Cerebral Valley: Daniela Amodei (Anthropic) and Hema Raghavan (Kumo) w/ Konstantine Buhler (Sequoia)

[Music] foreign [Music] partner at Sequoia capital and uh have been in the AI space for quite some time 13 years believe it or not like since undergraduate uh have been in the AI space compelled by it and this is a pretty exciting moment for everyone and so I'm very very excited to bring two brilliant Founders in different parts of the AI space doing very very important things in the AI space so Daniela and Hema are are truly brilliant in different ways and in different types of companies and so the first thing I'd like to ask you both is to please introduce yourselves just a 60-second intro who you are what your company is uh what you guys are doing there and then we'll go from there so Danielle oops okay I got that thank you uh okay I guess to start off uh I'm Daniella I'm the president at anthropic just a very kind of quick uh background about me I actually originally started my career working in global Health spent time working in politics both uh on a campaign and then later on Capitol Hill and I moved to work in Silicon Valley just about 10 years ago almost to the day I took a job at a very tiny startup at the time called stripe that was about 40 people and was there for about five and a half years really had the opportunity to grow and scale people and business teams in my time there and then move to work at openai in 2018 and did something really similar across research engineering business and people teams and I co-founded anthropic along with my six co-founders we started in very very late 2020 so we've been around for just over two years an anthropic's mission is really to build reliable steerable and interpretable AI system items that really have humans at the center of them amazing hey Ma I'm him and I'm a co-founder and I head engineering at kumo.ei I started my career in text natural language processing question answering a couple of decades ago and a very different world in the tech space back then over the last decade or so I've been working with graph and graph neural networks and Kumo is about a year and a half old company and our tagline is to query the future we actually aim to make querying the future that is building predictions as easy as it is to do SQL today so today you can go to your data warehouse and actually ask a question like how many users are going to churn or how many users churned last week and what kumu will makes it just as easy to do is answer the question about how many users will churn next week and you can ask many different questions in a very SQL like interface so we want to make AI easy and accessible for many different audiences very cool and now Hema mentions natural language processing this is the OG llm I think it's fair to say you've been looking at that for a long time I understand that when you worked in a lab back in the day there was a neighboring robot with the same name uh the same name as Hema and Hema can stand and I got a lot of flack for that so we've come a long way in the general AI space I was in the NLP lab and that was the robotics lab and Hema constant she stood for hierarchical embedded some I forget what the full acronym was amazing well this Haymarket constant you can also run a great AI company so it's a killer combo if both of you were not running an AI company what would you be doing Daniela I think if I weren't uh running anthropic uh two answers so the first is uh I actually really loved the work I did in global health I didn't have the skills then to be as impactful as I wanted to but after uh getting the chance to grow my skills in Tech I might go back and do something there um the other answer is that I would love to coach a high school football team a funny story about me I know a little ghost in Silicon Valley I'm a huge sports fan and I love coaching management and development and I would really love to do that in just a totally different context so the real life Ted lasso right here and and we will talk about that in the culture section that's great hey Mom um an AI startup yes so um if I weren't building an AI company I still I still think I'd be doing something with AI because I've spent two decades in the field I think there's just a lot more to be done and I wouldn't want to leave at this moment in time but uh just like you I love coaching I love mentoring uh constantly now selling I love mastering my headstand yes so maybe I'll be doing some of that but beyond that I think I'd still be you know taking AI to you know the next level it may not be building a company but in some other way very cool okay um I want to double click on your taglines as businesses so they mentioned it a little bit in passing um you guys are obviously one of the leading llm providers in the world but one thing that's different about you is this concept of constitutional AI has everyone heard this concept is this something we're pretty familiar with maybe you can rehash it and tell us a little bit about that and then Hema this concept of querying the future or predicting the future then maybe you can show us how these two connect as well sure yeah I think maybe even backing up kind of a step from constitutional AI really the the kind of way that uh the Safety Research and work that we're working on at anthropic we've really sort of seen in uh Claude which is the product that we push to Market is this kind of framework of being helpful honest and harmless and constitutional AI is really an approach to achieving that uh through a different sort of system of of training and you know normally reinforcement learning from Human feedback which is something that our team also sort of you know co-developed uh is is the sort of standard approach for doing it right now but constitutionally I really aims to use kind of you know what we call a constitution of norms when training uh you know large generative models for helping them to achieve that kind of HHH set of goals cool and HHH helpful honest harmless and the fourth age oh yeah I have a this is an internal thing but um I I kind of think that we should add a fourth age which is humor um Claude is really really funny if you've had a chance to play around with it um makes great jokes every you know mostly dad jokes but but very good dad jokes it is funny I hear that it has a large database yeah that was that was a dad okay that was really good there we go the claw joke right there that's a claw joke yeah so making uh so uh the goal of Kumo is to you know help you query the future right and if you uh take a look at how data sits in your Enterprise today it's largely in data warehouses and building machine learning pipeline today on your data warehouses is onerous it takes it could take three to six months to still build a uh you know initial prediction we want uh you know Enterprises to be able to try Ai and all different parts of their product so you you need to shrink that time to you know bring your first model out and that's what Kumo aims to do we you know with a very simple predictive query like interface we we enable you to be able to create a first model in a matter of few hours or a day and be able to stick it into your product right and we want you to capitalize all the data that's the petabytes of data that's sitting in your data warehouse and uh that's you know the Kumo promise I think I want to bring uh I said I spent about two decades in the you know text space and we used to do a manual feature Engineering in the early 2000s you know I was telling Constantine I was trying to fit part of speech tags into a 512 MB machine and we used to just have to manually squeeze the juice out of that right and where are we today so can we take that same Journey for Enterprise data that's what Kuma aims to do cool so you're taking uh massive amounts of Enterprise data anymore and you're helping you're helping one thing that's very captivating to me is you're helping Enterprises treat a database uh of past data the same way that they might treat the future so there's this Common Thread between both of you of prediction so maybe you can walk me through what that means and like spell that out for the audience so that they understand what predicting the future looks like for an Enterprise database and then talk a little bit about your customers who they are what they look like why they would be using this and then same for you Danielle so let's start with an example yes a customer I think it's always great to start with an example um an example could be a Marketplace a customer who has buyers and sellers and one of our customers is in that space and they want to be able to determine uh you know who's good what's the next best action right like and that could be a predictive problem that they want and based on that they might want to send a marketing campaign an email or be able to actually even uh you know personalize the the app experience all of those are predictions and we enable all of those predictions to be done in a matter of a day so so that's the Kumo product and the underlying technology is graph neural networks which we think is which basically subsumes a lot of the you know our uh the rnns cnns and lstm architectures and allows us to capture uh the relational structure that exists in these data warehouses we will visit that again a little later so I think it's super powerful concept Danielle for our part our customers really range so uh businesses of all sizes so everything from you know brand new startups to household name companies uh can make use of Claude right of these uh very powerful generative AI systems and the other is you know individuals so on the kind of uh business side some of our early users have included you know notion assembly AI Robin Ai and Juni and on the kind of individual consumer side if you've seen quora's po app that's partly fueled by openai and partly fueled by Claude from anthropic and then just today actually we released Claude and slack and so any workspace now has the ability to integrate it as an app so you can use Claude as a help helpful friendly assistant to do things like summarize threads or help answer questions click on links things like that I have been using Claude and slack for months now I hope I can say that and I can say it does live up to the hype it's very very good may or may not make its way into the memos I write but that's off the Record so this idea of prediction just want to double click on this a little bit more for both of your companies because as I listened this morning as I listened this afternoon and as I think about my own AI training a lot of artificial intelligence is based on this concept of predicting the future in different ways I mean what you've built with Claude what's being built in llms are Auto regressive models that are predicting the next token based on massive context Windows what you're building are basically predictions about user actions based on relational data in a graph maybe you can tell us a little bit about where we are in the evolution of the overall AI Revolution and where you think we're going to be going next each of you from different perspectives um maybe whoever would like to jump in first okay I go ahead okay so very polite group here so um I think in terms of uh I think the uh AI for multimedia data is in a journey ahead of where it is for Enterprise data today in the sense that we've gotten to the place that we have these Transformer models this scaling we've actually we we have enough evidence out there about and we have applications booming right like this I mean all of the talks this morning are all about applications that are booming um and I want to remind ourselves that this did not happen overnight it was uh several decades of work that actually you know brought us to this inflection point I think with relational data with Enterprise data where you know I want to say halfway through that journey in the sense that we've seen lots of wins in some so the kumoko founders are from Pinterest Airbnb LinkedIn and a lot of our founding team are from you know b2c consumer companies and if we've seen wins off this technology in these spaces right so we want to actually bring that to the the broader Enterprise uh We've made inroads into making compute cheaper for graph neural networks I think the myth was that you know they were not cheap and that was the case for Transformers maybe I mean they're still not cheap but at least you can scale them out right and now we've gotten to the place where we can scale them out so we're we're getting there and you know we're starting to see some pretty good wins in uh the GNN journey and I think the Transformer journey is you know I'll let you jump on that one yeah yeah thanks um yeah I think you know from from my perspective you know sort of similarly you know to hammer I think in some ways like the kind of generative AI you know Revolution is here right we um so many people probably you know in the audience would say oh we're at this inflection point and I think that's absolutely true right we're starting to see large language models be able to do um many things right that we didn't think were possible even a year ago or a couple of years ago um but I also think in some ways we're still really early in the journey right there's a lot of um things that these systems still still can't do right Hallucination is still a big problem um and I think there's I think there's really quite a lot of you know potential directions that these these these Technologies can can go in and I think that can happen both at the sort of foundation layer right of of developing more powerful models um you know our team is sort of known for doing some really interesting work um on a topic called scaling laws which is a paper that you know we published in in sort of early 2020 the that basically says that as models kind of get you know more resources right as they're given more compute they're just going to get more powerful and I think that's really been borne out kind of of in the in the past year or two years six months but my sense is that that's that's kind of the the sort of training layer and then there's this other layer of what do you do with these powerful models and I think we're still very very early in that journey and and unsurprisingly a lot of that Innovation is being driven by uh smaller companies and and you know early stage tech companies but also a lot of the larger tech companies I think have a really important you know role to play in this too and I'm I'm very curious to see how you know things like um you know major tech company assistance improve with kind of the development of these large models and having them kind of go to scale so let's talk about that David versus Goliath type of situation for both of you I mean in the case of anthropic you guys are are really competing against the biggest players in the world uh these are trillion dollar companies that are building llms in the case of UMO you guys are up against automl at also the biggest companies in the world who can throw a ton of resources against this problem why why are you going up against that what makes you think that you can win uh what's the end state of this whole AI problem five or ten years from now um you know what is that the the generative AI space is going to be really large right as kind of a segment of of the economy and so I think while the large companies will of course capture some of that value I would be sort of surprised if there were not room for quite a lot of innovation um and I think you know it's kind of the story of of any major company right depending you know people said the same thing about stripe they're like PayPal has has this unlock how could you possibly improve on kind of payment processing I don't think that that you know I think everyone here today would agree that that wasn't true at the time and so I think because we're still fairly early in this journey there's there's a lot of room for Innovative small Nimble creative hard-working companies just like in any industry to come in and and disrupt and innovate and I think in a lot of cases those are often the companies that are actually driving The Innovation right because there's more of um a desire to create and to innovate um versus sort of maintaining the status quo totally that's that is good news for all of you out there building companies yeah so uh I think let's talk about creativity right so I think uh larger companies uh may have Deep Pockets but oftentimes the boxed in terms of creativity because you know you have a product line you've created swim lanes and you're trying to you know uh streamline that production line for what's your Revenue stream right which basically means teams are just focused on something super narrow and uh creativity comes at the intersection of areas and boundaries and that that that's a large reason why bigger companies fail to innovate so creativity is with the younger players right and the scrappiness and you know sometimes scrappiness actually gives room for creativity right like at Como we actually uh you know someone talked about not having GPU resources early in the club you know early enough in the cloud that really made us innovate in terms of how we thought of our model training architectures and so on right your investors are cheap that's what I'm hearing about Buster so in that scrappiness also there is innovation but I think that's uh you know even if I compare it to about a couple of decades ago what happened was Innovation was restricted to universities and then uh you know there was research labs but now I think Venture funding has come is also realizing that hard tech just needs breathing room and space to build out and I think that change has been extremely refreshing so I think that I mean we've we've seen some of the players you know uh that that are talked about here just coming from the fact that you know some of the Venture funds are willing to take that big bet to you know because you need that horsepower that you know Computer Resources the teams to actually build that so I think uh you know it's it's good times to be a startup and solving hard problems Venture funding would recommend I'm kidding Uma are you able to talk at all about the cost advantages I don't know if you're able to or not but you guys have become world leaders in graph neural networks in short order and have significant cost advantages against some of those trillion dollar companies I don't know the anthropic benchmarks but I'm sure they're also incredibly impressive are you able to talk about any of that if not we'll move on to the next question I can talk in terms of at a high level about tech Innovation so you know uh the uh assumptions we used to make with graph neural networks in particular was you know graphs are just hard to distribute so when you think of text you can chunk up your text and put it on different machines and you can train a model on that and then you you know you can move data around but you know uh during gradient descent but graphs are just so inherently there's no natural separation so you would think that you have to stuff everything into memory and we've innovated we've come up with something called compute compute separation internally which just lets us separate out some of the CPU compute into the uh and the GPU compute and it just came from the scrappiness of you know uh you know AWS costs for an early stage startup are still pretty expensive and somebody talked about uh you know people having you know broke a deal so it took us a while to you know prove ourselves and get to that so I think so without revealing costs yeah I think we've been innovated to actually keep costs down it's very impressive and by the way you guys as well I was talking with your co-founder about this last week uh you guys have hit way above your weight class against these big players building great models uh kind of on a budget I mean yes it's expensive but on a relative basis on a budget any comments on how you guys do it or or how you plan to keep doing it yeah definitely um only in this context would you say on a budget but yes I hear you in comparison tons of millions of dollars but uh but uh but I think you know there's there's sort of the like you know the technical side which probably you know my research director would kill me if I shared too much there but there's obviously technical innovations that that we kind of lean into but I also think there's just kind of a um like a kind of cultural approach right which is we sort of have this theme at anthropic of um you know where it's a very scientific founded research team right folks who come from um you know physics phds Neuroscience backgrounds and my impression is uh is that you know there's this almost um like we have this you know internal culture thing right we're like do the the dumb simple thing that works yes and I think in a lot of cases there can be a desire for Innovation to be something extremely fancy and creative and often the biggest kind of Innovations really come from just doing simple things that work effectively that's true on the research side but really also on you know the how do we work together how do we communicate how do we partner together some of those things don't need to be fancy and complicated totally agree totally agree and both of these companies the teams are technically brilliant I do not use that lightly like technically incredibly deep across both of your founding Stacks like you guys are doing the closest thing to rocket science uh I guess SpaceX I guess does the closest thing but the the closest thing to quantum physics okay maybe not quantum physics either but you know what I mean uh in computer science out there it is truly amazing but one thing that you both have have said you have in common is that sometimes it is the simpler things done incredibly well and this is something that I've noticed over the past several years in AI it is sometimes the crazy non-obvious ideas that actually work really really well I don't think a lot of experts thought that Transformers would take them as far take us as Humanity as far as they have taken us so maybe you comment on that and then a comment on graph neural networks and where you think that those will take us as well so I can start books and where we're headed today right so um I think in terms of modeling Paradigm which is relevant across your Enterprise right so if you think about the average Enterprise you have uh you know people entities you have transactions you have uh uh you know the loans you or you might have clicks so you know most Enterprises have that kind of data it seems obvious I mean you store it in a relational database but it seems obvious when you visualize it it's all a graph right and what Kumo's done is actually made it easy to materialize that graph and actually run predictive queries on top of that um from a uh you know uh scientific innovation perspective where we uh where you know we've made inroads but you know we will keep pushing the bar there is actually scalability and uh uh you know the automl piece right like how do you actually you know do a model architecture search how do you keep compute costs cheap and we'll keep you know moving ahead on that I think on the Kumo side as well I must I do want to add that we do have one of the uh the best open source libraries we do maintain one of the best open source libraries for uh graph neural networks it's called python geometric and actually that is our you know a huge community of contributors that come in from researchers from all over the world so you know we have that so we're we're a Lean Startup but then we are actually leveraging the entire research Community by maintaining this open source platform and actually bringing in all the Innovation all the goodness that's coming in papers there and that that lets us leverage you know scientific innovation as well and bring it to the Enterprise so that's the way we see this all of this play out we'll keep innovating and we'll keep maintaining our open source to bring in scientific innovation it's amazing yeah I think maybe on our side I would point to the value of having a really incredible technical research team that communicates really well and effectively with very impressive product people and product leaders and and business folks and I think you know here it's like a little bit of a balancing act right but my impression is that um you know anthropic has very impressive set of you know researchers coming from a broad range of backgrounds and we've been really um sort of pleasantly surprised the degree to which having this kind of environment where researchers can Thrive and where you know Safety Research which I think is is sort of one of the areas we're best known for right we've published over 15 papers across a variety of different disciplines about how to make AI you know more reliable more steerable more interpretable and I don't necessarily think that that's unsynonymous with you know like technology company right a company that's also thinking about how to take that research and turn it into something very practical and valuable that that people can really use and so it's kind of a um a particular take on this kind of innovation question but I think something about the way we've sort of structured the organization seems to be really supporting a lot of our deeper goals well time flies when you're talking to brilliant Founders we are almost a time and we're going to open it up to everyone I'm sure a lot of people had questions for these two but I'm going to ask permission to ask one final question because I know this about them I think that it will be a really fascinating answer that's okay with everyone um you both are culture carriers you're both Founders and co-founders of your companies they're both amazing brilliant AI companies um and you have set yourselves apart within those companies as really important culture carriers can you tell me a little bit about what you've done in order to do that what that culture is like and how you how you Foster that sense of leadership in your companies you go okay um I think first of all always dangerous to ask a Founder about culture we'll talk forever but uh to uh on the on the briefer side um I really think a couple of things sort of about anthropic culture that that stand out to me the first is really this this sort of interdisciplinarity that I've been talking about so we just have folks who who are really talented and impressive who come from a wide range of different backgrounds you know we have sort of what we joke as the physicist pipeline of folks who have you know phds or we're professors in physics who are excited to get into ml we've trained them up and found them to be you know really incredible contributors but also you know former neuroscientists and biologists but also you know product leaders and software engineers and incredible developers who are excited to kind of turn this technology into something that people can really use and Business Leaders policy makers right we're highly interdisciplinary company but I think the thing that unites all of those people to want to be at anthropic is the mission right they're there because they really care about ensuring that the impacts of what we're building are positive right and that what we are doing is promoting something that's truly safe and I think the last thing I'll say there is that there's sort of a maturity to um to the company itself I think some of that comes from the fact that the founding team you know we're all in our 30s and 40s um many of us have families and so I think there's a way that there's just it's not the first company we've ever worked at before and we've learned a lot of the lessons about um what to do and not do at a quickly scaling company right at the beginning of the year I wrote a document basically for I originally actually just wrote it for myself as like shower thoughts and then I shared it with the leadership team and then they said actually you know you should share this with the whole company and it basically just said here's what we should expect is going to happen in 2023 from a growth perspective for the organization right I said you know things are going to start to get weird around 100 people 110 people right it's going to get harder to communicate we have to communicate more it's we're gonna these the following things are gonna feel thrashy right there's just lessons that I've learned from doing this twice before and going through that sort of phase of hyper growth that we don't have we don't have to reinvent the wheel right there's things about us that are unique but there's also things about us that really translate to any quickly growing startup and rather than making each mistake again my hope is that we can kind of LeapFrog some of those two so I think Daniela you said almost everything I would have said I think higher great people I think people aligned with the mission of the company um on the Kumo side as well we indexed on people oh you know with experience so people who had been had built things in their you know previous companies and you know and we had people new college graduates but we also had you know experienced people and we also wanted to create a culture that was in inclusive for that uh stage of life that they were in right so I you know we have a lot of babies I have two young kids so you know we created a culture where you can work hard but you actually even you know you're able to support that you know the surrounding system that comes with it right and so we created a culture we actually wrote Our uh you know value statement we love graphs so everything has to do with graphs so we just said we're like nodes in a network better together and that's our you know our tagline and we remind ourselves of that you know continuously right in decision making Frameworks and everything that you know what's the best decision for the company let's move forward let's take the shortest path first we're a team right and that's uh that's the Como culture you are two amazing nodes that I'm grateful to have in my graph uh just brilliant leaders wonderful culture leaders thank you so much for the time heyma and Daniella and last question are Kumo and anthropic open to Partnerships and customers yes okay there we go no surprise there thank you both please join me in thanking these amazing panelists [Music] [Applause] [Music]

Channel: Newcomer

Views: 1,922

Keywords:

Id: MRNKzKgjqx0

Length: 32min 16sec (1936 seconds)

Published: Fri Apr 07 2023