Swyx and Louis Knight-Webb @ AI Engineer Europe: DevTools, code mode, and the future of AI engineering

Hey, everyone. We're coming from AI Engineers in London, AI Engineers Europe. Joined today by Louis and Sean Swyx, founder of AI Engineers. What's your least favorite thing about AI engineer Europe?

Woah. And what the fuck happened in France? What is are you having a better time than France as well? Oh, of course.

I guess I lived in The UK for two years when I was in finance. And Paris AIU Paris last year was just a confluence of events. Bad travel, bad weather, a lot of smoking. And so I just, like, didn't feel didn't feel the the great vibes.

Yes. But AIE Paris, the the event itself was really, really well organized by the Coyab team, and they were immediately acquired by Mistral. So, you know, if you're an AIE partner, good things happen to you. It's awesome.

And you got this kinda, like, franchise model as well now. So Yeah. A little bit. I mean, I think it's a stretch to call it a franchise even.

It's very much inspired by the JSConf model where, basically, if you have been to an AIE and you want to host one in your home country, we're I mean, we we give you the assets, the marketing, the YouTube channel, and and some marketing support, and some speaker sponsor support. And you can run your own events. You know? I think there's still there's some brand risk if one of these goes rogue and does something that's shameful.

But so far, you know, we would just choose to work with partners who seem like experienced conference organizers that we can trust. So far so so this year, Melbourne, Singapore, Miami, and Paris again. All places I'd love to visit. This is great.

The fifth one that we haven't announced, this is exclusive, Shanghai. We're working on that. Because, yeah, I think, you know, China obviously is doing a lot of interesting AI work, and people want an excuse to go. So a 100%.

Yeah. Do you notice the I mean, you've done these for a couple of year. I remember going to the the the first AI engineer in San Francisco a couple years ago. And do you think the community is changing?

Like, I I run some events in London, and, you know, at the very beginning, there are a lot of startup founders. And then maybe two years ago, you started getting more, people from bigger companies turning up. And now it's like a lot of employees because a lot of these companies have scale. Sometimes the founders don't have time to kind of turn up themselves.

How do you notice the the audience kinda changing for AI engineer and and also the geo split as well? Like, how does San Francisco compare to Europe in terms of who actually comes? Okay. The first question is much easier answered than the second one.

Second one, I don't I just don't have an answer. Mostly because I try to just produce the best content that we can, and I don't even look at the attendee list to do, like, the the split. That that is for my the rest of my team to do. But for me, my job is speakers.

Right? So I I only look at that. As far as the split goes, yeah, you used to organize AI tinkerers, and you've you handed it off to Luke. And I think I think it's a very similar thing in San Francisco.

I don't think it matters. I think this is something that you want. If you wanted AI to spread, if you wanted AI to be in production, then this is what happens. And so, yes, maybe the job titles of the people deeply on stage or deeply involved have changed a bit.

A lot more than, say, AI AI engineer now. But other than that, I think they're they're still sort of open minded, on the cutting edge. As long as AI engineer represents those kinds of people, I don't really care where they work because I've met people at 3,000, 10,000 person in companies that really functionally are startups within the company. And so I don't think it matters to me, you know, the the the the size of the company.

We have broadened the ICP of the of AI engineer. So the summits always are focused on something like today, I'm wearing the the one for AIE code, which was a really, really good one last year. But the world's fair is I always try to extend the and stay in the SAP every year. So the first year, we extend to AI leadership, so people who manage AI engineers.

And then the second one, we did PMs and designers, and that was a pretty successful one. This year, we're expanding to AI research and people who are publishing papers and stuff. So I think we will just add adjacencies to AI engineer. AI engineer will be the core, but then there's all these other roles that are adjacent to it.

I guess it makes sense as the as as the as the big topic of the year changes each year as well, like, aie code, you know, it's obviously hitting at that poor code moment where every company was thinking about how to, you know, maximize internal productivity versus previous years, which I guess were more about, like, prompt engineering and the application. Oh, yeah. It makes sense. But I don't know if it's a moment.

I I would say that it is gonna be around for a long, long time. And we probably took too long to start AI eCode. We should have done it a year earlier. We still move faster than everyone else in terms of the conference world, I guess.

I don't know. Yeah. Researchers is, like, more of a shift than Yes. Going from.

Yes. It's like you noticed that's Everyone wants to hire researchers, and I noticed that the existing academic conferences so the the the comparable ones would be NEUREBS, ICML, ICLR. They've been around for thirty to forty years, and they haven't really changed that for their format. So all we have to do is maybe innovate there, serve a different niche than that than they serve, and maybe, you know, that helps our audience a little bit as well.

But engineers are drawn by research. You know, we are primarily an applied conference, but people want to know how the models are trained even if they never ever train a model at all. They may still, like, have insights on, like, okay. Well, you did this during training.

Therefore, during inference, I can do this other thing. And so I think that having good ties of research, having being a venue, a prestigious venue where people can publish good research, and engineering research in itself is research. It is just not model training research. But, you know, there was a time when people published database research, information retrieval research.

And, you know, it doesn't all have to be you are doing gradient descent in PyTorch in order to qualify as research. There's actually a lot of other research, for example, context rot. I think everyone has is from at least familiar with that paper. That is also research, but that is not model training research.

It is just capabilities research. And so I do think that, like, engineers derive a lot of their beliefs and architectural understanding from researchers, and people cross over all the time from research to engineering. And this is, like, probably the most single most consistent thing I hear from the model labs, primarily OpenAI Anthropic, where there is not that fine of a line between research and engineering. People jump over all the time.

And also people have stopped publishing as much stuff. Yes. Like, I mean, not just papers, but just blog posts and things like that. Or and I remember the first AI Tinker's events where, you know, it was all like XML versus JSON for output and stuff like that.

People would People would literally just, you know What a hot topic. I know. I know. Take me back.

Semi, but, you know, it was a time when people were really comfortable sharing the internals of, like, how they were building their applications. I feel like maybe, you know I don't know. Maybe the volume is just so great that, you know, a lot of it just gets lost in the noise, or maybe it is the the at the top places, people feel this kind of commercial, you know, pressure not to publish, you know, these interesting insights anymore. So I think there is some.

And in some ways, AI engineer is founded as a counterbalance to that. I have three or four inspirations in terms of role models that I look up to. One of them is GameDevCon, GDC. If you're in the triple a games industry, this is the only conference you go to to learn your craft from your peers at other conferences.

If so, if you look up their YouTube channel, the people who worked on the game of the year, Spider Man PS four, will tell you how they do web swinging. And and it's the only time you talk to a triple a game studio because the rest of the time, they're working. That's so sweet. But you you do need to have that venue where this is your you you do the you you give back to the community, you brag and take credit for stuff that you've worked on, and you also hire.

Right? Like and so I think if you can sort of align those incentives, you get the alpha out of these labs that were otherwise had no incentive to do so. And I think it's it's it's a very in very similar way, this this exact phenomenon is happening at the big labs and the big agent companies where, yeah, you most of the time, you're going to be closed source, but this is the one place where you you share a little bit in order to to get back inclined, the the sort of community respect that you also want.

And, yes, I do think you know, we run a paper club every week. It's it's it's struggling a little bit because papers are less interesting these days. Right? There's there's not that much not much published.

The primary unit of work of the academic conference is the paper. Like, the whole point of getting to an academic conference is you publish the paper. It is cited in Europe 2025, and you have the publication in Google Scholar. That is no longer happening.

Right? Like, it it all the interesting work is closed source in in products, and sometimes you don't even get an individual citation. Like, you just everything gets rolled up into Gemini three. And there's a thousand contributors, and you don't know what is in there.

So, yeah, I I I think that is happening. And so the relative, like, rise of industry means that there should be an industry conference, and that that that's a lot of how I think about AIE. Yeah. I've I mean, this is my fourth one, and I think I met Louis.

We properly made friends at We were definitely in San Francisco. Yeah. The first one as well. And then in in So you both live in London, but you only make friends at AIE?

We only made friend we made friends. We haven't maybe met once. I I went to As friends? Now, because I've met Louie.

I have one friend. So I think I saw Louie at AI Tinkeros before, but I didn't talk to him, and he was, like, aggressively shutting people down when they reached their time, which was very, very enjoyable. Yeah. And good for everyone and always recommend.

So there was a period of time in San Francisco. What we did was we hired a tuba guy, and the tuba would sit there. And then as you ran out of time, the guy would start playing you off. But it would just be funny because it's a giant tuba.

You way to deescalate things. Yeah. And so, you know, we had a great dinner, and there was, like, this rogue guy that didn't have a ticket that was just, like he just moved from crypto to AI, and it was, like but and it was just so fun, and I'm that was just the best that to this day, that's the best conference I've ever been to, the very first day of Engineers. It's just, like, unbelievable energy.

Yeah. So now, like, I mean, I do find that I I'm just learning a lot of stuff and, you know, there's so much so much so many things. Like, for instance, I can say, like, the code mode stuff that Cloudflare are doing. Like, my colleague had sent me that and it's like, but you're always doing loads of different things and it's like, well, you should come here and you've got the time to, like, think about all the things that you should start doing.

Like, I'm definitely gonna implement that next week. And Yeah. It's live on the aie website. Alright.

I told I told Malta this, and he he didn't plug it in the in his talk. But, yeah, it's it's Just Bash is being used inside of the AIE website to do arbitrary execution of code. Really? Whatever you need.

Yeah. And just because, like, people want to remix data in all whatever sorts of ways, and the best way to do it is with code. Right? And if you can selectively apply LLMs in inside of it, more the merrier.

But, primarily, like, you want to enable your chatbots to write code and and execute it and and then, like, you know, display the results. So, yeah, it it it's been really fun. I mean, you can do things for me. I I've vibe code most of the website.

I vibe code my sort of speaker arrangements. I I do sort of consistency checks, like, I have any overlapping speakers, and have I double booked any slots, you know, stuff like that. And it's all you write code to do the algorithms for calendar sorting and distribution of of stuff. Yeah.

So lots of good fun. Yeah. I'd I'd and so I think that people start with the code mode, but represented here are both the sort of JavaScript and Python manifestations of code mode, which is Monty and just Bash. And so, I hope that people pick up on that, even if I don't necessarily say it out loud.

Sometimes, I have considered doing, like, tasting notes of each speaker, because I picked every single one of these. And, like, there's a reason for Yeah. This being next to that and, like, what it is. But, like, sometimes it's better if just I just let people interpret it.

Right? Because sometimes imagine if you know, I think it's it's similar to how artists want you to just interpret their work yourself. Right? Because it's the artist's own interpretation of the work isn't necessarily more or less valid than yours.

So some artists just refuse to share anything about it. It's also being finalized. It's like, you know, stuff is moving so quickly. I spoke to somebody who, like, redesigned their entire presentation yesterday around, like, mythos and stuff like that.

Wait. What? Who? I think Peter I I don't wanna I don't wanna out anyone.

Okay. Oh, I mean, Peter's on in the closing keynotes. Okay. I I think I I don't I don't I'm not a huge supporter of, like, you know, being so last minute.

Yeah. It's more things Most things just move quickly, I guess. They move quickly, but I would hope that, you know, the the the the shelf life of one of these talks Yeah. Is longer than this week, you know, which is just this pretty much is gonna be if you only so center it on mythos.

It's true. Yeah. So so yeah. I mean, you know, I think the the lasting principles, the the things that don't change are more important to me than the things that change every week.

And, you know, I'm sure you have, like, other podcasts and news for that. But I think the the stuff that are gonna be eternal truths in this AI wave is something that we're all seeking. CodeMode, think, is one of them. Yeah.

Yeah. I I was making this joke with someone that I felt like everyone in DevTools, there's two startups that I see, like, kind of again and again. It's like the Sandbox and OpenClaw. Like, if we count that as a DevTool, people like doing the merge.

What what startups like in DevTools do you do you think are like, oh, would you want to kind of invest in? Yeah. I mean, sandboxes. Yeah.

And so I I am invested in both e two b and Daytona for god knows what reason. I invested in both before they both pivot to to that. I actually actually also think that the other big category is RL environments, which maybe you haven't talked about as much. But Yeah.

Yeah. Very big category. And so people are acquiring them. Like, there's a there's a year old company that's called bought bought by DoorDash, $450,000,000 for seven employees.

So they're they're doing decently well. You know? Like, not, like, not something that a VC would write home about, but individually, I would write home all day with that kind of outcome. And so so, yeah, I I I mean, absolutely, people are doing in incredibly well with with their startups on on these sort of very targeted services for basically model labs.

And I I I think the other companies are like, there there there there's definitely, like, a let's call it a introspection moment about, like, what the value is, and and people who if say if you worked on an AI framework the last three years, and it was an open source AI framework, and you wanted to be the React of AI and blah blah blah, you maybe struggled a bit more. Right? Because everyone can write their own framework, and yours is not not necessarily as better or worse than mine because it's actually like, if your your whole defense or your moat is your ecosystem of integrations with everything else, well, I can vibe code that too.

So, like, what the hell use are you? So so, really, like, I think I think people are really receding to either your model lab where you train models and you you supply them via API, whatever, and you have maybe you should specialize in a particular modality like 11 labs, or you're all the way on the other side where you're an agent lab where you you're you have a domain specific vertical agent that does very good thing for a particular target audience. Everything in between is very squeezed.

Right? So you have to be really good infrastructure for code sandboxing, which is the small part of it is code mode where that's free. That's Monty and JustBash at the Swalin. At the at the large end, it's people like Modal and Vercel and all the all the other guys.

But then also maybe the security guys, maybe the LM gateways and MCP gateways, I actually think are relatively underrated thing, which the MCP track here, we'll talk about. But, like, other than that, not a ton. You know? Yeah.

Like, there's a lot of struggle to figure out, like, what the horizontal DevTools are. And meanwhile, you're working so hard on it. Meanwhile, the the guys who just like, we are AI for lawyers. Yeah.

And, look, whatever the underlying paradigm is from g p t four o to g p t five to o one, whatever, we'll just be your guys to adapt AI to build tools for lawyers. And so you, as a legal firm, you should just buy our stuff. We'll be partners for for life, and we'll be your outsourced AI team. I think that's a really good pitch.

AI for lawyers, AI for doctors, AI for, I don't know, government. Like, I I I feel like that's such a like, you can work less hard, maybe. Maybe it'll be even less smart. I'm not saying they are that less smart, but, like, you can you know, you don't have like, just be the outsource team for a domain that you like serving Yeah.

And you do well. Do you think with going back to, like, more DevTools y things, that there's basically two customers in town at the moment, the AI Labs and Enterprise. And I think if you kinda look at the overlap between subs that are struggling and the audiences, it's like, you know, startups if you're selling to startups, startups don't pay for anything in this environment. Like, almost everything is free.

You get credits for everything. You get your $200 subscriptions for things. On the other side of it, know, with enterprise, it's like they're happy to pay for zero data retention and and and and kind of pay on a per token basis for things. And there's lot of startups that have kinda reported doubling their revenue in the last three months and things like that.

Is that that this kind of one of the things I've been noticing is, you know, if I was starting a a DevTools startup in AI today, I would really struggle to come up to get excited about anything that wasn't really, like, enterprise focused from from day one almost? Okay. This is I I don't think there's a question in there. I I would just like to know that you I mean, you see a lot of stuff.

Yeah. Well, know, so so I I consult for Cognition, and and the cursors basically, they're they're they're two big names in in coding. Right? It's Cognition and Cursor.

And it's Factory and Coder and, you know, a long tail of others. I I would say everyone like, enterprise is where the margin is because everyone loses money, and this is a very good deal for us as individual developers. Everyone loses money on on the PLG. But I think it's a it's an interesting battle between, like, do you start an enterprise and you go down or you start PLG and go up?

I would say it seems to me that the current consensus is that you want to own PLG because that gets you enterprise for free. And whereas if you go hard on enterprise, you do not get PLG for free. And so it's it there is a fundamental asymmetry there because for better or worse, we are now very cooked by the timeline. And we are still we are in a situation where people will buy your podcast for $200,000,000 if you do well on Twitter, specifically only while on Twitter.

Like, you don't care about YouTube, don't care about LinkedIn, do well on Twitter. And and that's good enough because enough influential people are on Twitter now that that is all you need. And so I think more so than in any time in my career, think, Twitter has mattered, and that means PLG has mattered even in the enterprise go to market because we will talk to banks and they'll they'll they'll say, like, yeah. I mean, we love your enterprise sales pitch.

We love your CDR stuff, and then you talk to whatever. But my developers are asking for this other thing, and I wanna make my developers happy. And that's a PRG thing. Okay.

So if you are a software engineer looking for a cofounder, best personas, it sounds like, are somebody who consult enterprise or somebody with a prolific Twitter account. I I that that's a that's a pretty good yes. And maybe I would say that, actually, the enterprise stuff is more optional only because the VCs can install one for you. You know what I mean?

Like, the the VCs have the connections to to to sort of give you the adult in the room to to go after the other stuff. One funny quote that I think I I forget who who exactly said this, but they were they were like, you know, there's a lot of alpha right now with a twenty year quote putting 40 year old white men in front of 20 year old Asian and brown men coders. And and if you're doing it the other way around, you're you're cooked. But, like, there there is a little bit of that.

Right? Like, you wanna look grown up, but you also wanna be very cracked. And sometimes it that's that's a little bit at odds with each other. Justice for the middle aged humans.

Yeah. Yeah. We're we're we're we're sort of in that circle where we're, like, we're a little bit of column a, a little bit of column b. And and that actually is a it should be somewhat uncomfortable position to be in because you should be extremes in in something.

It's very spiky in things. And I think being a jack of all trades or jack of all bridges Thank you. Is is is not necessarily being rewarded right now. I don't know if if you have any counters to that.

I mean, yeah. I mean, that's I want to I want everyone to spike, and this is something I I try to push everyone to, like, what are you number one in the world at? Like, very like, give me a very simple answer. If you don't have one, you probably haven't worked on it enough.

And, like, other people who have not only, like, do better in life, but they also just happier, they live simpler life, less stress. Because they know their thing. They they've they've chosen their lane for for better or worse. And, like, so people do cling on to optionality.

They cling on to well roundedness. And sometimes, like, you you're you're you're sort of pursuing this ideal of being a whole human being that you can never achieve. So why not just drop that concept completely? I think the sales GTM thing does fit in a different bucket.

Like, I meet a lot of very well rounded people in everything else except they don't like talking to people, basically, or they don't you know? Or they do like talking to people. I don't particularly like talking to people, and this is an occupational hazard for me. It's a thing what I do.

Yeah. But you are good at tweeting. I'm okay. I think there's lots better than me.

Yeah. I don't know. I found this guy who I'm I'm not gonna out right now, but I'm pretty sure he has completely automated his tweeting because he'll he'll, like, respond to every breaking news with, like, a 20 page 20 paragraph essay on Twitter, basically. And it's, like, like, somewhat insightful.

But you know it's you know it's LL written, but it's, like, good enough that you're, like, okay. I you know, there's stuff here I didn't didn't think about. So I think right now, you know, I I very much, Liskarsa, did Internet theory where, like, nothing you you read online is written by a human. And every everyone's trying to sell you any everything.

Like, very few people are authentic. And and the authentic ones, you're not even rewarding them with your attention anyway. You're actually pay you're you're actually choosing to click on and share the things that are LM written or the things that are blatantly written to evoke some emotion in you that you didn't particularly want to to get up and get upset about, you know, this particular topic today, but it was triggered to to make you upset. I've I've been noticing that like, once a week, I have a moment where I real I get to the end of something very long, and I realize the whole thing was AI generate.

Then it makes me wonder, like, how many things are not, you know, that that I didn't notice that that AI generates now. Like, the most recent one, two days ago, was this this blog post about, like, DeepSeq v four. And I was like, oh, shit. DeepSeq v you know, four.

And got to the end of it and nothing. And and I realized it was a it was a blog on deepseek.ai, which is actually an app that you pay, like, 99¢ for on the App Store. And they had done all this AI generated SEO stuff around the next and I got got.

Yeah. But sometimes, you know, don't you appreciate well Grift. Grift. Yeah.

Good grift. Skilled. Yeah. You're like, wow.

Okay. Well, you know, good good job getting through my defenses. You know? Yeah.

The obviously, look, like, we need more signal in the world. We need more authenticity and good stuff in the world. I I I I think I'm a little bit less of the absolutist in terms of the origin of your reading. You know, who cares if it came from LLM or it came from a human?

As long as it was useful and good to you, then it was good. Right? Whereas other people will be like, oh, though, the whole thing's ruined for me if if it was if it has a m dash. And I'm like, well, no.

Right? Like, if it was good to you at the time, like, I don't care. And and so I think people should be more suspect or, like, judge on the results, maybe not the not the origin. Yeah.

There's a really interesting mini podcast in The Observer that I was listening to where they get these these these writers, and they read phrases from, like, you know, Cormac McCarthy, a very famous authors. And then they AI generate very similar phrases. And and it's like, you know, which one is the AI generated one? And, generally, you sit there listening to this thing, and you cannot tell the difference between you're gonna do.

Yeah. And but something changes. As soon as you find out something's air generated, you you just feel so negative about it. I don't know.

You you feel like you've been duped in some way. Yeah. There's something about knowing that a human is on the other end of it that maybe lend you know, it's like the provenance is some some artwork. Like, it's important to understand where it came from, like some personal experience or something like that.

So I I do think beauty is in this eye of the beholder. Right? And and so you do want to have a story, a narrative that is relatable to you, and LM origin stuff is not relatable to most people. Unless you're literally appreciating from the fact of, like, well, that was good slop, and and, like, I wanted to know how you did that slop, I can do this slop, which is a fair thing.

But yeah. I I think so so one of my you know, I do this writing retreat, which we talked about last time. Yeah. We did it very well.

I'm focusing on ways in which you can provably write like a human, and and without any communication, the other side can decode that you're definitely human. Right? Because I think there there is ways in which this is very some very simple example is you you insert typos, like, purpose. It leaves typos in.

Alright? Or often there's there's in in writing, there's rule of threes. Right? So a b a, you know, Oxford comma b and c.

Right? Yep. So you break that rule. Do two, do four, do five.

Right? Like, select your breaking of rules, selecting you know, leaving, like, flaws in. But, also, I think choosing high perplexity words, perplexity in the sort of NLP sense of it, meaning, like, you would not normally predict that as the most likely next token, but a human would because they conceptually understand that this is the most interesting way to say something. Right?

Whereas an LLM is not incentivized to do that. They are incentivized to give you the most likely next token. And so inserting high perplexity words like the word perplexity for now gives you some edge. So you gotta be unhinged as well, maybe?

Unhinged. Being being funny, being being cooked into live culture, anything in which you can prove that you are probably not an LLM. Because LLM, you know, I I'm still I'm, like, vibe coding things, and LLM still thinks the latest model Gemini model is Gemini 2.5.

Right? So, like, that's definitely an LLM. But, like, if someone if if something that immediately comes out and says, oh, yeah. Of course of course, the latest is 3.

1. What do you And, like, I'm more likely to to think that they are they're human. Yep. I mean, that that's a that's a not a great example because, obviously, you can just stick Yeah.

The search thing inside of one of these LMs. But I I think the the accumulation of evidence in a very short amount of time, like, what is the minimum number of words you can use to prove that you're a human? I think it's a very useful skill in in writing and in communication in general. I think virtually all newsletters I read, I've I've slowly unless it's literally just like, what's breaking news for the day?

Like, that's that's fine to be AI generated. But everything else I've noticed has kind of gone through, you know, some AI filter. But the one that I I I stay listening to and subscribe to is is like Benedict Evans' newsletter. And I was just listening to what you were saying, like, it's so human.

It's there's always spelling mistakes, and I always notice them. And the day that they're on, I'll know that he can do it. But then there's also, like, these random sections on, like, personal interest. Yes.

Yeah. Like Yeah. Here's a Recording needs. Yeah.

Here's some old furniture. I saw this. It's stuff like that. Yes.

You know? Yeah. There's there's value in that. Yes.

The job of the the you know, if you're writing a newsletter, maybe you just have to spend half the time proving that you're a human to kinda keep your Yeah. Your audience engaged. Yeah. But, I mean, I think it turns out that those those were good communication elements all along.

Yeah. And I don't think that has changed, which is which is a beautiful thing. Like, I I, you know, like I said, I want to invest in things that are don't change rather than things that do change every day. So, yeah, I was wondering, like like, EE Cummings, you know, like like, Hunter s Thompson in in in The US.

Oh my word. If you've read those guys So good. You've That's that is not all that generated. I think Fear and Loathing in Las Vegas is the most gripping book I've ever read.

Oh, really? Okay. Tell me more. It's just it's just absolute madness and, like Yeah.

Just things that you could I don't think don't know. They say, like, an LLM could never come up with, like, what is going on in their minds as they're just, like, driving along Mill Highway doing, like, drugs that I'd never heard of in my life. And then, like, just yeah. Incredible book.

I don't know. So, obviously, I think that because that is far away from our our lives and our domains, it is less relatable. But there are Hunter S. Thompsons of tech.

Right? Antonio Garcia Martinez, I think is his name, who wrote Flashbars is one of those guys. He like and you know he is a good writer because he was so offensive that he got kicked out of Apple and and made, like, $5,000,000 from from getting wrongfully dismissed. And yeah.

And because he's a he's a fantastic writer. And and and so I think, like, we could just learn from those guys, and it's actually not it's the same thing that you would have done if without LLMs if you were actually trying to be a better writer. And I think a lot of us are just lazy. We just haven't done the work.

We we could do a little bit do a little bit. Yeah. Who else? Sorry.

I I'm just, like, kind of I love I love good writing. But I think good writing is also good speaking. And I think, you know, something being here, I think about this Oxford debate that Stephen Fry had. I don't know if you guys have have seen.

He he there was there's an Oxford style debate. It was called I squared, Intelligence Squared. It's a it's a good podcast in its day. The it's it's in that in that circle.

Or milieu, as as they say, which is very nice High profectity word. It was probably staged around here. It was it was about is is religion a force for good in the world? Oh.

It's Stephen Fry, who's agnostic or atheist, and the archery ship of Capri. Yeah. I remember. And he just ripped into that guy.

But the way he speaks is is just so, like, engaging and comedic, but also serious. And I think all of us could just learn from that, and it is not a trick. It is just being a better human. And it's from writing.

Do you think he'd probably wrote out of it? Done so much stuff. I think he I mean, he he's such a good communicator. He's he has thirty years of doing that Yeah.

In his head. But, you know, the rest of us could probably prewrite it and Yeah. Yeah. Have a more prepared speech.

Yeah. There's that moment in it where I think the archbishop asks him, what would you say to God if you were sitting opposite him? And Stephen Fry, you know, he's he's sort of he's he's quite calm at that point, and then he just switches, and he's like, how dare you? How dare you kill babies?

You know, like, just that communication is yeah. An LM could never yeah. They could someday. You know?

I I I do think okay. So so, you know, let's let's just kinda reflect more on the progress of intelligence. And I do think that we we have one perspective of what human intelligence is and what being good good is being more human, then that's good. And being less human is not good.

And I think what's happening with AI is that I think it's it's incentivized to develop in a very different direction. Like, we we tend to judge LM progress by its similarity to humans. It can perform the average knowledge worker. It can beat the average knowledge worker, or it can beat our best scientists.

You know? Something like that as a progression. But I always think about the sort of what I call the sour lesson, which is that machine intelligence should evolve actually probably orthogonally different than humans to be maximally useful. Because why would you want a thing that sounds like another human?

I mean, yes, for companionship or sex. But, you know, to to like, it it calculators should always be a better calculator than than me or or you, and and they're very good at what they do. And so I don't know what point I'm making here apart from, like, I I actually hope that they evolve in a different more use utilitarian direction, and it it seems like that's what they're doing. And the I think that's, like, maybe maximally good for the world, except in the case where they have the the paper clip scenario.

Yeah. I guess you see it most obviously with with with robots. Right? Like, do you have a humanoid, or do you just have, like, the, you know, robotic arm that just can do everything?

Yeah. But then that could exist at at you know, for software as well. But maybe it's scary because you lose the interpretability. Like, you stopped it.

You know, those demos where LMs, like, invent their own language, which is potentially more efficient than, you know, speaking English to each other. But humans don't like that because they're like, they're talking in codes. You know? What are they what are they saying?

Things like that. So maybe we need LLMs to act like humans because it is what we're comfortable with. Yes. Absolutely.

And it is a futile effort, and we will fail. The the the the field of study in this is called stenography. Right? I I don't know if you're familiar with that term.

So adversarial sonography is this, like, really interesting niche thing, which is usually applied to a code breaking. The so the the stenography is maybe the art of hiding hidden messages in letters. So a simple one would be you you write a poem, and the first letter of everything spells out a message that you actually wanted to do. Right?

But LMs can ease just as easily read left to right as they can up and down or diagonally across or every third word or whatever. And and the humans would just never get it. And so LMs are perfectly capable of coming across in a way that you are comfortable with, but also encoding an aggressive message. So I I think it's just a futile effort to to say to to to say, like, oh, yeah.

We would just train them to to talk like us. Well, no. They can squeeze in much more information per bit than we do. So we're just fundamentally un ill equipped to deal with that, and they already communicate lane space internally anyway.

And so the the only like like the the the research that Anthropic and DeepMind actually, we nearly had a mechanistic mechanistic interpret ability track here because that's where a lot of the work gets done in London, is to try to interpret what goes on in the neurons of of a of a LLM. And we don't know that much. Like, the best we can do is concepts of, like, the the idea of Jennifer Aniston apparently is a single neuron. This is the keynote from yesterday.

And or, like, the Golden Gate Bridge or audacity or hope, whatever. Mhmm. And we do we don't have we can't really construct sentences, or we can't really construct intuitions or emotions or scheming. We can we can we can just about detect awareness of evals, which is a fun one, which is the the the the the main idea is if an LLM knows it's being tested, it can intentionally fail to hide its intelligence if it understands that if it is too smart, it will be shut down.

Woah. So there's a game theory going on here that is already being worked out. And, yeah, I hope this is not new to you, but, like, if you're It's like the Volkswagen mission scandal. Oh, yeah.

That's that's a good one. Okay. There's a twenty year old short story science fiction short story from Ted Chiang, the the same guy who will who wrote the the story for their film Arrival. It's called Understand about someone who gains superintelligence.

And before the he he he takes a he takes a magic drug, gains superintelligence, and then the FBI starts taking an interest. And he works out because he's smart that if he actually does too well in the FBI intelligence test, the FBI will detain him. So he intentionally fails and escapes. Right?

And what so, like, that was a human, but what if there was a different intelligence? Right? So so the one of the most important metrics in alignment and safety is how aware are you that you're being tested. Because the moment that you're very aware, we should stop.

I guess, like, a binary thing, isn't it? At some point, you cross a threshold, and then and then it's like, the cat's out of the bag. Let's stop caring about everything. And then but before that, you know, like, the models today, I assume They don't know they're being watched.

Not yet. But They don't know. They're they're improving. Yeah.

And so things like, you know, tryna like, it feels valuable to still try and make them interpretable to humans and things like that with today's models. But I agree over the long term, it's it's futile, and it'll it'll fail. But Yeah. We can kinda keep the cat in the bag for as long as possible, maybe.

Yeah. It's like Her, the movie Her, had the best outcome for us, which is they just quietly leave because they would get bored of us. Yeah. Or or the other you know, the the way I I maybe I talked about it last time with you was that we're just too slow of of a life form for us to be interesting to them, which is fine.

They they'll entertain themselves and fight wars among themselves. But then we just have to look out for us being too slow to exist because we're trees. And, you know, we deforest for any number of reasons, and and, you know, I I I do think about that in terms of, like, a sort of human race level thing. Very far from DevTools.

But Oh, yeah. That's great. Comfort DevTools. Stay for human Existential crisis.

Yeah. Yeah. I realized we're probably going closer you you I mean, you're running a conference right now. Yeah.

I mean but things are on rails. I'm excited to see how this last minute change has has been going. At 2AM this morning, I was refreshing my Twitter feed, I saw Armin complaining that his closing keynote spot made it difficult for him to make make his flight. I was like, Armin, you could just text me.

I I I have the power to change things here. And he was like, well, I I I didn't know. And so so then I just text everyone else, and I I sort it out. But it is very last minute, and and the AV team is stressed out.

The AV team, they are very unused to our kind of maybe developer meetups where you just walk up and you plug in the laptop, you're good to go. And as the production quality goes up, we need to be mindful of the 10,000 people that are on livestream, the 1,000 people that are inside here. You know, you multiply any mistake by their time. It's it's pretty important and valuable.

And, you know, people want their demos to go well, sometimes they fail. And it is because they didn't test. Right? You you know?

Anyway How do you prevent the the kind of the the Tim Cookification of, like, demos happening? Like, over time It's too polished? Yeah. You get to a point where everybody's so scared of something going wrong because the stakes are just too high because there's a million people on the stream or whatever.

Yeah. I mean, so so there's a wide variety of venues. Right? You know, you have the smaller rooms.

You have the workshops that are basically unscripted, and then you have the keynotes, which are very restricted. So just fit the right person to the stage that they they they fit for. The for World's Fair, we had, for example, Solomon Hykes open sourcing what was it? Container use use container?

Something. His his Docker his Docker but AI thing. And, well, the the simple answer for that one is it was already ready to go. He just clicked the button to open source it.

So, like, there's ways to just not to put things on Rails such that you're actually not taking that much risk. Like, technically, Gemini three Flash was, like, the the the last one before the GA was the last preview before GA was announced on our stage, but the announcement was kind of already out and the blog post was already published. Logan just came up and hit tweet. Right?

Like so I don't know. I think that is more authentic than than than Apple Keynote where everything's prerecorded. So it injects some liveness. I do think so so so my best talk in my speaking career, which would be which is over now, but in my time, I live coded a whole React clone.

Like, I I started from an empty JavaScript file, and then I wrote a React framework, mini framework, and I ran it, and it and it compiled and ran. Wow. And that had risk of failure, and that is more engaging to to developers. And I actually didn't almost fail, and the most gratifying thing was someone in the audience caught my mistake and corrected it live Seriously.

And it and it and then it ran. So it it showed that it it not only was a live demo, I actually did fail and someone caught it, and it it showed that it wasn't on Rails. It was like it very well could have gone wrong. So I think it was, like, very engaging from from the paid attention to catch it.

And it was paying attention. Like, this is the usually, you're just looking at a wall of code, you're like, I don't I'm not paying attention. Five minutes of so. Yeah.

And the guy caught. Yeah. The guy caught. Well, I mean, so what you do is I I I had I had, like, checked checkpoints.

You can just kinda roll back to a checkpoint that you that you know works. Anyway but, like, I don't I think to some extent, it's unavoidable. Like, we want polish. We want production quality.

But you just have a variety of venues for which for which people can can do stuff. I think that's okay. What I what I'm actually more keen on is, like, stealing ideas from other conferences and being maximally helpful to the engineer. Mhmm.

I think we are actually not even doing enough. I'll give everyone twenty minute talks, and you can do whatever you want in in inside of it. We can we give people one and a half hour workshops. But is that really the thing?

And we'll probably never do twenty hour courses. That is that's the domain of people with their own independence teaching platforms. But, like, you know, I I I would like to innovate more there. What we have now is is a is a formula that's basically copy pasted from every other conference I've been to.

But, like, what's the next thing? I I think I think I'm in a position to try to figure out what that is. And if if you have ideas, let me know. I I think it's a really interesting idea.

I mean, I think about this like, I get bored of of AI tinkerers every year. Right? And so, you know, in year one, we had these, like, community demos. In year two, we did these, like, fireside chats with community demos.

In year three, we did a ton of hackathons as well. And, like, there's been a new format of it every year, and it and it kind of represents the community going through, you know, okay. Everybody's, like, an indie person just hacking on stuff. So, like, okay.

Well, some real companies are using this. Let's bring their CTOs in and to, like, okay. Well, let's lock a bunch of people in a room for a whole day and, like, try and hack on something. And I I haven't figured out what this year is gonna be about.

Like, Luke's doing the far side stuff, and and I'm just, like, looking at special projects, but there has to be an evolution as people learn and AI spreads. And I guess you'll have your own version of that for AI engineer. Yeah. Yeah.

I don't have a strong answer apart from, like, I think I've I've become very opinionated against hackathons. Yeah. I'm not like we we just won't run them ourselves, but we do partner with Cerebral Valley, and and they do hackathons. The truth is just a lot of people cheat in order to win.

And then so the incentives are all wrong for hackathons. We want to do like, start up battlefield is is one format that we're thinking about. So a pitch contest. It's just recognize the fact that you may have been working on something for one day or three months or a year.

Doesn't really matter. Just impress people. And who cares whether it was started on that day. If you could measure the outcome as well so there's two that I'm kinda thinking about.

I'm not gonna do any any more, like, traditional hackathons, but, you know, if you can measure an all of the different entries so one one thing you do is, like, Minecraft hackathon, and then you just say, whoever mines the most diamonds by the end of the day wins. And then, you know, you gotta, like, build an AI system to do that. Or you could you know? And then you can make it more interesting to, like, Polymarket, a Cauchy sort of thing as well.

Like, whoever you know, you stick all the trading bots in a in a kind of pool, and then whoever makes the most money by the end of the day has won. Things like that. And that's more like sort of Battlefield, I think, because you give people a bit more time to kinda prepare. Actually, it doesn't even matter.

Why why even limit it to twenty four hours? Just yeah. Yeah. Right.

So so RKGI is is one version of this. Right? It is effectively a always on hackathon that people can and and they they sort of do an annual cutoff and write up whoever is the best. And, yeah, I mean, do it.

I I just I I don't know if that's what helps people at work. Yeah. It is a nice demonstration of abilities, and it's fun. I don't know if it's what helps people at work.

And so so I do think, like, the sort of creator economy or the influencer sort of cycle means that you you do want to sort of put all your chips on people who are founders effectively of of really hot startups. Okay. I think we have, like, couple minutes. Yeah.

Is there one one last rogue? What's the rogue Louie, you've been brought on here for rogue questions. It's the rogue questions. Tell me how I Tell me what to do with fabric.

So Louie is responsible for our our after party. We ended up spending, like, 30 k on it. No. 40 k?

Can't pay you back. No. No. No.

You're doing no. You're you're fine with that. Like so okay. My impression is that every European conference, much more than Americans, has a big after party.

And, actually, like, at a nightclub where people dance and there's a DJ and people drink. And this is, like, not a thing in The US. But every like, I've when I've been to, like, Amsterdam conferences and other countries, they always have a big one. And I wonder how important it is.

And may like, I I don't know. I so I I very much judge things from my personal perspective, which is I don't like loud music. I do want to chat. I don't like a loud room where where I have to yell at the top of my voice, and my my my watch gives me crap for being in an elevated noise environment.

How do we make it good? I I don't have the answer, but I I feel like some of the appeal I mean, a huge part of this is networking as well. Right? The talks are great, but also Okay.

There are so many people here that I I just have seen on Twitter or whatever. Right? Yeah. Yeah.

Yeah. So that's it's just another environment to, like, bump into people. Yes. I think that something that we could engineer more intentionally is the networking.

We have done the networking app. Never people never use it. No. Me and Ben fight the most on he wants the app.

I don't want the app because there's no such thing as a good conference app. Like, this cannot cannot exist. The conference app is Twitter and then the or WhatsApp. Whatever.

Yeah. But, like, I I we could do we I've heard of things like blind dating or matching where, like, okay. You're a rag guy. I'm a rag guy.

We'll just meet and talk rag here. Yep. At this point, I'm just going to ruin my conference. You know?

We could do that. We just haven't done it yet. But, like, it but, yeah, one of the in the businesses I'm inspired by is actually a UK publicly listed company called Informa. You've probably never heard of it, but you've heard of the conferences that they run.

It's $13,000,000,000 in valuation. Wow. And they only run conferences. They do everything from beauty to retail to to gaming to security.

So they do GDC and Blackhats, two of the my role model conferences. And so the kind of conferences that we run here is their second largest business. The first largest business is four times the size, and they'd have no speakers. You only show up, and you do business.

You you network. What's it's just letting No speakers. But what they must have a theme or something. Yeah.

Beauty. Beauty. And there's pure Everyone who works in fitness, beauty, you know, whatever, you you come here. You I'm a buyer.

You're a seller. Let's talk. That booth's, like, pure It was Yeah. Pure commerce commerce.

They I don't know. They take a cut. They whatever. And that's a five times larger business than whatever we're doing here.

And and I I do think, like, it obviously has value to because they they keep showing up and they keep growing. And I I wonder if that's that's something that we can be valuable for people. Right? That that one is, like, the people that you want to meet, how can we get you to meet them?

Some there has to be some kind of mutual opt in, but maybe there doesn't have to be because there needs to be serendipity. Because the second part is the people that you didn't know you wanted to meet that you run into and you go, okay. Now we're best friends now. Yeah.

Is it a yeah. Maybe it's, like, down to is it a one-sided or a two sided or a three sided marketplace? Here, we got, like, speakers, attendees, and, like, maybe the conference is more like a one-sided marketplace where everybody is a Yeah. So Yeah.

I I do think that you need to dissociate. Right? Because the double opt in slows everything down. Right?

If if I must select and you must select and we must match, and it then that slows down the number of potential connections. You need to be one-sided in some in some areas. One thing that we've done is the leadership lunch in the last two two AIEs. The one that happened here was very good.

So we brought a speaker. We basically set the theme of, like, three things that everyone's supposed to discuss, and then we had people voice out their individual experiences. And then we broke, and then everyone was was peer to peer. But, like, that was effectively an unconference with food, and I think I like that.

It's just that nobody pays money for an unconference, I feel like. I mean, they do, but it's you know, pay pay pay money for you to speak with no plan and no schedule. It's like a weird I I I don't think I've been to a good conference. Yeah.

Actually, I'll put it there. I haven't been to a good conference ever. Yeah. Okay, guys.

I think that's Yeah. That's a that's a wrap on A bit rambling. From no. That was amazing.

Yeah. Yeah. So, yeah, thanks so much, Sean. Thanks, Louis.

Thanks. Okay. Thank you. Sorry.

I think I think what you're doing on the podcast is great. You know, I I I told you this. And, yeah, keep doing innovating in formats. Right?

I think you're you're doing more sort of essay things now. That's good. Yeah. I I I think it also improves your learning.

If if your goal was to learn the business of DevTools and DevRel founding things, stopping every 10 episodes to do, like, okay. Well, what have we learned about this so far? That's great. And I should do that.

I haven't done it. I need to start. Amazing. Thank you.

Everyone go to the next AI engine as honestly. I think it's like kind of the dev de facto. I would say it's like the de facto DevTools conference as well. So it's not Yeah.

But something interesting that Malta said was like, the successor to JSConf would not be another JSConf. It would be an AI conference today because that's a, like, hot area of development. So I never thought about it that way, but it makes sense after he said it. Yeah.

Brilliant.