S3 E3 Scale Founder CEO Alexandr Wang: in AI your data writes the program

[Music] Our Guest today is Alexander Wang if you look him up online you'll immediately learn that Alexander became the youngest self-made billionaire in May 22. you'll also find headlines he's the next Elon Musk when you meet him in person you'll quickly learn he's one of the most passionate AI Visionaries and entrepreneurs working really hard to help realize the benefits of artificial intelligence Alexander is founder and CEO of scale AI which he founded in 2016 back when he was only 19 years old in fact he dropped out of MIT to found the company scale AI is now a seven plus billion dollar company AI capabilities are driven by data in fact ultimately most often AI capabilities are bottleneck by data klei resolves his polymeric by helping you get your data labeled and more generally helps ensure the data you use to train your AI systems is of the highest possible quality scale also enables leveraging Foundation models that is very large pre-trained models and then train or really fine tune your own models with your own data on top of these Foundation models skills customers include Microsoft Etsy flexboard GM instacart sap Square Toyota open AI Adept cohere stability and the list goes on and on and in fact not only companies but also the US Air Force US Army are customers of Scalia Alex so great to have you here with us welcome to the show thanks for having me those are very uh a very nice intro thanks so much well you've done so much it's easy to make a very nice intro for you I'm so impressed always when I see what you've been up to whenever we catch up now Alex before diving into today's conversation I'd like to thank our podcast sponsors index Ventures and weights and biases index Ventures is a venture capital firm that invests in exceptional entrepreneurs across all stages from seed to IPO with offices in San Francisco New York and London The Firm backs Founders across a variety of verticals including artificial intelligence SAS fintech security gaming and consumer on a personal note index is an investor in covariant and I couldn't recommend them any higher and in fact Alex I think index is also an investor in scale yeah that's exactly right uh Mike wolby who I think is a board member of both of our companies has been uh has been a star supporter of ours and index Ventures has been an incredible supporter of all of our success since we were I think when they invested we were about 20 people to now about 600 people so um I really couldn't say enough kind words about them and their support I couldn't agree anymore um in fact part of why we're so excited to bring them on board is because recommendations we got from from youth to bring them on board so um our other sponsor wits and biases is an ml Ops platform that helps you train better models faster with experiment tracking model and data set versioning and Model Management they're used by openai Nvidia and almost every lab releasing a large model in fact many if not all of my students have Berkeley and colleagues at covariant are big users of weights and biases now Alex let's Dive Right In scale has the mission to accelerate the development of artificial intelligence and is doing so by providing a data Centric end-to-end solution to manage the entire machine learning life cycle that's a lot could you unpack that for me yeah so um so we we've been working on this kind of as you mentioned uh started the company in 2016. so we've been in the AI game for some time now um and uh and you obviously have as well so when we got started we're really we really got our start was in helping companies build really high quality data sets and so the the you know our business has two parts the first part is what we call scale data which is again all around helping companies and firms get the highest quality possible data for their for their algorithms and this is this has sort of changed over the years you know in the early days maybe there's a lot more on images um then it moved to a lot more around video and now it's sort of changed to be a lot more on language as the sort of like um as the major data types have changed over time but that's where we work with you know innovators like open Ai and meta and Microsoft and um and a lot of the the largest labs in the world then we have a second part of our business called scale applied AI which are basically a a suite of AI models and product products that can be applied to specific business use cases across you know any industry or across industries that we that we built these solutions for so this includes um like we work a lot with uh as you mentioned the Army in the Air Force on defense and intelligence problems that's one big part of our business we also work with large e-commerce companies um comments like instacart or grab in Southeast Asia um to help them use better AI for better recommendations and search we also work with large insurance companies and the list kind of goes on and on but that's where we go to companies that don't necessarily have ai teams or do have ai teams but sort of have a long laundry list of business suitcases and we go and rebuild um or we have these AI models that we can deploy to them they'll immediately Drive business value so these sort of two halves of the world for us are really how we try to serve you know really every business out there whether you're an innovator that's trying to build the most Cutting Edge most highly performant models or you're somebody who really wants to leverage the benefits of artificial intelligence now I think what you're describing there also highlights a very interesting Trend and uh in the early days back 2016 when you started very few people were building AI models to to do something with and it was Specialists building in but there were specialists in neural net design and so forth not in labeling and collecting data and so forth and that's where you saw an opportunity to really help everyone with their efforts but these days almost every company everywhere uses AI not everybody wants to run their own in-house operation and that seems where your second push at scale comes in where you provide the full service um I'm really curious when you think back I mean originally you are not doing that you were just focused on data quality data annotation what was the moment where you decided actually that's not enough that's not going to be enough to do the things we want to do and help people build AI that they want to build you know as when we were in this sort of very focused on scale data very focused on data annotation data labeling Etc um it was uh it was in it was a great place to focus because as you mentioned you know one thing that we we pride ourselves on at scale is that we're always willing to do this sort of like dirty jobs that maybe other people aren't as willing to do and so you know in AI it was back in 2016 what that was all around data you know data was like the the sort of like ugly work and maybe the Glamorous work was getting the neural network and trading the neural network and then ultimately applying that neural network um but one of the the limitations that we found pretty quickly like you mentioned was that the the number of of sort of companies that had the resources and human capital and and sort of um and and Cat an actual Capital you know cash to be able to meaningfully build their own algorithms and actually have a play at building the Best in Class algorithms was a really small list and so as a business you know what we did is we we went on to all those companies you know there were a few hundred of them we went to all these companies uh we we did our best and sort of showing them the the quality of our products and quality of our Solutions um a lot of them got really excited and as more more and more of them got excited because it got captured more and more of that market we realized hey you know this is this is great and we're having we're definitely having a big impact on the AI industry but if you looked you know this is probably 2019 2020 when we're sort of having this sort of like realization um if you look at the entire uh economy where you look at like you know every company out there everywhere his organization out there we're only going to address such a small chunk of that overall Market if we stay focused on these innovators who have the human capital and the resources who will build their own Cutting Edge models that we need to we needed to figure out how do we serve you know in service of the mission how do we serve the rest of the sort of the rest of the world and that's really what sort of motivated um the the applied AI vision and it was something that like you know in the early day that's something we would always talk to our investors about and we would talk about how like hey this is something that we could do over time if we are so lucky to be successful in the first business and uh and it was a great you know sort of with great Fortune we we were able to start um start making that a reality yeah it's really impressive the trajectory the company has has been on and going back to the very early days and I want to wake make our way from their at 19 you dropped out of MIT to start the company um that must have been a big decision what did you see in in labeling as at that time the starting point as being so promising that this was worth uh dropping out so there was sort of um you know there's kind of like a personal story here so I was I was studying AI machine learning at MIT and this was um the year when uh Google came out tensorflow um you know deepmind released alphago it was maybe the first um you know you you probably actually remember the multiple ways but it was one of the very early sort of like deep learning hype waves and um and I remember that so tensorflow came out and I wanted to use it um and so uh I wanted to build like a a uh camera inside my fridge that would tell me when my roommates were stealing my food and part of that is like a you know an image recognition model that would tell you what the foods were and then sort of like help recognize when they were sort of like a disturbance in the force uh so to speak and um and all I did you know I needed to train like a like an object recognition model for that and I remember pretty vividly like all I did was like I took um I took like the uh like one of the uh examples like one of the tutorials that Google had released with the with tensorflow and then um I just swapped out like a bunch of constants so like the number of classes and the the sort of like size of the images and all that stuff and then I had to go and spend all this effort in collecting a data set um so I had to collect a data set of like different conversions food different foods in there you know all that kind of stuff and then like painstakingly had to go and figure out how to get them labeled um at that time was mostly just hand labeling them because it was sort of like everything else was just too much of a pain in the ass and um I was a college student so I didn't have that much money to spend on other things and then um and then and then uh lo and behold you know after labeling like tens of thousands of images I started to get something like kind disorder you could you know you could see like kind of worked and um I kind of with this and then again the rest of the code was all the same you know and I sort of had this realization that um in the in sort of like this deep learning era the code stays mostly fixed and the thing that varies from application application it actually is actually data or or put another way the sort of like the data is what actually sort of Congo did the programming of of the algorithm itself um and that felt like a sort of like one of these profound but in sort of like you know almost like pretty realizations and then you sort of like think about that a little bit more deeply and it's like oh wow that means that there there probably should be companies focused on helping people build great data sets because that's what you know again if you think about the variation where they the sort of like competitive um playing of most companies that's going to be um around data data and data sets and so that's going to be like a key piece of key piece of infrastructure and you know I probably didn't I couldn't articulate all of that in those terms at the time um but that's sort of like roughly this sort of like train of thought that led me to uh to start the company to focus on data and I remember like for a long time it's sort of like um a lot of people ask questions uh around like hey this actually like that interesting of a business or is that actually that interesting of a place to to focus most of people with machine learning could recognize that it was an important problem um and and it needed to be solved but then there were still all these questions around hey is that actually like a you know is that is that a cool business and I would say for us it's never been a cool business um or or at least like uh as compared to like a lot of the other businesses you could start but I think that that's that's been one of our sort of like uh cultural traits or one of our advantages is that we we've never been too focused on being the the cool kids but even more focused at being the useful the useful people I remember in the early days when scale was started I was thinking okay we we as a community within the past ship data for labeling let's say to Amazon Mechanical Turk and a scale came about was more specialized so people working at scale doing labeling are more trained and ready to do labeling and the software is more specialized to to support it but I remember also thinking like how defensible is this really going to be is this something that you know can keep standing on his own as a business or will there be you know so many parallel efforts and just in some sense a race to the bottom rather than an actual business but you you know you've proven that this is a real business I'm curious when you think about back then how you thought about defensibility and looking back what do you think makes scale sticky and people come back to them won't go to somebody else who might have a slightly lower price but then maybe can't offer everything you're offering yeah so I think the first piece is um competition is kind of a reality of business it's maybe the like one of my macro realizations um over the you know since running the business and um you can even see this now with with Google you know we've sort of like lauded Google in search for you know decades for being this such a defensible business and then now um with uh with Bing and Bing chat and whatnot the sort of like um the threat which we'll see obviously if it comes if it sort of happens it's like hey we're gonna have Bing chat and we're going to price our ads way lower and you know Google you're gonna have to respond because this is gonna be this is gonna get really tough um and so I do think competition again is always a reality business and then one of the things that sort of like I I really sort of um almost instructively from from Amazon um learned or took away was that um the marriage of technology and operations uh applied in sort of a in a in an Innovative way and with sort of continuous Improvement could actually become or is is actually in and of itself remote and so you think about Amazon as a business for a second um they have two businesses two major businesses one is um uh e-commerce and and sort of being the everything store and the other is AWS where they sort of are this incredible Cloud platform and in both cases you actually think about what their what their core mode is and in both cases the mode is actually it's not pure technology it's actually a combination of technology and operations so in the um on the uh on the e-commerce and the everything store side that like at this point Amazon's number one mode is their Logistics it's the fact that you know I buy something off Amazon and I can get it in one or two days which is absolutely incredible to be able to support you know millions and millions of products delivered to you within one to two days that's an insurmountable value prop compared to anyone else and you think about what actually happens when you click you know the buy button you know it's it's actually very operational and sort of like made in the Intensive process like somebody goes into a warehouse figures out where one of these you know obviously love is like more automated now with robotics but um which I'm sure you know all about but someone goes in the warehouse finds the object finds whatever object somebody wants to buy puts them into a box packages it up mails it out and then that this box gets um gets delivered and Amazon now owns a bunch of the delivery process and so when it gets delivered it goes into a truck and then somebody takes in somebody drives it you know it's just like is this very manual processible out of steps but they've been able to apply technology to each of these steps successively and quite quite obsessively over the course of decades to result in sort of this like insurmountable value prop um the same thing is true on the cloud side you know in that case it's more around devops so how can you like deliver all these services that just massive scale with such high availability at low cost and it's the same thing just with devops instead of sort of physical operations and so when I thought about labeling you know I didn't think about it as like hey it's true that it's operational you know humans are involved in the process of labeling and so as a result it's not this sort of like it's not a pure Tech problem where it's not a sort of one of these clean problems but I you know I always had a lot of conviction in the same way you know if you were able to marry technology and operations incredibly well and sort of take the process of of getting a piece of data labeled and getting humans to chunk over it and verifying the quality and you were to sort of scrunch that down or optimize each step with technology and and our own algorithms and and our own systems then you would result in something that was like impossible to catch up with and so that's really how we've thought about the business and I think what's played out in reality which is that you know like like as you mentioned back in 2016 2017 um there was a lot of competition I remember we would go into um uh sales pitch meetings and people would say you know I've gone in the past month I've gotten emails from 30 labeling companies um and and so so so sort of tell me how you're different and I think it at first though like maybe a little daunting but over time we sort of prove that we prove exactly what I just mentioned we could achieve higher quality um higher scale um you know best pricing and and uh and continue scaling in that way one thing I've always been curious about um and maybe the information is available online but since you're here I'm just gonna ask you all right what's the role of learning models to accelerate the labeling process the best the best ways to think about this are to that like um you know I think I think Tesla has some great presentations on this actually around how they build up their data engine and where they apply learning models in the process you know I think they're sort of like um our philosophy is really that like you know the output or the outcome that we want is really high quality data that matches as well as possible the data that you want your the sort of like thing you want your model to predict um there's a lot of sort of like judgments in that process that it might be easier for a human or a machine to make assuming that they have like enough of a prior that prior needs to be usually given by you know that prior might be hard for the mission to produce but but the actual result might be easy for the human to produce so one example is like in producing in outline of an object um machines are actually pretty good at producing elements of objects but especially if you can give them a like a tight bounding box around an object they can produce that outline extremely well then if you ask a Human Instead to present outline of that object that's pretty painstaking and takes a while I mean there's just a lot of clicks involved and so a lot of the process of learning models like how can you reduce the problem to how to reduce one step that's very expensive for a human to do in two two steps where the first step is efficient for the human to do and the second step is efficient for the machine to do or you know vice versa efficient for machine and then efficient for human but it's sort of like um P sifting to get like sifting through these sort of like complex and manually intensive and sort of like long tasks into uh the sort of like almost the sort of like eigenvectors of tasks or if you will or like what are sort of like the principal component tasks that that you can Farm off some of them to humans some of them into machines and therefore generate um really high quality data using you know dramatically less effort that's that's sort of like one big part of it and then the other big part of it is that um humans actually make a lot of mistakes um uh that's that's sort of like a a trait of humans that we're we're not perfect and um Quality so quality control becomes really really important in any sort of uh manually intensive process whether that's manufacturing or software develop development or uh or or um data production um and so a lot of is how do you apply learning models to help catch from the humans are making mistakes and vice versa so that ideally these sort of like errors that the humans of the machines are making are non-overlapping uh yeah not overlapping so that we get overall much higher quality data it's very interesting because even though in some sense you start out from labeling as or high quality data a lot of it initially labeling being the key thing to provide and at first it's in some sense the Dirty Work nobody wants to do but soon enough you realize then to do that work efficiently you actually get to do machine learning in-house to partially automate not fully automated obviously obviously otherwise there's no real labeling involved it partially automate that process and in fact you probably built some of the most advanced models in-house like marginal net models for vision language as a way to Aid the labelers I imagine yeah totally and this is actually like I mean I think in the early days we actually it was hard to um you'll remember these days well um it was hard to hire um the best machine learning Engineers because you know they could work on our problem which is a subset of the overall sort of like these like um these sexy AI problems or they could go work at let's say a self-driven car company um and work on like you know arguably one of the sexier problem statements of our time and um and so that was it was one of the harder things but I think it's we showed you know as time went on and we sort of showed what the like actual problem set that we had was all of a sudden became a lot more attractive and again I think I hate to keep referring to Amazon but but I do think there's such an inspiring company in many ways because you know um Amazon is probably the like at this point the largest scale uh a player of Robotics to uh to many of their processes at least in definitely Advanced robotics um and a lot of us because they they focus on these like very operationally challenging problems and that gave them sort of like the playground um and the sort of sandbox in which all these very interesting technical problems emerge and so then maybe one of my one of my other learnings which is that um technical uh technical Intrigue is is somewhat fractal fractal in the sense that like you know almost any problem you can find collapse you know sort of can expand into a suite of a very interesting technical challenges I love that observation also it really resonates I feel like wants to dive deep enough into something you care enough about it a lot of interesting things pop out it's never as boring as you thought it was going to be sometimes you wish it stayed boring but uh it's always uh immediately much more interesting now I'm curious um you have many companies as customers are there any companies that you think back to in the early days like who were the initial ones dad uh you know you got on board and are there anyone that still remained today from those early customers in the very early days you know um a lot of our customers were uh the same self-driving car companies and it was sort of again it was because it was sort of The Perfect Storm of they were they had they had to build their own algorithms um they uh had huge amounts of data at that point sort of like um almost like in in um uh unprecedented amounts of data because they sort of like these cars going around collecting terabytes and terabytes of data and they've come back and they have to start doing machine learning on them um and they had just um these sort of uh these very ambitious goals and and huge amounts of resources and so those were these were a lot of our our early customers and many of them still remain you know we still work with the folks of folks at like neuro for example or a lot of the early startups from the ecosystem and um they were the first ones not neurospizzically but other self-driving car companies were the first ones to really take a chance on us because I think from their perspective it was sort of like what do we have to lose you know we're in this um we got to build we gotta build self-driving cars and this is like one of you know 50 problems that we've to solve so we might as well um work with a sort of like smart enterprising company to solve one of these 50 so we can focus on the other 49 and um and this was like the birth of I think a lot of the incredible Tech that we were able to build and the birth of a lot of um a lot of the business up until you know even until you know 2018 2019 this still was like the The Lion's Share of what we were doing was still related to autonomy autonomous vehicles self-driving cars in some capacity and um and it was a it was a very you know it was incredible problem set because a the sort of like the scale of data and the scale of challenges and sort of the the like that was where at the time a lot of the sort of like most Cutting Edge techniques in machine learning were being applied and so the opportunity to sort of like be close to and aligned with The Cutting Edge of technology I think was like incredibly exciting but then also we got to work with some of the most um demanding uh machine learning Engineers um that that were out there and uh and I guess another thing that I've learned and I'm sure you've learned that covariant is that working with demanding customers is one of these sort of um in the moment quite unpleasant but like in the long run very rewarding processes where you know the demanding customers they make your life um pretty unpleasant and and and uh and sort of push you to do things that you probably wouldn't normally do but then in the process you get you have the opportunity to sort of improve your product improve your processes and really sort of um out of all that pressure builds sort of like these incredible incredible things now one thing that stuck with me for a long time for the self-driving car companies is catching up with undercar Posse a couple years ago uh it was still a Tesla at the time heading up the autopilot and AI efforts there he was saying um essentially the majority of time has to be spent on effectively the labeling Playbook which is very interesting because you know people talk about you know self-driving car maybe it's a sexier uh application to work on turns out when you go there he tells you actually the real thing we need to be get getting really right is the labeling Playbook and in fact it's a book that's like 80 well it's not a book literally but it's you know it's 80 Pages 100 pages and you know training the people who um have to then execute upon what's in the book and have to flag when something in the book maybe isn't right then we need to update it like that whole and something's operation as you call it is is the thing that matters which is so interesting because that's what you specialize in and have played it seem play out across many Industries yeah I mean I think it's sort of like um I do think of of Andre um as quite Visionary in general and I think that like one of the things that he sort of you saw he like wrote the short story I can't remember what exactly it was called just it was um it was along these lines but it's sort of like I think the way that you know he really thought about the sort of like the labels and the data sets as like a compression of the human insights and intelligence and knowledge that you were trying to like um you know simulate and emulate and learn from your model and so and so therefore these like the like you know what are the lossy steps here lossy Step One is like what is your what is your labeling guidelines um or your labeling playable because you mentioned lossy step two is like how do people perform against that against that labeling um against that Playbook and then once you have all your labels velocity step three is is the model and like you know the neural network and the architecture of the parameters and all that stuff and how that predicts against uh answer a problem and so it's actually really I think that one of the like Visionary things that Andre really did was that he he focused on you know there's so you know I don't think any any one of these three is like intrinsically less important than the others but uh but there were so many people focused on the last step there are a lot of people focused on trade on building the best and real Network architectures and training those up and building them up um and so the first two lossy stabs there were sort of met much fewer people focused on and that was really where maybe there's more Alpha and more competitive advantage to be to be squeezed out of the situation yeah it's very interesting when when it's something somebody Dives so deep into the problem and then comes back out with insights that everybody else kind of instantly agrees but wasn't necessarily thinking that way um ahead of time now uh a little while ago I saw you at the fortune brainstorm AI Event in San Francisco and you emphasize the importance of artificial intelligence for the military a topic in fact many Tech folks like to avoid how do you see AI play a role for the military and why do you like double clicking on that topic rather than like many others who are evading that topic most of the time totally um so I guess to maybe to start on this I'll give a little bit of my own background so I I um I grew up in Los Alamos New Mexico which is uh where the atomic bomb was first built the Manhattan Project was sort of housed there it's in the 40s you know they you know hundreds of the most brilliant scientists in the world physicists and Engineers um sort of convened and and solve the sort of series of real technical hurdles to ultimately build the atomic bomb which was this piece of technology that I think um thankfully we haven't had too many of them detonated um uh in in the past you know 70 80 years um but uh has been a really you know sort of like Monumental technology for how much of the world and geopolitics and all that stuff has sort of played out and when you grow up there every year you sort of relearn the story of of the Manhattan projects just to like get an appreciation of where you're from but also it's it's this incredible story it's really crazy in a lot of ways and uh and and there's actually soon to be a Christopher Nolan movie about Oppenheimer about this this set of circumstances but so I grew up in this environment both my parents um are physicists they worked at the National Lab they worked on National Security problems and so um I unlike most people uh that I've seen in in sort of like the tech ecosystem in the Bay Area um I always sort of uh was pretty deeply ingrained in the sort of like marriage between technology and natural security and um and so you know as the sort of like events of the past few years have been playing out and you know maybe the most uh Landmark moment was uh sort of like the Google project Maven um sort of like con Scandal so to speak when when Google should refuse to work with us military um due to you know activism from from employees within within Google um it really um a really made me uncomfortable and sort of I think it it uh part of the reason why it caused me to you know honestly Focus even more on National Security is I think that we we kind of don't have a choice I think that one of the maybe differences in how I think about it versus versus how a lot of other people think about it is I think the way that um and the way that I've heard a lot of technologists think about it is um hey we're building the small technology we can all agree this technology would be it would be better off for the world if we never used AI for you know autonomous weapons or um or some of these sort of like violent applications the technology so we're not so we're not gonna be focused on building which I think is you know I can I can really appreciate that train of thinking but the issue is that um this is a coordination problem it's like fundamentally like a you know a tragedy of the commons or or prisoners dilemma or whatever or you know your whatever favorite framing you have is which is that we can have that belief but then there's gonna be bad guys in some other part of the world or some other part of the um um you know somewhere else in one of these other countries who are not going to have those uh those principles and are going to very willingly apply it to um through you know National Security applications autonomous weapons or or the like and I think you saw this play out in practice over the past few years where there's there's a whole Chinese cottage industry around facial recognition so AI applied to facial recognition for you know National surveillance in China and uyghur detection for the suppression of you know that that racial minority um within China and so I think it's sort of like it's very clear to me that we're going to see bad guys um uh around the world utilize AI for uh for autonomous weapons I think there was even there's an example I think in Israel of sort of an autonomous turret that was used um uh or Israeli Tech I believe an autonomous turret that you know whenever some it you know any anything or anyone entered like this like a particular bounding box in the in the sort of video it would sort of be shot and um and killed and so okay so now so assume that that's like a belief that you have that we're in a world where these things are going to be built so so people are going to build these things then um then you get into I think the sort of like the fundamental question which is like okay do you do we believe that America do you believe if someone's gonna build them it's important for America to have them and I think that you know a little bit of of sort of world history and elements for like um I'll uh you know people can certainly argue about this but I do think that the sort of era of relative peace that we've had since World War II is is pretty incredible if you think about like human history pre-World War II it's it's punctuated by War and really like literate with war and in the past you know 80 years have been um surprisingly uh peaceful at Pax Americana I think is what what it's like often called and I think a huge part of that is because America has been the clear superpower of the world we've really LED and this is a sort of a combination of military dominance technological dominance economic dominance you know all these factors together and I think we risk falling into an era of a lot of chaos if at any point America doesn't doesn't lead it on these dimensions and so um and then I think maybe the last thesis here is that I think that the that whatever country utilizes AI most effectively in in their National Security for defense intelligence Etc is going to be militarily dominant I think that you know it's it's really it's not hard to imagine the scenarios in which autonomous weapons or or fleets of fleets of autonomous weapons could massively out compete and sort of outmaneuver um a human coordinated or human coordinated uh warfighting force and we're seeing a lot of that you know the glimpses of that play out in the war in Ukraine of Ukraine versus Russia and they think that that trend is only something that accelerates so I don't talk a lot but like um but that's that's really the core of why I believe what you know what I do and why I think it's important to work on these things and I think I think it is everyone's rights sort of like choose whether or not this is a mission that they want to get on board with but I would I would sort of um I would encourage people to think about the sort of like broader Game Theory uh rather than just sort of the question of do I believe this technology should be applied to this problem right I mean it's very complex I think the game theoretic formulation gives a very different perspective than the just immediate direction to the specific technology what it might be capable of um it's a much bigger picture way of of thinking about it now not sure if you can comment on this but are there any kind of things you've done with the government with the military that you can say something about that are interesting from an AI perspective yeah well I think one of them one of the big problems is just around um you know it's all around geospatial intelligence or satellite satellite image recognition and this is I think one of the the first use cases that the government really started applying which is that you know there's a huge amount of satellite data satellite imagery data that's collected at any moment um from from all those sort of like satellites orbiting the Earth they're costly Imaging Imaging the Earth and um a lot of these images are really high def you know I think a lot of people have used Google Maps or Google Earth and sort of have seen the sort of like quality of satellite images Imaging and um I think one of the craziest things that I learned about before um before really sort of like thinking about this problem is that um the government collects a huge amount of satellite imagery and then really um is quite bottlenecked in the analysis of this imagery you know most of the images go totally unseen by any human and without you know any sort of deep learning methods it's very hard to actually if a human doesn't look at an image then there's not that much value that you get out of the image and so um if you thought if you think about it it's like there could be all sorts of bad or relevant or sort of like interesting stuff going around on in the world that we just don't process because you know we don't have the right we don't have the right automated techniques so one of the examples that we deployed was um in the war in Ukraine we built these algorithms that that in all the major cities could detect um the level of damage associated with um basically on a building by building level associated with the war so that in a somewhat real-time basis you could basically assess like oh there's a lot of damage in that neighborhood or that that Corridor or that area let's coordinate a humanitarian response but also potentially coordinate a conflict response and um and this is a technology that you know I think for most people machine learning sounds pretty direct and and you know doesn't sound too complicated but has immense benefit and impact in the actual sort of like application and roll out and productionization of the technology so that's one example I mean I think there's all sorts of other examples in the you know in applying it to different kinds of data that the government has access to and applying it to a lot of the really manual processes two that the government uses I think what another thing that I like to point out to people oftentimes is that um the government has a huge amount of of really manual processes that don't need to be many um and uh and and those are the kinds of things that I think maybe frustrate citizens or slow us down or you know um and a lot of those show like Legacy problems but I think that part of that is like it's been hard for the government to to keep a pace with just like sort of the the ways in which the world has been scaling um which are which are quite complex and unique and so there's all sorts of really boring processes that you know thing about like Medicare Medicaid or anything about um you know a lot of uh Health and Human Services you know there's just a lot of things that are sort of um Antiquated that that need to be revamped um and uh and we're trying to build an algorithms sort of help call this the exciting thing is this is becoming possible and that brings me to our shida the next topic I want to want to discuss with you which is we've gone from an era of AI that's been amazing and surprisingly good at what it was doing which was essentially building a marginal Network specific to the one task you want to solve and just collect a ton of data for that one task to a new era where um very large models are trained on a very wide variety of data um often called Foundation models and language-based large language models that somehow have the ability to be good at everything somehow and even being good at niches um very often even though they're trained in a much more General way and I'm curious how that shift in AI how you see that kind of just from a pure AI Evolution perspective but also how does that affect the scale business of providing data that might often be more specific in the past to specific applications you know you have to acknowledge how just amazing what's happened is in the sense of like both the sort of like I think the paradigm shift that you're describing in terms of you know the shift towards Foundation models but also just like the you know the foundations of all of themselves are capable of so much more than um many people imagine I think it's really like I think it's been impressive and surprising to me um having sort of like been an observer um of the field and sort of been been in the field I think the sort of like the things that we can now see and are capable of these Foundation models are capable of is like it's totally shock you know I think one of my sort of like litmus tests for this has been I think for a long time you know let's call it um let's call it uh 2016 to 2020 uh to 2020. um a lot of times people who didn't weren't super in the field without asking a question was like can AI do insert blank right or like you like can I use AI to uh do and then whatever is in the blank and then most of the time the answer would be like oh no not really like it's pretty limited technology like you know um uh but it can do maybe like some very small set of that problem and then now you look at it and um and you look at these sort of like large language models and chat gpg like systems Etc and all of a sudden the answer to a lot of those questions is actually kind of like you know it's like yes but it's not very reliable um rather than just like a flat out no which is pretty um which is pretty insane and I you know personally I think that this sort of like last Model reliabilities wouldn't be incredibly difficult to to to to to um to to tackle and I don't you know I don't know I may be more skeptical of our ability to tackle that as a community as a whole but the um it's been a very it's been a whole shift from uh from an ecosystem perspective and I think it's created um I think it's rap it's dramatically reduced to barriers of Entry of being able to use Ai and so that that I think is one of the exciting things all of a sudden for so many problems you can now you can now build Solutions or build products to solve some of those problems with without requiring um all the sort of like pain and hurdles that you needed before which um I think on the whole mean just like a massive growth to the AI industry and I think it's been one of the main certainly one of my main sort of like concerns with the industry which is like hey if it used to be that you needed 100 million dollars to do AI effectively in any way then like how the heck are we ever gonna how many people have 100 million dollars to spend um but now that now that you don't I think it's sort of like this incredible um you know it it this the true democratization moment of AI but I'm actually curious to hear your perspective obviously you um you think about this stuff as well and you've sort of seeing the same transition happen well I've seen it happen on multiple fronts and kind of the most intriguing thing to me from a covarian perspective where we build AI for robotic automation pick and place typo pressures and warehouses is that it's one of those things where maybe five years ago definitely 10 years ago people would have said you should build a specialized model for grocery picking versus a model for apparel picking versus a model for yet you know electrical supplies picking and it turns out that building a single model for all of them is better for every single one of them it's it's this funny thing where specialization doesn't help you it's the opposite it's by training on everything that you get a level of generalization I think that's what you're alluding to is that these models have a level of generalization that's different from what was possible before and allows you to kind of get to much higher levels of reliability because generalization is a lot better than it used to be and so training is very large model and literally all types of objects to then go pick just groceries or go pick just apparel has been to me very interesting has been something we've been doing for a while but to us big part is we didn't realize how much was part as a big Trend until it also started happening in language and we're like wait this is actually very general this is something that is the new wave of AI is that you should probably in general train in a very general way now I think one of the big differences and I imagine this will affect what scale does a lot is I think in the language space one of the big challenges today is that the models just generate statistically plausible patterns of text that's the default and that's not usually what humans want sometimes sure if you want to be entertained it's pretty much ideal but if you want to get truthful writing done or truthfully investigate something this this disconnect between statistically plausible and truthful is a very big one and it kind of for the open-ended question but I'm kind of curious the role scale could play in somehow what's the new kind of annotation that will inject a notion of truthfulness into these models totally yeah and I think that the the um you know if you if I go back to um some of uh you know when you ask how do you actually automate how do you apply technology to the labeling process I think the the the key thing is that and this is what I sort of like my general philosophy is that you want to invest the human effort towards the parts that you know humans are especially good at Android decompose the problem into into these pieces where you know machines can do what the machines are great but the humans can do what the machine the humans are great at and I think in the era of large language models we're seeing that same Paradigm play out just in a slightly different you know form where now all of a sudden the things that you really need the humans to do are kind of exactly what you're saying which is like you know how do you make sure the model responds in like a useful way how do you make sure the model responds truthfully how do you make sure the model um you know actually cites proper evidence or uses proper evidence how do you make sure the model um you know uh it resp like use understand the question in a way that isn't just sort of like you know isn't in line with like what would be out on the internet but there's a lot of sort of like how human like converse with with another human or whatnot and so all these sort of like um all these dimensions in which I think people have you know I think one of the one of the cool things that's happened in the past six months is that gpd3 has been out since 2020 right like it's it's it's been out there but then it was Chachi and the sort of like that form factor and that model that all of a sudden really um blew it up at least in the eyes of sort of like the the average consumer and um and the you know the the common understanding is that the biggest difference between um chat gbt and hgbt3 was the application you know there are a few differences but the biggest one was the application of reinforcement learning with human feedback um and sort of uh doing exactly what we're talking about where you have human experts look at you know what are all the ways in which the model could finish the sentence or finish the or respond to a query and have them have the human sort of say that's actually the best one because of X Y and Z reason because of um because of this and um and uh and it was through a lot of that you know to the tunes of like tens of millions of dollars of that that um that the model actually becomes dramatically more useful and even it was actually really shocking to me but even um I was talking to a journalist the other day who was sort of like um they were describing this process and sort of the the layman's terms that in in layman's terms and the way they sort of articulate was like hey it's almost like gpd3 was a genius but it was sort of like you know it was impossible to communicate with and then all you through using rlhf just sort of like you know taught it how to communicate um in a way that was actually useful to people and I don't know if that's the Perfect Analogy it's certainly um but it's like a relevant knowledge it's a way to think about it which is like hey all of a sudden um the ways in which you uh the ways in which you apply human Insight is to make is to make the models more useful which is I think in line with The annotation on the whole was maybe moving more towards the concept of you know the broader concept of alignment which is you know now that you have these powerful AI system how do you align them with human intents and what humans want rather than sort of um rather than being isolated in some way from the from sort of like um from what humans care about so so that's one way I think about it but then and then this trend of rohf revers are learning with human feedback we've seen as just a is is um is like a very big and growing part of our business it turns out that um you know uh at least right now looks like this is the main way in which people think about building really really incredible high performance models and so it's become a really big part of of what we're what we're what we're providing to a lot of these AI companies and and I think it's great Paradigm it's a it's a it's a pretty efficient Paradigm for getting the models to improve um and it's one which really um utilizes and emphasizes which humans are really good at so I'm excited about it just to maybe make sure I fully understand it in your case do you see the would somebody provide to scale let's say a large language model that they are reasonably happy with and then you would run reinforce learning with human feedback meaning it would just generate things and you must would rank and that's the new effort the humans do ranking responses and so forth to to then improve that model um now what I'm curious about is when you run this right different humans will have different responses what they like what they don't like and of course we have some Extremes in the ecosystem where it's not so much humans feedback but where there is things train on specific humans sound like Elon Musk sound like this person that person it just seems like this human feedback is a pretty big open question there um and sometimes what your customer would want it to be what what type of feedback they would want seems like a whole lot of Playbook to be built there possibly customer specific yeah no I actually think this is like the to your point around when when um Andre talked about how uh a big part of of the work in autonomous vehicles or building on tops here called Tesla was the uh with the labeling guidelines I actually think this is that kind of on steroids which is that you know and this is let's get some more philosophical things but I think over time what's going to differentiate these OMS is is you know in part you know who is able to apply more compute to it or who has slightly better techniques or who has like slightly better set of tokens upon which they trained upon but um those things are are very hard to have as sort of like you know real decisive Edge on relative to anyone else and so a lot of the I think then you get to this like new phase which is much more nuanced which is like what are your what are your subtle design decisions that you that you build into the model and where are the design decisions get made I think is that the step of like how are you defining the rlhf guidelines how are you how are you defining the sort of like how you want these human rankings to be produced and um and I think this is pretty akin to you know almost like interface design for for a a UI right it's like okay this is where you're sort of like this is where you're deciding if somebody asks a you know a purely informational question what should the response kind of be like or if somebody asks a more opinion based question how do you want the model to respond or someone asks um something that is about coding what's you know what's the right structure and format you know it's actually all of these are sort of subjective decisions or many of them are subjective decisions and this sort of like design Paradigm is how you build out the rhf principles and so um so I think it's I think it's really important and then and then you know to the church to the point before and around like what are the lossy steps you know lossy step number one is okay how do you distill what you actually want into the set of guidelines velocity step number two is um you know how do you uh how do you make sure that people are actually how are people performing against that how are Layla's performing gets that set of guidelines in lawsuit step three is sort of the magic of the models you know how's the model actually rock this and then and then kind of um uh make sense of it all and so um and so under this Paradigm I think it's sort of like it's it's it's really really important and it's sort of the you know it's it's undeniably really important it's it's very interesting to me personally as a reinforced learning researcher how you know reinforcement learning went from this thing that is this promising thing always you know it'll it'll be critical in the future but it's maybe not all the way there yet it's not being made part of products yet sort of you know contextual Bandits for for advertising and recommendations to all of a sudden it's just staying down seemingly makes the difference between as you said tpt3 that's been around for almost three years to chat DVD coming out a few months ago being essentially the same but but that subtle difference of that component and completely taking off and taking the whole world by storm not just AI researchers dangers but everybody being interested it's it's so uh it's really intriguing to see it have play out um and I actually another cool part of this has been that um you know the uh this is like one of these these funny things I mean this like the instructor GPT Paradigm or or in general sort of like this reinforcement learning learning with human feedback because it's something that open ai's been working on since um since 20 2019 I think was probably the first paper maybe 2018 was the first paper they published with this sort of like idea and so it's really not actually a new idea at all either it's been this idea that's been out there for a while and I think that the maybe one of the insights of of chati was like okay you applied in huge measure and all of a sudden you get like this crazy new performance I think this may be one of the kind of exciting things about machine learning and Ai and research is that you could is probably the application of old existing idea that actually when applied in great measure actually are going to result in like incredible uh incredible outcomes yeah it's a very good point I remember the first uh RL with human feedback paper from open AI um was actually even earlier I think maybe even 2017 or 2016 and it was to teach your robot to backflip and the reason was because normally when you want to do reinforced winning with robots the typical Paradigm would be you ride some piece of code that is the reward function and then it tries to optimize that reward that's equivalent to the score in the game but for a backflip how do you write a piece of code that says this was the quality of the backflip it's like it's not clear how you how you do that part so you don't have a scoring mechanism instead you must would watch two backflips and say which one are attempts really early on because the agent doesn't know how to do it say which one of the two is a better attempt and over time it would actually learn the reward function which is also what's happening of course now with with these language models learning the reward function to be optimized because it's hard to specify by hand which it's so intriguing how RL kind of in that way which many people I think at the time saw is like a very specific version of RL with human feedback much more tedious but it's actually the one that that makes its way into the most prevalent application today is as you'll acknowledge like pathing in AI I think is always really funny and and kind of like uh you know it's it it uh it never exactly matches sort of uh sort of what you think is gonna happen and so even the emergence of language models I think was sort of like I think there's sort of like two reactions you know I think I when cheap G3 came out that a lot of people have or even gpd2 came out but it was sort of like you know the reactions were oh this is like wow this is like really something or or like a lot of reactions are like oh yeah it's cool but you know it's you know this this and this and this are always going to be limitations or like it doesn't demonstrate like real reasoning or you know whatever it might be and I think this sort of like one of the coolest things that the AI Community has shown is that um I think it's really surprising honestly that as you just keep scaling it up a lot of those criticisms like slowly slowly um uh go away and so you know it is it is this crazy thing where it's like we're at this point where we have this sort of like Paradigm that you know we keep scaling it we see crazy new things who knows what happens if we keep scaling it forever talking about crazy new things um I recently I saw you have someone at scale with a job title called Advanced prompt engineer or at least that's how the advertised sells uh people have been joking about this right uh this prompt engineering as a new job but sounds like this is a real real thing now uh prompting as a job yeah no I think that um you know one of the one of the illustrations of this that I S that that um that I thought was like really interesting was that if you um if you went to uh if if this was like on Twitter recently but there was people who figured out how to like jailbreak out of the GPT chat or sorry not to be a big chat um if you're gonna know how to get the Bing shot to reveal the entire prompt that's used to sort of like um to sort of like set up that chat and um the problem that sets us up is is uh it's quite long quite detailed and there's a lot of like different nuances to it and I sort of imagined the sort of like the Bing team the Microsoft team before they launched a big child sort of like a b testing different problems versus one another and trying to figure out which one was sort of like what are the right sentences to include the prompter what not but I really think that is the Paradigm in many ways of the future I think that's sort of like one way to think about these llms is that they're in many ways like a new computer and um and figuring out the right ways to sort of program this computer this sort of like cognitive computer um in ways that actually produce what you want and and sort of like Get It engineer the right outcomes from it is going to be um is giving a new way of um a new way of programming or certainly like this like new frontier of how to get machines to do what we want you know I think another um I think one of the this this uh this prominent engineer that's on our staff that you referenced uh his name is um Riley Riley good side and I remember when he showed me some of the sort of like prompts that he built and and you know some of them were like he would sort of like coax the model into the exact kind of reasoning or the exact kind of sort of like train of thought that you want to make sure the model had before doing some like very complex verb analysis or or he would you know there's one this was actually maybe the most incredible where he sort of um taught the model how to when it didn't know the answer ask the python shell for like you know for the you know these models aren't good at doing arithmetic for example so when the model doesn't know what it wants to do teach it how to use the python shell to be able to to ask for an answer and these are I think are our crazy illustrations of like how how you can sort of like use this new computer in sort of like interesting ways and um I don't see that going away I think that like we're gonna need to you know need to know how to use these these new computers just in the same way that you know there's coding which is which is the art of making computers do stuff well and there's also sort of like um there's all these Arts around getting humans to do what you want to do you know management is an arch or um or uh you know sales is an art or or marketing an art you know you know so I think that um I think it's maybe somewhere in between these of sort of like somewhere in between management and coding but um but there is is sort of like a um a study around how to get uh these models to sort of do what you want and they think it's sort of a um I don't know it's it's one of the exciting fields of human knowledge that I think has yet to be explored so and is highly valuable you know Bing chat you know that's a valuable thing that's produced it's very interesting that the prompt that you you give it affects so much the behavior of course it makes sense the example you brought up is very interesting to me because it sounds like a whole new kind of security research area where you know which prompts are such that they're also secure against being revealed because if you spend so much time engineering this really good prompt if somebody chats with your Bot and then the bot just reveals it that's it and then you lost your advantage maybe so it seems like there's a lot of interesting things that will play out in that space yeah no I think it's uh I think the the there's so many security um I don't know if there was flaws but sort of like uh implications of these large language models that I think that's gonna be one of the bigger you know I think everyone's trying to Grapple with it but it's one of the bigger things we have to think through as we roll out the Technologies like how do we do so in a way that um you know if you think about you know if we go back to think about the government example right um we really you really don't want your adversary knowing about any of the vulnerabilities of your model um the issue is all these models that are out in the open they all have huge amounts of vulnerabilities and it doesn't take you very long to find the vulnerabilities that they have and so it requires for this this like secretive Paradigm where I think we're gonna all have to you know build models in secret that that um that other people don't have access to and therefore they can't identify the vulnerabilities but then they're still sort of pretty fragile or sort of um pretty uh pretty vulnerable so it's sort of a you know I don't know if you're I'm sure you remember this but like there was an ERA where adversarial examples was sort of like the hot thing in AI research and there was sort of this example of like a sticker you could put on uh on anything and all of a sudden the like an image classifier would would think it was a turtle or a God or whatever it was um but that that kind of thing is like you know just an illustration of again how fragile the models are and and so therefore it creates all these risks going from the models to the impact they'll have right um in one of your recent talks actually the title of your talk was applying artificial intelligence to redefine every industry um of course that's a lot to cover so are there maybe some industries that you see in the next few years uh will really see a big revolution uh thanks to AI one of the analogies I like to use with folks especially folks outside of the world of AI is I think um what people see which is Chachi BT and stable diffusion and um and maybe the music LM and Bard and stuff like that that really is just like kind of the tip of the iceberg of um of AI uh of sort of AI impact and a lot of the the sort of like real impact is going to be under the surface and it's in sort of like bleeding and then flowing through all these sort of like major industries and sort of driving just like very meaningful um economic and sort of like uh an automation of value sort of throughout the biggest industries in the world a few that I'll reference um that I think might be the fastest to change I think that advertising um is gonna fundamentally shift where we're going to move towards a world of fully automated advertising um today um you know if you think about how advertise advertising is done there's actually a very small number of ads that are out there ads in this form of like you know what's the actual content that is in the ad and um and companies spend billions and billions of dollars of producing and sort of Distributing and showing people this content but fundamentally that's you know it's a pretty imbalanced equation you know we there's so much user generated content out there but there's still so few ads um so I think that we're gonna see this sort of like arms race of Brands and companies um moving to build you know millions and millions of variations of their ads using generative Ai and then assessing them independently and seeing okay what is going to be um what's going to perform the best and so the way this is going to play out for each of us is we're going to get ads that you know seem highly more personalized um or potentially more creepy and are sort of like much better tuned to be the exact thing that's gonna get me to click and sort of like um and and drive a purchase so that I think is one area that changes pretty dramatically um another that I think we always talk about but I hope I hope I'm right about is uh is Healthcare um both in the sort of like service of healthcare as well as uh as well as health insurance um health insurance for what it's worth is this massive massive industry trillions of trillion dollars to spend but is ridiculously inefficient and very very manual in how fundamentally you know behaves so that's a clear area where chat should be like systems or general large language models can produce huge amounts of automation to actually drive Improvement there but then um in addition to that so is um so is uh so it's Healthcare itself you know globally speaking there's a massive shortage of doctors um you know most individuals most people when they most care people who get care even in developed countries they can only really afford one opinion even though the sort of like consistency of doctors you know diagnoses or responses is quite variable so so we're in this sort of like very constrained world of of how Healthcare is delivered where because we're constrained to one opinion from one doctor and there's not enough doctors um you know we have a lot of sub-optimal outcomes whereas I think we're going to get into a world where like as an LJ just mentioned humans are going to do what humans are good at machines are gonna do what machines are good at and you're gonna get dramatically better medical diagnoses globally so um those are some examples I think we talked about defense and and National Security that I think is another really huge example and I think the sort of like you know you can keep going but the examples are everywhere this is really exciting to to kind of hear your vision of the future of AI and the impact will have on all our lives especially with through things like Health Care um if you have the time I'd like to ask you um one last question which is you're very busy um what do you do to take your head off of things and to relax this has changed uh through the years I think that uh that um one of the newer hobbies that I've liked a lot is is hiking which um I didn't used to like very much I didn't used to like physical activity too much but uh you you sort of you learn to love this sort of like meditation um and the the sort of yeah meditative component of it which I which I really appreciate um and then I I love uh I love consuming content of various forms so whether that's reading or watching TV or watching movies um I think it's sort of like one of the Great joys of of you of life is that you can sort of like experience these stories that are they're like really built to to sort of like elicit these like interesting responses out of you and um and then and I love history um which I think is is common among a lot of people that I know but I think that's sort of like the one of the like most interesting sort of like conundrums of of humanity is that like on some on some level humans haven't changed that much and so sort of like you know um human Tendencies haven't changed that much and how humans behave is not too fundamentally different from from that of you know um thousands and thousands of years ago at the same time uh our world has changed so dramatically Society has changed dramatically the technology we use has changed dramatically um how we relate to one of those change dramatically and so there's sort of this like dual component where on the one hand you know maybe some people think that human nature is extremely predictable and if you study enough history you know exactly what's going to happen on the other hand we're in unprecedented times with unprecedented Technologies like Ai and so it's like impossible to know what exactly is going to happen and so I think that you know studying history is one of these like it's almost these great puzzles which is like you know are there actual scalable lessons that you can learn across human history or or we sort of like always barreling into into the unknown and um and I find that super interesting and therefore love learning about you know about what humans have just asked that really resonates and I'll I'll say a big reason why I enjoy hosting the podcast is because I think it's so interesting to hear the Back stories behind the people who do the Innovations in Ai and to see I should it's quite a wide range of backstories and motivations and paths to get where they are today and see what you know hopefully other people can learn and be inspired from it and definitely your story Alex is a really amazing one um I really appreciate you making it time to share that and also to share your vision for the future of AI thank you so much thank you so much Peter it's a lot of fun

Channel: The Robot Brains Podcast

Views: 7,726

Keywords: The Robot Brains Podcast, Podcast, AI, Robots, Robotics, Artificial Intelligence, LLMs, foundation models

Id: BNqb3Xuv-0A

Length: 73min 1sec (4381 seconds)

Published: Wed Mar 29 2023