# Podcast Crossover: AIE, AGI, frontier lab strategy with ​ ⁨@matthew_berman⁩ and @swyxtv Latent Space · 2026-07-10 Shawn 'Swyx' Wang, founder of the AI Engineer conference, tells Matthew Berman how he seized the industry shift by buying ai.engineer and partnering with a veteran conference organizer, crediting Andrej Karpathy's early endorsement. He argues Etched's ASICs are a natural next-gen bet for transformer inference, not an NVIDIA disruptor, and that Fable 5's slowness and cost signal the end of the LLM scaling era—making model efficiency the next problem. Swyx interprets OpenAI's reported 5% equity offer to the US government as a pragmatic multi-turn negotiation, likening it to Singapore's Temasek model, but warns against premature utility regulation. He pegs his own P(doom) at ~5% over 50 years, rejecting near-term doomerism as egotistical. For founders, he advocates building 'agent labs' that solve specific customer problems (e.g., for lawyers, dentists) rather than betting on model routing, which he dismisses as a marketing line that fails to exploit a single model's full stack as deeply as frontier labs do. ## Questions this episode answers ### What is Swyx's probability of doom (P(doom)) from AI, and how does he frame the timeline? Swyx attaches a timeline to P(doom): over 50,000 years (roughly the span of Homo sapiens), he estimates a ~90% chance that AI leads to human extinction or displacement. On 10 years, near zero; on 50 years, about 5%. He believes it's egoistic to assume we live at the exact moment everything ends, but that the default is low alignment probability, so we should engineer safety. [20:59](https://addtry.com/c03a27ca-94e1-4b82-9353-d0421137bdcc?t=1259000) ### How did Swyx come up with the idea for the AI Engineer conference and what was the hardest part of the first event? Swyx noticed trends: frontend, cloud, data engineering all got dedicated conferences, and AI would be next. He bought ai.engineer and planned it with partner Ben; the kickoff was endorsed by Andrej Karpathy. The hardest part was convincing people to show up with no history, and also getting a big lab (OpenAI) to attend, which established the value of a neutral ground where labs compete. [0:17](https://addtry.com/c03a27ca-94e1-4b82-9353-d0421137bdcc?t=17000) ## Key moments - **[0:00] AIE Origins** - [0:17] Shawn Wang bought ai.engineer and started the AI Engineer conference after seeing AI engineering would become a professional field like front-end and data engineering. - [2:25] AIE's value proposition is getting frontier labs like OpenAI to show up and compete on a neutral, even playing field. - **[3:06] Chip Startups** - [3:13] Etched is a next-gen ASIC optimized for post-transformer workloads that didn't exist when Cerebras started, says Shawn Wang. - [4:13] Shawn Wang: 'It's a pretty good bet, man.' - **[4:38] Fable Review** - [6:16] Anthropic's Mythos and Fable rollout delays were not due to chip shortage but planned months ahead, says Shawn Wang. - **[7:41] OpenAI Equity** - [8:01] Q: What does Shawn Wang think about OpenAI reportedly offering 5% equity to the US government? - [9:54] Regulating AI as a utility in 2025 is too early, Shawn Wang argues, comparing it to trying to regulate electricity while Edison was still inventing. - **[10:27] AI Regulation** - **[12:39] AGI Safety** - [12:40] Shawn Wang's p(doom) over 50,000 years is ~90%, but over 10 years near zero; over 50 years he estimates 5%. - [13:34] Shawn Wang: 'We don't have a right to exist any more than anything else that went extinct before us.' - [14:24] AIE positions itself politically between EAC (effective accelerationism) and deceleration movements, with engineers in control using guardrails and evals. - **[16:21] LLM Overhang** - [16:35] LLMs enable recursive self-improvement but stay within known distributions; true unknown discovery still needs human researchers, says Shawn Wang. - [19:07] Shawn Wang predicts Fable may be the end of this LLM era due to slowness and cost, requiring new architectures like thinking machines or SSMs. - **[21:42] Agent Labs** - [22:08] Shawn Wang advises founders to become 'agent labs' for verticals, as the outsourced AI layer that adapts new capabilities rather than getting wiped out. - **[25:02] Model Routing** - [25:02] Q: Is this the era of model-agnostic companies? - **[27:24] Wrap** ## Speakers - **Swyx** (host) - **Matthew Berman** (guest) ## Topics Hardware, Agent Platforms, Language Models ## Mentioned AWS (company), Anthropic (company), Cerebras (company), Cognition (company), Etched (company), Grok (company), MatteX (company), NVIDIA (company), Novas (company), OpenAI (company), xAI (company), Claude (product), Cursor (product), Fable (product), GPT (product), Opus (product), ai.engineer (product) ## Transcript ### AIE Origins **Matthew Berman** [0:00] I don't even know how to start this. You have 30 minutes,right? **Swyx** [0:03] Yeah. **Matthew Berman** [0:07] 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** [0:17] Yeah. I'd 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'd 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. And so I bought the domain ai.engineer and planned this conference, and I wrote the blog post, and it really kicked off when Andrej 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** [1:09] 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** [1:14] 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 React to Thought, which is the best React conference in 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** [1:35] What was the most difficult part? First, during your first AIE, and then this one? **Swyx** [1:40] Yeah. I think the first one was just convincing people that this is even worth showing up to, because, like, first of anything, you don't have any history. Yeah. Like, now people have word of mouth, they have the YouTube channel, it's all good. I think the very first one, you got nothing. And the people just had to take on a leap of faith of like, well, my reputation, Ben's reputation, come together, let's do something interesting. I don't know,right? And it was also 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 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 sort of like an even playing field, but you're just not going to get that in, like, a dev day or a code with Claude. **Matthew Berman** [2:49] 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 way. **Swyx** [3:02] I just did a one-hour chip session with Etched, which just came out of stuff this week. ### Chip Startups **Matthew Berman** [3:06] So good. Yeah, yeah, yeah. I just saw that. What do you think about Etched, and are they going to disrupt NVIDIA? **Swyx** [3:13] 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 be continuing to build GPUs. They're going to. **Matthew Berman** [3:29] Is Etched more akin to a Grok or a Cerebras? **Swyx** [3:33] Yeah. I have been pitching them as, or not pitching, but framing them to friends as next generation Cerebras, next generation Grok,right? Cerebras started 10 plus years ago, not their IPO. That process in form factor is relatively well established. What Etched and MatteX and the other new gen chips are doing are much more dedicated to, like, basically everything that we know post-transformers, post-ChatGPT, and optimizing for that workload. So it completely makes sense that the Cerebras, Groks, and Novas of the world didn't optimize for it because it didn't exist back then. **Matthew Berman** [4:05] Yeah. So, you know, when you make custom chips, you run the risk that the model architecture changes drastically due to some kind of innovation,right? **Swyx** [4:13] Yeah. They don't care. **Matthew Berman** [4:13] They don't care. **Swyx** [4:14] Yeah. The ChatGPT, like GPT-3.5-ish 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 4.0 is still being used,right, by some folks. Because they just don't move. Like, once the thing works, it works. Like, don't touch it. ### Fable Review **Matthew Berman** [4:38] Well, allright. I want to talk about something new. And you've been obviously quite busy this week, but Fable 5 is now back. Have you had a chance to. **Swyx** [4:46] They released it for us. **Matthew Berman** [4:47] They did. They released it just for you. I don't think. Maybe you tell me. Have you even had a chance to. **Swyx** [4:51] Yeah, yeah. So there's an Easter egg on the website that is built by Fable. It's a game that's you. **Matthew Berman** [4:58] ai.engineer. **Swyx** [4:59] Yeah. **Matthew Berman** [5:00] Okay. You want to give any hints as to what to look for for the Easter egg? **Swyx** [5:04] Look for a globe icon that bounces when you hover over it. **Matthew Berman** [5:08] Allright. So you've played around with it. Does it feel the same as those couple days when it was first released? Because a lot of people are talking about it being, you know, nerfed, and there's kind of a lot of floating around on Xright now. **Swyx** [5:19] I see, I see. Yeah. People are very quick to complain about Anthropic. I think the, I don't know if, look, I don't know if it's nerfed since the first Fable launch. I haven't, I don't, you know, I don't have hard numbers on that. I know that there are false refusals happen and still happen a lot, and those are really annoying when they happen,right? The downgrades to Opus. **Matthew Berman** [5:43] I haven't had that happen to me once. **Swyx** [5:44] Okay. **Matthew Berman** [5:45] I haven't had a reroute happen to me once yet, which is. **Swyx** [5:48] Maybe you're not spicy enough. **Matthew Berman** [5:49] Maybe. I'm definitely ripping tokens though. So if it were going to happen, it would be there. **Swyx** [5:54] But maybe you need to talk about chemical weapons or something, then it'll go. **Matthew Berman** [5:58] Not a lot of my projects need that. **Swyx** [6:00] So, no, I think, 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** [6:16] A lot of. **Swyx** [6:16] I think maybe the other part about, you know, there was a narrative like a month ago where, like, people were like, oh, Anthropic is limited on chips, limited on compute capacity. That's why they're not rolling out Mythos and Fable. That's clearly not true. **Matthew Berman** [6:31] Well, now it's not true,right? I mean, they. **Swyx** [6:33] It probably wasn't true a month ago. It didn't change that much. Like, they know their roadmap. **Matthew Berman** [6:36] Maybe not a month, but like, you know, the Cerebras, sorry, not the Cerebras, they just partnered with, I think it was AWS, and they landed a bunch of chip deals, a bunch of compute bandwidth deals. **Swyx** [6:48] Yeah. But you think they, like, weren't planning this for, like, months before? Like, they knew this. **Matthew Berman** [6:53] Oh,right. **Swyx** [6:54] I was just saying, like. **Matthew Berman** [6:55] XAI, that was the big one,right? **Swyx** [6:57] Yeah. Anyway, I'm just saying, like, these things don't happen just overnight. And, like, they probably were already, like, the Mythos 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. **Matthew Berman** [7:13] 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. It's happening again, to be fair, with Fable, because it just uses so many more tokens. But two months ago, it just seems like they could not get enough compute to serve the models that they have. **Swyx** [7:40] Yeah. Well, only they know. ### OpenAI Equity **Matthew Berman** [7:41] Yeah. Allright. Let's talk about the other frontier lab, one of the others. 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 is offering 5% equity stake to the US government. **Swyx** [8:00] Yeah. **Matthew Berman** [8:01] Like, what do you think is going on there? **Swyx** [8:03] I saw that headline. I don't know where, I don't know anything about the sources or anything. 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, like, factual to say. **Matthew Berman** [8:21] Yeah, yeah, yeah, yeah. **Swyx** [8:22] And I also, but I also just generally do think the people of the United States or the people of any country where, you know, like, all such, like, frontier intelligence is created, they do need to have a share or a say in some upside. Otherwise, you have, like, a permanent underclass and 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 ChatGPT Pro subscription or something. But, like, giving the government 5% to have the government have a stake in the success of OpenAI as a leading lab in the US does tend to make sense. 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 government's, like, nationally critical, important companies where the government owns a stake. What's so different? **Matthew Berman** [9:31] Do you think they're going to treat it more like a utility? Because. **Swyx** [9:35] Oh, yeah. **Matthew Berman** [9:37] So there's, like, a few different kind of 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, obviously, we're speculating now. What do you think is most likely to happen? **Swyx** [9:54] 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 lighting and all that, and you immediately try to regulate it. Like, no, you probably wait 50 years first. So it's, like, a little unfair to be, like, in 2025, you know, like, four, three, four years after ChatGPT to be like, no, like, we should regulate it now, like, a utility, because 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** [10:27] So 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 models more quickly? You're going to step back on regulation? ### AI Regulation **Matthew Berman** [10:37] Speculate with me, Swyx. **Swyx** [10:38] Yeah. So, like, I'm not a citizen, so I do tend to be careful about what I say politically. I do think that it is interesting to think through people optimizing for this administration versus, oh, there's an election every four years. And, like, actually, this is a multi-turn game, not a single-turn game. So maybe don't only think in one turn. **Matthew Berman** [10:58] Yeah. 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 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 is going on? **Swyx** [11:26] I think the White House thinks it is consistent, and it is probably better than what the states tried to do themselves a year plus ago with SD47, all those things. So, yeah, I mean, yes, on a national level, do it. And beyond that, I think the 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** [11:54] Yeah. **Swyx** [11:55] But I think, honestly, I want to read the book. **Matthew Berman** [11:58] An initial gambit that is quite extreme, then you roll it back, but you still. **Swyx** [12:01] It's happened before? **Matthew Berman** [12:02] Yeah. I mean, it's literally he wrote about. **Swyx** [12:03] It works? **Matthew Berman** [12:04] 40 years ago, however long it was, 30 years. **Swyx** [12:06] It works? **Matthew Berman** [12:07] It works. **Swyx** [12:07] And now, wouldn't it be funny if, like, this is actually how humanity is saved from Skynet, is the art of the deal? Because there was no other, there was no other mechanism under capitalism to control companies. **Matthew Berman** [12:25] But do you think the government is more worried about Skynet, or do you think they're more worried about other countries doing distillation hacking or using the model for cyber warfare themselves? What do you think? **Swyx** [12:36] I think there's multiple reasons that all align. So. **Matthew Berman** [12:39] What are you most afraid of? ### AGI Safety **Swyx** [12:40] When they all align, it's hard to tell motivations because they all point to the same answer. **Matthew Berman** [12:44] Yeah. **Swyx** [12:44] So, like, there's no, like, there's no point ranking them because they're all aligning. Well, I am a doomer in terms of, like, people talking about PDUM of, like, I've met people with PDUM of zero, PDUM of less than 1%. Mine is, like, closer to 90 on a scale of 50,000 years. I think people, PDUM should be attached to timeline. **Matthew Berman** [13:08] Yeah. **Swyx** [13:09] And so I think. **Matthew Berman** [13:10] 50,000 years. **Swyx** [13:12] Which is a rough approximate time of the Homo sapiens,right? 50,000 to 400,000 is, like, our span. And, like, you know, like, we don't have aright to exist any more than anything else that went existing before us. And if we're birthing a new life form, it is reasonable to expect that by accident, you know, we wouldn't, we may not do so well compared to it. I mean, how are trees doing versus human civilization? And, like, when we move much faster than trees, these things move much, much faster than us. I think it absolutely bears some thought. **Matthew Berman** [13:44] Yeah. I mean, what you're saying is basically the chance of actually aligning these models in the long run is very low. **Swyx** [13:52] I think our default assumption should be that it's low, and if we miss in theright direction, we will have missed in the safe direction anyway. **Matthew Berman** [13:58] Yeah, that's fair. **Swyx** [13:59] So I think one thing that I often don't comment about is where AIE is in relation with EAC and the decel movement. And it's, like,right in the middle. So I talked about this at the last World's 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 EAC is. **Matthew Berman** [14:23] Yeah. **Swyx** [14:24] And that's why putting guardrails and systems and fine-tuning and doing evals are important and an AIE versus with an EAC, you just talk about AGI timelines all day long. **Matthew Berman** [14:35] Right. **Swyx** [14:35] Because you don't care. I do think that some measure of having the engineers in control of things, watching things, monitoring chains of thought probably is theright path for humanity. So, like, the light cone of it, like, you know, we live, as far as you know, we're the only life form in the observable universe. Life is very fragile, and, like, we shouldn't assume that by default we get theright to continue existing unless you engineer it. **Matthew Berman** [15:02] If you had to shrink the timeline from 50,000 to, let's get your PDUM on 10 years and 50 years. **Swyx** [15:09] Oh, yeah. I mean, PDUM in 10 years is near zero. **Matthew Berman** [15:13] Well, that's good. That's good news. **Swyx** [15:14] Yeah. 50 is, like, that's, like, the end of our lifetimes. And, like, I think, I don't know, I'm going to just throw out 10. No, 10's too high. 5%. And, again, I think on the timelines, more of what you see in Foundation, you know, the TV show? **Matthew Berman** [15:36] Yeah. **Swyx** [15:38] Or Dune. **Matthew Berman** [15:38] Great show. **Swyx** [15:39] I think 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 LLMs 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 egoistic. **Matthew Berman** [16:13] That's pretty good. **Swyx** [16:13] That's very egotistical. **Matthew Berman** [16:13] Yeah, that's what I said. **Swyx** [16:13] But you're the main character, really? **Matthew Berman** [16:15] Yeah. **Swyx** [16:15] Like, you know, who's to say you're not, like, you know, in prehistoryright now as far as the future people are concerned? **Matthew Berman** [16:21] Yeah. Do you think large language models ### LLM Overhang **Matthew Berman** [16:27] are enough to lead to recursive self-improvement, thus leading to some. **Swyx** [16:31] We have some form. **Matthew Berman** [16:32] Next architecture that might be AGI? **Swyx** [16:35] Oh, I see. That is a very nuanced question. I think actually my answer is no on that one. We, yes, LLMs enable RSI. RSI is really just, like, is it recursing? And yes, it is recursing. We have a whole auto-research track yesterday covering that stuff. 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 distribution of stuff we already know. **Matthew Berman** [17:04] Yeah. **Swyx** [17:05] So discovering true unknown unknowns, having real innovation, having a real sense of old 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 DoorCash is saying is absolutelyright. Like, current learning paradigms. **Matthew Berman** [17:31] What part? **Swyx** [17:32] Data efficiency is the next problem. We had a data track on the same thing. 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. **Matthew Berman** [17:58] So do we have to try to fit large language models, the compute that it requires into kind of the human box? **Swyx** [18:05] No, no, no. **Matthew Berman** [18:06] Why does it have to be? Well, if it's not exactly how humans learn, then it's not theright way. **Swyx** [18:11] No, I absolutely agree with what you're saying there, which is, they call this the sour lesson,right? Like, every time you try to make a human analogy to machines, you probably fail because machines develop very differently from humans. **Matthew Berman** [18:20] Yeah. **Swyx** [18:21] So, no, 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 it is super inefficient. Like, that is something that we know is an unmitigated negative. So let's make it more efficient and better. And that means that, you know, in order, instead of 2,000 examples to learn one thing, what about 20 examples? What about two examples? And that scales a lot more. And that means, you know, we can actually get to a point with continual learning that we can actually have agents that adapt and build up a real-world model. Otherwise, we're always stuck to the pre-train, post-train paradigm that is probably hitting some kind of limitright now. **Matthew Berman** [19:07] Yeah. **Swyx** [19:07] Like, I genuinely do expect Fable to be the end of this era of LLMs because you can't, like, I already told you about the slowness. Do you agree with the slowness, by the way, that Fable's slow and expensive? **Matthew Berman** [19:19] I've actually, so I've been using it in Cursor, and it feels a little bit faster than when it first came out a couple weeks ago. **Swyx** [19:24] Yeah. **Matthew Berman** [19:25] But it is slow. It is slow. **Swyx** [19:26] I mean, look, Anthropic's working on better inference. Yeah. No, train through constraint, but it is slow. And, like, therefore, you just won't use it for some things that you know don't require Fable. **Matthew Berman** [19:36] Yeah. You also don't want to pay Fable prices for everything. Why would you? **Swyx** [19:38] Yeah. Like, even, like, I'm trying to live in an infinite budget world, but the budget I don't have is time,right? Like, there's the limiting budget that everyone has. I don't care if you're working at a frontier lab. You still have time as a limiter. So, 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 it is, maybe it's, like, thinking machines and stuff. Maybe it's together AI's SSM stuff. Whatever it is, we don't have it yet. **Matthew Berman** [20:18] I've seen a lot of, and I've thought a lot about this. The model capability overhang still seems very real, even from the previous generation, Opus, GPT-5.5. The amount of value that can be extracted from those models still seems, at least to me, to be critical. And now we have a whole new generation of model that we're even going to get more model overhang from. Like, what do you, do you agree with that? Do you think. **Swyx** [20:46] Oh, yeah, it's great. **Matthew Berman** [20:47] We just need to keep building the tools around them? **Swyx** [20:48] 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. So I'm very pro that. 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 built out, I don't need it anymore because the next model has got it from just, like, a single prompt. So I'm going to throw it out. But, like, 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. Yeah. **Matthew Berman** [21:42] 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. ### Agent Labs **Matthew Berman** [21:53] 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** [22:08] Yeah. My answer to this is two words, agent lab. I like the two-word solutions. I like AI engineer, I like agent lab. So I have a piece on this if people want to see. It's on late. **Matthew Berman** [22:18] Where do they find it? Latent space. **Swyx** [22:20] Yeah. Latent.space/agentlabs. **Matthew Berman** [22:22] Okay. **Swyx** [22:23] 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, whatever the new model is, I will be the AI guy for dentists or for lawyers or for finance people or for coders, whatever. Then you will just build that lasting brand of, 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** [23:02] Isn't that betting against the generalization of the models and the capability of the labs, frankly? **Swyx** [23:08] No, yeah. So let me think about this. So no, but maybe yes. So 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, literally, like, I am the AI layer for lawyers. I'm your outsourced tech team for lawyers. Okay, my old business model's 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** [23:46] Yeah. **Swyx** [23:48] So, like, there will always be capability overhangs. They may not stay still. And so you got to be nimble. But the CRS of the world, the cognitions of the world, the cursors of the world, the Decagons and 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 frontier labs on, like, the latest models. But in terms of solving user customer feedback, you know, I've been inside of Cognition for the last six months. There's no one else. 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 need the Microsoft Teams Zero integration? Got it. You don't use GitHub? You use this, like, weird org thing that is like a fork of Atlassian's Bitbucket when it was open source? Got it. Claude's not going to do that. **Matthew Berman** [24:40] I mean, let's continue on that. So you've been at Cognition for six months. **Swyx** [24:45] And you had the food. It scores well on Foodbench. **Matthew Berman** [24:48] It is a very good food. **Swyx** [24:49] Foodbench. **Matthew Berman** [24:49] High on Foodbench benchmark. **Swyx** [24:51] We didn't score better on other benchmarks, but food is a good start. **Matthew Berman** [24:54] But you just have to invite me back. **Swyx** [24:55] That's how you got to run the benchmark again. Do you think, sorry, I lost my train of thought. Oh. ### Model Routing **Matthew Berman** [25:02] Agent labs. **Swyx** [25:03] Yeah. So a lot of discussion lately has been around token budget and really token maxing being unattainable to most people. Do you think that's actually a huge value to the non-frontier lab, agent lab companies? So the cursors, the Cognitions of the factories of the world, 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 kind of the model agnostic companies? **Matthew Berman** [25:34] Yeah. I think that is a reasonable conclusion to make, which doesn't mean that I think it. I think that is a lot, that is currently what everyone is saying because of this current debate. I do observe that in general, the big, big wins have just been going all in on one thing. And I'm old enough in tech to know, have seen this before with the CloudWave, where there've been companies that were all in on AWS or all in on GCP, and there have been other companies that were multi-cloud. They were always, you know, like, I'll use 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. **Swyx** [26:23] Right. **Matthew Berman** [26:23] So it's a really good way to be lowest common denominator of every model out there and completely miss all capability. **Swyx** [26:30] Yeah. You also don't have the economies of scale that an AWS has going all in, owning the entire stack. **Matthew Berman** [26:35] Yeah, yeah. So, yeah, 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 from Anthropic versus what Anthropic does with model labs or with agent labs. And I think, like, that's also an interesting discussion on that front too. Anyway, all wish to say, I do think that this routing thing is a marketing line. I wouldn't necessarily believe it so much as when I talk 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, exploit everything they give you. Or do you want to just stay superficial and only do check completions across 100 different models? **Swyx** [27:21] Yeah. **Matthew Berman** [27:21] Right? Which is more likely to win as an agent lab. ### Wrap **Swyx** [27:24] Allright. Well, Swyx, I want to say thank you very much. **Matthew Berman** [27:27] Yeah. **Swyx** [27:27] I appreciate the conversation. **Matthew Berman** [27:29] It's good to chat, man. **Swyx** [27:30] Yeah. **Matthew Berman** [27:30] I'm glad you could see me on my biggest day, on my biggest show. **Swyx** [27:34] Hell yeah. **Matthew Berman** [27:34] I think you've seen me in, like, the podcast, the office, and all those things. This is, I'm an in-person community guy. **Swyx** [27:41] I am genuinely impressed with what you've built here. You have a bunch of great, excited, enthusiastic people, not only giving talks, but also learning a ton and, yeah. **Matthew Berman** [27:51] Making connections, getting co-founders. I need your help on our YouTube. Our YouTube site. **Swyx** [27:55] Awesome. Let's do it. **Matthew Berman** [27:56] So, yeah, lots of learning to be had. **Swyx** [28:00] Thanks, Swyx. **Matthew Berman** [28:01] Thanks. **Swyx** [28:02] Very nice. --- This library is powered by PodHood (https://podhood.com), the podcast website platform.