I have seen demos that are 10,000 [music] tokens per second as well as hundreds of thousands of tokens per second. I think that we'll still have the 10 [music] 20 trillion parameter model runs. People who are somewhat preaching small models, they're coping because they don't [music] have GPUs anyway. So, uh cry harder. Yes, you're spending a billion tokens a day, but you goddamn better be actually having real humans be using your app because it's super super easy to code a lot of slop that makes your app completely unusable and you will just waste a lot of money and waste a lot of time or your startup will fail. All right, Swick. So, you're the founder of the biggest conference for AI engineers in the world and you just had a massive event. What's the main thing you've learned about how people are building with agents? I think people are doing a lot of token maxing but also very cautious about the ROI or the efficiency with which they are doing it. We have this token billionaire program where if you spend a billion tokens a week, you get this like fancy gold card and special lounge and everything. We expected there to be 50 people out of the 7,000 that we had and we got something like 300 [laughter] token billionaires. So that gives you an idea of the level of adoption. I mean you obviously work at Cognition as well. We're going to get to that. But I've seen people like from cursor post that they're doing like 8 to 10 billion tokens per day. So is that like your experience as well that people at the you know at the labs or at the cutting edge AI startups are doing like billions of tokens per day? >> Yeah, eight is very high. Um I don't know if I've met anyone that actually does eight. I would say yeah I I do think that obviously if the tokens are free, you're going to try to use it more. I think it's a good career bet because there's no downside. The company pays for it and the career upside is that you you are the first to find some interesting innovative use of agents and a lot of it will be slop but the people who have the stomach for slop I think are still very underrated. Most default engineers would be biased towards efficiency. They hate slop. It's actually non-conensus to embrace slop and try to figure out okay what part of this is useful. >> Yeah. I mean you know when everybody can generate 100 landing pages or 100 designs for a single feature then really the judgment comes there. I guess what's what's some of your like things that you found yourself needing to change as you're managing you know more agents or burning more tokens like for me like it has been the consciously staying focused on one task rather than you know running like 30 agents in parallel working on 30 different tasks. So I'm wondering like what working habits do you notice yourself adopting? >> I definitely don't do only one task. I think that you should have some sort of add. I think you have one let's say high focus high concentration task but then you can have many background tasks that are just doing a lot of things that are repetitive or what have you or or researching or prototyping whatever we work a lot in Slack we use Devon a lot obviously as a I'm a customer as well as an adviser to cognition right so um I get to see both sides and when I find a bug uh the nice thing is I have a direct line to the CTO to go and fix it and uh you see Mox a lot uh which is kind of like the sort of terminal uh multi- code thing. They're actually in the the co-working space next door to us. >> Oh, nice. >> Um and I think mostly I I mean I still use cursor a bit. I would say that the the need for the IDE has really declined to just being a file editor or like you know the way to review your files. And I think the sooner we move towards that future the better. But I do like occasionally jumping out to the cloud codes and then uh if I need anything computer use computer uh the best computer use agent in the world is Codex? >> Yeah, I mean the Codex app is really polished. They've done an amazing job there. Do you think there's a need for a graphical user interface or will we just like stay in the terminal? >> Yeah. Uh we have a speaker that does agent craft uh Ido Salomon. He has spoken twice now about agent craft and he used a super graphical interface of the Warcraft uh type simulation, right? Every orc that is building a building is is a is an agent. It basically just depends on your kind of brain. If it fits your brain that you really want to see the visualization of the things moving around, I think that can be very helpful. Most other people are just in sort of independent threads. So there's no benefit to visualizing anything apart from just a flat list of things. I think that I think somewhere between like flat list maybe comb board which is more like the conductor approach um or devon desktop approach and then full graphical interface like uh intent or agent craft. I mean Devon has been very early to slack right we now see cloth trying to do the same. So I think that there's a real need there. Also, um I want to say like congrats on SW 1.7, right? I just saw the announcement like literally an hour ago. On that topic, like do you think there's going to be more models coming from like not you know open anthropic Gemini like this is a fine tune of Kim K 2.7 but still you know the very impressive benchmarks, very impressive speeds on Cerebras. How do you think about like the future? Do you think it's going to be dominated by Opia or Enthropic or will we see more custom smaller models? >> Uh you know Kimmy is not that small of a model. Um and also composer 3 as well. Um I think that the frontier will always be a frontier lab. Um and but most work is not frontier. Most work is just boring routine stuff. So there is space. AI is very very big. Um you can have multiple winners. The way I phrase it is there will be the model labs who uh need to accumulate all this GPU research and cutting cutting edge researchers and are trying to go after AGI and then there are agent labs like cognition um that go after domain specific problems like enterprise coding that will be able to train their own models uh for those those coding use cases and I do say train instead of fine-tune mostly because the model compute at this point in post- training uh is at least uh equivalent to amount of pre-trained compute uh which is the the sort of u kimk 2 2.5 checkpoint in some cases like the publicly disclosed number from composer is that they're four times uh the compute of pre-train as compared to post train so I think then you're just doing continue training [laughter] it's not a it's not a thin layer anymore >> it's not shifting the underlying you know allocation of compute like you know less less time in pre-training more different specialized fine tunes and post trains like how how should people think about this? >> Yeah, I think it's a golden age for people in post training for sure. I think the interesting question is well is that a temporary phenomenon and actually the the Fable class and the stuff that is coming next year after Fable um you know people are just ramping up these GPU clusters and actually we are still probably two three cycles away from topping out the current gen of LM design and and um code design with hardware. So I think that a lot of it is still up for grabs. I think that we'll still have the 10 20 trillion parameter model runs that just destroy everyone else. Uh which you're going to see tomorrow with GPT 5.6. Uh that that is that is very true and I think that will keep going you know. So, uh, people who are somewhat preaching small models, usually one, they're trying to sell you their startup. Two, they're coping because they don't have GPUs anyway. So, uh, cry harder. Uh, but I do think that there is some way for them to win, right? Like, it's it's a it's a big world. >> So, you kind of mentioned, you know, this latest generation of the biggest models, Fable, GBD 5.6. Has that changed how you work and how you think about AI? um somewhat I think so. Um I think the problem is it's very subtle uh how these how these things changes and if you don't explore and if you just use the same prompts that you use for GT5 for GT5.6 or for Opus versus Fable I think you will just get slightly better quality and you won't be that impressed. Um, I think the opportunity comes from really trying different kinds of prompts because that is what big models smell sound sounds like and looks like. And so for me, I think the interesting uh talk on this is the field guide to fable that Tariq Shipar uh did the AIE. Uh we published it. We we we speed tracked it because Fable is very hot right now. But um I I like the idea of like trying to find your unknowns, right? try to um I think when you are familiar with a model, you know what it can do, you know what it can do reliably and you go to it with the same prompt, same workflow, same everything, then that's a candidate for distillation and it's a candidate for small models. Uh and you you know what to expect. But I think whenever you deal with a bigger model or a new model, just ask it weird things. Ask it for unknowns. Ask it to brainstorm with you. ask it all these kinds of open-ended questions and I think you get some interesting answers that you wouldn't have otherwise and I think the people who know how to explore capabilities will do much better than the people who use AI like a slightly smarter tool. Yeah, my experience was that, you know, Opus and 5.5, they were like kind of like tools, right? Like they were really good, but you still need to tell them what to do, how you want it to look like. With Fable, you need to be like you can be more abstract. Obviously, it will still do a specific task, but you can really feel the model understands to like more degree like what you're doing, what's your intent and stuff like that. So, you really don't need to be as specific. So, has that been your experience as well? And like do you find yourself you know trying to think more higher level problems more abstract problems? >> Yeah I I I do think so. So um I really like what Boris Churnney has been saying around um thinking in terms of loops and going one level higher in loops like you start to create loops that generate loops. Um like you know for him you should not be writing prompts you should be writing a loop that creates prompts. Uh and that's part part of that is you know your first taste of that is the goal uh when you set a goal uh that generates prompts or that basically sort of repeats the goal again and again to the bot until you know some condition is reached and I think that is one instance of a broader category of examples that include loops and then you need loops that generate loops right so I I think that's what a lot of people are exploring these days of how to sort of maximize your subscription or token span or what have you. Um I like I really like the idea that a lot of us are building from the loop up uh which is called loopcraft. Uh but also we should actually start from the top loop and try to work our way downwards like anything that we do manually at first try to intermediate it with an agent and then figure out how to generate loops from that. And I think so there is a skill to that that is being developed today. >> I mean you had a keynote just about loops right? So what makes a great agent loop and what are people missing? >> Very simply, I think a great agent loop has enough specification in it to know what you are looking for for the for the agent to figure out and some verification of how do you know when you're done? How do you know what good looks like? And I think on a a third piece of that that is often omitted because just for convenience but it helps a lot is what you should not do um because you just know the flaws of the agents. It's like ah my a my agent or my model keeps writing 10,000 line code files and that is very hard to maintain. That's very hard to parallelize with multiple agents. So every now and then do some garbage collection. every now and then refactor every now and then reduce dup look for look throughout the codebase and reduce duplication stuff like that. Performance and design is also a very interesting one especially when it comes to anything visual or front end. You want the the bot to take a look at its own output and often it doesn't do that. Often it forgets mobile uh versus desktop. Just like very simple things like that. Um, people still rely too much on oneshotting things and so they they want the default alignment to just work, but it's very very unfair to just expect everything to work the first try um without even looking at your own output. Just just write the code and like hopefully uh yolo and figures it out. But I I think I think we're we'll eventually get there. Um we just need to help it along while this AI is being developed. So say someone's starting a new project and they you know want to be as productive as possible. What what's the like the first couple of loops they should be thinking about right? Like before they have like customer tickets and then you can you know do a bug fix automatically. What if you were starting a project today? How would you think about like implementing loops so that you're not doing everything manually? >> Yeah. Um I think I really like the uh a couple things. one the the first one that everyone this is not really a loop but you can put this on a periodic like weekly basis which is the grill me or the interview me loop >> literally have literally it's flipping the role of you versus the model ask the model to prompt you with questions so that it can extract from you what you're trying to do uh because we're actually very bad at prompting we're actually very bad at expressing what we want we don't even think about like forks in the road until we're presented with the forks in the road then they were like oh yeah actually that's very important we should clarify on that to me that's what you know hiring an intelligent employee does uh which is like I tell them to do something they were like no I'm not going to do that first of all I want you to answer these two questions right and then they do it uh and I think that's uh very important and to encourage the model to do I think secondly the other thing I really like about loops generation is researching and brainstorming which is essentially free, right? Like go look at my competitors, go uh research and brainstorm uh three ideas from me every single week and then uh prototype them and implement them. >> Uh that's like a super simple loop. You can throw away that code. Uh most of it will be trash. Some of it will be free, just free ideas that you can just get, right? Um I think logging that feedback. So I think closing the SDLC is what they call it. So what does that mean? It means doing things like hey does your website have error logs and and production logs. Have you connected that to your agent so that it can periodically read the a read the logs and also um uh you know respond to them or read them. Um and so the ideal situation that you want to get to is the self-healing app, right? Where if you have a bug, um it will the uh the loop will already detect it and already propose a fix. Um so I I I think those are the baseline loops. I think the overall thing that you should then work towards after you set up your full SDLC is tell figure out what goals you want. Like I want more sales. I want more conversions. I want people to spend more time on the site. Um and that is a very very long like multimonth goal that you can sort of work towards and collaborate with the AI on on working towards uh and that just means basically treating it as a full employee like uh check all these research goals you know that to me that's what auto research looks like right you auto research you have a loss function and you're trying to optimize against loss function try whatever you want and so it's easy to roll back it's easier to experiment but you you find wins over time and you sort of start to do some gradient descent And I think that applies to a lot of goals that we have and it's essentially a sort of goal- driven loop that hopefully you can start with some basic intuitions and and go towards there. Um I do think that you don't you can get into trouble with very big projects where um you don't understand the codebase anymore and you may have to you may run to a point where you spent two months on the thing and then you have to throw it away which is very sad. Uh and so the way to avoid it is to make sure that you always understand the data structures. Uh because the the that's the clear separation of like what data get gets recorded uh and what gets reported and what the UI can show and what the what the workflows can work on. If the data isn't there, you can't do anything, right? So make sure make sure you fully understand the full everything that gets logged. do a lot of interviews and a lot of self introspection about do I really need this uh data, where do I keep it, how how frequently accessed is it? uh these are for the more serious projects that you really want uh to do to to keep in production because I think a lot of code is not that serious is throw away like you you'd use it for a day or use it for a week but if you want to use it for more than a couple months uh you should invest the time in understanding the data layer >> that's actually a great answer I have like five new notes just from that okay so on that last point when somebody is building you know aggressively with AI you mentioned the data structures What else should people never give up, right? Like what else should you be still fully in charge of, still fully aware of as you're building with AI, even when you're doing like close to billion tokens or more per day? I think user testing um models are still really bad at basic human intuition things like hey when this resizes like this button should be here and like or like this validation obviously you know should be there because models just don't live in the same context that we do of using similar apps all day long every day you know for our work for our personal lives everything right so we [clears throat] just have this like unspoken expectations of what a good application looks like. And it's not just looks, it's also how it works in the back end as well where you just expect like where's my audit log like where is my off, where's my orgs versus teams versus all these things and you can trust the model to do that, but like you should probably uh specify that behavior or just like keep testing that behavior and so um I think that is still really not going away. you. Basically, what I'm saying is that yes, you're spending a billion tokens a day, but you you god then better be actually having real humans, especially your target users, be using your app because it's super super easy to code a lot of slop that makes your app completely unusable and you will just waste a lot of money and waste a lot of time uh or or your startup will fail. Um it's just as simple as that uh because nobody's going to use your Um uh because you don't care. So if you don't care, then why should your users care, right? I think that the care is still very easy to convey. Uh the taste and the selectivity of not shipping everything just because you can I think people really start to value as a premium. Uh even the minor details of design, there's so many ways in actually in which models are uh not good enough. And I think people who are high standards can see this. But many people uh experiencing LLM psychosis where they're like, "Oh, models are better than me and everything. I'm [snorts] just going to sit on the beach and have the coding agent code for me." Uh those guys are never going to make it because they don't have taste and and so there's going to be a billion people like them and all of them will fail. Um I think that uh the it's it's comp comparatively easier to stand out by just finding ways in which uh things are still made with craft and human taste and I think it it shows in in the products that people make. >> So just use like manual literally dog fooding manual testing clicking through every button and making sure you have like your hands dirty every single day. >> Yeah. and asking yourself, is this what my users really want >> or is it just what the model felt like coding today? Because you have all these loops and they're incentivized to have as many tokens as possible. >> Obviously, they're spending ton of money, but is this what your user wanted? Are they are they asking for it? Are they actually going to use it? Or are you just throwing 100 different configuration options to them and they're just actually going to use two? Well, then they delete the delete an IDA and just only have two, >> right? But like it it's hard to say no because >> code feels like it's never been harder. It's never been harder because it's easier to build anything >> than you know. >> Yeah. So like you know like I feel like doing it today. Okay, just add it add add and then you have this slot that nobody uses. And I think this is one of those things that's like it's not new to AI. I think it's a design problem in general. I think one of the most insightful comments I've heard about this is Bejian Stostroke the creator of C++ um who always evaluates the proposals to change C++ and people want like I want my datetime type here I want my long int here here I want my short uh float here whatever um all these things he's like his statement was basically each of these proposals individually makes sense together they are madness meaning oh if I if I say yes to you I also say yes to that guy that guy that guy I'm going to end up with 200 different data types that nobody can use and nobody people stop learning the language altogether and it is my job as the overall head of the language or head of design or whatever to say is there a broad solution here that like one thing solves 200 things or are those things small enough that they can do it in user land or they can they can live without it and I only keep it And I think that is still at a premium maybe more at a premium these days. >> So really having deep understanding of what the customers want of that domain really knowing it better than anybody else and developing that taste as you said. >> Yeah. >> So what does it take like what's the difference between the founders that have taste and don't have taste? >> Oh that's tough. I don't know if I have taste. I think for me AIE worlds fair is is my greatest expression of taste, right? like I I choose the humans that choose everything else. And I think uh the difference between people that have taste and people who don't. First of all, don't lie. Um and it's very simple, but a lot of people lie. >> Um what is an example of a lie? Oh, my small language model 4B beats it is fable class, right? you see a lot of this a lot and some of it might be somewhat legit, but in general people know you're full of >> And so the the kind of founder that would willingly lie like that um is just doesn't have taste, right? Like because they they don't care really for the magnitude of their claim to match up to what they actually can deliver. And so I think um a lot of people just fail that out of the bat. Very simple to weed out. Beyond that, I think the kind of people with taste also are more involved with their problem than their specific company or solution. Right? So I get a lot of people who come on AIE or the inspace and they just want to show their company like uh we are the best uh we put so much work into this like we we're the best inference provider for blah blah blah. Okay, cool for you bro. Um they don't tell you anything about how how the problem is being solved. They don't tell you any they don't teach you anything about how we're better than uh we as a field we as an industry are progressing. It is all about you and not about us. And I think we are looking for a story of us uh progressing as a civilization or humanity or industry. Um and if you can connect that to a broader goal that everyone wants like everyone wants faster inference. Yes. Okay. Cool. Um tell me what are the road maps or the the road blockers or like the the key insights that you've developed over the last two three years of your operation in order to get there. You don't have to open source everything. Yes, you're close source. Yes, you need to make money. But teach me something, man. Instead of just saying like here are my benchmarks. We're better. We say on our benchmarks we are better than everyone else and everyone else sucks. Okay. But like again, how do I know or like why do I care? You know, I think um the best people with taste share something because they know they have a hundred trace secrets. They're only sharing two, right? Um and so and the two are relevant to the discussion and also not the full story, but they know how to guide the industry discussion they and stay relevant. Whereas a lot of people without taste don't. they're just followers. >> I mean, it's super common. It happens to me as well, you know, when you have somebody and they just push their thing above like contributing to the discussion. I mean, we can talk about like why that is, you know, maybe, you know, they're burning a lot of money. They don't have a lot of round left and it's kind of desperation. But like, if you think about it, you know, Elon Musk, he doesn't like try to sell you a rocket. He like talks about why we need to get to Mars, you know, it's about the bigger mission and kind of explaining his thinking and explaining from first principles why he's doing what he's doing. And then that's why people get like excited about it. You know, they understand the mission and like you said, it's contributing to the civilization rather than just like buy my thing or you know implement our inference. >> Yes, I agree with that. Uh and so basically what he is is is a missiondriven person, right? And so if you have a good mission, everyone will want to join you. Everyone is cheering you on. Everyone wants to support you. If you have a shitty mission like I want to win at the cost of all my competitors, uh then nobody g nobody cares. And I think people don't understand what what a good mission looks like. Um I think the other thing [clears throat] that people uh do a lot is basically they're they're not trained to do this, right? Like a lot of founders are technologists. It's not because they're running out of money. So they don't know any better. They don't see that many positive role models. They don't know how to translate positive role models into what they can do. they don't see themselves as possible doing capable of doing it. So, um I I uh I I advise people um I wish I could do some kind of training course. I maybe I can I don't know. Um of like no actually everyone can do this. Everyone can pitch their story better, tell their story better. Uh, and actually it just starts from having a decent enough mission that's ambitious but doable like and like you know make us want to root for you instead of just saying how much money you raised, how quickly you raised 100 went to $100 million. No one really cares uh apart from just investors. Um involve us in the project the broader mission right and then we care. Um people care about all sorts of things including like the performance of a database. You just have to make it cool and technical and tell people why why you are the best people in the world to do this because you you have done nothing else for the rest for your entire life except this and it's great like I you know I why do I care about Vitz and planet scale? I don't really like I'm not in that high performance database world, but I care about YouTube and I think you know making sure that YouTube runs well and and performantly and understanding how that also applies to Uber and LinkedIn and you know all these other companies. Yeah, I care. But like you got to tell that story that way of like why people should care instead of just saying you know my benchmarks are like 50% better than my competitors and the competitors suck. Um that's very negative. So what advice would you give to new AI startup founders today? Because there's a lot of you know talk like SAS is dead this and that comp like AI labs are going to replace you you know don't do this hire your person stay solo until you're a billion dollars. There's so much advice floating around like what are the main things you would like to communicate to somebody who wants to start an AI startup? >> First of all I'm not in the startup founder advice business. I'm not a VC. I uh I just interview founders. Um I do think that there are two things I offer to people when anx question comes up. The first very simple one is that the application layer is still very rich for building. Uh so literally look up my agent lab thesis or agent lab playbook uh which I wrote on latent space. Um basically building an agent lab involves being the domain experts for vertical. Um whether it's coding or it's healthcare or it's legal. Uh just own a vertical and be the AI guys. Uh why does that help? First of all, you have much more dedication than the generalist uh model labs. And second of all, you can pivot whenever the next model comes along and that's very important, right? That this model independence uh layer, right? Like Sam Alman always says u you know you the kind of company you should build is the kind of company that gets happier every single time we release a new model not uh uh fearing for their for your life and I think that the that kind of company is the agent lab um that said I would also point out for more ambitious founders uh so that's like kind of like the the lower risk like almost guaranteed to win if you're just like actually good at execution. The one that is high risk high reward uh that requires research is the domain specific model lab. Uh so we're talking about black forest labs, we're talking about thinking machines, we're talking about uh even the more specific ones like core automation or flappy airplanes. Um all those are basically some part of the model tool chain. Um and I think ba I think in some extent like uh people are very willing to fund new research. Uh engram uh is is another example there. Very willing to fund new research but uh and you can raise like you know orders of magnitude more money for the latter than the former but uh you know obviously the risk of failure is high and I don't think that you should necessarily care. I think you should uh you know uh take the high risk because that's what investors are taking anyway and if you at least have somewhat of an honest coherent vision of what to pursue go for it uh people will back you and and I think it's a it's a beautiful age for genuinely new research like people know that at some point that the LLM transformer autogressive paradigm will end so they are funding the next thing already um I think anything that is more capital intensive um is good and I think founders who are on the VC path should understand that you are paid to deploy capital with some kind of ROI and you're not just here to build a profitable business. Obviously you want profits but people want moes and modes involves putting a lot of capital down in order to secure some kind of technical advantage or distribution advantage. And a lot of founders are very timid. They're like, "Let me build this open-source orchestration agent framework and release it and get a lot of GitHub stars and then I will do some consulting and make some money and then dot dot dot I will be a billionaire and they will never get there. They're not going to make it. Uh they're too timid. They there's no capital involved. You call yourself an infraound but you don't actually run any infra. You're just a layer on top of AWS." Um so the exceptional founders I see there are actually building infra like they are figuring out the data centers they are buying the GPUs they're they're buying the data building the data collecting the data whatever that is real infra everything else is games >> all right let's I need to go deeper on that a lot of founders being timid you know obviously it doesn't apply just to infra but uh like how do you see that you know like how how do people become timid is it like they get their first taste of success and they don't want to let go of that, you know, like Elon when he made 180 mil re he reinvested that into like the next two companies. >> As hell. >> Yeah. So, like how do you notice yourself being too timid and you know, how do like how do people think bigger? How do you force yourself to think bigger constantly? >> Oh, I'm I'm super timid. Like I know I I'm very good at content. [snorts] I do content. You know, I'm very I I know how to do a conference at a conference and it's it's profitable. It's it's growing. Uh this is not the most ambitious idea I have, right? Uh so I I am not don't take my example uh of me as a successful founder because I'm an average founder I would call myself you know um still successful um you know it's going to be worth something but um I I think the the really really good founders that I I I'm privileged to know and and and to interact with and be friends with are way better than me at at being ambitious and and all these things. I think again it comes down to do you have a good mission and then once you have a good mission do you have a good way to get there and can you recruit the people along the way to help you to get there right and recruit also means employees but also also customers also partners and even competitors uh if they respect you enough they will do it um and so uh I think Elon obviously has great mission uh and I think the the beauty about having great mission also means that if you fail people will give you a second shot. Uh it is not over. Um if you have to delay your deadlines, doesn't matter. The mission still stays the same. So I think that mission is is really helpful. I do think that ambition when it comes to ambition, people should just think about what are people obviously going to need in 30 years or maybe 10 to 30 years and build for that because most people will think on a timeline of two to five years. And so um your ambitions automatically increase when your time frame increases. And if you're like, man, I don't know if I can do this, but someone should do it. Someone will probably do it. And it just takes a lot of work up front. That kind of fear is why you should do it. >> I see. So, who are some of the best founders, you know, and what like traits do they share in common? >> I mean, so many. Depends what I depends what you mean by know. Uh, at this point, you know, I've I've had Mark Zuckerberg on the pod. I've had Greg Brockman on the pod and uh they've all been very generous with their time as well. Mike Kger I just interviewed at at Voltzare uh another extremely uh incredible founder uh who who you know had a failure which is uh artifact in between Instagram and anthropic labs and so I think it's very interesting to sort of talk to that class of people uh closer to home I would say the you know the more successful founders that I know are like Mo Rous from Versel uh Paul uh Kus also for Superbase those those are like maybe like the they started the companies 10 years ago um is it 10 has it been 10 no superbase is at least is five years uh but Versell is 10 um anyway so like that cohort and then there's the the the even even shorter cohort of uh you know two to three year old companies uh many of which uh you know are like the Maddx's of the world they're they're starting a new uh new chip, you know, so um and etched as well that just came out. Um these guys are taking on Nvidia, right, which nobody thinks they can compete with. But I think on a on a first principles basis, they they see some optimizations that a GPU, literally a general purpose um graphics program processing unit, um doesn't optimize for transformers and they can probably do custom silicon. that thesis is probably correct. The main risk is well is OpenAI going to use your third party chip or just develop their own. Um and so that's obviously an open question there. And I think there's enough Neolabs. I think that the the importance of an ecosystem is important enough that uh some independent chip supplier like a MedX or an edged will will get there. Um so yeah and competing with Nvidia is very ambitious. Um I think uh engram competing with uh like trying to solve continual learning is very ambitious. Um you know I can keep going like Black Forest Labs like you know uh image and video uh and and I I think what's interesting is a lot of the most ambitious companies um just have a general domain. They don't have an end point. You're there. You ask them things like where is this going to be in like 5 years? They don't have a real answer. They're just like better, you know, uh my thing but better. They don't really care, right? Like they're just in this for the long run. And so either you're with them on the mission or you will not invest or you will not really be friends with them. And you know that's probably as it should be. Maybe their mission is bad. But if you understood the importance of voice then you would have invested in the 11 Labs. Uh and 11 Labs is you know one of the top startups now. If you understood the importance of nontechnical people making marketing pages then you would have invested in lovable which is also another uh deacord now. Um and then even same thing for sandboxes like modal and E2B and um uh what's the other one um and Daytona as well. uh all these are very ambitious companies because they they're basically reinventing serverless or cloud for agents and again that's another whole innovation right less less like sexy uh you know it's still infrastructure but I think uh you know cloud has these like cycles of every 5 to 10 years there's like sort of a new wave of deployment paradigms that basically always just means you want computers to spin up faster and faster and so they're doing it um they they were first to do it and so they're getting ridiculous valuations now because of it. >> Do you think that like everything that was designed or built before you know the CHBT wave when the modern wave of AI explosion has opportunity to be disrupted or optimized because like from first principles if it existed before you know agents and LLM became popular then it probably wasn't optimized for them. >> Yeah. Exactly. So uh they you asked the easy version of the question is it has the opportunity of course they have the opportunity everything should be uh has the opportunity to be reinvented uh but things are sticky you know so like git was invented before the inter before like really as the internet was kind of taking off um >> and so will git be reinvented um people are trying but you know chances are they will not succeed because protocols are very sticky uh data formats are very sticky So what will be reinvented? Uh I think things will be reinvented and and you know MCP versus REST and GraphQL and all those other sort of backend front-end communication protocols are are being reinvented or integration protocols let's call it. But uh I don't think that that much of the fundamental underlying layer of the internet will change. Uh I think it's mostly just the surface level trends. >> So what's the harder version of this question? You said I asked the easier one. >> Oh, the easier the harder version is what will what will get reinvented and like making predictions rather than >> I see I see. >> So like you know git will probably stay but like GitHub is getting disrupted from you know multiple angles. >> Yeah, >> I see. I see. Okay. Should people like when building new stuff should they think about building for humans as well or like purely for agents? cuz like I you know I had some like cool ideas that I wanted to build and I was like man if I just like define a /go loop and codex well like this this whole product is obsolete. So like how would you think about that? Should people even care about like web UI SAS or is just like how can we make the agents more powerful? >> Um how can we make the agents more powerful? Uh I see. Um I think that is a good question that I don't really have the answer to yet. I think both are important. they're, you know, roughly 50/50 in most usage charts of any documentation site or developer tooling. So, you should just do both. Why choose? I think what's good for humans is also good for agents. Um, so this is kind of a false dichconomy as far as I'm concerned. But I do think that probably agents first stuff is a matter of scale. And when you start something from literally zero, you probably start from humans first because you're going to take a while to get registered in AEO anyway. So you might as well I mean I'm going to give you my example to like explain why I asked it. So basically you know I was like well that would lead me to the next question but I was looking for like who's the best uh engineers and you know developers in the area. So right now for context I live in Poland and I was like okay why not just like build some you know some mix of obsidian and you know proactive reach out you know deep research plus agent and then I'm going to have like a like a visual graph to see which people are the best for specific role and then I was like as I kept building this I realized when I was doing this with Codex those was you know when Fable was banned I was like man Codex in these tests as he was testing this it already produced my result like it already found the best people in the area just through like two or three/go loops and then I scrapped that whole idea of like you know building the nice visual graph interface together. So like that's why I'm asking this question because I find myself doing more and more of my work through codeex or cursor or close code like through talking to the agent and then agent does the tasks. >> What's the question? [laughter] Well, okay. So, yeah, the question is, I guess, if you see yourself, you know, doing the same pattern where like you do individual tasks less and you do more work from these main agents and if you think this will like like what what impact will will this have as people like like you you mentioned, you know, the correct app is super polished and it's only going to be better. You know, now tomorrow's coming GB 5.6 the these like there's a few main agents that people are interacting with more and more and or they have their own personal open claw her agent or pi agent as more people interact with their agent and their agent does everything else like what implication does that have on the internet on you know what products to build stuff like that >> uh yeah very broad society question I don't know if I can answer that uh comprehend >> just start with yourself just start with yourself >> yeah yeah so so I think the def definition of an individual task changes right like so my goal posts have shifted And to me a task is just a prompt in codeex for doing something like uh you know pay my medical bill um and I'll just give the website and that's about it. You can just go and do the task. So that the individual task has gone up in terms of the level of abstraction. I think that's really good. Um yeah I think you know what that means for society. I think hopefully a lot of routine white collar stuff goes away. Um and it allows people to be more creative in in flow uh and to lower the activation energy for doing um you know anything exploratory and and creative. I think that's the positive side. The negative side is people will use it for spam and they have they they are constantly they have the wrong incentives to do anything else. They they they don't have a bigger goal apart from uh let's try this quick marketing hack. And so they suck. uh they're killing the internet and um so people are trying to build you know bot-free internets um they will mostly fail uh but you know like Sam Alman has this world coin thing that he's doing where he scans your eyeball and like you know you have to sort of prove you're human I think proof of human is something that people do talk about but building a new social network or building new internet that is u segmented away from the bots is an idealistic thing that will probably fail because we do want bots in our lives and we do want bots to take limited autonomy and like sometimes reading bot generated slop is useful I read it all the time because it doesn't take a human to do like a basic routine summary of something so why not right and so I think we need to find a balance as a society and maybe we take 30 years to get there rather than three and sometimes people are just too impatient >> so what does it take right now to be on the cutting edge of AI and Again this is very broad question so you can tackle it from any any angle but I'm talking like you know work ethic educating yourself number of tokens burned what type of skill and knowledge you need to have like when you think of the people who are not necessarily founders only but like engineers you know designers whoever is on the most cutting edge of AI what are the things that it takes to be there >> I think being informed obviously is is important because you have to plug in to the news flow uh that is somewhat self-s serving because obviously I write AI news and I curate a conference that people go to to keep up on these stuff but I don't see how you can get around it like you you do want to uh be at least somewhat aware of what other people are doing and then other than that also giving up on being cutting edge of everything because that is also just impossible like that you are >> converging on being on the cutting edge of the the world um so you can be on the cutting edge of memory and saying like I'm the memory guy like I will do everything regarded to memory. If you do a diffusion model don't talk to me go go to the other side of the room. Um and likewise for generative media people they only do generative media and that's great I think. So like specializing there is is great. uh a lot of people basically should be cutting edge by being specialists and then a certain small group of people probably like myself who are who have ADD and are addicted to being broad uh they should be the sort of broad service layer for you in order to for you to be narrow and deep right so um I do think doing deep work is a great way to do cutting edge work um there is a essay I like to refer to people called leadership at solitude. Uh it is a by a military general that uh is advising people who are graduating from the officer candidate school and he says like sometimes in order to lead you should not be among people you should be alone uh and where where you can actually think quality thoughts and actually reflect on where the people need to go and then you can lead them. You cannot lead if you're always following the crowd. if you're in the crowd because then you only do what the crowd wants. Uh and that's not leadership. That's uh you know that's pandering or whatever. uh and I think that is a very good reminder that the role of a a leader uh you know in in a field is to you know keep an eye out for the important problems to be solved uh and d and direct the industry that way rather than always be keeping an eye on the latest headline of this guy raised $100 million that guy raised 2 million whatever doesn't matter You mentioned like you know mission is essential and one of the other pieces is hiring. How would you hire for people in this age? Would you look some people say you should look for seniors because they have you know fundamental understanding and they can leverage AI better. Some people say you know just hire like 20 year olds who are like at the cutting edge of AI and understand everything. How would you think about hiring? >> Having a mix of experienced and talented people is best. Um, I'm the kind of person that doesn't care too much about paper experience. I just care whether you can do the job. But I think experience does help you do the job. And so having experienced supervisors be uh be around the managers basically um is very helpful. And then having younger, less experienced folks who are more capable, more creative or spiky, I think is good for overall team performance. And I think the common answer is basically just do a work trial with them, whether it's like a contracting thing for a month, three months, and then you hire. Um, and also I think you should be relatively quick about firing as well if you need to do that. Uh obviously try to give people a chance, try to give people frequent feedback, but um you know like keeping a bad hire affects your other hires as well. So I think uh the sooner people like people understand that they are on a high performance team then they they'll also refer their other high performance people. If you have a team that is loosey goosey and low standards, then your overall company is not going to do well anyway. And so, uh, it is unfortunately the the nasty job of the the the leader to say things like, "Hey, you're very really valuable here, but you don't work well in this team. Therefore, we need to find someone else." And that's a very tough call to make because anyone who is willing to help you at least in the early days is an asset. Um but sometimes they are net negative. Um even with all the positives that they bring. So I've had to make some of those calls in the past year and it's it sucks but uh I think my team is much happier for it and you know we're doing well there. Um still a lot of like in-person hiring. I think we have this rule like we always must meet people in person face to face. Uh because obviously people are always on Zoom. They're looking at other stuff on the screen while they're on Zoom. Uh or there's just subtle things that you don't really pick out until you're you meet in person. So there is still that very nonI element of hiring. Uh I think people are exploring other things like Juicebox is doing some AI interesting AI recruiting things but I don't really have much experience there. Let's end on a positive note. What's some things people should be looking forward you know in the next 6 to 12 months given how fast AI is advancing? >> Probably super fast inference. Um people are used to maybe 50 to 100 tokens per second inference from the major model labs. Um but you know anything powered by Cerebras including uh su 1.7 that was just released today is on the order of a thousand tokens per second. Um much faster is coming. Um I have seen demos that are 10,000 tokens per second as well as hundreds of thousands of tokens per second. So those orders of magnitude improvements um they are already in demos and now it's just a question of productionizing for for massive scale. >> And so uh when you have 100 times faster inference your the type of products that you make will change. >> All right. Awesome. So where should people go? Should they go to Twitter? Should they go to latent space? AI engineer everywhere. latent space is probably the most frequent newsletter. Um, so if you uh yeah, if you if you like the kind of stuff that we discussed here, we discuss this all the time on on the newsletter and the podcast. Just just go to laton.space and like check it out. >> Awesome. Appreciate it, Twix.