day in and day out in the hope that you don't join the prominent underclass. And that feels like very like plainly misguided in some way, right? Like I think that that's that that's where I would for sure put a flag.

In the ground and say, like, no, no, like this is like the best time ever to build beautiful, new, different things. Like let your creativity run loose. Like, do that at this point in time. Don't don't like grind slop on something that like I think like that's the weird thing, right? Like if if the grind slop or if your 996 routine is more important to you than like the product you're putting into the world.

Matt (24:54.385)
Yeah, and have

Wilhelm (25:06.946)
That feels really weird. Like you should have some vision, you should have some goal. You should have some like something you want to make happen in the world. And then sure, work hard at it, but that but that should come first.

Matt (25:11.139)
Nah, dude. I agr

Matt (25:16.505)
I I

No, but I I I okay, I agree with you. I I agree with you in on so much on so many levels there. But I just want to caveat this with like if you look at other industries, especially in somewhere like London, which is kind of funny because there's a lot of finance going on there, the amount of people who I can't leave work until my manager leaves work. I can't I can't I can't have dinner I can't have lunch away from my desk because that would like look like

I d I don't care, you know. I must find something to do here. I must be up. Like I know people like constantly working till one, two AM on stuff that they really don't believe in because that's what they think they grind like this for ten years and or fifteen years or twenty years. I don't know how long it is. I've no idea how much these people make. and then they'll be like financially free and you know, they can do whatever they want with their life afterwards. And this is a plan for some people. Like I think we we have the we have a really good

Wilhelm (25:55.106)
Yeah, yeah.

Yeah, yeah. Right, right, right.

Matt (26:19.754)
We're like like as developers w always in this like special case where traditionally you could earn quite a lot of cash without working a huge amount of hours. and that was like a that was like a thing that I think what these people are saying and what it really winds other people up is that nah nah nah, that is not a thing. Like we are the same as everyone else. Go and you you gotta put your twelve hours a day in every day of the week if you wanna do something like

Wilhelm (26:31.852)
Yeah.

Wilhelm (26:44.622)
Mm.

Matt (26:49.39)
like that actually does d has those impacts. And I don't necessarily believe that because I think the internet is a massive accelerationary factor and obviously AI is even more. and like you know, I just I just think it's sad. I just think it's sad to be projecting like that so much on the internet. But that it's the fact what we're talking about is insane, but it's Harry Stebbings's brand. So we let we let him have his brand.

Wilhelm (27:00.973)
Mm-hmm.

Wilhelm (27:05.058)
Yeah.

Yeah.

Wilhelm (27:12.14)
Yeah.

Wilhelm (27:16.864)
I think it's I think it's weird that yeah, and to be fair, I I think he's like a nice guy. Like he seems like a nice smart guy. I think he just has this he's figured out some engine for his social media like approach that works, which is in part rage bait and part like it's just like what he does, I guess. yeah.

Matt (27:38.883)
Yeah. It does wind a lot of people up, I think me included. I still haven't quite worked it out.

Wilhelm (27:43.062)
It's sad it's sad that like mean you look on any of these social media things and it's it's it is sad how much how much of it is like rage bade, right? I mean how well it works. Yeah. That's it's a bit it's a bit weird.

Matt (27:54.317)
Yeah. Yeah, it's just it is bait, but but it the weird thing is in this case, I don't think it's just bait. Like it is fundament it's true. Like his like and when he talks about the culture inside 20 BC, holy shit, I wouldn't want to work there. Like that is not how I would want to do investing. That is not how I'd want to run a company, that is not how I would want to work in a company.

Wilhelm (28:08.908)
no, sure, sure, yeah, yeah, yeah.

Wilhelm (28:15.48)
But like but but the point is

Wilhelm (28:20.536)
But I'm sure the real version has caveats and like is more nuanced and whatever, right? But like if you if you write a more nuanced version I feel like it's if you write the more nuanced version of like the twenty VC culture, whatever it is, like I feel like it's it's just not gonna get as many clicks.

Matt (28:26.677)
Nah, but you must have heard horror stories.

Matt (28:36.503)
Well, not that culture specifically, but like you no, but you must have heard horror stories. You must have friends who work in like weirdly like weirdly demanding environments. Okay, so so this one. Just just this like egotistical shit. Just just f put in perspective. Nah, dude. So you must have no, you definitely do. Okay, so I'll put in perspective. So I've a f I have a friend who used to work on the podcast of a very prominent

Wilhelm (28:51.436)
I just have you, man. You're my friend.

Matt (29:05.007)
podcaster. Let's put it that way. I'm not gonna say his name because it's a bit ridiculous. but like top couple of podcasts, at least in the UK and globally. and they had this thing. Like it's quite s podcast teams are quite small, maybe 20, 25 people. this included some other company stuff that that they were doing. and they had a thing where the the the boss person, let's call him Derek, he he

Wilhelm (29:15.116)
Mm-hmm.

Wilhelm (29:31.598)
Mm-hmm.

Matt (29:35.049)
He had a spreadsheet where he would invite employees, like it was a special like Google, Google, Google Sheet. he would invite a maximum of seven employees to it. And this spreadsheet was where you would document your fitness activities. And in this spreadsheet, if you were invited to the spreadsheet, you had like a free subscription to Third Space or something like that, or a fancy gym in London. and if you were in the bottom

Wilhelm (29:58.199)
Mm.

Matt (30:03.598)
What was it like two every week on this spreadsheet? You got kicked out. So, and it was like a secret spreadsheet inside the company. And it was basically so he could keep his employees, like the ones he liked, doing exercise, so they would all look really good. And and I'm just like, yeah, like get people doing exercise ex exercise that's good for them. But like that is not a stress that you want in your life for your boss to

Wilhelm (30:09.379)
The gym.

Wilhelm (30:32.162)
Ha ha

Matt (30:33.293)
Like I go do exercise to get away from work. I don't wanna be in the gym being like, I have to register n like ten more pull ups, otherwise I'm gonna be on this like lower class of employee status. It's I just insane. yeah.

Wilhelm (30:45.112)
Yeah.

Wilhelm (30:49.41)
That sound yeah, that that that that sounds quite entertaining.

Matt (30:52.844)
Like I I mean you can kinda see where it came from because it just sounds like a funny gag in the moment and then you're like, my god, like as a manager of people, that is horrific. Like just let people have their space. Let let them enjoy their letting them enjoy their life as well as like work is a part of life, but it's not like the whole thing. Like especially as an employee. Yeah, if you wanna rate your your your workouts, then you be you, but don't bring your employees into it.

Wilhelm (31:00.556)
Yeah, yeah, yeah.

Wilhelm (31:04.813)
Yeah, yeah, yeah.

Wilhelm (31:22.624)
Yeah. yeah.

Matt (31:24.984)
So anyway, that's like that's like my story of like how intense and weird work in these like small things can get. And so I am not surprised when Harry Stevens like, two thirds of my twenty VC employees have a twenty VC tattoo.

Wilhelm (31:40.91)
Is that also true? Everyone's getting tattoos, man.

Matt (31:42.636)
I think that's also true, yeah. It's just weird to have a tattoo of the company you work for. I'm sorry. Like it's weird. It's

Wilhelm (31:48.11)
I'll let the people have some fun, Matt.

Matt (31:53.868)
Yeah, I you do you obviously.

Wilhelm (31:55.214)
Yeah, yeah. You you exact that's the thing I feel. I to be honest, that that's my biggest issue with the six monthly debate. It's just like people telling other people what to do.

Matt (32:06.817)
Yeah, so.

Wilhelm (32:07.746)
There's no there's no one way to live a life. Many ways to live a life.

Matt (32:11.905)
Don't tell me to work seven days a week.

Wilhelm (32:14.742)
Yeah, or or don't join a company w that works seven days a week if you don't want to work seven days a week.

Matt (32:21.313)
That is also exactly accurate, yeah. Yeah. No no this is true. This is true.

Wilhelm (32:23.222)
Yeah. wait, you put lots of other stuff on the agenda. should we talk about the George Hotz thing?

Matt (32:32.635)
this is fun. Yeah, so George Hoss is like a very sort of prolific hacker, really. he was one of the first people to jailbreak the iPhone, or the first person to jailbreak an iPhone, I think. Then he d he he almost went to prison for was it a PSP that he jailbroke or something like that? Like there was some there was some he had some crazy lawsuit for a while. But he was like very young when he was doing all of this stuff. and

Wilhelm (32:40.758)
Mm.

Wilhelm (32:45.25)
All right. Mm-hmm.

Wilhelm (32:53.314)
Yeah. Yeah yeah.

Matt (33:01.973)
Now he runs a company called, or he definitely runs a couple of companies, but one of them is called TinyGrad. I think he also runs something else called Comma, which is self-driving driving cars. But Tiny Grad is the one that he's more known for at the moment. and that is they put together like GPU-enabled computers for local LLMs. So they're still very expensive computers. They're about like 30 grand. Yeah, and then they also sell.

Wilhelm (33:15.223)
Uh-huh.

Wilhelm (33:25.518)
they they they actually sell a computer, right? Yeah, yeah, yeah.

Matt (33:32.092)
All of the like the software as well. Well, I guess the software's for free. Like TinyGrad is like a is an inference engine that runs well on AMD GPUs. Yeah, so Tiny Box, sorry, Tiny Box is the is the machinery. But he so he's like been on a few podcasts and stuff over the years. He he's he's a interesting character. Like people can like look up some of the things that he said about.

Wilhelm (33:42.902)
That's called tiny box the their hardware. Yeah.

Matt (34:01.856)
various topics after this. So don't really want to go into it that much. but yeah George Hotz and he's quite interesting on Twitter because he has this very like pro AI and anti AI stance at exactly the same time, which I'm like kind of there for because I also feel the same where he's like super pro he's super pro AI in in terms of like self-driving is going to be super cool and he actually runs like an open source self-driving company.

Wilhelm (34:13.772)
Run it.

Matt (34:31.824)
he's selling like machines to do local inference. And then he's super anti-AI in some of like the the harder, like the more in innovative things that people want to do. And his did I did I put his did I put his blog post in? But there was a line in it that really the last bit I I'm just gonna like I'm just gonna read it. So the blog post is called The Eternal Slopt Tember.

Wilhelm (34:46.701)
Yeah.

Wilhelm (34:50.176)
Yeah. Yeah. I just skimmed through it. Yeah.

Matt (35:02.172)
And he says the first line is I'm c I'm gonna give the first line and the last line, and then you can go and read it yourself. It's not very long. But I'm calling it now the adoption of AI agents to software development will be one of the most costly mistakes in the field's history. Agents cannot program, it's taking longer and longer to realize that they can't. They're highly sophisticated statistical modeling designed to mimic the distribution of programming. The output is broken, but in a way that's getting harder and harder to detect, which is exactly what you'd expect from an increasingly accurate statistical model.

And then the last like line is, well, paragraph is without endorsing all their ideas, I'm now in the Lacoon Marcus camp on LLMs. I don't think models like this will ever be able to program. I think the process matters. I think that deep learning is still a solution, but the real programming agents will need world models, not some RL VR shit that comes out the failing test that comments out the failing test and tells you all the tests are now passing. The real story of this era will be who manages to avoid harming themselves in their AI psychosis.

So

Wilhelm (36:01.41)
Bombshell. Yeah. That's obviously quite different to how we feel about it. So how yeah.

Matt (36:06.003)
Yeah, well, I I think that's two things that are kind of funny. Like the output is broken but in ways that are getting harder and harder to detect. So to me, that means the outcome is becoming less broken because from a like a scientific like I I don't have this like statistical no, sorry. A as an as as a I I studied in mechanical engineering and as a mechanical engineer you look at a lot of like empirical data.

And if there is something that is undetectable, you kind of don't really care anymore. So it for an example, everything has a tolerance. when you when you make a part, you put it in CAD, and you're like, this part is gonna be 10 millimeters or let's call it 100 millimeters long. It's a hundred millimeters long and like ten millimeters deep and I don't know, twenty millimeters high. And you put it in CAD and then and then and then you get it back from your machinist and the

Wilhelm (36:41.878)
Mm, right.

Wilhelm (36:54.232)
Uh-huh.

Matt (37:04.703)
the C and C that they used to machine it was well even if it's imagine it's not even a C and C like the the cast the cast that they used to like that it was forged or something and it's like a hundred and one one millimeters long and you go well okay a hundred to one hundred and one we could sand it down maybe get it back down to a hundred sweet maybe we could like route it or something and then you're like okay let's we need to get painted now. So let's let's go get it let's go get it painted and you like a couple options. You've got a powder coating

You're gonna slap some paint on it, or like you're gonna do something, you're gonna spray it, you know, whatever. and you're like, okay, we'll just do it like we'll just do it the we'll go powder coat it, we'll go powder coat it. That's really strong. And then it comes back and it's like 101.1. And you're like, what? We've gone up like 1.1 mils now. And so like you always have a tolerance, and then you have like your your surface roughness is now like it's got to be within tolerance because otherwise it might like break against another moving part or something.

Wilhelm (37:56.022)
Right, right, right. Yeah, yeah, yeah.

Wilhelm (38:02.99)
Interesting.

Matt (38:03.412)
But but everything is within tolerance and the point is that you try and get something to within tolerance where the the maximum error something can be is like within your within your stated limits for like the part. And I guess what I'm trying to say about this is like if something approximates to the correct solution, even if I can't like tell it from the correct solution, then to me it's the correct solution.

Wilhelm (38:16.023)
Yep, yep, yep.

Matt (38:30.581)
and so I'm kind of okay with stuff getting better and better to the point that I can't tell. He's obviously not, because he wants things to be correct. Unquotes correct.

Wilhelm (38:38.602)
Yeah, yeah. I mean, do you think it's just because he does such like low level stuff that like I mean he he's writing like CUDA kernels and stuff, right? Like I mean this tiny like it's like it's customized kernels depending on what you're doing with the thing. Like that's like it doesn't really get more low level than this. So I can imagine that models aren't very good at this. And also he really cares about like the depths and details of like how it works. It's yeah.

Matt (38:48.5)
Yeah.

Matt (38:57.608)
It's brutal and

Matt (39:06.024)
Yeah. He cares about correctness. And I think there's a there's there's a l there is a massive difference between like computer science and and like compilers and all of that stuff and like thinking about this stuff from like this is silicon on it this is a silicon chip to a lot of the stuff that we end up looking at, which is so many layers removed that people like him have done a lot of the harder stuff at that point and we just have to move

move things around a page or send some JSON data to this place and try not to make too many requests to a database, you know? Like like there is th there there are many different gone.

Wilhelm (39:43.884)
Yeah. Yeah. Ex like do you think s i say that they for their company they need to like some internal dashboard to sol show like a bunch of orders and make it like easy to refund them or something. Like that's like a support workflow that they want to build. Do you think he would be like, No, we can't vibecode that because agents can't program, AI can't program, LMs can't program? Or would he be like, Yeah, okay, no we can do that?

Matt (40:08.989)
Yeah, one hundred percent he probably doesn't even consider that programming.

Wilhelm (40:12.18)
I see, I see, yeah, yeah, okay. Yeah, I mean then maybe we don't disagree with him.

Matt (40:16.668)
Yeah.

So so yeah. there's so many different levels to this field, man, and different areas of the stack. there's always different levels to this game. I I find this really fun when I go to Chamonix and actually like when I hang around with some of the sports people of a previous life where y you think you're good at something, like for instance, like I'm a decent climber, not bad in any climbing gym in London. I'm climbing

Wilhelm (40:24.568)
There's levels to this. Yeah, yeah.

Wilhelm (40:44.291)
Mm-hmm.

Matt (40:47.707)
Some of the hardest stuff that people are climbing, like I'm not great, but I'm not that bad. And then you go to you go to Chamonix and you're like you get having a chat with like a random dude in the pub and you're like, yeah, I like climbing. He's like, yeah, cool. I like climbing as well. I'm like, what do you climb? I don't know. Seven C, eight A. It's like, wow. Okay. So we're climbing, we're climbing, we're climbing different levels, basically. And

Wilhelm (41:01.857)
Yeah.

Wilhelm (41:06.208)
Right, right, right. Yeah, yeah, yeah.

Wilhelm (41:12.075)
Yeah, yeah, yeah.

Matt (41:13.189)
And it I think that is that's the whole pitch of why you went to San Francisco, right? You like with an area in an area where the talent for what you want to do is like concentrated and you turn around to someone and you need to raise some Sequoia and you turn around and the person behind you is raised some Sequoia. Like I think I I think there are there are always levels to this game and I guess a lot of the time hearing George Holt speak, you you realize how many levels there are.

Wilhelm (41:18.807)
And

Wilhelm (41:29.294)
Yeah, yeah, yeah.

Wilhelm (41:38.294)
It's it's hard like just from like looking at a crowd or whatever to see how many levels there are to this. Or like if you are somewhat of an outsider to a thing, it's hard to see exactly like how many levels there are to this. Yeah.

Matt (41:49.64)
And there are founder levels as well. I I was listening to an A16Z analyst, or XX A16Z analyst, I forgot his name, Benedict something. He's a British guy. Benedict Evans, yeah. He was on Lenny's. and he was quite he was quite funny. He was some remarks that Yeah, go for it. So some remarks that Dario made, he just like kind of shot down really based on

Wilhelm (41:52.048)
yeah, yeah, yeah.

Wilhelm (42:01.452)
And then like Evans. yeah, he's total legend.

Wilhelm (42:11.246)
I'm just gonna run to the toilet, but I can still hear you.

Matt (42:20.208)
I guess quite a lot of research. and that was interesting. He was he was kinda like, yeah, so I would listen to Darry a bit on anything to do with chips, like chip forecasting, anything like that. but you know, in this field, I think I know more than him. And this is all the research to back me up. And I guess it's like people have lanes. Not gonna ask George Hotz how to rig rig a windsurf sale, but in this regard, he probably knows a bunch more than me.

Yeah, right. Well while Wool's gone, I'm gonna stop monologuing, but I'm gonna look at some other stuff. What did we have on the agenda? Well s okay, so while he's gone, a guy called Sergio Sergio, I apologize if I'm not saying your name correctly, listened to our last podcast where we were talking about we would like a NYX replacement, a NYX-based replacement for Brew, and

They built it, which is absolutely epic. It's called Root. So we'll leave a link in the show notes, but essentially, if if anyone wants to have a look at it, it's Root is a safer brew for developers and AI coding agents powered by Nix with no Nix knowledge needed. he's they're up to well, just giving them a star on GitHub. And it looks really, really interesting. So yeah, it's called Root. Did you see this one?

Wilhelm (43:51.126)
I yeah, I I saw the tweet. Yeah, very cool. thanks for listening to the pod, Sergio. What's what for so I've never done anything with Nix. If I wanna like wet my beak, what's what do I do?

Matt (44:04.977)
Well, I guess go and have a play with Root. I mean, it says you don't need any Nix knowledge. So you should just have a play. You can Yeah, and it's ironically it's published to Brew. Shouldn't Root be published to Root? I don't know if that that's even possible. But but yeah, you can install it with brew, install root C L I and well CLI Dunno what happened then. and yeah, go for it. Like he's

Wilhelm (44:21.688)
Mm-hmm.

Wilhelm (44:32.642)
But what's like something I can do with Nix and root that like I would want to do?

Matt (44:37.127)
so Nix Nix is a tool for deterministic builds. And you can use that to build your OS, your like you can go as far as you can go as low down as like s like build your package, build like your development environment, and you can go as high up as like build your whole OS, like everything you install. and you even use like Nix Nix OS to install like a whole Linux distribution. And you can go from anywhere in between.

I think with Nix. It's a huge ideology more than anything else. I haven't like played around with it hugely, but a lot of the time when I think about a problem, it's normally to do with packages being out of date or different or like wrong. And in some regards that's been fixed on on project specific stuff, but where it hasn't, Nix is the thing to use. in my opinion. I

Wilhelm (45:11.191)
Yeah, yeah, yeah, yeah. Okay.

Matt (45:35.404)
A lot of people still just version their dot files and that's also fair.

Wilhelm (45:39.438)
Should maybe my my Mac mini should have a entire Nix based setup so that when that eventually crashes.

Matt (45:42.49)
Your Y your your Mac Mini Yeah. Your Mac Mini should be fully nixified, prob most likely. Or at least the stuff that you use on it should be so that if you did have to wipe it, it wouldn't be the end of the world. Yeah.

Wilhelm (46:02.402)
Makes sense. Okay. I mean that sounds like a project. But sure will happen eventually. can we Yeah.

Matt (46:07.248)
Yeah.

So yeah, I want to see what this this this this root thing is. Anyway, it's it's built in rust, looks kind of interesting. I wanna have a look into it a bit more. but yeah, obviously we won't do that now. But yeah, definitely have a little look in the future. Thank you for building that. I mean it's cool that someone listened to the pod and then decided, I'm gonna go build something that we talked about, which is sick. And if you do do that, tag us on Twitter because we love seeing this sort of stuff.

Wilhelm (46:33.773)
Yeah, very cool.

Wilhelm (46:39.95)
can we talk about Opus 4.8? Came out on Thursday. Came out on Thursday. So we've had a few days to play with it. It's interesting how quick the vibe shift was on this because I I feel like it was announced, and there was the initial reaction seemed like overwhelmingly positive. And I'm thinking mostly about the like Dan Shipper everything, everything.

Matt (46:42.509)
Of course we can.

Wilhelm (47:06.572)
As in like the every blog post, which was like Anthropic is back, 4.6 and 4.7 were hard to love, 4.8 is incredible, like better than GPT 5.5, it's so good. Like it's the go-to now for like them, the only thing that's holding Anthropic back is like the harness or like you know the codex app is better or whatever. and then people started playing around with it, and then there was a pretty, pretty quick like swing back from people saying, like, it's just like

Matt (47:07.525)
Mm.

Wilhelm (47:36.498)
a bit weird it like pushes back on you a lot even when like it's not as much a collaborator they've they've tried to nerf it in like a bunch of ways that feels strange and and different I had like one or two like weird interactions with it and this is kind of interesting like I I'll mention the one thing like I had like a I did a bunch of blood tests like last week mostly to feed them into back into my agents back at a chat and

And tell me if my supplement stack is working well and and if I'm healthy and all this. And there was like one thing in there where it was being weirdly coy. Like it clearly wanted to flag something in my data, but it like it was talking around it. Like it didn't want to initiate the flag. Like it was just like I don't wanna go into details of like what it actually is, but it was like, like have you noticed.

Matt (48:23.861)
yeah.

Wilhelm (48:35.83)
Or like i it was just very strange. And then I'm like, bro, like what are you talking about? Like why are you being like this? Like just tell me what you want to say. Because it didn't wanna say. So and I've never had that before with a model. That felt very strange. So yeah, I don't know. I I I don't know how I feel about this model. I also haven't used it a ton. I use it a bunch on Thursday and then on Friday, but not over the weekend.

Matt (48:59.033)
Yeah, wow. Yeah, I think I I mean to me it spells bigger than previous ones. I think 4.7 was a little bit like the best way I can describe it as like circular.

Wilhelm (49:08.078)
Mm.

Matt (49:15.683)
Like it didn't explore any opportunity any possibilities, it like just it just carried on on the direction it was in and never never found anything interesting. and then I've heard four point eight being compared more to five point four, I think it was. Like GPT five point four where where it in e any scenario it's like right, I need to like see what's going on here and then it like looks up it finds the context itself, which which is in my opinion a lot a lot

Wilhelm (49:16.397)
Mm.

Wilhelm (49:22.77)
interesting. Yeah.

Wilhelm (49:32.014)
All right.

Matt (49:45.345)
Easier

A lot easier to like deal with and easier to work with because you don't have to supply everything up front. The model can actually go and like look for context a little bit better. So so I I think I think that's my only good take. That's my only like strongly held take so far on 4.8. what I really enjoyed was they have this like new ultra code mode in in I I tried it earlier today and it's really nice in Claude Code.

Wilhelm (49:58.35)
Yeah.

Wilhelm (50:11.969)
yeah, yeah. I I haven't tried this yet.

Matt (50:19.162)
I and like I feel like the measure of harnesses is kind of in is interesting because on some benchmarks clawed code actually comes out really badly as a harness for the opus for the opus models, which I find funny, or maybe just shows that the benchmark is broken or whatever. I I really don't know where I really don't know what that means because surely Anthropic is testing their

Wilhelm (50:28.462)
Mm.

Wilhelm (50:36.216)
Mm-hmm, mm-hmm.

Matt (50:45.336)
They must have like internal benchmarks of on their harnesses. And you would hope their harness comes out best on their internal benchmarks. And then other people have benchmarks on harnesses. And and the and the claw coat harness doesn't come out best on their benchmarks. So maybe they're hitting like weird cases. You can never like the benchmarks are so broken. I don't know how I have like benchmark delusion now. And I ha the models are so close now, I can't even

Wilhelm (50:46.125)
Yeah.

Wilhelm (51:09.784)
Totally.

Matt (51:12.568)
This is another thing that Benedict Guy was talking about. He's like it remains to be seen if the AI models actually or if the model labs end up having pricing power. like or if it's going to be some I hate to use the word commoditize, but some like commoditized intelligence, like y or not intelligence, but a commoditized infinite word generator. I don't know. like

Wilhelm (51:34.593)
Right.

Wilhelm (51:40.364)
Yeah, interesting.

Matt (51:41.494)
It i is the model is the model gonna be commoditized? Like, is there gonna be any meaningful difference between a GPT-6 and Opus V? Like, is there gonna be any meaningful difference? Like, will we will let the average person be able to tell? Because if they can't tell, then I'm sorry, George Hoss, I don't believe in what you're saying. Maybe it's the case for like

You but if the average person can't tell, then to me they're commoditized. And then and then I think that could be a top for the market, dude. We could be we could be entering like the SPAC generation of of of of AI models, like where where everything was gonna about to go public and then about half of them got public before the floor fell out.

Wilhelm (52:12.702)
Yeah.

Wilhelm (52:18.902)
Так

Wilhelm (52:31.118)
Hm.

Wilhelm (52:34.72)
Interesting. Yeah, I feel like this feels so hard to predict. w wait, can we talk about the ultra code thing? What what does it actually do? Is it just like a mode beyond max? Okay.

Matt (52:42.765)
I have no idea. It just has some fancy colours. Yeah, it just has fancy colours. That's that's all I could it you you go effort level, which I actually been struggling to find in Claude Code because I'm so used to Shift Tab, which I have in Pi, to flip between effort levels. I like completely forgot how to do it in Claude Code, but it it came as a prompt and so I was like, I'll try it. Effort level ultra code and then everything goes like sparkly and purple and you're like, whoa, it's like Pokemon.

Wilhelm (52:57.077)
yeah.

Wilhelm (53:08.782)
But the idea is that it dries harder than even max. Like it's a level beyond max, or is it something okay, okay. Okay, okay. But it's it looks cool, which to be fair, that's a good reason to

Matt (53:15.543)
Yeah, yeah. I have no idea. I didn't even read that part of the blog post. I

It just looks cool. And as I said, when in an undifferentiated market, you gotta look cool. That is your differentiation.

Wilhelm (53:27.522)
You gotta look. The thing that I have really enjoyed the past few days is the workflow, like the dynamic workflow feature. So if you type workflow, then because I've been doing this like quite like wildly detailed like data warehouse work, basically like structuring

Matt (53:34.69)
Mm.

Wilhelm (53:50.318)
Like it's almost as if like Simple Paul had like a data science, data engineering team. And we can use all of our like product analytics event data in like an actual useful way and do things that we just couldn't really do before, like a proper cohort analysis of like different eras of use, patterns of usage, like all the things that, yeah, previously I think you would have had to have like a team to tailor like your

like understanding your users and usage in a way that's like tailored to your product as opposed to like a generic mixed panel or something like that where you just can't go very deep. So we've been setting that up like a proper data warehouse thing. And there's a lot of kind of work that can happen in parallel. Like, this schema needs to be deployed and then we need to check like if it worked and like at the same time we need to make this change. And the workflow thing

I don't know how much time it actually saves, but for it for sure is super cool that it's like smart like sub-agent dispatch and then it'll reconcile the things and it will like loop it back into the main thread. So that's been like that's been very fun. Or I get the I get the idea of it. I get the the vision behind it, and it's fun to use, so I'm here for it. And for context, this is how Jared rewrote Bun in Rust in a week. It was using

The workflow feature, as I understand it.

Matt (55:15.925)
Yeah. We have Yeah, we have like these two kind of similar threads going on at the moment with the harness, with harness orchestration, I'm calling it, which is you have workflows, which are essentially like a big prompt that you hand off to another model and say, Do this, do this, do this, do this, do this. and then

Wilhelm (55:16.696)
So massive parallelisation.

Matt (55:42.335)
A and maybe there's some determinism on like whether they actually do the thing or not, and then it comes back to the main thread and you do those in parallel. We have that, and then we have like the single threaded maximalist, which is kind of like the auto research goal, like forward slash goal style thing where it's like this there is a goal to do something, and every every time you do something, check against the goal, if it works, s like

Wilhelm (55:53.87)
Mm.

Wilhelm (55:59.446)
yeah, yeah, yeah, yeah, yeah, yeah, yeah.

Matt (56:11.891)
Ship it in coach. If it doesn't work, revert and let's just keep going. And let's just keep going forever and ever and ever, just gradually pruning context at the in the middle of the window. And I I'm not sure. I I have been really bad at predicting like which patterns are gonna last the test of time. So I don't wanna try and predict it, but I do find it interesting that there are these two schools of thought. There's like massive paralyzation.

Wilhelm (56:38.83)
Mm.

Matt (56:39.071)
which is what dynamic workflows aims to be. and then there is this like single threaded token, like infinite long running background token max, basically. yeah.

Wilhelm (56:46.06)
Mm-hmm.

Wilhelm (56:51.702)
Yeah. This actually reminds me of something. because so I feel like the slash goal thing works very well in codecs because I trust codecs to compact, auto-compact like well and ca carry the intent forward. But I don't trust cloud code and and opus. But maybe it's just maybe it's just like my maybe I'm missing something. But like for me, like whenever I get in cloud code, I get to like like max like say thirty, maybe thirty five percent context limit use.

I'll be like, write a handoff prompt, clear the session, paste the handoff prompt, that's it. Is d what do you do? Like or when do you clear the context window in in Cloud Code or in in Pi, I guess. Or do you trust it to auto compact at like a one million context window?

Matt (57:36.86)
No, I would never go to one min I could never go to one million. I normally go to like two hundred and like two hundred odd K mark with the GPT models and then five hundred K with Opus. I think is like kind of okay.

Wilhelm (57:46.858)
Okay. So you go to like fifty percent. And and the GPD models, they have a smaller context window or are they also one million?

Matt (57:53.632)
I think they're meant to be one million, but I feel like it ends up getting pretty garbage over over their actual con over the their trained context window, which I guess is two hundred and forty-odd tokens, or their like base model context window. I the something has a context window of 240 odd tokens, and so I think it gets quite a lot of gets quite sloppy after that. But I I think that's actually a harness approximation, the thing you're talking about. because have you tried Pi Auto Research?

Wilhelm (58:01.304)
Yeah. Interesting.

Wilhelm (58:11.511)
Yeah, yeah.

Wilhelm (58:18.71)
Mm.

Wilhelm (58:22.805)
I have still have not, no.

Matt (58:24.572)
Yeah, so y y you need to try that and then you need to try it with the opus models and then tell me that it's not just a harness thing. Because

Wilhelm (58:33.614)
So you you you think that auto research works equally well in Opus and GPT?

Matt (58:39.508)
Yeah, it's great in Opus. Like, I've had no problems at

Wilhelm (58:42.304)
Okay. I think what I'm basically just saying is that like by default in Codex, like I feel like it whatever it does behind like in I when I'm using Codex, I don't worry about the context window. It just keeps going well. Like it just the way the compaction works is good. But and I and I feel like in codex the default compaction is way below a million. Like it happens really quickly that they like figure out the compaction. Whereas with Opus, I can see it doing garbage things.

as I get closer to like thirty, forty percent context window use. And there's no automatic compaction for me. But I mean, these things are so complex and maybe I'm missing a flag or something like that, or I'm missing some setting or I don't know. But

Matt (59:25.726)
Yeah, I don't I think it's just a harness thing. I think it's just a harness. And like if you look at some of the compaction prompts, they are very different. Like, some try to keep the current thread of where you are, some try to keep the whole vein of the whole conversation. some try to keep the original intent as much as possible. So like Pi will tell you like what the original intent was.

Wilhelm (59:45.089)
Mm-hmm. Yeah.

Wilhelm (59:49.676)
Mm-mm.

Matt (59:54.664)
like the next steps and also what files will che what files you might be interested in. was there that I feel like what Claude does is like this is the current state of where we are right now. And then it doesn't the maybe you still get the last message, but you don't get much more about like where you're trying to go to. So so I think the Claude code harness like kind of sticks you in time when you do compaction.

Wilhelm (59:55.022)
Mm-hmm.

Wilhelm (01:00:00.6)
Mm mm mm mm.

Wilhelm (01:00:15.296)
I see, I see. Yeah, yeah, yeah.

Wilhelm (01:00:21.59)
I see, I see.

Matt (01:00:22.601)
But this is just like a feeling. I haven't measured it properly. I don't even know how I would measure it. But y you should read some of the prompts, the compaction prompts. They are really interesting.

Wilhelm (01:00:31.576)
We're all about the feelings. Okay, wait. So in Claude do you do you run slash compact in in Claude code? Okay.

Matt (01:00:37.329)
Yeah. So in Claude I in Claude Code I think you have to be much more like active about context management. Yeah. And you do have to like be like, I'm at four hundred K now, I'm moving on to like a different idea, but I c still kinda want some of the context of what I la of what I did last. I do compact and then I actually provide like details of what I'm gonna work on next and so it knows which bits of data to keep. Yeah. I think they're probably doing something a bit fancier there.

Wilhelm (01:00:43.51)
Yeah, that's what I feel as well, yeah.

Wilhelm (01:01:03.085)
Yeah, yeah, yeah. Yeah.

Matt (01:01:06.909)
to like gather the right context for what they consider the task to be. I don't think they carry on with the intent very well. Where their pie is like all about the intent. It's like this is what we were doing. These are the follow-ups that we're going to look at next. These are the files that we modified.

Wilhelm (01:01:25.826)
Yeah, yeah, makes sense. I have to run, I have a meeting. I want to shout out one thing, which I just saw on Twitter and I'm keen to test out, which is called herder.dev. One terminal, the whole herd. So I think it's basically like a wrapper around your terminal that helps you manage all your different like agents, harnesses.

Matt (01:01:27.465)
And it's very specific there.

Matt (01:01:33.033)
Gone.

Matt (01:01:52.998)
It's a wrapper like Tmux. It's like Tmux. Or it's a wrapper around Tmux.

Wilhelm (01:01:56.822)
I think it Yeah, exactly. Yeah, yeah, yeah. Which is cool. my current setup could definitely use some work. It's it it works for me for now. So what I actually do is I have VS Code open and then I just have like the terminal dragged to like almost max size because I sometimes want to look at the files, but most of the time just look at the terminal. Then I have like two terminals side by side, and then I might have like

I don't know, eight or well, actually I guess eight would be like the the the top, maybe more like four or five different tabs, which each have one or two terminals in them. and that kind of works so far, although sometimes it like the whole thing just crashes in this weird way and it stops rendering. And I have to like do the command shift p reload window trick to like get it to reset.

Matt (01:02:42.43)
Dude, sack off, sack off VS Code as a terminal. Dude, you should. Okay, so I'm I'm pretty much only in Ghosty now. Like I have like maybe like 20 terminal tabs open at any one moment. And I'm very active about pruning them. I feel that they're kind of similar to like Chrome tabs. When I have more than like 20 or 30, my brain starts freaking out. And so I'm very active about pruning them. I don't have any like like session management system.

Wilhelm (01:02:46.926)
So wait, what do you use for this? What do you what what's your setup?

Wilhelm (01:02:53.943)
Okay.

Wilhelm (01:02:57.772)
Yeah.

Wilhelm (01:03:02.669)
Yeah, yeah, yeah.

Matt (01:03:11.74)
I'm looking to add it. I was looking at CMUX, but I don't know what I'm how I'm gonna do it. But then for diff viewing, I use Pi diff viewer. So I was I was using yeah, and then if I manually want to go in and like write my own stuff, so like I want to go and like write my own docs or write I don't know, like a README and be more specific about things, or like a change set, then I will open like a then I'll open an actual

Wilhelm (01:03:16.182)
Mm-hmm.

Wilhelm (01:03:21.966)
Okay, so it's in the terminal.

Matt (01:03:41.999)
editor or I'll just like nano or Vim i i in my terminal, yeah. Because I don't do it very often anymore. Like very rare that I open a full well I mean today I think twice I opened a full editor. and like I was

Wilhelm (01:03:45.804)
Really okay.

Wilhelm (01:03:55.458)
Yeah, yeah. I find it just quite helpful to see on the side like, you have like a hundred dirty files in git and like see the files or like see the see the git tree. And sometimes I look at the file tree and I like it's nice to know w what files exist and occasionally I like open one. But I think mostly I have the git the git thing open.

Matt (01:04:10.14)
Yeah.

Matt (01:04:14.511)
If I want to do that, I I do that a lot in Vim, basically. If I want to do that.

Wilhelm (01:04:18.348)
Okay, so you look at the as in like you run git status, or is there some like plug in?

Matt (01:04:24.879)
No no no no I look at files. No no no. So I look at I look at not like the files that have changed, but I if I wanna look at my files, I look at them in Vim or in Nano. depending on whether I can remember the key bindings in Vim and how I'm feeling. It the I had a thing for NeoVim where I could see the file tree. and that was kind of cool, but I don't really want to I could kinda get bored using like trying to set all that stuff up.

Wilhelm (01:04:38.008)
But if you want to look at the tree, the file tree.

Wilhelm (01:04:45.558)
Mm.

Matt (01:04:52.431)
So if I really want to see like I mean I have my own like Git wrappers, like CLI wrappers. and also I just use the git shade, like whether it's been changed, like how many files have been changed. But if I'm honest, I very rarely do that because my coding agent is like constantly using git. I can just see in the in the thread like what how many files are changed because I see it all the time. And I'm always trying to commit in between each changes. So I use like git diff viewer.

Wilhelm (01:05:15.586)
Yeah.

Matt (01:05:20.358)
to then look at the individual commits and I can see like how many files have changed, like what's been changed and all that stuff there. So

Wilhelm (01:05:25.42)
Yeah, yeah. I probably have worse git hygiene than you and I have like hundreds of dirty files all the time.

Matt (01:05:31.258)
Yeah, I rarely have that because like most of the things I'm looking at is kinda kinda focused and when they're a bit more free form, I tend to open a PR, draft PR quite quickly, which keeps me honest, I would say. Yeah. But I wanna try something for orchestration. Yeah. I wanna try something for orchestration.

Wilhelm (01:05:44.396)
Yeah, makes sense. Okay, Matt, I have to run yep. Let us know. I'm gonna try this harder thing. It looks cool.

Matt (01:05:52.612)
Yeah, it does look cool. Anyway, it's been a pleasure, boss. We haven't we've we've talked about a lot of stuff and also not very much. I feel like this was a very gossipy session.

Wilhelm (01:06:03.118)
Big gossip sash. Gotta be done occasionally. I'll be in Europe next week so we can have more of a lap as well.

Matt (01:06:05.424)
Very gossipy session. Yeah. Well lovely to see you. You yeah, well you actually. What what days are you gonna be around?

Wilhelm (01:06:13.918)
well well we're going to this festival this week. Flying out tonight, tomorrow night. And then I'll be in London the following week. So I'll be in London all of next week.

Matt (01:06:17.142)
yeah dude you're going to London aren't you?

Matt (01:06:24.56)
Yeah, and I'm in I'm in Dorset, but I'm not coming to London 'cause I a wedding.

Wilhelm (01:06:27.82)
You should just come through, man.

Matt (01:06:30.438)
I'm not coming. I'm not coming. I've told myself I'm not coming. There was like so many things that I could have done in London if I'd have decided to go to London next week. But I rejected all of them. I'm not coming. I'm going to Dorset. I'm going to my wedding. Not my wedding. My friend's wedding. And it's gonna be great fun. shit, we forgot the jingle. Yes, let's do it.

Wilhelm (01:06:31.566)
What?

Wilhelm (01:06:41.59)
Rough. Rough.

Wilhelm (01:06:46.542)
Should play the jingle as an outro?

Wilhelm (01:07:01.036)
And goodbye. Peace everyone.

Matt (01:07:04.229)
Bye.