Matthew Berman: I don't even know how to start this. You have 30 minutes, right?

Swyx: Yeah. Yeah.

Matthew Berman: This conference is incredible. There are so many people here. I want to start with the origin. How did you think to start an AI engineering conference? How did it come about?

Swyx: Yeah. I've been a speaker for many years before that and I also had seen people seize a moment where you can see like an industry shift. And I thought it was pretty clear. I had seen basically front-end engineering become its own professionalized field with dedicated conferences, dedicated influencers and tech stacks and all those things and I seen the same thing for cloud engineering and data engineering and all that and I was like it's very obviously going to happen to AI. So I bought the domain AI.engineer and planned this conference and I wrote the blog post and it really kicked off when Andre Karpathy also endorsed it and I think it's gotten more true over time which is the kind of way I like to see it. Gartner put it at the peak of its hype cycle two years ago. It's still hypy, still going.

Matthew Berman: Had you started a conference before? Had you ever done a conference before? You were just like okay I'm just going to do it?

Swyx: Well I didn't do it alone. I partnered with a friend who had done previous conferences. So basically from my conference speaking career, I just took the best of every conference I had been to prior and Ben was running Reactathon which is the best React conference on the West coast. So I just had him start AIE and I would do the content, he would do the logistics and that's how it goes.

Matthew Berman: What was the most difficult part during your first AIE and then this one?

Swyx: Yeah, I think the first one was just convincing people that this is even worth showing up to because like a first of anything like you don't have any history. Like people now people have word of mouth, they have the YouTube channel, it's all it's all good. The very first one you got nothing and the people just have to take a leap of faith of like well my reputation Ben's reputation come together do let's do something interesting I don't know right and it was also down not that much in terms of like the tech right like at mostly we were still talking about prompt frameworks and maybe a bit of RAG but there just weren't any agents like it it was very very primitive at the time. I think also the last thing was getting a big lab to come. And for us it was Logan at OpenAI at the time. And I think a big part of our value proposition now is that we can get all these labs to show up at a neutral place where they all compete on the same grounds, which is maximally beneficial for engineers and most competitive for the labs. And it turns out the labs like it. They like winning on a in on sort of like an even playing field. But you're just not going to get that in like a dev day or Code with Claude.

Matthew Berman: You know, you know what I really appreciate? There seems to be the entire stack of complexity here. Everything from kind of really practical everyday user kind of standard user and then all the way down to highly technical deep in the weeds.

Swyx: I just did a 1hour chip session with Etched which just came out of stealth this week.

Matthew Berman: So good. Yeah. Yeah. I just saw that. What do you think about etched and are they going to disrupt Nvidia?

Swyx: I don't think they want to disrupt Nvidia. I think the better framing is that all of inference is just so goddamn big that of course you're going to have ASICs for inference. You're going to and Nvidia is going to continue to build GPUs.

Matthew Berman: Is it is etched more akin to a Groq or a Cerebras?

Swyx: Yeah, I I have been pitching them as or or not pitching but framing them to friends as next generation Cerebras next generation Groq. Cerebras started 10 plus years ago. Now they're IPOed. That process in form factor is relatively well established. What Etched and Madrona (MatX) and the other new gen chips are doing are much more dedicated to basically everything that we know post transformers post RGBT (GPT) and optimizing for that workload. So it completely makes sense that the Cerebras of the world didn't optimize for it because it didn't exist back then.

Matthew Berman: Yeah. So, you know, when you when you make custom chips, you run the risk that the model architecture changes drastically due to some kind of innovation, right?

Swyx: Yeah. They don't care.

Matthew Berman: They don't care.

Swyx: Yeah. The GPT 3.5ish architecture has mostly stayed the same this entire time. It's a pretty good bet, man. Like even if there is a new model sometime in the future, the current workloads on existing models like 4o is still being used right by some folks cuz you just don't move. Like once the thing works, it works. Like don't touch it.

Matthew Berman: Well, all right. I want to talk about something new and you've been obviously quite busy this week, but Fable 5 (Claude 3.5) is now back. Have you had a chance to...

Swyx: They released it for us.

Matthew Berman: They did. They released it just for you. Have you even had a chance to...

Swyx: Yeah. Yeah. So, there's a there's an Easter egg on the website that is built by Fable. It's a game that lets you...

Matthew Berman: AIE engineer.

Swyx: Yeah.

Matthew Berman: Okay. You want to give any hints as to what to look for for the Easter egg?

Swyx: Look for a globe icon that bounces when you hover over it.

Matthew Berman: All right. So, you have you've played around with it. Do does it feel the same as those couple of days when it was first released? Because a lot of people are talking about it being, you know, nerfed and there's a kind of a lot of floating around on X right now.

Swyx: I see. I see. Um, yeah, people are very quick to complain and about Anthropic. I don't know if it's nerfed since the first Fable launch. I don't have hard numbers on that. I know that false refusals happen and still happen a lot and those are really annoying when they happen. The downgrades to Opus (older models). I haven't had that happen to me once.

Matthew Berman: Okay.

Swyx: I haven't had a reroute happen to me once yet.

Matthew Berman: Maybe you're not spicy enough.

Swyx: Maybe I'm definitely ripping tokens though. So if it were going to happen, it would be there.

Matthew Berman: But maybe you need to talk about chemical weapons or something and then it'll go.

Swyx: Not a lot of my projects need that.

Matthew Berman: Exactly.

Swyx: So no I think like look when it works it is extremely smart. But it's also very slow and I think definitely I don't see myself using it for everything because of that right like you only use it for the smart problems and that's probably the way they want it anyway.

Matthew Berman: I think I think maybe the other part about you know there was a narrative like a month ago where like people were like oh like Anthropic is limited on chips limited on compute capacity that's why they're not rolling out mythos (Opus) and fable (Sonnet). That's clearly not true.

Swyx: Well now it's not true right I mean they...

Matthew Berman: It probably wasn't true a month ago. It didn't change that much like they know their road map.

Swyx: They just partnered with I think it was AWS and they landed a bunch of chip deals a bunch of compute bandwidth deals.

Matthew Berman: But you you think they like weren't planning these for like months before like...

Swyx: They knew this.

Matthew Berman: I was saying like...

Swyx: xAI that was the big one right.

Matthew Berman: Anyway I'm just saying like these things don't happen just overnight and like they probably were already like the Sonnet rollout was genuinely safety related. That's why it was like limited. It wasn't because there's some secret conspiracy to like limit or ration compute. I still do think they were bandwidth limited or they were compute limited because if you look at their quota and how aggressive in reducing it and using it. It's gotten much better but especially two months ago, I mean you would burn through your quota in a matter of minutes. Um it's happening again to be fair with Claude 3.5 because it just uses so many more tokens. two months ago. It just seems like they they could not get enough compute to serve the models that they have.

Swyx: Yeah. Well, only only they know.

Matthew Berman: Yeah. All right, let's talk about the other frontier lab. Okay. Let's talk about OpenAI. I woke up I think it was this morning to the news, I guess rumored. I haven't I got to actually read more deeply into it, but OpenAI's offering 5% equity stake to the US government. Like what do you think is going on there?

Swyx: I saw that headline. Uh I don't know where I don't know anything about the sources or anything. I I think it kind of tracks like it's not within the outside range of possible things to do. I think OpenAI has been relatively more friendly with the White House than with other Frontier Labs. That's that's like factual to say.

Matthew Berman: Yeah. Yeah. Yeah. Yeah. And um I also but I also just generally do think the people of the United States or the people of any country where um you know like all such like frontier in intelligence is created they do need to have a share or a say in some upside otherwise you have like a permanent underclass in significant social distress and I don't know you know I've played around with the idea of like universal basic AI like everyone gets like a chat pro subscription or something but like giving the government 5% to have the government in have a stake in the success of OpenAI as a leading lab in the US does tend to make sense. Uh I come from a country where this is normal. I'm from Singapore. Temasek and GIC own significant portions of the Singapore economy and it works just fine. And actually a lot of our pensions and our savings and our insurance and whatever is all invested in governments the like nationally critical important companies where the government owns the stake. What's so different? Do you think they're going to treat it more like a utility? Because there's like a few different flavors of how this plays out. The government owns 5%. they give out distributions to US citizens or maybe they treat it much more like a utility and they start heavily regulating and can step in at any time. What do you think is most likely to happen?

Swyx: I think this thing is too volatile to treat it as a utility, right? Like imagine if like Edison was like, you know, working on his like electrical stuff and light and lighting and all that and he immediately tried to regulate it. Like no, you'll probably wait 50 years first. So it's like a little unfair to [laughter] like in 2025, you know, like four 3 four years after ChatGPT to be like no, like we should regulate it now like a utility because it's there's no more innovation to be had and you should just leave it to natural monopolies and charge like cost plus like I don't think that's we're there yet. I think we need to wait another few decades.

Matthew Berman: So what what is it? Is it pay to play? Is it bend the knee? Is it hey, we're going to give you 5% and you're going to allow us to release our our models more quickly? You're going to step back on regulation? Speculate with me, Swyx.

Swyx: Yeah. So, like I'm not a citizen, so I do tend to be careful about what I say politically. Um, I do think that uh it is interesting to think through people optimizing for this administration versus well, there's an election every four years and like actually this is a multi-turn game, not a single turn game.

Matthew Berman: So, maybe don't only think in one turn.

Swyx: Yeah.

Matthew Berman: Um, okay. So, then like let's continue down this track a little bit. The government, especially in the last few weeks, has been very hands-on deciding what models are publicly available, the timing of models, right? So, GPT 5.6 (5) was just announced, but they have to wait a few weeks to actually release it. Like, do you think this is the new standard? Do you think they're kind of scrambling to come up with a framework that's consistent across all of the AI labs? What do you think's going on?

Swyx: I think the White House thinks it is consistent and uh it is probably better than what the states tried to do themselves a year plus ago with SB 1047 (SB47) all those things.

Matthew Berman: So yeah, I mean yes on the national level do it.

Swyx: Um and beyond that I think the best I best I can say about this is this isn't new in terms of tactics in politics and this is the art of the deal.

Matthew Berman: [laughter] Yeah, it's Yeah. Honestly, I want to read the book and it's

Swyx: an initial gambit that is quite extreme, then you roll it back, but you

Matthew Berman: Yeah. I mean, it's literally he wrote about 40 years ago, however long it was, 30 years.

Swyx: It works.

Matthew Berman: It works.

Swyx: And

Matthew Berman: would it be funny if like this is actually how humanity is saved from Skynet

Swyx: is the art is the art of the deal? Because there was no other uh there's no other mechanism under capitalism to control companies. But there do you think the government is more worried about Skynet or do you think they're more worried about other countries doing distillation hacking uh or using the model for cyber warfare themselves? What do you think?

Swyx: I think there's there's there's multiple reasons that all align. So

Matthew Berman: when they when they all align it's hard to tell motivations because they all point to the same answer.

Swyx: Yeah. So like there's no like there's no point ranking them because they're all aligning. Um well I I am I am I am a doomer in like in terms of like people talking about pdoom of like le I've met people with pdoom of zero pdoom of less than 1% mine is like closer to 90 on a scale of 50,000 years um I I think people pdoom should be attached to timeline

Matthew Berman: um and so I think 50,000 years

Swyx: which is the rough approximate time of the homo sapiens right 50,000 to 400,000 is is our span and like you know like we don't have a right to exist anymore than anything else that went existing before us and if we're birthing a new life form um it is reasonable to expect that by accident you know we we wouldn't we may not do so well compared to it um I mean how are trees doing versus human civilization and like when we move much faster than trees these things move much much much faster than us I think it absolutely bears some thought

Matthew Berman: yeah I mean I mean what you're saying is basically the chance of actually aligning these models in the long run is very low.

Swyx: I think our default assumption should be that it's low and if we miss in the right direction, we will have missed in the safe direction. Anyway,

Matthew Berman: yeah, that's fair.

Swyx: So, I think um one thing that I often don't comment about is where AIE is in relation with EAC (e/acc) and the the DEL (alignment/decel) movement and um it's like right in the middle. So I I talked about this at the last WS fair keynote where politically we want to be pragmatic. You want to be optimistic, but you also don't want to be unconstrained optimistic, which is what EA (e/acc) is.

Matthew Berman: Yeah.

Swyx: Um and that's why putting guard rails and systems and fine-tuning and um doing eval are important and an AIE versus with an EAC, you just talk about Agi timelines all day long, right? Because you don't care. Um I do think that some measure of having the engineers in control of things, watching things, monitoring change of thought probably is the right path for humanity. So like the lyon of like you know we live as far as you know we're the only life form in the in the observable universe. Uh it it's life is very fragile and like we shouldn't assume that by default we get the right to continue existing unless you engineer it.

Matthew Berman: If you had to shrink the timeline from 50,000 to let let's get your pdoom on 10 years and 50 years.

Swyx: Oh yeah. I mean pdoom in 10 years is is near zero. Uh

Matthew Berman: well that's good. That's good.

Swyx: Yeah. 50 is like that's like the end of our lifetimes. Um and like I think I don't know. I'm I'm just throw out 10. No, 10's too high. Five%. Um and uh again I I think on the timelines more of what you see in Foundation, you know the TV show.

Matthew Berman: Yeah. Great show. Great show.

Swyx: Um or Dune. I think um when you when you think about the long evolution of life forms and history, you shouldn't be so sort of near-term focused because yeah, probably LMS are going to run out at some point and they're not AGI and okay, we have maybe another 30 years of AI winter or something and then like the next paradigm really is actually the thing or it takes another 3,000 generations like what like you know who are we to say that we happen to live in the exact moment that everything ends. Like that's very that's very egotistical like you're the main character really

Matthew Berman: like you know who's to say you're not like you know in prehistory right now as far as the future people are concerned.

Swyx: Yeah.

Matthew Berman: Do do you think uh large language models are are enough to lead to recursive self-improvement thus leading to some next architecture that might be AGI.

Swyx: Oh I see that is a very nuanced question. I think actually my answer is no on that one. Um we yes LLM enable RSI. RSI is really just like is it recursing and yes it is recursing on it's we have a whole auto research track yesterday covering that stuff. Um but it is limited in its recursion because it probably just explores things that have been explored before. So it's like not that far off from like the dist distribution of stuff we already know.

Matthew Berman: Yeah.

Swyx: So discovering true unknown unknowns uh having real real innovation having a real sense of world models that is still the domain of research and yeah we probably need something else. We did have a world models track here. It is not very well developed. Everyone knows it but clearly this needs to be fleshed out more because the stuff that Dorcash (Dwarkesh) is saying is absolutely right. Uh we like learn current learning paradigms.

Matthew Berman: What part?

Swyx: Uh data efficiency is the next problem. um we had a data track on the same thing. Um and you know like learning over a trillions of tokens in order to get to some form of human equivalent labor is very very inefficient versus what humans do. Humans learn on the order of like millions and they can already do like very useful things billions and you're a full working adult. So, do we have to try to fit large language models, the compute that it requires into kind of the human box? Why does it have to be well, if it's not exactly how humans learn, then it's not the right way?

Matthew Berman: No, I I absolutely agree with what you're saying there, which I call this a sour (bitter) lesson, right? Like every time you try to make a human analogy to machines, you you probably fail because machines develop very differently from humans.

Swyx: Uh so, no, I agree. I I agree with that. And I also think that it can still be an alien form of more efficient learning. Doesn't matter. We just know that it's super inefficient like that. That is something that we know is a unmitigated negative.

Matthew Berman: So let's make it more efficient and better. Uh

Swyx: and that means that you know if in order instead of 2,000 examples to learn one thing, what about 20 examples, what about two examples? Um and that scales a lot more. Um, and that means, you know, we can we can get we can actually get to a point with continual learning that we can actually have uh agents that adapt and and build up a real world model. Otherwise, we're always stuck to the pre-train postra (post-train) paradigm that is probably hitting some kind of limit right now.

Matthew Berman: Yeah.

Swyx: Like it like I genuinely do expect Fable (Claude 3.5 Sonnet) to be the end of this era of LMS because you can't like I already told you about the slowness. Do you agree with the slowness by the way that Fable's (Sonnet) slow?

Matthew Berman: I I've actually so I' I've been using it in cursor and it's it feels a little bit faster than when it first came out a couple weeks ago. Um but it is slow. It is slow.

Swyx: I mean look at topic (Anthropic) is working on better inference. Yeah. No train through straight but it is slow and like therefore you just won't use it for some things that you know

Matthew Berman: don't require Fable (Claude 3.5 Sonnet). Yeah. [laughter] You also don't want to pay Fable (Sonnet) prices for everything. Why would

Swyx: like even like I'm trying to live in an infinite budget world but the the budget I don't have is time, right? like this the this is the limiting budget that everyone has. I don't care if you're working at a frontier lab, you still you still have time as a as a as a as a limiter. So, um yeah, if that's like if a 20 trillion model is it, there's no 200 trillion, there's no quadrillion scale model, then that's kind of it as far as usability is concerned living in an infinite budget environment. So therefore, you just need something different. whatever is maybe it's like thinking machines stuff maybe it's together AI is SSM stuff whatever it is we don't have it yet

Matthew Berman: I've I've seen I've seen a lot and I've thought a lot about this the the model capability overhang still seems very real even from the previous generation Opus uh GPT 5.5 (GPT-4) the the amount of value that can be extracted from those models still seems at least to me to be uh uh critical and and now we have a whole new uh generation of models that we're even going to get more model overhang from what do you do you agree with that do you think need to keep building the tools around

Swyx: AI engineer exists in the white surface area between the peak capability and deploying it everywhere else right so the more model research peaks and spikes capabilities in one domain but it's not evenly distributed in all products yet that's where engineers have a job forever basically Um, so I'm very pro that. Um, I I I think capability overhang will exist for a long time. I think it does keep have waves of consolidation where you're like actually all this stuff I build out, I don't need it anymore because the next model has got it from just a single prompt. So I'm going to throw it out. But we do spring cleaning every now and then, like that's normal. And we build it up on previous gen model assumptions that then go away. And I don't think we should feel any attachment to the code. At the end of the day, we're all just trying to like serve customers better, do work cheaper, faster, easier.

Matthew Berman: Yeah. Let's let's talk about that a little bit because I've built a bunch of cool kind of one-off projects that one model generation later I was like, "Oh, well, that's kind of useless now." And then you you know, you have anthropic moving in different directions. They're moving up the stack to the application layer. If you're a founder, what are you know, what advice would you give to founders who are trying to build at the application layer? Where should they be spending their time? How should they think about the expanding model capabilities in each subsequent generation of model?

Swyx: Yeah, I my answer to this is two words. Agent lab. I like I like the two-word solutions. I like AI engineer. I like agentlab. Um so I have a piece on this if people want to see. Uh it's on where they find it laten.space/agentlabs. Um and basically the idea is that you always want to be the AI guys for your customers. Actually don't pick the solution, pick the problem. And if you're like okay I'm like whatever it is in AI whatever the hot thing is whatever the new trend is what the new model is I will be the AI guy for dentists or for lawyers or for finance people uh or for coders whatever. Um then you will just be build that lasting brand of we like we will build the products that do the last mile for you and fold in all the new functionalities and features that people discover into that. That is the sustainable thing.

Matthew Berman: Isn't that isn't that isn't that isn't that betting against the generalization of the models and the capability of the labs frankly?

Swyx: Uh no. Yeah. So let me think about this. So no but maybe yes. So I I don't know where I stand with regards to your question because the models do generalize and sometimes wipe out entire product categories. But I do think that if you are smart enough to like I literally I like I am the AI layer for lawyers. I'm your outsource tech team for lawyers. Okay, that my my old business model is gone. Fine. I'll make a new one with like whatever new capability overhang is created. What you're betting against is capability overhangs ever existing in the future. And I think that's a pretty safe bet to make.

Matthew Berman: Yeah.

Swyx: Um so like there will always be capability overhangs. They may not stay still. And so you got to be nimble. But the Sierras of the world, the Cognitions of the world, the cursors of the world, the decagons and um Harveys, these are all agent labs for their field. They can be trusted brand names to always apply the latest AI to their thing. And maybe they're a little bit behind the the frontier labs on like the the latest models, but in terms of solving user customer feedback, you know, I've I've been inside of cognition for the last six months. Um there's no one else. the the labs do not have 200 people dedicated to like you know being on call with you with Goldman Sachs going like okay guys what do you need we got it you you need the Microsoft team zero integration got it you don't use GitHub you use this like weird org thing that is like a fork of Elastian's Bitbucket when it was open source got it claw's (Claude's) not going to do that

Matthew Berman: I mean let's continue on that so you've you've been at uh Cognition for 6 months and you had the food it scores well on food bench

Swyx: it is a very good food bench high on food bench benchmark. Um

Matthew Berman: we need we need to score better on other benchmarks but but food is a good start.

Swyx: You just have to invite me back. That's how you got to run the benchmark again. Um do you do you think um uh sorry I lost my train of thought. Um

Matthew Berman: oh yeah so a lot of discussion lately has been around token budget and and really token maxing being unattainable to most people. Do you think that's actually a huge value to the nonfrontier lab uh agent lab companies? So the cursors, the cognitions of the fact (factories) of the world are isn't this kind of like prime time for them because first of all they're model agnostic. They actually have incentive to do model routing really well whereas the frontier labs don't. Do you think this is the era of uh kind of the the uh model agnostic companies?

Swyx: Yeah, I think that is a reasonable conclusion to make which doesn't mean that I think it. I think that is a lot is that is currently what everyone is saying because of this current debate. Uh I do observe that in general the the big big wins have just been going all in on one thing. Um, and I'm old enough in tech to know have seen this before with the cloud wave where there have been companies that were all in on AWS or all in on GCP. And there have been other companies that were multicloud there always, you know, like I'll use the the the Terraform, the whatever. And the simple argument is that if you route, if you pride yourself on routing, you will never exploit the full capabilities of one because you're not all in on one. You're not. Right.

Matthew Berman: So it's a really good way to be lowest common denominator of every model out there and completely miss all capability.

Swyx: You also don't have the economies of scale that an AWS has going all in owning the entire stack.

Matthew Berman: Yeah. Yeah.

Swyx: So so yeah, Asian (Agent) labs get discounts from every model provider. And that's also very interesting when people compare public pricing of like a discounted cloud code (Claude) from Anthropic versus what Anthropic does with model labs. Oh, with with agent labs and I think like that's also an interesting discussion on that front too. Um anyway, all all which to say I do think that this routing thing is is a marketing line. I wouldn't necessarily believe it so much as when I talked to the top tier agent builders because they're all about maximizing and fully exploring like the complete prompt surface area and optimization area and tool use and caching and god knows what else in there, right? Like like exploit everything they give you

Matthew Berman: or do you want to just stay superficial and only do chat completions across a 100 different models? Yeah. Right. Which is more likely to win as an agent lab.

Swyx: Yeah. Right. Which

Matthew Berman: All right. Well, uh, Swix, I want to say thank you very much. I appreciate the conversation.

Swyx: It's good to chat, man. I'm glad you could see me and on my biggest day on my biggest show. I think you've seen me in like the podcast, the the office and all those things. This is I I'm an inerson real community guy.

Matthew Berman: I I am genuinely impressed with what you built here. You have a bunch of great excited, enthusiastic people, not only giving talks, but also learning a ton.

Swyx: and and uh yeah, making connections, getting co-founders. Um I need your help on our YouTube, our YouTube stuff.

Matthew Berman: Let's do it.

Swyx: Uh so yeah, lots lots of learning uh to be had.

Matthew Berman: Thanks, Swix.

Swyx: Thanks. Very nice.