Transcript of Alexandr Wang: “This is a Once-in-a-Civilization Opportunity”
Y Combinator
0:07All right, [music] full rockstar0:08treatment for Alexander Wang, everyone.0:11All right,0:13[cheering]0:13[applause]0:16so why don't we start out uh backstage?0:18were saying, you know, one of the cool0:20ways to think about this event is like,0:22you know, this room is actually full of0:24people who are just like us, but when we0:26were 18 or 20 or, you know, there's some0:2916 year olds in this audience, you know.0:32Um, let's jump to your story. I mean,0:35you got it came up always really smart0:37like math olympiad like jump us to, you0:40know, the Alex of that time like what0:42were you feeling? What were you0:43thinking? And what drove you down this0:46road?0:47Yeah, I um uh well I grew up in New0:50Mexico, Los New Mexico um which now0:54Oenheimimer famous, but um it really was0:57the middle of nowhere. And uh I remember1:01I did all these math competitions, all1:02these um computer science competitions,1:05but then um I knew I wanted to do really1:08big things and it was like not exactly1:11clear how or what the exact paths to do1:13that would be. Um, and I had a friend1:16who was really into programming. Um, and1:20you know, after high school got an1:21internship in the valley, worked, I1:23think his first internship was at1:24Palunteer. And, um, and he, you know, he1:27was kind of this, um, influence for me.1:30And so, after I finished, uh, high1:32school, I ended up working at Quora, um,1:34here in Silicon Valley. And then, um, I1:38worked there for a year. I took a gap1:40year to work there. Um and then I went1:43to MIT. Um and this was I was 19 when I1:46worked at Quora. I was 18 when I went to1:47MIT. And I was 19 when I started um1:50scale. And I remember this period from1:52like 17 to 19. It was um uh I felt like1:58I was constantly changing like you know2:01exactly what I want to do was constantly2:02changing you know I was learning so much2:05just from the people around me and it2:06was just like I felt like I was drinking2:08from the fire hose pretty constantly2:10during that time. Um and um I would2:13definitely recommend you know the two2:15things that were really important. One2:16is I think working at a company was2:19really valuable because like I think2:21from the outside in you have no idea how2:22companies work. You have no idea what it2:24looks like to actually build something.2:25You have no idea what it looks like to2:26iterate on something. You have no idea2:28what it looks like for groups of people2:29to make decisions. And so I thought that2:30was really important. And then going to2:32school at MIT was actually really2:35important because it just gave me a lot2:37of um opportunity to explore what was2:41interesting. And so it was at MIT that I2:43started training my first models and um2:46that I like played around with2:48TensorFlow which had just come out that2:49year at MIT and where I like ultimately2:51came up with the idea of scale. And then2:54after one year of MIT, I applied to YC.2:57You know, it felt like kind of like a2:58miracle to get in at that time. And uh3:01and YC was was really critical to my3:04entrepreneurial journey. Like I don't3:05think um like YC is this amazing blend3:08of uh you know, they're very supportive3:11and they obviously want you to succeed,3:12but they also give it to you very real3:14and they tell you when you're being a3:16dumbass. Um which I think is uh you3:18know, that's what we all need in life.3:20So um yeah that was I think the story3:22till then I was 19 started scale and uh3:24the rest is history3:25>> I guess with uh you work with Jared3:27Freriedman at the time and um you came3:30in with actually a very different idea3:32than what ended up becoming scale.3:35>> Yeah. Yeah. So, we wanted to build um3:37like an AI agent, funnily enough, for uh3:39for doc for to help people like get3:41medical care. Um and it was like the3:44right it was a great example of an idea3:47that I think will ultimately exist. Like3:50I think we're even seeing it now like AI3:51agents to help people get medical care3:53are very real, but it was the wrong3:55timing. Um, and uh, and we worked on it3:58for about a month or two before Jared4:01pulled us aside and we're like, "Guys,4:03this is I don't know if this is going to4:04go anywhere." [laughter] Um, and uh, and4:07that's exactly what we need to hear. And4:09it was at that time when like, you know,4:12I had studied AI at MIT. I like trained4:14models and we thought we sort of went4:16back to the drawing board, thought4:17deeply about where the opportunity was4:19and came up with scale. I guess selling4:22data at the time, you know, large4:24language models were not even had had4:26not really come to the four yet. Um, but4:29self-driving cars were sort of coming up4:31and and computer vision suddenly became4:33so that was sort of the first market. Is4:35that right?4:36>> Yeah. So, the the story here is that4:38like I was when I was at MIT, I did a4:41bunch of projects to like train train4:43models of various forms and these were4:44like, you know, by comparison today,4:46they're like little toy models. And um4:49and I remember to train a model uh I4:51needed three things. I needed a uh GCP4:55account like I need an account on some4:57cloud service to get compute. I needed5:00um the code to run to actually train the5:02model. And I needed data. I needed a5:04data set. And uh for two out of these5:07three things, you could just press a5:10button online and get them. And then for5:12the last one for data, there was like no5:14effective way to get data for training5:17these training these models. Um and so5:20it felt incredibly obvious that this was5:23going to be the future that there was5:24going to be a way to um you know press a5:27button so to speak and get data. And uh5:30and it was very funny because in the5:33years that followed like in the first5:35many years of scale data was very unsexy5:38still. Um, every time we would go out to5:40fund raise, even though our numbers were5:41great and we had great revenue, you5:43know, VCs and investors would always be5:46very skeptical. They'd be like, "Oh, I5:47don't know if this is a good business.5:48Does this have longevity? Is this5:50durable?" Um, and uh, it was really5:53weird to me, but, you know, none of the5:54investors had ever trained a model, so I5:57guess they didn't really get it. Um and6:00uh fast forward to today, you know, we6:02we managed to raise money, we managed to6:03keep going, managed to keep growing the6:04business, but um the very same investors6:07who passed on us and were um were very6:11der on the potential of AI are writing6:14think pieces today about how data is so6:16critical and is one of the biggest6:17business opportunities um in AI. So, uh6:20it's very funny to see that whole whole6:22thing come full circle. I it seems like6:24that's actually a real good um case6:26study in first principles thinking,6:27right? Like you can't start a company by6:30opening the pages of the Wall Street6:32Journal and saying,"Well, this data is6:34hot. Like we're going to go work on6:36that." Like you literally couldn't have6:37started scale that way. You had to start6:39from uh think like simple statements6:42that are about the world that you know6:45to be true and then sort of building6:48something for that.6:49>> Yeah. I think the the key thing is you6:51need to develop conviction in a set of6:54beliefs that nobody else um agrees with.6:57Like I think if you look at all the most6:59successful companies in the world um7:01they were started at a time long before7:06the sort of like core idea was popular7:08and they work on that. they toil in7:10obscurity for years and years before,7:13you know, the the idea or the space or7:15the concept or the business, you know,7:17becomes consensus. And the only way7:19you're going to be successful is if7:21you're able to identify these truths7:23about the world early, long before7:25everyone else. And I think that like I7:28mean, one of the most surprising things7:29like, you know, scale, we've been7:30working on AI for a decade. you know,7:33you just you can't base your business7:36decisions based on what everyone else is7:38saying around you. Like if you go too7:41much with the herd, you will get7:42immensely confused and you will end up7:45um nowhere. And so you have to develop7:47your own compass of what you think the7:49future's going to look like um because7:51everyone else will just confuse you. It7:54seems like one of the things you got7:55incredibly great at was, you know, you7:57start with this kernel of like we7:59believe X and nobody else believes it,8:01but then the mechanics of building the8:03business are talking to investors and8:05convincing them and not letting them8:08demoralize you, talking to customers who8:11I mean should just get it and then8:13especially like convincing people to8:15come work for you.8:17>> Yeah. I think that the the8:20these early mechanics of building a8:21company like the these are things that I8:24think you might have some predisposition8:25to be good at, but like nobody is good8:28at starting a company when they start a8:30company. Um and uh and I remember8:33talking to a lot of the investors who I8:36met very early on and they, you know, a8:37lot of them would say like, oh, like,8:39you know, you just grew so quickly and8:41you changed so quickly and like I8:43didn't, you know, I didn't see it at the8:44time. And I think that's probably true8:45for literally everyone who starts a8:48company. Like nobody is nobody is good8:50at something they've never done before,8:52right? And so um I think for all8:55entrepreneurs, you start out pretty8:57shitty at everything. And um the whole9:00game is how do you develop yourself to9:03continuously improve to get better and9:04learn quickly? Uh backstage we were9:07talking about this is actually a really9:08lucky time to start a company because9:10you know obviously you can come do YC uh9:13you you know the people in this room9:15have each other which is kind of wild9:17but uh not only that now you have uh9:19ideally a personal AI that's going to9:22tell you you know hey these are some9:24ways to do it. Um, do you think that9:27would have helped you like accelerate9:28even faster? Like, you know, talk what9:31do you think it's like to start a9:32company today with with AI in the age of9:35AI?9:36>> Yeah, I mean, I really think I think9:39we're at this like amazing moment in the9:42world where the bottleneck is not the9:45progress of the AI models. the9:47bottleneck is diffusing that through the9:50rest of the world and and helping the9:52world adapt to this amazing technology9:54that already exists. Like I think if the9:56models didn't improve at all from today,9:58there would still be like decades and10:00decades of like total upheaval and10:02change in the economy and how the world10:04operates and and everything around us.10:06And um so I think it's as a result it's10:09like one of the most incredible it's10:12probably a like once in a civilization10:15opportunity to be a dreamer and to have10:18a vision and to have ambition and to10:21impose a view of how the future world10:23should look by building something10:24amazing. Um you know one of the things10:27that we were we were chatting about um10:30uh you know backstage is you know when10:33when I started scale or you know 1010:35years ago if you start a company you had10:38to be um you know it was like David10:41versus Goliath and you had to be clever10:43and you had to find like an angle into10:45the market and you had to sort of like10:46you know figure out um a way to compete10:49even though you had much fewer10:50resources. And now I actually think with10:52the power of agents um and AI broadly10:57speaking it's much closer to Goliath10:59versus Goliath like I think but maybe11:01the startup is like a mecca Goliath that11:03is like vastly enhanced by the power of11:06agents and AI and you know the the large11:08companies are the sort of like more11:10traditional Goliath so to speak. But I11:12think that startups now like if you11:15properly embrace AI agents and um figure11:19out the way to leverage their strengths11:21in the most like ambitious ways, you can11:24easily out compete incumbents.11:27>> So let's talk about super intelligence11:30because that's clearly that's even in11:31the name of your lab. Um what does super11:34intelligence mean operationally inside11:37meta right now?11:39Yeah, I think that you know we a year11:41ago Mark wrote this um uh memo about11:45personal super intelligence which I11:47think actually is very similar to your11:48concept of personal AGI but you know we11:50believe that everybody in the world you11:53know all the billions of people in the11:55world are going to have a super11:57intelligence that is adapted and11:59tailored to them that is enables them to12:01accomplish their goals knows their12:03context and ultimately is an expander of12:05their own agency like I think the thing12:07that we think a lot about is is agency12:10expansion. How do we help people12:12accomplish things that they couldn't12:13have ever dreamed of before? And what12:15would everyone in the world do if12:17everything was just easy? Um, and we12:20think about this in a in an ecosystem12:22way um as well. I think uh you know12:24Patrick mentioned it, but you know we12:26don't believe in this totalizing you12:28know totalitarian view of you know AIs12:31that control the world. We believe that12:33these are going to enhance this very12:35broad ecosystem. So you know we believe12:37in billions of people all around the12:40world all having their own personal12:41super intelligence and we also believe12:43in you know an explosion of12:44entrepreneurship. There's 200 million12:46businesses that uh are on meta platforms12:50today. We think that number should go to12:51billions with this explosion of of12:54creativity and using AI tools. And12:56ultimately we think that you know it's12:58going to be this like dynamic ecosystem13:00of business agents working with you know13:02personal agents and developing this sort13:04of like uh complex ecosystem that is13:08fully AI supercharged. So I was really13:11psyched to see Meta Spark uh 1.1 my my13:14open claw absolutely loved it. Um how13:17you know how how has running a frontier13:20lab been? Um you know the metas-park13:23level is sort of the opus level. uh13:25what's coming down the pipe and also I13:27think that you're uh you know13:29increasingly looking at open source13:31which uh I think this audience really13:33loves.13:33>> Yeah. Yeah. So I think it was um it's13:36been you know I've been at Meta for13:38about a year now and it's been um quite13:40a year. I think uh you know getting in13:44and um you know meta we we've talked13:46about it publicly like law before wasn't13:48on the trajectory that was needed for um13:50for meta and so I got in there and we13:52kind of did a zerobased build of how do13:56you um you know build an entire frontier13:59lab uh you know in some ways kind of14:01from scratch obviously using a lot of14:02what we had um and move as quickly as14:05possible and so within nine months of14:07that moment we launched Museark 1 and14:09then two months Later we launched Muse14:11image and Musepark 1.1 and um you know14:15there's a few things that I think have14:16really struck me about this. You know14:18the first is talent density was14:21incredibly important that was the the14:23core thing to bet on and um like talent14:26density is something that compounds14:28naturally like the more talented people14:30you have the more of the most most14:32talented people want to join you. um you14:34know and and I think it's it's kind of14:37um amazing to see on the inside but you14:39know frontier AI work is research like14:43we are it is scientific work we're14:45exploring what can you do with these14:47models how can you push these models14:48what is the what are the reaches of what14:50can be accomplished with these models14:51which requires a totally different14:53mindset and operating model than you14:56know existed for internet companies or14:58internet products and whatnot there's a14:59lot more about experimentation about15:01science about scaling And everything15:04ultimately is about how do you develop a15:07lab, an operating model, a system that15:09will just um be able to compound with15:12all of the exponential growth that will15:14happen in the ecosystem. Both the15:15exponential growth in capabilities, the15:16exponential growth in compute, um the15:18exponential growth in adoption and15:20usage. Like these are all um we are on15:22this like very very steep exponent15:25across maybe every dimension of the15:27ecosystem. And um it's important to15:29develop like a like an organism. That's15:32how you think about the lab that that is15:33able to sort of grow with that. Um, no,15:36it's been it's been very exciting and15:37we're we're going to be shipping a lot15:39more. So, um, I think uh, you know, we15:42will we just launched Morg 1.1 which was15:45a great model. We're going to continue15:46to have updates on the Muse Spark line.15:48Um, we'll also have bigger models on the15:50way that I think will be uh, much more15:53competitive with even the very best15:54models that are out there today. We're15:56going to be launching a harness um soon15:58and have been working on a harness to16:00help empower all the developers and16:02agentic developers out there. Um and uh16:06and then we're also um you know as you16:08mentioned we're working on open source16:09models and we want to kind of as I16:11described before like we believe in a16:15decentralized world of AI capability and16:18progress and development. We want to we16:20want to empower the broader ecosystem16:23and everyone in the world to be able to16:25build and develop using this technology.16:26And so um we have a lot of exciting16:29things on the way and I think we want to16:30be um we want to empower the ecosystem16:33and developers as much as humanly16:35possible. I mean it sounds like one of16:37the ways I mean certainly when I was16:39using u muspark with my openclaw like it16:41it became clear that it was as good as16:43opus especially for that sort of agentic16:46flow with skill files but it was like 8x16:49cheaper actually. [laughter]16:51Yes. Well we I think this this goes to16:53it like I don't you know we don't16:55believe in a world where these models16:57are so expensive that you know they get16:59rationed only for the most wealthy of17:01developers and and companies. um it's17:03important for everyone to be able to use17:05the technology to um and to build17:08whatever they want to build with it. And17:10I think that you know we take a view I17:13think the best AI products haven't even17:16been developed yet. You know that if you17:17look at the AI ecosystem and everything17:21that's happened like every wave is 1017:23times bigger than the past wave. So you17:25know when I started scale the first wave17:27was maybe self-driving cars.17:28Self-driving cars are really awesome.17:30They're like really really cool. But17:32that was like pales in comparison to17:35large language models and chat bots. And17:36like you know chat bots became this17:38thing that was like probably 10 times17:39bigger even than um than uh you know17:42self-driven cars. And then there were17:44coding agents which came a few years17:45later. And coding agents are probably 1017:47times bigger than than um chat bots. And17:50I think we're just on this steep curve17:52like we're going to keep seeing these17:54new modalities and form factors and17:56developments of the AI paradigm that17:58will each be dramatically bigger than18:00the last. And so um you know our point18:02of view is like let's let's unleash the18:03ecosystem. Let's explore and let's see18:06um let's build you know kind of the18:08future of the world together.18:09>> So what's the best way to actually take18:11advantage of the coding model uh for18:14Muse Spark? It's open code, right?18:16>> Yeah. Today um the the easiest way is to18:19use open code. We have like onboarding18:20on the website and then uh soon we'll18:22have a harness of our own and um18:24ultimately I think we want great models18:26that plug into all of the available18:28harnesses and empower as much you know18:31uh sort of combinatorial innovation the18:32ecosystem as possible. Yeah, I know the18:34harness is uh you know under wraps18:36still, but like can you tease us with18:38you know I mean I still use Open Claw, I18:40still use Hermes agent. You know it's uh18:43you know these things are I call them18:44Ferraris that break down on the side of18:46the road all the time. Like is this a18:48Ferrari that won't break down? Like you18:50know tease us a little bit.18:51>> Yeah. Hopefully hopefully it doesn't it18:53doesn't break down. I mean I think we're18:54really focused on speed. I think speed18:56is um you know for anyone that uses18:59these tools speed is probably the you19:01know one of the most critical things. I19:03think also reliability like you19:04mentioned we want to be extremely19:05reliable. Um we want to be very19:07extensible and to scale to as complex19:11and interesting of a multi- aent setup19:12that you that you want to have like I19:15think there's so much innovation that19:16will occur even above the harness19:18frankly um in terms of like how to19:20orchestrate and set up loops and and19:22develop like you know very complex19:23ecosystems of these agents working19:25together. Um uh we want to be really19:28extensible and and um ultimately we want19:31to just empower people to harness this19:33technology because harness19:36[laughter]19:37no pun actually pun not intended but um19:40but there's like I truly believe these19:43these models are already just incredibly19:45powerful like they should they should be19:48so powerful to fuel you know um many19:51many points of expansion of GDP growth19:54and I think it's like up to smart people19:57with vision and ambition to make all19:59that happen.20:01>> Let's see. So, one question. I mean,20:03when you look back on the decade, um,20:06what do you think they'll say was20:07obvious in hindsight about AI that20:10people are just missing in real time20:11right now?20:14>> You know, so much of the debate that20:16happens these days is around, oh, how20:19good are the models actually getting and20:21can the models actually bridge this20:24issue? And you know when are we going to20:25get super intelligence? Is that in like20:272 years or 5 years? And you know are we20:30going to hit a wall? And you know, so20:31much of that debate is like I think um20:35in some ways uh a little bit of a waste20:38of time because you know I think it's20:40inevitable that we're going to have very20:42powerful models and um you know rather20:45than I think we'll look back and say oh20:47all this arguing around like when20:49exactly was going to happen was sort of20:51um was shortsighted because the reality20:54is we are just as a entire human20:57civilization on this incredible20:59exponential Like you cannot look at the21:01progress of AI over the past decade and21:04not just be totally aruck by how far21:08it's come. Like a decade ago, the best21:10AI models could recognize cats in21:13YouTube videos. And now, you know, we're21:15talking to um you know, a digital god21:18that can, you know, I mean, we've all21:21seen some of the hacks and some of the21:23some of the things these systems are21:24capable of. And you just can't help but21:27be aruck. And and I think this trend21:29will just continue like these these21:30models are going to become more and more21:32powerful. And so I think a decade21:34looking back it'll it'll be obvious that21:38intelligence became abundant and that21:40agency became abundant like the current21:43trends we're on are just going to keep21:44continuing and um this will be very21:46strange. I mean I think for the history21:50of humanity um you know groups of smart21:53people getting together towards a shared21:55goal was was the bottleneck of progress.21:57You know us the United States of America22:00in some sense was an example of this22:01like the United States of America is22:02formed from a smart group of very smart22:04people getting together and having a22:06vision for the future that they wanted22:07to enact. And that's the story of nearly22:09every company um in America and the22:11story of every YC company. Um, and22:15that's going to change. Like all of a22:17sudden the scarce resource isn't going22:19to be intelligence or agency. I really22:22think it's going to be vision and22:24ambition. It's like do you have a clear22:26view of what you want the world to look22:29like in the future? What is the like one22:31way in which you want to put your finger22:33on the scale for how the future of the22:35world will develop and how the how the22:37world will look like in 5 to 10 years22:39that it does not look like today and you22:41have the ambition and drive to like go22:43through all the crap to make that22:45happen. And AI will make that easier22:48like agents and AI makes that maybe 1022:51times or 100 times easier than it was a22:53decade ago. But the flip side of that is22:55then you all of a sudden you can dream22:56bigger. Like I think and the world is22:58like um you know there's so many things23:01that need to evolve for us to be able to23:05fully embrace this technology. Um you23:07know the world is like really just you23:09know barely even ready for this23:11technology today. And I think you know23:13as a builder we have a responsibility to23:16prepare the world right like we have to23:18help enterprises and governments you23:20know to adapt to this new technology. We23:23have to help figure out how we secure23:24the world from a biocurity perspective23:26or a cyber security perspective. We have23:28to figure out how we um how we're going23:30to to manage all these risks that we see23:33with this new technology. But on the23:35flip side, it's also the time of like,23:37you know, unprecedented opportunity for23:39humans. Like we can develop new23:41sciences, we can solve problems in23:43health and biology that have been23:44forever unsolved. We can build new23:45businesses that you couldn't have even23:47imagined before. There's like new23:48creative opportunities that couldn't23:50have existed before. So, it's like it's23:52just this incredible cradle of of23:55opportunity and risks that uh that I23:58think makes it like no better time to be24:00someone who's a builder and um and has a24:03strong view of how the world should24:05change.24:06>> Do you think the path has changed? I24:08mean, one of the things I saw I think24:09Stanford uh the amount of computer24:11science majors actually dropped by some24:13double-digit percentage. people sort of24:15worried like which is sort of insane to24:18me like you still sort of need those24:19skills to even create agents that are24:22that good. Maybe that won't be true. I'm24:24not really sure. How you know have you24:26ch you know what do you what would you24:28say to people in this audience right24:29now? like this is sort of a real24:31question that people are sort of facing24:33like should they become more word cell24:35and less shape rotator like what you24:37know what's the move and you know has24:40that changed um the kind of people24:42you're looking to hire and you know how24:44you manage your teams right now at Meta24:47>> I think systematic and rigorous thinking24:50are still incredibly important because24:52you know the abstraction layer I mean I24:54didn't used to believe that this is how24:56this was going to play out but it really24:57has like the abstraction layer just24:58keeps changing So, you know, when I25:01started a company back in my day, we25:03wrote code. Um, and [laughter]25:06and now, you know, I'm sure nobody here25:08writes code anymore. That's ridiculous.25:09But, um, but now it's about how do you25:11orchestrate the agents together? And25:13then it's like how do you develop these25:15organizations of agents? Like, how do25:16you get like a million agents to work25:18together? Well, and then it'll be how do25:19you get like a trillion agents to work25:21together? Like I think that there's25:23going to be this continued um uh need to25:26figure out how you structure uh25:29workflows at the abstraction layer that25:31we're going to be operating at. And that25:33form of like rigorous systematic25:35thinking. I mean traditionally the way25:38this would work like in my era of25:40starting companies is you would start by25:41writing code and then you would have25:43organizations of humans and you'd figure25:45out how you how to organize those25:46humans. Um, and that required systems25:48thinking and now maybe it's like much25:50more much closer to first you you25:52orchestrate the agent then you figure25:53out how to orchestrate like these armies25:54of agents but um but I think systems25:58thinking is never going to go out of25:59style. So I think it's definitely a26:01mistake to go all in on Word Cell like I26:03think you need to you need to shape26:05rotate. Um but then I think the sort of26:08like um much more of I think what's26:11necessary going into the future is26:13having um a deeper sort of compass and26:17philosophical view on how the world26:20should develop because I think there are26:22there are many many lessons um from26:24human history around um how how we think26:28civilization go through this period and26:31um you know humanity will change more in26:34the next decade than it has in the last26:36h 100 years probably. And um and so I26:39think like the imperative for us to have26:42positive visions for that and have26:44coherent articulations of how that26:46should develop are are more important26:48than ever. Um let's get a little more26:52concrete. I mean, one of the things I'm26:54curious about is like, are there sort of26:55applications of AI that you're seeing26:58among your friends or internal to meta27:00that you can talk about that are, you27:02know, there sort of obvious near-term27:04maybe people haven't figured out yet? I27:06mean, give us some alpha. [laughter]27:09Um, I mean, I think there's still just27:12like astronomical opportunity in uh,27:15agentic looping and and figuring out how27:18you develop systems that enable you to27:21spend like 1,000x more or 1 millionx27:24more on tokens to drive an outcome in a27:28in a continuous feedback loop. Like if27:30you think about most companies,27:31companies are just these like large27:34scale feedback loops where humans are27:35operating each of the edges. Like you27:37know companies they um they get27:40customers and they figure out how to27:41make those customers happier and if27:42customers are happier then they spend27:43more and if they spend more then you can27:45hire more people who can then go figure27:46out how to get more customers and make27:47those customers happier and that's like27:49this you know that in some sense is the27:51feedback loop of uh of every startup or27:54every business and you know these within27:57that there are micro feedback loops that27:59exist and I think developing agentic28:02systems that can operate and optimize28:04these feedback loops is there's like28:07just huge amounts of of alpha there like28:09I think we've seen internally at meta um28:12cases where if you can develop the right28:14agentic loop and you have the right eval28:16or the right metric for the agents to28:18optimize you can have a swarm of agents28:22accomplish more than like a team of 10028:24engineers in you know uh very very28:28handily actually very very easily and so28:30I think figuring out what the world28:33um looks like with lots of uh sort of28:36this like um these agentic coordination28:39problems. I think that is like one of28:40the most interesting problems today.28:42>> So mechanically speaking I mean markdown28:44files cron jobs is I mean is it that's28:47and then basically pointing the agent at28:50enough data so that it can figure28:53something out that you know maybe isn't28:55in distribution.28:56>> Yeah.28:57>> Yeah. Yeah. mechanically figuring out28:59what the metric is and then yeah it just29:00comes down to skills markdown files29:04wrong jobs29:05>> goal29:06>> yeah slashgoal like I think I think it's29:08always funny how uh mundane everything29:10is once you really dig into it but um29:12>> so it's not magic you know I mean some29:13some people put a lot of magic there's29:15like some LinkedIn threads out there29:17about some magic stuff29:18>> top advice ignore LinkedIn29:20>> LinkedIn is where you get customers29:22[laughter]29:25>> so I like to end on this which is um you29:29know you get a telegram to send to the29:3218-year-old version of yourself. You29:35know what do you say to that person29:37right now given all you know I mean29:39thank you for coming back and sharing29:41your wisdom with this audience. I mean29:43you know what would you send in a a29:44message in a bottle to the 18-year-old29:46version of yourself right now? Yeah, I29:48think the29:50I think it really boils down to develop29:53your own internal compass for how you29:55think the future will develop and have29:58strong conviction in it because you know30:01you will get so you will get inundated30:04with noise and people telling you30:06and like you'll it'll be very confusing30:08and it'll be very hard and especially30:10when you're young and you don't have30:12experiences like it can feel very30:14difficult to um have true conviction in30:18what you believe and and what you want30:19to do. But I think that's the most30:20important thing kind of as we talked30:22about, you know, um it took a deep deep30:26conviction in what we were building to30:28be able to weather the the sort of30:30storms of many years of um of uh chaos30:34in the market, in the industry, and the30:36people around us. And so um and and the30:39other piece of advice I would have is30:40try to identify what is the what is the30:44exponential in the world that has both30:46the steepest curve and will go the30:49longest and you know many decades ago30:52this curve was uh was Moors law and that30:55probably was you know that was at the30:57time like clearly the right thing to30:59invest on. I think right now it's AI31:00progress but there will be more of these31:02very steep curves in the future and it's31:04fine if these curves start um you know31:07the starting point is very boring or31:09like it doesn't even seem that31:10interesting like it you know when we31:12started when I started working on scale31:15um you know we had cat detectors in31:17YouTube videos and that felt you know31:20it's hard to say explain the story that31:22that's like the most important31:23technology of our time but it was on31:24just this like unbelievable exponential31:27um and I think I have one last thing I31:29got to say.31:30>> Yes.31:30>> Yes. Which is uh we are uh Meta is proud31:33to offer everyone in this room $1,000 of31:37free credits for the new Spark API.31:40[applause]31:40>> Fantastic.31:44And uh31:48and we're going to keep making the31:49models better. And uh right now Spark is31:52I think 8x cheaper than Opus. So, uh, so31:55if you convert that to opus dollars, uh,31:57[laughter]31:58it's a lot more, but, um, no, everyone32:00here will will work to get everyone the32:02details on how to get, uh, how to get32:03these credits. And, uh, we're really32:04excited to see what everyone builds.32:06>> Alexander Wang, everyone.
6,262 words · 867 lines
