Hype, bubbles and AGI, Karpathy on Dwarkesh, Claude Code is a revolution and MCP fixing humanities problems

"chat to your mouth Claude and grow a new tooth"
"I think I would have like a huge ego now for sure"
"have you seen the blimp"


Takeaways

All of these people in this industry have suddenly become important at work.
Their work suddenly became important to a lot of the world.
If I had worked at Anthropic five years ago, I would have a huge ego now.
The growth in the AI industry feels unevenly distributed.
The current time feels strange due to the rapid changes in AI.
Ego can be a significant factor in how individuals perceive their roles.
The importance of work can influence personal identity.
The AI landscape is evolving in unexpected ways.
Understanding the psychological impact of industry changes is crucial.
The conversation reflects on the novelty of the current AI growth.

Creators and Guests

Host

Matt Carey

agent and mcp at Cloudflare

Host

Wilhelm Klopp

building @kolo_ai

What is You've Been a Bad Agent?

Wil and Matt discuss tech, startups, and building really cool things with AI. Sometimes joined by (actual expert) friends.

Wilhelm Klopp (00:00)
What up?

Matt Carey (00:01)
Yo, dude, I'm having a tensing.

Wilhelm Klopp (00:02)
You look totally

fine, man. You said you're looking not pretty today, but I disagree.

Matt Carey (00:07)
I need a haircut. Really badly. Yeah. Yeah, I'm a hun.

Wilhelm Klopp (00:08)
Nice jumper too, I need a haircut too. Are you at home?

I didn't know you had exposed brick at home. That's very San Francisco.

Matt Carey (00:15)
Mmm.

Yeah, it's like a converted warehouse thingy in... Yeah. You've not been to my house. You've been to my house,

Wilhelm Klopp (00:24)
And do you know when the

docks your location? It's always important to practice operational security.

Matt Carey (00:29)
Yeah, I don't know, crossed my mind just then. was like, no, I'm just in East London. No Whitechapel.

Wilhelm Klopp (00:38)
I know, I also said the address of a London development where a friend lived ⁓ in a cafe. Like, ⁓ you live in something Royal Mint Street. ⁓ And he works in security and he looked at me like, dude, you just compromised my security to this whole cafe. I think it was, ⁓ what's?

What's the London cafe that does these really fancy desserts and also other fancy things? It's like Otto Lengi. Yeah, it was like an Otto Lengi. So I don't think anyone in there is gonna break into his place or whatever, but he was not happy with me.

Matt Carey (01:09)
I was gonna say the Otolango Cafe, yeah. It's really good, that place.

also break into his place while he was at the cafe. That's some pretty good teleportation. How have you been,

Wilhelm Klopp (01:19)
True, yeah.

And there's more... Yeah,

not too bad. Yeah, pretty good. I've missed chatting to you. We've had an epic run with tons of incredible guests. I feel like it just keeps getting more fun. But I've missed chatting just us as well.

Matt Carey (01:35)
Yeah, me too. We, I still haven't published the last one. I need to publish the last guest.

Wilhelm Klopp (01:37)
and

I mean man, I feel like you can do whatever you want since you do all the work. You get to decide how it works.

Matt Carey (01:47)
Not like, I mean, at this point AI does all the work. I just press yes on Riverside and then go push. Yeah, it's pretty good.

Wilhelm Klopp (01:53)
No way. Yeah, that's great to be honest.

We need to figure out a way to get the jingle in there. actually, let me find the file. I'll make a note of that. I'll do that later.

Matt Carey (02:02)
Yeah, I have no way of doing that. Yeah.

Wilhelm Klopp (02:05)
Like it's just not a thing in Riverside.

Matt Carey (02:07)
No, I'm sure it is. I just haven't been bothered to learn how to do it yet. Yeah. I was actually looking at like whether we wanted to get an editor who could use Riverside, but like actually edit properly. know a guy, I asked him, but yeah, need to work that out. Cause that could be cool. If something that enables us to get some time back and it just happens, we have to do anything. That would be sick.

Wilhelm Klopp (02:11)
⁓ Okay, maybe I can look into it.

Mmm.

Nice. I mean...

It's the dream. It's the four hour work week. know? Get back the time. Well, I mean, we've done quite a number of episodes now, right? So I think at this point you deserve an editor. That's just the rules of the pod. The pod biz.

Matt Carey (02:31)
The dream. The fall. Jesus.

Yeah, maybe we should do it. I was looking into it before Sri Lanka and before the whole job change and everything and then it's kind of been on the back burner since then. But I'll look back into it.

Wilhelm Klopp (02:54)
Okay, good

thing you mentioned this. Yeah, so we've done 14 episodes. So that's a lot actually. I think there's like an 80-20 type rule where like most podcasts don't make it beyond three and then most of those don't make it beyond nine. And then I think the next milestone we've yet to hit. ⁓ yeah, congrats dude.

Matt Carey (03:00)
That's insane.

Is it 15 the next one or is it like, does it start getting exponentially longer? Like if you make it to 100, if you make it to 100, is it the likelihood that you'll make it to 200 then? Is that like, or yeah.

Wilhelm Klopp (03:16)
Yeah, I don't know. I could be making this up.

I probably yeah, something like

that. But also the number of podcasts has shrunk dramatically, right? Because every time it's like 80 % of those who didn't make it to three or sorry, 80 % of the podcasts who made it past three don't make it to nine. So now you're left with like 20 % of 20%. So it's like all this.

Matt Carey (03:40)
Yeah, so.

Wilhelm Klopp (03:45)
advanced maths. Anyway, how was, okay, so can we talk about, you've done some incredible things and some big things. I'm keen to chat about all of that because I don't think we really have. I think we touched briefly on your new job in the last episode, which isn't out yet. But I wanna know everything. Are we going to have like,

Matt Carey (03:48)
Ha ha ha!

Wilhelm Klopp (04:07)
breaking Cloudflare news on this podcast.

Matt Carey (04:10)
⁓ yeah, I mean, we can try. I'm probably going to get told off or something. yeah, I'm pretty... Yeah. Yeah, no, no, I don't.

Wilhelm Klopp (04:14)
You

I'm just trying to put you on the spot. Our earnings call is tomorrow. Any tips? No, I'm just joking. For the federal

regulators, that's a pure, humorous, ⁓ yeah.

Matt Carey (04:28)
⁓

Yeah, yeah, I won't be able to talk about that but ⁓ I can talk about like the stuff that I'm doing mostly which is quite cool because it's like all in the public it's all open source, which is nice ⁓

Wilhelm Klopp (04:42)
Nice, nice, Yeah,

I'm not even sure how we should start about this. Okay, but you had like a offsite recently, I think in Lisbon. And all I know about it is that you got some sick t-shirts made. what can you tell me about like what was the offsite? Was it like, did it have a specific purpose or like what did you get up to?

Matt Carey (05:03)
Yeah,

yeah, I'll talk about it. So yeah, like about a month ago, I joined Cloudflare to work on the agents team. Yeah, like my colleagues are awesome and Kate, our engineering manager, ⁓ when we went on offsite this week, it was like the first time the whole team as it stands now had got together. ⁓ We have Sunil who now lives in the North of England. We have like two people in London.

and then we have one person in Europe so it was the first time we'd all been together, the team in its current formation really and it kind of felt like a real team I think I wasn't around previously before that but I think it's been quite light on people until before then and now we've got to like critical mass which is pretty cool so yeah Kate

Wilhelm Klopp (05:47)
So did someone join at

the same time as you?

Matt Carey (05:50)
No, it was

me and then like a month before someone else joined and then a month before someone else joined and then a month before Kate joined and then pretty much Sunil was being by himself and since then when he was like the sole member of the Agents team from when he created the Agents SDK. I don't know when it was, around Christmas time.

So anyway, yeah, Kate made these awesome t-shirts that say, so we all have matching t-shirts. She didn't make one for herself though, so she has to go make one for herself, because I think she forgot. they all arrived and yeah, she didn't have one for herself. And I was like, oh, this sucks. But oh, she's got, it says Agent Zero, which is a reference to something that I can't quite remember now. I think it's a sports reference.

Wilhelm Klopp (06:16)
Hehehe. Hehehe.

What does it say on the t-shirt?

Mmm, very good.

Wait, isn't it like a patient zero reference from like World War Z or something like that? a...

Matt Carey (06:39)
Yes,

I initially thought it was, but then...

There was definitely something else. ⁓ okay. It's also a character in X-Men. I really should have looked this up beforehand. But yeah, anyway, Agent Zero, ⁓ it says on the front in really cool embroidered letters. And she got us all ⁓ custom Tony's Chocoloney, which was really sweet as well. That's so cool. I know, really amazing. And really pretty as well.

Wilhelm Klopp (07:05)
That's great. Yeah. Damn. Who doesn't love Tony Stoccoloni? Okay. So

agents team, that sounds huge. I feel like the industry has kind of wrapped its head around what an agent is, or we've all maybe come to the same understanding somewhat of what an agent is, but

Yeah, what does that mean? Agents team. It could mean like a bajillion things, right? I imagine applying agents to Cloudflare is like a long, could be a very multifaceted project, but I assume it means something more specific than that.

Matt Carey (07:37)
Yeah, I mean, I still don't really know what agents mean. I actually posted on Twitter, which I think was a post, you know, when people put like an A in the word post, ⁓ like P-O-A-S-T. It's like basically like a shitpost. ⁓ Yeah, I posted on Twitter like, I'm pretty sure that's what that means. I thought it was a misspelling of boast for a while and now I think it's a shitpost. Yeah, post. But so I probably need to Google that, but anyway.

Wilhelm Klopp (07:40)
Yeah. ⁓

⁓ Is that what that means?

Gotcha. Okay. Interesting.

Matt Carey (08:04)
Yeah, no, so I put on tour a few days ago, like, I'm still not really sure what an agent is. ⁓ And I was kind of taking the piss, but also, yeah, I still don't know. I've seen so many different definitions, even recently, people are coming out the woodwork with new definitions. Like the open code, open code team, they wrote some docs and halfway down they were like, and you define this and this is your agent. And I was like, what? So.

Wilhelm Klopp (08:10)
Mm.

Mm.

Mm.

What was their agent? Was it not something sensible?

Matt Carey (08:31)
Yeah, I think

previously people have been like, it's a loop with tool calls. And they were thinking of it less from like the implementation detail. And they were like, it's a prompt plus allowed tools, which I think makes more sense. I think that's what they were saying. I think that that makes more sense to me than it being like an implementation detail because yeah. Although I did ask someone commenting that it was, it's...

Wilhelm Klopp (08:35)
Mm-hmm.

Mm.

Interesting.

Matt Carey (08:55)
an abstraction that web devs have put in because they're silly or something, something along those lines. And yeah, which I thought was, it was very helpful, but I did at the first time I felt really like gleeful, but I did the, Grok, is this true? Who, who coined, who coined the term agents and were they a web dev? And no, it was a Belgian computer scientist in the 1980s. And she is my new hero. ⁓ she coined the term

Wilhelm Klopp (09:00)
Let's go with that definition. That sounds really helpful.

No way.

Matt Carey (09:24)
yeah she coined the term agents or like autonomous agent and artificial agent yeah so i thought that was that was pretty cool

Wilhelm Klopp (09:30)
Interesting,

This just makes me wonder, obviously agent existed pre computers, you know, we had James Bond. What does he do if not just pursue a goal with some tool calls? With a very wide variety of tool calls. Like if you think about James Bond, ⁓

Matt Carey (09:42)
Pursue a goal of agency. Yeah, what do you call a group of agents? What do you call a group of agents?

Wilhelm Klopp (09:51)
⁓ Fleet I like fleet

Matt Carey (09:54)
Okay. Yeah.

Wilhelm Klopp (09:55)
Wait, are you setting this up for a joke? Am I supposed to say, I don't know, what do you call a guy?

Matt Carey (09:59)
No, no, no, I'm not saying

self-return, I'm genuinely interested because if you have a group of agents, what do call them? What's a word that makes sense? So immediately you see that your art's a group of agents.

Wilhelm Klopp (10:07)
Yeah.

A gaggle of clothes. Do you have an answer?

Matt Carey (10:11)
I think it's a,

well, the working answer at the moment is an agency of agents.

Wilhelm Klopp (10:17)
I mean, that just sounds like you're trying to confuse people. An agency of reading.

Matt Carey (10:20)
Yeah, maybe. Maybe we just call it collective.

Yeah, I don't know. don't know. Anyway, agents team, I'll tell you a bit what we does. Okay. So we work on the agents SDK. which contains all of the MCP stuff on Cloudflare. And we work on sandboxes. And that's like our very opinionated wrapper around containers. So it's all very public.

Wilhelm Klopp (10:26)
Okay, yep.

Nice, nice, nice.

Matt Carey (10:44)
It's all like the highest level of abstraction away from the bare metals of the Cloudflare Dev Platform. ⁓ Yeah, it's really good fun.

Wilhelm Klopp (10:52)
the agent SDK in brief lets you build agents on Cloudflare.

Matt Carey (10:56)
Yeah, it lets you build stateful applications with ⁓ user and human interaction ⁓ distributed across the whole of the Cloudflare edge network ⁓ on the developer platform. Like the best of durable objects, the best of workers. ⁓ Yeah, it's pretty cool. And there's a lot of changes happening soon. Like I think we're almost getting to the point where we know what it should look like.

Wilhelm Klopp (11:12)
Mm-hmm.

Yeah.

Matt Carey (11:23)
and we're about to make the next step of implementing that. So it's going to be exciting.

Wilhelm Klopp (11:28)
Nice, very cool. I feel like I saw a really cool tweet from Naresh it was going a little bit or calling back to like the iconic iPhone keynote where Steve Jobs was like, it's an iPod, it's a phone, it's an internet communication device. Are you getting it? It's one, it's one device.

Matt Carey (11:43)
Yeah, it's a web browser. Yeah.

It's an iPod. It's a phone. It's a web browser. It's one thing. And I was, yeah, I commented underneath him actually being like, it's an agent. It's a sandbox. It's MCP. And that's our agent of STK.

Wilhelm Klopp (12:02)
Wait, I'm just trying to find

his tweet. What did he actually say? But it's... Just bring it up. So he's on your team as well, right?

Matt Carey (12:06)
He said something similar.

Yeah, yeah, he works on sandboxes. So it's like, he put agent sandbox and then I can't remember his last one.

Wilhelm Klopp (12:14)
file system? No.

Matt Carey (12:16)
Yes,

file system, was, it was file system. Yeah, an agent of sandbox at file system. Yeah, I didn't understand the file system one, but I'm sure he'll have a very, I'm sure he'll have a very good explanation for that.

Wilhelm Klopp (12:19)
yeah, nice.

I

see. Yeah, I guess you can have a sandbox without a file system or whatever. But I guess the file system is kind of like a key part of Claude, right? Once you have a file system, you can search through stuff that's on disk.

Matt Carey (12:37)
File system

I have another colleague, like another member of our team who made this mad agent ⁓ based on a file system that was like sort of, it was based on R2. So like the CloudFlare S3 API, the object storage. And he like emulated a file system with just like a bunch of CLI commands. And it meant that you could, he exposed this terminal in a UI. ⁓

Wilhelm Klopp (12:57)
Interesting.

Matt Carey (13:03)
that was like hosted and you could just like write like you could write to files, read from files and then specific files had specific properties of the agent. So one of them was like an email allow list file. ⁓ And it said like what emails he liked he would accept in his inbox. And this agent was like basically monitoring his inbox. And one of them, one in his disallow list was like any recruits, any recruiters asking about asking him to join his crypto project.

Wilhelm Klopp (13:16)
No.

Matt Carey (13:31)
And it was like, tell them to fuck off. And then he showed us, and these are all of the logs of this agent telling all these recruiters to fuck off. And I was like, that is sick. And he was like, yeah, it's super easy to change because you can just write to the file through a UI. And he's basically just writing prompts. And I think...

Wilhelm Klopp (13:38)
wow.

That's fascinating.

Right, yeah. So you can update

the R2 files like without having to shell into the...

Matt Carey (13:53)
Yeah, I think that's like the gold standard. Yeah, and the agent just

reads from R2 every time it wakes up. ⁓ And yeah, like you don't have to redeploy. You can change the agent behavior without redeploying, which like, I think that's clean. It's really neat.

Wilhelm Klopp (14:08)
That is cool, Yeah, interesting. The way you actually get a file system on Cloudflare right now, right, is that only through containers?

Matt Carey (14:16)
Yeah, so as sandboxes you can get a file system, I think.

Wilhelm Klopp (14:20)
And then wait, what is a

Matt Carey (14:20)
quote me on that. can have a

a sandbox is a wraparound, the containers product. yeah, so the sandbox, essentially they're just like, we want to use containers first, particular things. And those particular things are highly opinionated. And because they're highly opinionated, basically everyone using the dev platform to do agenticky stuff on containers is going to have to do the same thing.

Wilhelm Klopp (14:26)
⁓

Matt Carey (14:47)
And so on our team, they were just like, right, we just make a sandbox SDK because the container platform is super well designed for it. It just needs a bit of fiddling around the edges to make it work well ⁓ and give a good dev experience. And that's what the sandbox SDK does. It's super clean. ⁓ Craig, one of the dev advocates, he posted a video about running Claude Code in a Cloudflare sandbox. And yeah, I think we're in the Claude Code docs for running it.

running coding agents in the sandbox is pretty, it's pretty cool. And yeah, it won't just be coding agents at some point. I think it'll be everything will run in these little sandboxes.

Wilhelm Klopp (15:16)
no way.

Yeah, yeah, I guess

Claude's doing a big push that like it's, think that they renamed a bunch of stuff, right? Kind of starting to plot a future where Claude code is not just a coding agent, but a generic agent. And there's a bunch of like marketing collateral ⁓ you start seeing around like, Ooh, actually it's not just coding agents that need bash. It's like every kind of agent needs or will benefit from bash. So it seems like they're setting up like that kind of future.

Matt Carey (15:50)
think it makes sense from their point of view because they're like, it's very hard for us to train a model that's very good at all of these weird things that people might want to do with it. It's very good at interacting with HR systems. It's very good with interacting with, ⁓ I don't know, all the stuff we were trying to at StackOne, right? It's like, it's very hard to train a model to do that if they're not like, if that's not the only thing they're doing. And imagine you might need to...

Wilhelm Klopp (16:00)
Mm-hmm.

Mm-hmm.

Matt Carey (16:16)
Like the solution might be you actually train a model for each individual different type of interface. So like one for an HR system, one for an ATS system, one for Google docs, one for, ⁓ like making PowerPoints. You might have like a, like a model that's like super fine tuned for those different things. But for Anthropic, that's kind of a tough bet, right? Because their bet actually is if we go bigger and we go make our models better at everything that like emergent behavior will come.

Wilhelm Klopp (16:25)
Right.

Matt Carey (16:43)
And so it's half branding, but half just like like standardized so we can generate training data on standardized interface rather than doing what some other people were doing recently and everyone was generating, creating RL environments, like reinforcement learning environments to train on specific use cases. ⁓ So it's...

Wilhelm Klopp (16:43)
Mm-hmm.

Matt Carey (17:00)
It's we'll see if time plays like that.

Wilhelm Klopp (17:01)
Yeah, yeah, yeah. And it just does seem like it also

works, right? I guess Claude code showed us that that model fundamentally works in some way. I'll tell you actually a fun little thing. ⁓ I had quite a, in general, this is actually kind of a very GitHub reminiscent week for me. I bumped into an old colleague in a coffee shop.

who'd actually read my blog post about San Francisco. And we just kind of recognized each other and started chatting about stuff. It's also, I kind of, ⁓ I'm slightly jealous of you, of all of the like corporate ⁓ gossip and rumour mill stuff you have, I guess a new rumour mill you have access to now. Because we were just chatting about like old gossip and it's just kind of fun. Like there's just clearly a human... ⁓

enjoyment portion of gossiping or whatever. ⁓ so we had a good chat then yesterday it was actually the sort of GitHub universe which is on at the moment and I went to this event last night actually an MCP event at the GitHub office because I hadn't been back to the GitHub office since the start of this decade and it's changed in like wild ways. ⁓ Wait, where was I going with this? I was setting us up for something here. ⁓ yeah, I actually used Claude Code yesterday. ⁓

⁓ for like I don't use it that often to build like completely random one-off apps, but ⁓ I wanted to like, ⁓ like in the GitHub Slack, we had so many incredible, incredible emojis, like custom emojis. ⁓ Like people curated that shit over like, I guess more than a decade now.

Matt Carey (18:38)
Mm.

Wilhelm Klopp (18:42)
There is a ⁓ like ex-GitHubber community where someone I think made available a ⁓ majority of these ⁓ custom emojis. And it's crazy. It's like a gigabyte zip file. And so I think the custom emojis are like... ⁓

sorry, the standard emoji set is like a thousand five hundred or something. I have maybe two thousand, like the ones that ship by default and like every phone. Right. This get up custom one has 45,000 emojis in it. So it's huge. But then it's like, how do you look through all these little images? Like, it's not a very pleasant experience looking through it, like on the file system. So I asked Claude to just build like a little one off app ⁓ to like, let me search through them, look at them all. And it built it all like in react with like virtualization. It's like.

Matt Carey (19:16)
Nice.

Wilhelm Klopp (19:35)
reasonably fast, the search works. ⁓ And it's like, wow, this is just a thing we have now. I'm still amazed. So I guess the beautiful thing about what Claude Code has taught us is that this tool calling model just really works. Or really works for ⁓ if you, guess, I don't know how they trained that model. I assume they gave it some access. Or as in gave it some sense of that tool calling is a useful thing.

Matt Carey (19:56)
Yeah.

I will be interested when someone brings out ⁓ like a good book on what, or even just a really good blog post of what made, what was the standout moment where, or the standout feature or the standout thing that made Claude code such a, a use like such, such a breakout product. Because I think it was like before Claude code, I do think of coding as before and after Claude code also, like before Claude code.

Wilhelm Klopp (20:11)
Mm.

Wow.

Matt Carey (20:28)
you had

like the cursor agent was quite good, was decent, but it was, it wasn't, it was, it was hamstrung inside of cursor. Like it's just, just inside of cursor. You couldn't take it out and put it anywhere. And like, I don't know how often you find yourself doing this, but I find that I start Claude code in my like GitHub repos folder on my laptop. And because I have like multiple different repos and doing different stuff.

I might like look up from some docs in one repo, search the web and then they helped me with an implementation in another repo. And that just like fully wasn't possible before in cursor, or maybe it was, but you had to open it in the, you had to open just like the route and then it all got pretty messy. I don't know. I think more people just like that you, like small UX change, it allowed more people to do this sort of thing naturally. And like to have the embedded

Wilhelm Klopp (21:09)
Hmm. Right.

Matt Carey (21:20)
I don't know. I'm really interested when someone smarter than me comes out with like, like all the reasons why I called code was awesome and having it in a terminal and having it embeddable and having it like, like just being a, a bun process or just that you can run anywhere, like on any Unix machine. Yeah. I mean, maybe those are all the reasons, but it seemed mad that we, no one ever thought of it before.

Wilhelm Klopp (21:26)
Mm-hmm.

Totally.

Totally. ⁓

⁓ Just briefly back on the agent SDK, I feel like I should build something with it. Can you give me a challenge of something to try and build with it?

Matt Carey (21:54)
Yes. ⁓ you love Slack, right? Simple Poll loves Slack. You love Slack. You should build something in Slack, like a Slack agent that can...

Wilhelm Klopp (21:59)
Indeed. Yep.

Hmm.

Matt Carey (22:05)
Do something funny. I always really like the ones where people connect them to uncensored models. Maybe, okay, so we're starting this repo called Awesome Agents. And we're gonna put like very, like useful examples in that repo. So I think we've already had a WhatsApp agent contributed from my little brother actually. He contributed, it's pretty cool. Just like.

Wilhelm Klopp (22:27)
No way.

Matt Carey (22:27)
here is the agent's SDK working, or it's just a durable object actually, but here's a durable object working ⁓ to like hold a stateful conversation across multiple different messages. ⁓ And it's just like, yeah, here's a web, you host a web hook with a worker. Like that's so slept on as a use case. And I think I posted that on Twitter and got like the most amount of bookmarks I've had in ages. So I think that repo is gonna be really popular.

Wilhelm Klopp (22:52)
haha

Matt Carey (22:54)
for people wanting to get started because we've always had our examples but I think it's different seeing an examples and then seeing something that actually works like now you can go and just deploy it and use it and I think you got a more ideas it's called Cloudflare forward slash awesome agents

Wilhelm Klopp (23:02)
Totally.

What's the repo name? I'm just trying to Google for it.

Okay cool, it's not showing up on Google yet. But I found it. nice, yeah cool.

Matt Carey (23:14)
I it's pretty empty right now. It's completely empty.

You have to look at the PR. We haven't shipped it yet. shit. There's actually... ⁓ wow. I think we made it yesterday. So yeah, cool. So...

Wilhelm Klopp (23:21)
Nice. It has 28 stars, but nothing. Yeah.

Nice, nice, nice, nice. That's exciting. Okay, that's

a good prompt for me.

Matt Carey (23:34)
Anything you

build, ⁓ yeah, stick it in there or like, like keep an eye on that and then find something interesting that comes out of there.

Wilhelm Klopp (23:42)
So

any agents are good for that or agents built on Cloudflare?

Matt Carey (23:46)
So agents built on top of Cloudflare. They don't have to specifically use the agent's SDK, but over time will probably migrate everything to like an opinionated way of doing stuff. ⁓ Yeah, so I really didn't answer your question on what you should build. The thing that I actually wanted to build was like...

Wilhelm Klopp (23:57)
exciting.

Why you said

hook up ⁓ Slack to some kind of ⁓ sexy open source model, think is what you said. I don't know. What's your favorite sexy open source model?

Matt Carey (24:05)
Yeah, hookups like but.

yeah, don't, yeah, that, maybe that's actually not the best thing to shout about. No, the thing, one of things that's in my backlog of free ideas,

I'm only saying this because my brother did this on WhatsApp and it's really funny. You can just drop it in groups and it can just abuse people. Like it's actually just funny. You know how, have you tried poke by interaction? I think I was shouting about it on the, on the one we did with Luke.

Wilhelm Klopp (24:24)
⁓ nice.

⁓ sorry. Yeah, yeah, yeah. It

was abusing me a little bit. I had some FOMO for sure, but I haven't come back to it yet.

Matt Carey (24:42)
Did you get past the point where it abuses you to the point where it like sucks up to you to the point where it like gives you a price that's outrageous?

Wilhelm Klopp (24:47)
No.

⁓ I think I tried to be hard to get in my messages with it and it just shut me down immediately.

Matt Carey (24:56)
Yeah, it does start replying if you're too hard to get.

Wilhelm Klopp (24:58)
Yeah, it said, that's rich. The human wants me to prove my worth. Hate to break it to you, Will, but that's not how this works. I'm the bouncer here, not you. And that's it. That's the end of the conversation.

Matt Carey (25:07)
Ha

Okay, their onboarding flow is wild. I don't know how they got Claude, because I'm pretty sure it is Claude to be mean. I don't know how they did it. Maybe it's not Claude, maybe it's something like some fine tune or some uncensored fine tune, because it can be quite mean. I don't know. People share their screenshots of cursor saying like, fuck you, fuck you, I got this wrong, fuck you. And I haven't managed to get cursor or any of the models to do that.

Wilhelm Klopp (25:21)
Mmm.

Ha!

Any other first impressions of CloudFlare or anything else you want to share on that whole? mean, it's kind of a big new chapter, right? Exciting new chapter.

Matt Carey (25:49)
It's wild. ⁓ Cloudflare's awesome. It's a company that I've admired for a really long time from the outside. And when you get inside and you can read some of the internal blog posts, there's a big writing culture, actually, which is cool. Although apparently I have some words to write for Friday on our new agents blog. Yeah, that'd be fun. yeah, there's all these internal blog posts about people that just...

Wilhelm Klopp (26:01)
Mm.

exciting.

Matt Carey (26:13)
find something interesting. And so they write about like what we could do in that space. And they are, some of them are very unhinged and like really out there and they're an amazing read. Some of them are like, my God, this is such a idea. And some of them you're like, wow, this is too smart. This is too big brain. I have no idea what's going on here. But then if you read back a little bit, can find somewhere, can spot the whole arc where someone proposed an idea, then a little bit afterwards they did it.

Wilhelm Klopp (26:21)
Hehehe.

Hmm.

Matt Carey (26:41)
and it's been a mental success. Or there's one example I'm thinking of is someone thought of an idea, they wrote it at length about it and they basically forecasted the growth of another company that Cloudflare waited like a year and a half and then bought. And now the people who were working for that other company, they're working at Cloudflare now. yeah, some of the internal blog posts were really, cool. I think that writing culture has been...

Wilhelm Klopp (26:58)
Wow.

That's incredible.

Matt Carey (27:09)
one of the standout things.

Wilhelm Klopp (27:11)
I also am just looking at the homepage of the Cloudflare blog for the first time and I can't really believe my eyes because there's like five blog posts published today or seven.

Matt Carey (27:20)
A what? Seriously?

Wilhelm Klopp (27:20)
Like, there's

clearly... I remember...

Matt Carey (27:24)
Okay, so one of the ones that was published

the last, I don't if it's today or yesterday, but it's the tracing blog post. ⁓ So workers tracing. So this is actually insane. So this is the baseline team that got acquired ⁓ So Boris and Thomas and Max, they built this and they built all the observability inside workers. So you don't really have to instrument anything. It just, you had all of your logs.

Wilhelm Klopp (27:32)
yeah, I see it.

Mm-hmm.

Matt Carey (27:53)
But now you actually have traces. So if you just do observability true inside your Wrangler or you do it in the dashboard or whatever, when you deploy your worker, you can see every single function call that's being made like inside the worker. Yeah, it's so fucking cool. ⁓ Yeah, like it's just insane. And you can see all of the like the timings of how long stuff was and yeah, man, that's instrumented out the box in workers.

Wilhelm Klopp (27:55)
Mm.

No, that is so cool.

That's incredible.

Mm-hmm.

Yeah,

that is wild. That is incredible. Because I remember workers used to not have the reputation of being the most observable like some five years ago or something. ⁓ Yeah, that's a huge win. That's really cool.

Matt Carey (28:30)
Yeah. And they're also that like Boris is a big hotel believer, like open telemetry. And so they're actually just emitting hotel like compatible, like traces and spans and stuff and logs. So you can actually like plug it into a third party provider and send it to like Datadog or whatever if you wanted to.

Wilhelm Klopp (28:35)
Mmm. ⁓

⁓ interesting.

That's cool. ⁓ I should play around with that. He's coming to SF, I think, in like a few weeks or something. Should catch up with him.

Matt Carey (28:55)
Is he? Oh.

Yeah, you should convince him to move there.

Wilhelm Klopp (28:58)
Yeah? ⁓ Him and not you? We're working on you.

Matt Carey (29:01)
You should.

I'm not

coming anytime soon. Yeah, I was going to come in November, that whole thing just demo days is still happening, but it's just taken me being out of the loop for, I don't know, a month or something in between jobs. And then it's just, it's just knocked everything back. And, and I didn't really have any solid plans before. And so, yeah, it's going to be pretty tough to swing big travels around.

Wilhelm Klopp (29:10)
Sad, sad, ⁓

Totally, yeah.

So there'll be an AI demo days in November in London. Nice.

Matt Carey (29:36)
There'll be one in November in London, and

then there'll be one like definitely in January in London, but then at some point I wanna go back to the US and do one. I think it'd be really cool. And yeah, there's some good appetite for it. So might as well give the people what they want, you know?

Wilhelm Klopp (29:55)
That'd be cool. That'd be cool. ⁓ Okay. I want to talk for a second about just like the craziness of the hype at the moment. So I went to an event recently and it's just like, doesn't feel uncommon at the moment to see someone reasonably high up at a well-respected company to...

speak for 30 minutes about things that if you just know a little bit about like how something works, just makes zero sense. And it's unclear like what problem they're solving or like it's, ⁓ or let me back up. I guess it's.

Matt Carey (30:28)
You

Gone. You can't just say that in the abstract. You have to actually give the example.

Wilhelm Klopp (30:35)
No, I don't want to give the example. ⁓ But I think it follows this pattern of

Matt Carey (30:43)
I need some more details. You can't say, oh, they just spoke for 30 minutes and everything they said was bollocks. About what did they speak about? Like, it Accenture saying like, we're going to automate the whole of the world workforce by like 2028? Because I don't know.

Wilhelm Klopp (30:44)
Hehehehe... Hehehe... Hehehe... Hehehe...

No, no.

No, no,

it was a more reputable company than, than Accenture or like one where you would credit them with more of an idea to know what they're talking about. But like, I think what people are doing is like, there's a lot of like, ⁓ the world is about to change. It's going to be like so crazy. And like, soon we can use MCP to solve this problem. And you're like, wait, that's not how MCP works at all. You don't need MCP for that. MCP is not the unlock for that specific problem at all.

Matt Carey (31:05)
That's a, okay.

Ha ha ha!

Wilhelm Klopp (31:27)
But it's like you need to keep...

Matt Carey (31:27)
It's the unlock for all of the world's problems. We're gonna cure all diseases. We're gonna

bring everyone out of poverty, starvation, all with MCP. That, I promise.

Wilhelm Klopp (31:39)
I mean, I

do actually believe some weaker version of that, but I just feel like I've not seen the industry. ⁓

Matt Carey (31:49)
f**k's sake, Will. You're in the right place, Sam

Fran, you go to a coffee and Sam Fran, hear the next door being like, oh, so how are you planning for AGI? How are you prepping for AGI?

Wilhelm Klopp (31:57)
⁓

I love the impression, that's great. ⁓ But yeah, I just feel like there's a weird sense where people aren't too honest with what's actually possible today. And everyone's projecting what will be possible soon, but the projections are just bit strange. don't know, like obviously there's a bubble. I think bubbles aren't necessarily all bad.

I think they're maybe necessary or whatever, but I think it's just a It's a weird it's a weird place to build a product as well at the moment I think if you're a startup because it's like do you build something that works today? Like do you ship a working product that drives value to your customers or do you build some strange derivative of a harness that like shoves more shit into an LLM and then maybe the LLM will get smarter in the future and then you have a working product, but that's not a way to build products like

Matt Carey (32:39)
Yeah, dude.

Wilhelm Klopp (32:51)
That's just not fair.

Matt Carey (32:52)
Yeah,

I have thoughts. Dude, it's really interesting you're saying this now because I thought six months ago, this was like the craziest it's got. ⁓ And I was having these thoughts all the time about whether you build the product that works now, whether you build the product that works in the future. Luckily, I don't really have to have these thoughts anymore, which is nice. But the startup thing, it is crazy. And think Sunil put it really well about he was trying to build a startup with like,

Wilhelm Klopp (33:01)
Mm-hmm.

Mm-hmm.

Matt Carey (33:20)
the whole fabric of computing shifting under your feet. And I do think about that. But the thing about ⁓ the weird zeitgeist at the moment is I actually think it's going the other way. Did you listen to Kapathi on Dwarkesh?

Wilhelm Klopp (33:24)
Mm, right.

Yeah,

I was just going to say like, think that Andrea Caparti is like the big industry antidote to the hype, which we're we should be so grateful to him for this. Yeah. So go on, because I was going to bring the him up as well as kind of like, OK, this is.

Matt Carey (33:50)
Well, just for anyone who hasn't listened to it, and I haven't

like, want a caveat, I haven't finished it yet, but he basically gives a lot of a, he gives a lot longer time projection for something that might be termed as like AGI or something like that. And he's just very grounded in like what we can do now, what's needed to do in the future. And that these are like massive problems that even throwing like...

Wilhelm Klopp (34:11)
Mm-hmm.

Matt Carey (34:16)
I don't know, 30 billion, 100 billion at the problem is not going to make it any faster. And yeah.

Wilhelm Klopp (34:22)
Yeah, he's saying a lot of good,

like it's not the year of agents. It's the decade of agents. I think that's really like nice to hear, comforting to hear, right? He's saying, yeah, stuff like self-driving, like it's not solved yet. Like, you know, we need to figure out more stuff. Even with Waymo's, like there's teleoperation often still involved, right? And that's not a bad thing, but that's just like reality right now. And obviously we want to get to a place where like that's not necessary anymore. Or he's saying like in his own work, like he built this,

⁓ to this, I guess, learning project ⁓ called NanoChat. So like a very minimal version of like ChatGPT showing you like how to train it from a data set all the way to having like chat in the browser. ⁓ And he said the code for this, like, I think there are some stuff that like agenti tools like Claude Code were useful for, but most of what he found useful actually was ⁓ like Cursor is like super complete or whatever it's called, like the tab complete, which I think in Cursor is just like really, really good.

And he was in the driver's seat for most of it, which I think is just like a bit of a reset from like, yeah, of course I have my fleet of 10 clothes working in the terminals and they're like, I don't ever go for a lunch break without sending my agent off on three tasks that I can review when they're back, right? I think there's just like a bit of a disconnect there. So I think he is doing an incredible job at just because he doesn't have like, he doesn't work at a big AI lab anymore. He doesn't have like a particular narrative to.

to pedal.

Matt Carey (35:49)
Yeah, no, I-

Wilhelm Klopp (35:50)
Well, another

thing I've noticed is, sorry, go on. I feel like ⁓ when you work at an AI lab at the moment as well, it is just absolutely wild. Like I think they are seeing like unprecedented growth, wild use cases. You couldn't even imagine hyperscale problems that only come if you have like a billion monthly active users or however many chat GBT has. But I think it makes the... ⁓

Matt Carey (35:53)
No, no, no, go for it, go for go for it.

Wilhelm Klopp (36:15)
It makes the day-to-day experience ⁓ of someone who doesn't work at one of these companies very different from if you work there. And I think there is a bit of a, ⁓ not a rift, but just like you live in different worlds. I think the party is clearly happening at OpenAI, at Anthropic. And if you don't feel the party happening day-to-day where you are, I think you just kind of live in slightly different realities.

Matt Carey (36:29)
it's... yeah. Definitely a disconnect.

Wilhelm Klopp (36:43)
And it's different from what I've seen before. I think the SaaS boom that we had in, I don't know, 2014, 15, 16, 17, et cetera, felt just a little bit different. I think if you met someone at Stripe and you were like, oh, I tried to use your thing, but there was a bug there. And they'd be like, oh, cool, let me try and fix it. But you talk to someone at OpenAI now, and you're like, oh, yeah, there's actually this bug I encountered. I don't think they're.

really care that much? Like, that's just too much growth, too much stuff. The conversations internally are like, how can we procure a bigger data center that has nuclear capabilities to power all this growth? It's not like, there was a bug in this thing. That's not interesting. I don't know, maybe I'm ranting a little bit here, but ⁓ I think there is just like, yeah, it's an interesting time right now.

Matt Carey (37:34)
Yeah, no, I do think you're right about people are very, like not grounded at all. And it's refreshing to hear someone like Karparthy be like, he comes from a mega research background. He's done a lot of really cool stuff. He worked at Tesla, worked at OpenAI and he's like, ⁓ I know the state where we currently are. All these things need solving. Come on guys, let's buckle up pretty much. ⁓ Rather than like...

Remember when we did the podcast with Sunil and it was like around the time that these meta researchers were getting like billion dollar offers or something ridiculous. ⁓ 100 million yeah dollar offers. And he was like guys is it like last is it like last chopper out of NAMM like it is would they be leaving if GPT-5 was this good or are they leaving because they need like

Wilhelm Klopp (38:20)
Mm-hmm.

Matt Carey (38:26)
private island money because they're terrified of what's going on. And I don't know, man. I think they're all on an ego trip. I think like all of these people in this industry, they've just suddenly become important at work. Well, their work suddenly become important to a lot of the world more and it's gone to everyone's head. Definitely.

Wilhelm Klopp (38:38)
Hahaha

Yeah,

I mean, I think like, I wouldn't, ⁓ like, I think if I wasn't the situation, I think I would be very similarly affected. Like, I think if I happen to have lived a different life where somehow I was working at like Anthropic like five years ago, I think I would have like a huge ego now for sure. But I think what's just strange about the current time is that like, it feels like...

the growth or whatever in all of this AI industry is so like ⁓ unevenly distributed in a way that feels kind of novel to me. And I don't really know why or if it's gonna change or whatever, but like in the we work that ⁓ I work in, there's tons and tons of startups, right? And so many conversations I have with founders, like they don't really know what...

to build and like the clearest goal that I am hearing articulated is, I want to be aqua hired by open AI on topic, which is so like, that's not usually why we get into startups. Yeah. So it's like,

Matt Carey (39:39)
That's brutal. Yeah, that's brutal. If that's your goal, I think

at that point you should probably like cut off and quit because like you don't have the bit between your teeth there, do you?

Wilhelm Klopp (39:46)
But it's interesting. I think

it's indicative of how much stuff is shifting, how it's not really too clear what to build or how to participate. If we're in the midst of this incredible AI revolution, why?

Can't you or I spend a weekend building some use case for some, or sorry, building some product for some niche, badly in demand use case, and it makes a million dollars by the end of the year? Maybe we could do that if we're lucky and have, ⁓ you know, do the right thing. Okay, you think we could?

Matt Carey (40:19)
We could. No, no, but we could. But we could.

I think we could. I think we could. think if you wanted to become a startup founder and you wanted to put that on yourself, I think you could. It might not be by the end of the year. By the end of the year, you might have some sense of what you want to build, but that's always been the hard thing, right? Knowing what to actually build, like how are you going to...

Wilhelm Klopp (40:25)
Okay, I-

Mm.

Matt Carey (40:43)
Satisfy customers how you're gonna actually solve a problem like all of that stuff is the hard stuff we've just made the building bit a lot easier and Yeah

Wilhelm Klopp (40:53)
Sure, sure, sure.

I think, no, I mean, I think I agree with you. I think it's just like the opportunity space seems strange at the moment. ⁓

Matt Carey (41:05)
The thing that people

are scared of is like, build a feature, I build a whole product that OpenAI makes a feature, right? Like that is, that's the classic thing. But if you get away from dev tools, that becomes like only tangential. Like for it to be a feature of Claude code or to be a feature of Chai GPT and what you're doing is, I don't know.

Wilhelm Klopp (41:14)
Yup.

Yep.

Matt Carey (41:27)
selling bulldozers in Germany, it's it's a very, that's very hard for them to cross that boundary. But if you're making, I don't know, there was a big craze a few months ago about making reinforcement learning environments for everything. If you're building a reinforcement learning environment that at some point, OpenAI could just one shot the code for, then like,

you're probably not needed and you're not building for a lasting problem. I don't know. I've been listening to a lot of the product market fit show and I think it's really good. but maybe it's maybe pilled in the other direction because like they're all pretty much all success stories. ⁓ And they're like about how people, that people like fell into a problem basically. Like there's a really good one with the guy that made Mercury. And he was like, yeah, I was like thinking about this for a while. I went round, I like,

Wilhelm Klopp (41:50)
I think this is kind of... okay,

I see.

Yeah, yeah.

Matt Carey (42:13)
I did some customer research basically and by the time I raised money, I knew that this was gonna be a problem and as soon as we built the first iteration of the product, as soon as we were feature complete, we had product market fit. I mean, that's always the dream, right? But I don't know whether people are rewriting history or what they're doing.

Wilhelm Klopp (42:27)
Totally.

Yeah, I mean,

I think like I would, ⁓ like I wouldn't be ⁓ sharing this complaint if like Mercury was like, I guess some kind of like AI powered or cloud code powered thing that is like a good B2B business, right? Like it's a bank, right? That's the same Mercury we're talking about. Yeah. So I feel like if you could, if you had like,

Matt Carey (42:48)
Yeah. Yeah, yeah, yeah.

Wilhelm Klopp (42:52)
a couple, like I say, a dozen success stories of people who've built like a Claude Code powered agent that fits like some kind of B2B need. I just think that's the sort of stuff that I'm surprised or you just not seeing it yet. Hope like I imagine we will be. I asked like a friend about this. ⁓

Matt Carey (43:09)
Yeah.

Wilhelm Klopp (43:12)
I ask everyone, where do you think the party is happening at the moment? Because it's happening at the labs, the party is happening at 11 labs, the party is happening at Cursor and Lovable, et cetera. ⁓ And people have interesting different answers about it. ⁓ One friend told me about this website called leanaileaderboard.com. I'm just going to share my screen actually for you here. But anyone can go, it's literally leanaileaderboard.

⁓ And what this basically just shows is like how much revenue do you have and how many employees do you have and then what's the ratio?

Matt Carey (43:45)
⁓ yeah, shit. This is the one where like,

this is the one where like OnlyFans are like the highest, but they're not, they're not on this one actually.

Wilhelm Klopp (43:53)
Oh yeah, they're not here. Maybe they got moderated out. So think the most interesting, or well, there's a bunch of telegrams on top here with $33 million of revenue per employee, then mid journey with $12 million of revenue per employee. Exactly. So mid journey is interesting. Yeah, the cursor team.

Matt Carey (44:06)
Yeah, but these are like famously small teams, right? And like the cursor team is not 20 anymore.

Wilhelm Klopp (44:13)
Right, the cursor is up there, that makes sense. think that's where the party's at. Bass 44.

Matt Carey (44:17)
Base 44, I've

seen some adverts for Base 44, but I've never used them. Whoa, okay, they say it's one person. Yeah, so I think the number of employees here is probably out of date. But the Cal AI one, okay, so this is interesting. I wanna talk about Cal AI and Merkle underneath. So Cal AI was, yeah, isn't it like an 18 year old kid who made a calorie tracker app and basically just beat my fitness pal just with a slightly different UX?

Wilhelm Klopp (44:21)
It looks a little bit...

Yeah, yeah, That's interesting, exactly.

Yeah, yeah. Have you heard of Cal.ai before? I have not until seeing us. Okay.

⁓ it is a- yeah, I don't know the backstory, it seems like that could very well be true. It is a, ⁓ yeah.

Matt Carey (44:48)
Yeah, I'm pretty sure it is. I'm pretty

sure they used, he used vision models to like track what was going on and just made a better UX on what is solving essentially the problem of MyFitnessPal. So you take a picture of your food instead of like writing everything in and like completely just destroyed the market. And I'm pretty sure like if I remember the story, it's like an 18 year old kid who built it and it's insane. ⁓ Super, super cool. Yeah.

Wilhelm Klopp (44:52)
Mm-hmm.

Mm-hmm.

Matt Carey (45:15)
actually relatively similar to what's underneath as well, but we can talk about this like a while.

Wilhelm Klopp (45:16)
So I think.

Well, so the interesting thing, think with Cal AI or with a couple of these here. So I asked ⁓ one friend, like, where do think the party is happening? He goes, well, the car party is clearly happening in code generation. And every company that's like growing a lot is like, ⁓ like a lovable or whatever help generating code. But you're selling to people who can then not build businesses. Like there's no big businesses build, build unlovable or whatever, which I think is a bit of a slightly cynical take, but it is interesting, right? Like clearly there's a lot of like code.

companies in this list here. ⁓ But then the other interesting take with CalAI is like, you actually have these consumer first companies, right, taking off. CalAI, tracking calories is very much like a consumer thing. It's not like, ⁓ it's not what I was describing earlier, like a business, a ⁓ cloud code powered B2B tool or whatever. That's not... ⁓

Matt Carey (45:47)
Yeah.

Wilhelm Klopp (46:08)
We haven't seen those make it super big or whatever. So I think that's an interesting one. people have said this about the AI revolution, how lot of technology shifts. They start at the government and enterprises and then diffuse to consumers. Whereas ⁓ with LLMs, it's kind of the opposite. You start replacing all these consumer use cases, and then because the enterprise or whatever needs more guardrails, it kind of goes the other way.

Matt Carey (46:11)
Yeah.

Yeah, I

think if you were looking at like, it's hard because you have to be very unrational to look at it and think, oh, Cal AI is going to be a great business. Like if you're starting from the outset two or three years ago and you were like, I'm going to, I see GPT-4 got announced. I'm going to spend six months building an app that allows people to take pictures of their food. You'd have been like, oh, this is a demo, right?

Wilhelm Klopp (46:39)
Yeah.

Matt Carey (46:54)
Like you'd never been like, this is a serious commercial product. So I think there's really easy to be cynical, but actually they took a whole new interface on what is a problem that a lot of people have and they made it really useful. The one I'm really interested on is below, is Mercor,

Wilhelm Klopp (46:56)
Mm.

Mm-hmm.

Mm-hmm.

Matt Carey (47:13)
There's so many recruiters, man. There's so many, like there's online recruiters, there's consultancy recruiters, there's all of these types of people. And yet Merkle managed to become a billion dollar company and they're like really young, the founders, like what happened there? And I guess they just, they found a niche, right, that was overlooked and they managed to use, okay, so to the best of my ability, Merkle.

Wilhelm Klopp (47:32)
So explain briefly what Merkle does.

Matt Carey (47:40)
is like an AI powered recruitment platform. Not really AI powered, but like they use a bunch of ⁓ like fancy quizzes to test people's knowledge and particular use cases. Basically trying to see if they're better than GPT-4 or anything, or GPT-3.5 when it came out. Like, ⁓ are you better at this than anything? And then they categorize you. And then when one of the labs wants an expert on a certain topic, they just hire them out, day rate. And so...

Wilhelm Klopp (47:55)
Mmm.

Matt Carey (48:07)
they've managed to scale a consultancy model essentially, which I might say someone probably going to DM me after this and say I got the I got it completely wrong. But this is the way I think of it. And I think it's a really cool business model where you're like, oh, AI labs need data. Oh, they're just going to get it from the internet. No way. Maybe they can't get it from the internet. Maybe they need like 11. Like, it's also really funny because it ties into the fact that like LM is a judge and evals.

Wilhelm Klopp (48:07)
I see.

yep, yep, yep.

Matt Carey (48:33)
just haven't been solved yet. Like this whole evaluation system, where are all the eval companies? They've just died or like they've disappeared or they raised too much money. like, I think that, yeah, the eval thing is so like also super funny because you still have companies like Merkle who were doing like human powered evals basically. Every time you want to run a benchmark, you have to make some hacky Python script work. can't just sign up for a SaaS that runs your benchmark for you.

Wilhelm Klopp (48:34)
Mm-hmm.

Mm-hmm.

Yeah

Mm-hmm.

Mm-hmm.

Matt Carey (49:02)
You have to make some

hacky Python script work. there is no, the eval thing still doesn't exist apart from hiring someone to go through all your data and run it. Like maybe it does inside companies are very specialist tasks, but no one's managed to make the, no one's managed to make the blank eval system work, which I think is, there's a big surprise on this list is that there isn't an eval company and I'd have thought by now we would have had one. So I'm calling Merkle there and Surge also above that.

Wilhelm Klopp (49:19)
Mm-hmm.

Right. Yeah.

Matt Carey (49:30)
I'm calling those the eval companies.

Wilhelm Klopp (49:32)
Makes sense. they, the human party about companies. Yep. So they're kind of feeding into like the AI labs needs. ⁓

Matt Carey (49:34)
the human powered eval companies.

Yeah, and it's a mad,

like they've just taken a whole different niche on, like they've taken a whole different look on what was a cool niche. Yeah, there's like all of them, there's like Jenny is here, AI for writing, GPT-0, I remember when that one came out. That one's like AI detector, I don't know if it works. Photo room.

Wilhelm Klopp (49:53)
Totally.

Yeah. Yeah, I don't want to

be painting too much of a pessimistic picture. Yeah, yeah. I don't think this is up to date. ⁓ It's just, yeah, guess what I'm just interested in here is there is something funny happening and I haven't really seen it before. ⁓ One thing that, completely different note that I wanted to ask you about, which got me excited, is Claude Skills.

Matt Carey (50:06)
I think this isn't up to date.

Wilhelm Klopp (50:26)
Did you see this?

Matt Carey (50:27)
Yeah, I haven't used them yet, but I did read the blog and I had a look at the docs. ⁓ They're just mock-down files, right? There's nothing much more than that.

Wilhelm Klopp (50:38)
Well, I think the

cool thing is that you have this brief summary of the skill, right? And only that gets included in the context. And then until Claude decides to use the skill, and then more of the...

Matt Carey (50:55)
Okay, and it can request all

of it, yeah.

Wilhelm Klopp (50:58)
Or it automatically is included. So I think it gets around that MCP issue around polluting the context with too much stuff. ⁓ And to me, seems like a... I've often had this issue with Claude Code that I wanted to use various things I've made, ⁓ but maybe it considers them briefly, but then it just pursues some other approach. Whereas I can be like, hey, this is the debugging skill, and I want you to use the... When you have a debugging-shaped problem,

Matt Carey (51:06)
Yeah.

Wilhelm Klopp (51:27)
look at this for how to do that. I think in the Claude examples, it's like how to make a pitch deck with like the company's brand, ⁓ like following the company's brand guide, like for fonts and colors and whatever. it's like ⁓ Claude can reach for these skills when it happens to embark on a task that needs them. And then like it gets access to the full brief.

Matt Carey (51:30)
Yeah.

Okay,

I can say like one thing about Anthropic is that they are really good at distilling what is seemingly like distilling an idea that is native to LLMs into something that humans can reason about because that whole like, do I include it in the context or not? If you've ever built an agent, that is like a common debate you have. Like, do I need this in the context? Can I just use a summary? Like, where do I need like...

Wilhelm Klopp (52:07)
Mmm, right.

Matt Carey (52:19)
The classic one is ⁓ I made a little package called like cursor rules to Claude, where if you've written all your cursor rules out and now you're using Claude, how do you include them in your Claudemd file? And mostly it was just like, it included the summary and then it said, if you want to read the full copy of this thing, if you want to use it, you can just open this file and it gave a file path. ⁓ And so I feel like that's like a similar thing, but like,

Wilhelm Klopp (52:26)
Mmm.

Mm-hmm.

Matt Carey (52:47)
I just called it cursor rules to Claude, whereas rebranding it skills, kind of like, it's kind of great, right? Like it's awesome branding. Yeah, I think they're very good at that.

Wilhelm Klopp (52:56)
Yep. Yep, yep, yep. yeah,

I guess you're right in that sense that cursor rules already could attach to specific files, right? Or specific file types. Yeah.

Matt Carey (53:05)
Yeah. And

I think what's also smart is instead of calling it rules, because they could have just called it rules files, right? by saying skills, you're like acknowledging that the model might be bad at it. Like, like if you, if you give the model a skill, you're, you're, you're going to be happy if it does most of it, right?

Wilhelm Klopp (53:17)
Totally, yep.

Matt Carey (53:27)
was that as soon as you say this is a rule, you're gonna be super unhappy if it ever deviates from the rule. So it's just like excellent branding.

Wilhelm Klopp (53:28)
Mm-hmm.

I agree, yeah, they're good at this. They're very good at this. ⁓

Matt Carey (53:39)
Yeah, I'm very

impressed with how they've done this. And I think this is one of the beautiful things about Claw Code, why the UX is so nice, why it feels like quite intuitive to use. ⁓ It's the product behind it, like the product team on it is amazing at Anthropic. Really, really, really impressive.

Wilhelm Klopp (53:50)
Mm.

I have one last question for you, is, yesterday there was a really cool spectacle over San Francisco because there was a blimp, a massive blimp floating around. Like an airship? Like the helium filled, do you know what I mean?

Matt Carey (53:59)
gone.

What's a blimp?

Yeah.

I like the big silver ship things with like a, I don't know if they have to be silver, I kind of know what mean, yeah.

Wilhelm Klopp (54:23)
Like they were used quite heavily in like, I don't know, World War I, I think. Like a zeppelin. Yeah, yeah.

Matt Carey (54:26)
Yeah, yeah, yeah, yeah, yeah, No, I know what you mean. Yeah, zeppelin, yeah, yeah, yeah, yeah.

A blimp, is that what you call them? Man, when you said blimp, I just thought, this some like, is this like some mesh between another word and gimp? I was like, what, what, what is this?

Wilhelm Klopp (54:33)
That's a name for them, yeah.

Make all your blimp is AI gib. Make all your images. ⁓ Well, and but it was cool because there was a there was a James Bond movie like with Roger Moore in like the 70s or something, which was also a blimp over the Golden Gate Bridge. And that's exactly like I'll send you a picture. ⁓ That's exactly what it looked like. Yeah. And I looked into it and it turns out it's this research company called

Matt Carey (54:57)
Yeah, with the zeppelin, yeah.

A what?

Wilhelm Klopp (55:09)
Pathfinder, funded by Sergey Brin.

Matt Carey (55:12)
Wow.

Wilhelm Klopp (55:12)
who want to basically bring back blimps and use them as like alternative. They're called blimps. Yeah, they're called, well, you can call them that. Let's ask Lord, shall we? Where does the word blimp?

Matt Carey (55:17)
I still can't get over you calling them blimps. Are they actually called blimps? I've only known them as zeppelins.

Does blimp stand for anything or does it like?

limp. It's an

awful name. No wonder they died out.

Wilhelm Klopp (55:35)
Why does it feel awful to you?

Matt Carey (55:37)
Why do you think?

Wilhelm Klopp (55:38)
Does it sound sexual to you?

Matt Carey (55:39)
No, it just sounds like GIMP.

Wilhelm Klopp (55:41)
Maybe you've spent too much time on the computer as a young child.

Matt Carey (55:49)
I don't think so. yeah.

Wilhelm Klopp (55:51)
You don't think

so? You spend too little time on the computer as a young child? I think I spent just the right amount of time.

Matt Carey (55:58)
Yeah, I think just the right amount of time. I will go with that.

Wilhelm Klopp (56:01)
⁓ okay. Okay, wait. I mean, I have no idea if Claude is making this up. The most popular story is that comes from a British military airship designations during World War I. Airships were classified as either type A, rigid, or type B, non-rigid, limp.

According to legend in 1915 a British officer named Lieutenant A.D. Cunningham flicked the fabric of a Type B airship with his thumb and it made a sound like blip. ⁓ No way this is true. And the name stuck. ⁓ Yeah, I don't know about that man. ⁓ But anyway, my question for you is gonna be.

Matt Carey (56:30)
haha

That's a

much better justification. I like that. I'm happy with calling it blimp now. Yeah, go on.

Wilhelm Klopp (56:44)
If you had Sergey Brynman, if you had Sergey

Brynman, what research project would you fund to revive some concept of a bygone era or something like that?

Matt Carey (56:56)
That's so tough.

I don't know,

I was thinking about like film colours. You know, like Kodak used to make really interestingly coloured film and there was all of these different like film manufacturers and they all had like different colours and those colours were very particular to those manufacturers. Well, with digital, you can kind of get everything now, ⁓ but you really have to like, I think there are still people who spend a lot of time working on digital colours to make them look similar. But I think that's like a very good...

Wilhelm Klopp (57:02)
Mmm.

Mm-hmm.

Matt Carey (57:24)
Like that's not Sergey Bred Money, that's like, yeah, no, I think Sergey Bred Money, you have to go a little bit larger. No, you have to go a bit bigger.

Wilhelm Klopp (57:27)
could do that tomorrow. Yeah, no, I guess that's

a good answer, not that you have to...

Matt Carey (57:33)
Go on, something about colors with photography is something that springs to mind, that's like older and then we changed it and maybe it got worse for a while. But maybe if that was 10 years ago, it's a good answer. Maybe now it's a really bad answer. Probably something like battery storage or yeah, like solid state batteries or something like that. Like they're still not quite there. I don't know.

Wilhelm Klopp (57:47)
I like it. I like it.

Mm.

Seems like it would be good

for the world.

Matt Carey (58:00)
Yeah, or the mad one that people haven't been able to work out is that wave power. You know this? Or at least maybe they have, my uni days are out of date, but from what I learned at uni years ago, basically some small wave devices work, but you can't make them big enough because if they get too big, they just get destroyed by the waves. But wave power is some of the most untapped energy that we have. It's a huge amount of kinetic energy. Yeah, I don't know.

Wilhelm Klopp (58:09)
Mmm. ⁓

Interesting.

That's cool. I like that. Well,

let's make you a billionaire and then let's figure that one out.

Matt Carey (58:31)
Go

on, what would you do? You can't leave on that.

Wilhelm Klopp (58:33)
⁓ Yeah, I also, I don't know if I have a great answer. I think something around flying I would be interested in as well. Because so many of the like best airplanes like that you fly for fun, like the like hobby planes, they, ⁓ it's all designs from like the 60s. And you would you would like Cessnas and stuff. Yeah. And they are so ⁓

Matt Carey (58:51)
like Cessnas and stuff.

Wilhelm Klopp (58:57)
beautifully tuned in like minor ways. Like for example, on the one I fly the Piper, like the wings are like slightly different on the left and right. So that makes them harder to like stall and spin. So it's like these safety features really built into like the airframe. Like it's just really impressive, but it's just fundamentally like these designs that are like 50 plus years old. I think there was some guy on Hacker News who was trying to make like a modern

Matt Carey (59:10)
What?

Wilhelm Klopp (59:24)
⁓ like hobby, like Cessna plane that doesn't, that yeah. And also there's all these controls that like just feel like strange. Like you have to manually deal with like the carburettor heating when it's a certain temperature. And it's just adds to like the workload. And I think he was trying to get rid of a lot of that. ⁓ And it's a really interesting Hacker News thread with people commenting about it and all that. But I think something like that would be interesting.

Matt Carey (59:28)
Yeah, yeah.

I'm not gonna lie,

that sort of hacking sounds like people are messing with their coffee machine, except their coffee machine is like, it's like the types of threads that I read is like, oh, I put a new like PID controller in my Gaggia Classic. And it's like, ah, cool. But that's not 12,000 feet or 17,000 feet above the ground. Yeah, that's wild. I thought you were gonna say like bring back supersonic flight, like Concorde.

Wilhelm Klopp (1:00:00)
Yep.

Yep. Yep.

Ooh,

yeah, I mean, think for sure that as well. Like imagine. Imagine.

Matt Carey (1:00:14)
They were

talking about bringing it back actually. So the problem that I understand with Concorde is it's not actually a a like a pricing problem. It's not actually a technical problem. It's just that people live near airports don't want Sonic Booms.

Wilhelm Klopp (1:00:29)
Yeah, so Boom Supersonic figured out something on that front, right? Like they figured out a way to not fly over land ⁓ and cause sonic booms. They made some progress figuring out...

Matt Carey (1:00:31)
And that makes sense.

I did read something.

But the other one, okay, this is it. But these are all things that have happened. need to be better at thinking in the future. I think this is actually a really hard task because another thing we'll be dreadful, another thing that happened recently that does sound like sci-fi, and I'm gonna leave people on this vision, is that, ⁓ do you know you could grow a third set of teeth?

Wilhelm Klopp (1:00:51)
Yeah, clearly. We'd be awful sci-fi writers.

What? No.

Matt Carey (1:01:07)
Yeah,

yeah, seriously. There's this Japanese experiment that just finished some sort of trial and is going into the next stage of trial, I've no idea what it is, where they had an enzyme that activated your tooth root, being that you could grow, technically you have enough capacity or some raw ingredient inside your body to grow another set of teeth, but you only grow two. And you do that because there's an enzyme that's blocking the growth of your tooth root.

Wilhelm Klopp (1:01:17)
Mm-hmm.

Yeah.

Matt Carey (1:01:36)
All of this is very pseudoscience because I'm not a biologist. But yeah, theoretically if you block that enzyme and you can do it very acutely, so on individual tooth root, you could grow a completely new tooth. And so we might be making dentists ⁓ maybe a lot more useful. I don't know. How cool is that?

Wilhelm Klopp (1:01:39)
That's fascinating. Uh-huh, uh-huh.

Right. No.

⁓ You just talked

to your mouth, Claude, and you were like, I think this tooth, I could use a new one. I'm like, all right, let me block the enzyme, grow.

Matt Carey (1:02:08)
Yeah. So I reckon,

I don't know like whether you'll only be able to have one new one because like a dentist go to is like, there's a problem with this tooth, right? That's coming out. That is their like go to is like, that thing is coming out of you right now. that was, no, dude, that is always the thing that I was like struck about dentists. Like we're engineers. We try and fix things. Most of the time, if you're a dentist and there's a problem, it's like, it's coming out.

Wilhelm Klopp (1:02:19)
I feel like you're doing the dentist dirty on this one.

Matt Carey (1:02:37)
Like my brother dated a dentist for a while. I met a bunch at uni and all they learn at uni is how to pull teeth out. That is it. Like it is crazy. I mean, I'm doing them absolutely dirty right now. They learn obviously way more than that, but yeah, it seems like pulling teeth out is like a large part of the job description. But imagine if you could, yeah, if you could pull teeth out. Yeah, they seem to do polishing and pulling teeth out, but yeah, they could actually regrow them. I think that would be wild.

Wilhelm Klopp (1:02:38)
Hahaha.

Hahaha

Yeah, class is in session another day of pulling out teeth. And they could regrow. ⁓

Matt Carey (1:03:04)
Especially for old people, it's a mega quality of life change having good teeth.

Wilhelm Klopp (1:03:09)
Totally. I can see that. Okay, that's a cool one to research. I like that.

Matt Carey (1:03:15)
Anyway, right, we should leave this, otherwise it's gonna be a pain in the arse to review. But I'm literally gonna stick this one uncut. I think we had a good chat.

Wilhelm Klopp (1:03:16)
Alright, man.

Yeah. Really nice.

Yeah, this is awesome. I'm excited. Yeah, it's really fun just doing it two of us as well. I mean, it's fun to have guests. All of this is fun. I like it.

Matt Carey (1:03:27)
Yeah, we should do the

next two just us. We try to get back into a cadence. When it's just us, don't feel that bad about not reviewing it and not looking at it and not editing it that much because I know what we said and if I can remember that we didn't say anything that stupid, it's like, it's fine. ⁓ But I feel like when it's guests, I wanna get their best side across, you ⁓

Wilhelm Klopp (1:03:31)
Yeah, I agree.

Mmm. Yep.

I agree, I see. That's totally

fair. That's something I hadn't considered in the past. That makes sense. yeah, ⁓ Sweet man.

Matt Carey (1:03:52)
I feel like there's a bit more

of responsibility. Anyway, dude, it's been a pleasure. I'll catch you in a bit. Big love. Bye!

Wilhelm Klopp (1:03:56)
See you, bye, big love.