Bringing AI to the Masses with Adam D'Angelo, CEO of Quora

Hi guys, I'm Sarah Wang, General Partner on the a16z Growth team. Generative AI has initiated a paradigm shift that is already transforming our world. In our AI Revolution series, we talk to the people who are really building the technology to understand where we are. let's go and the big open questions in the field Our guest on this episode is Adam D'Angelo Adam has created and grown companies that have connected billions of people around the world. He was the CTO of Facebook until 2008 before founding Kora in 2009 later. Adam has been interested in AI for decades and joined the Open AI board in 2018 and is now using his experience scaling some of the world's largest consumer companies to build a platform that brings AI to the

masses

in this conversation Adam talks to a16z General my partner and my colleague David George about building AI infrastructure for creators the multi-model, multi-modal future of AI how AI will shape knowledge sharing on the Internet and much more.

There's a lot to dig into here, so let's start at the beginning of Adam's journey in AI. to take us back to the university of him Years in 2005 I was actually very excited about AI at the beginning of my career. I remember trying to build some AI products in college and it was very difficult, the technology just wasn't it. It wasn't there, it wasn't at the point where you were going to be able to make something that was ready for consumers and in the meantime, I just saw how social media started to grow and you can actually see a lot of social media technology is almost an alternative to AI, so instead of trying to get the computer to do everything, you could just connect people to other people over the Internet who could do those things in the same way that, like globalization, It can be a substitute for Social Media Automation, you know?

You can think of it as allowing people access to everyone else in the world for entertainment, fun, communication, whatever you want to do. I think it was an incredibly powerful technology and since AI hadn't gotten to that point yet, that was pretty much like that, that was the main thing you had to do to apply all the technology, so I first got interested in networking. social and then through my experience in Kora. We started with a completely human-powered product so that people would come and ask questions and give them topics and other people would sign up to answer questions and tell us what they knew by tagging these topics and we would try to direct the questions to people who knew the topic in particular and everything was manual, but we knew that at some point we would get to the point where the software would be able to generate answers We did some experiments using GPT three to generate answers and compare them with the answers that humans had written in Kora and many times gbd3 couldn't write an answer as good as what the best human answer was. that had been written, but I could write an instant answer to any question and the limitation for cor had always been the amount of time that high quality answer writers had to answer the questions, so what was really new about the movies was the ability to, at an extremely low cost, generate an instant answer to any question through that experience.

We found that a chat-type experience where you can type a question and then get an instant response from the AI ​​was most likely to be the best paradigm for interacting. with AI instead of this kind of similar post Paradigm that that cor had, yes, of course, and so based on everything we did, we came to build po as a new chat-oriented AI product. I think many people will be familiar with po, but I'll explain. just for us, how does the product work? How do you find it in the first place? How do you interact with it?

In the same way that Kora aggregates knowledge from many different people who have knowledge and want to share their knowledge. I want po to be a way for people to access AI from many different companies and many different people who are building on AI, so you can come to PO and use it to talk to a very wide variety of models that are available today and then we have all these other products that people have built on top of these models and we have an open API where anyone can connect, anyone who is training their own model, as well as any of these research teams, anyone who is doing fine tuning. , they can take their model and put it in PO and what we allow them is to reach a large audience quickly, so, so that they know, we think about Kora as a company, what is the role that we are going to play. playing in this new world with AI and what are the strengths that we had and what have we learned over the last 10 years building an operational Kora and there's actually a lot of this kind of consumer Internet knowledge that is important to get a product for mass market, so it's things like building apps on iOS, Android, Windows and Mac, interface localization, subscribing to AB testing, all these other types of little optimizations that you need to create a good product. consumption.

We want it to be a For anyone who's building AI, whether it's one of the big labs or an independent researcher, we want it to be a way to get that model and make it available to mainstream users around the world. There's a lot of things you just said that I'd love to go into more detail on one of the things you said, which is that you sort of listed all the models that you make available. There is a theory, which is that you know that one model, one company, is going to provide for everyone. You already know the solution they need for everything, there is another theory that says there will be tons of different models for different use cases.

The world will be multi-model and multi-modal. The theory behind PO is that the future will be multimodal and multimodal, why do you think that is the case? I think no one knows how the future will develop, but we think there will be a lot of diversity in the type of products that people build on these models and and and in the models themselves I think there are a lot of trade-offs involved in making one of these models: you have to decide what data are you going to train on it, what kind of tuning are you going to do, what kind of What are the instructions that the model expects you to give as a user?

What kind of expectations do you want to set with your users about what to use the model for? And I think in the same way that the early Internet had this huge explosion of different applications. I think we're going to see the same thing from PR AI, so at the beginning of the Internet the web browser came along and made it so that anyone who was creating an Internet product didn't need it. to create a special client and distribute it to people all over the world, they could simply create a website and this web browser could visit any website safely and in the same way we want it to be a single interface that can be used so that people can use that to talk to any model.

We are going for diversity only because there are so many talented people around the world who will be able to adjust these models. You can tune open source models today. There are also openi and anthropic products and I think Google is close to having something where you'll be able to tune all of these models and each one has their own data sets, each one has their own special technology that they can add to the models and I think that Through From the combination of all this, we will see a very wide diversity of things you can do with AI.

There are two things I would perhaps like to delve into further. One is that you know the idea of, for example, what constitutes the product itself, what it is today, and what it will have to become. And then secondly, you know the idea of ​​the long tail, like going for the long tail, incentivizing them, giving them an abstract platform. a lot of infrastructure that they don't know how to build and really leverage what they're great at, so first of all, what is the product, is the model, today is the AI ​​model that a lot of people probably say it is. largely the product, what are the advancements that you anticipate seeing that are going to change the way that people interact with them, allow new types of products to be created, you know, one way to think about it is whether the suppliers themselves Of models are going to be the ones who build all the products, if you are a great model builder and you know that you have dozens of employees that you can assign to the creation of a consumer product and you have the culture to do it, then you can do it.

I know you can go directly to the Consumer and you can create a good product. I think most people who train these models are not in that position. If you want to take this, take your model and bring it to Consumers around the world. the world you have to think about you need an IOS app you need an Android app you need desktop apps you need a web interface you need you need to invoice in all these different countries you have to think about taxes and um there's just a lot of work and you could spend it , you know, you raised some venture funding, you could spend some of that funding on hiring a full team and building all those competencies or you could spend it on further improving your model.

Yes, and I think different startups will choose different paths here, but I think for many of them the right path will be to simply set up an API or connect to the PO API and use it to reach a lot of consumers, very, very quickly, yes, they talk about the role that that kind of long tail of creators plays and like how they want to interact with them and what the incentive is for them to want to build on Po instead of others. Places, yeah, so we have a revenue sharing program that allows people to get paid as a result of people using their bots on PO and you know it costs a lot of money to provide inferences for these models and So, almost no other platform offers this kind of revenue share today, so if you have a model that requires a lot of GPUs to do inference, this is really your best place to come and have a real business.

Can you know that you can cover your inference costs and earn more? We believe that a lot of innovation will come from these companies. There are other companies that are building things on top of some of the big models, for example, from Open AAI and in that case, they have to pay an open inference cost, which is another kind of source of need for money, so in the PO revenue share model works the same way, where it will allow you to pay your cost which you are then paying to any other inference provider, yes absolutely, what are some of the really fun and interesting things that the creators already have built on Po?

A lot of people right now are excited about the imaging models we have. uh there is a stable SC cxl spread and then we let the users go and do some prompts to customize it to provide art of a particular style, so there are anime style sdxl bots on PO, those are popular, there is a company called En the playground, they are creating a product for people to edit images, but in the process they have created a pretty powerful model and they have that model available on PO and that is, it has become quite popular recently, yes, it is so cool that You can have a long line of these creators who create their own opinionated style of these basic models, but I think there's something to that when you provide the kind of infrastructure and support, and then you let the users or the creators, you know, do what What they know best, yes, and that's all you know.

I think it's very early right now, but I think what we'll see over the next year or two is going to be incredible. going from being something useful for some people right now to being something that is simply critical for many different tasks that anyone will try to accomplish. Yes, there is a very good analogue company that you and I know very well. um, which is Roblox in its early days, um, you know, the creators were there making games and they were pretty basic in the early days and they were a lot of kids, uh, you know, they learned how to make games and then eventually they graduated. . people who were able to make a living, so I think the ideal thing for you would be to create enough scale, um, that they can create large enough audiences to be sort of professionals of what they're doing, yeah, and we're, we're.

We're already spending millions of dollars on inferences, right now it mostly goes to the big model providers, but we want to let as much of that go to these independent creators as possible. Great, I want to switch gears and get maybe a little bit more like a conceptual AI, your CTO at Facebook at the time when social media was emerging and then right when the platform shift to mobile was happening, like that I would love to know your opinion on what the similarities are between the change of mobile devices and this AI. wave and what are some of the big differences, yeah, you know, I think it's very hard to say, I thinkWith Kora we were a little slow to adopt mobile, you know, mobile was one of the things on our list of many priorities. and it had to be the number one priority and we had to make more difficult trade-offs to prioritize it, you know?

We needed to do things like hire a set of different people who would focus on it and really take some, you know, have a period in which that we didn't release new features and we were just simplifying things because the mobile UI called. To have a different experience, when you have such a critical change in the platform structure, you need to rethink so many things that will only happen if you have this kind of leadership very strong from top to bottom, so this time you've done it differently. Yeah, yeah, sure, so yeah, let's talk about some of the organizational changes and you know what you've done to refocus on the important thing that's right in front of us here.

Yeah, I mean, I think the first thing is you know how to just identify this. trend and then start doing some experimentation from the beginning, um just to learn, and that didn't require any kind of strong, decisive leadership, but it just required paying attention to the market and then, but then, from that experimentation that led us . The enough conviction that, in our case, we said, "Hey, a lot of the cor product has been built around this publishing model" is fundamentally based on the idea that expert time will be scarce and AI, movie time, it is not. scarce in the same way so we need to rethink this was in uh I think in August 2022 we came to the conclusion that chat is the right paradigm for this and we need a new product that we don't want to try just try.

To modernize everything in Kora, we thought we were going to move too slowly, so we had a small team that started working on PO based on that talk about maybe the relationship between Kora and Po and how. Do you really imagine that will change in the future and then maybe there will even be an extrapolation of "ok cor and Po" and "human experts" and you know AI experts answer questions? Do they do it in the same place? Is it a different way of interacting? Yes Yes. I know we would love to have all of this as integrated as possible.

I think if you think about the relationship between Facebook and Facebook Messenger, these are two products created by the same company, but they share a lot. I think PO. and Kaora could evolve into a similar type of relationship. We would love to have more human aspects of Kora in po. We would also love to have the entire core data set in the PO Bots and we are working on that too. We've already released some of this to get some PO AI to generate responses that are available in Kora as these models continue to increase, the quality will increase more and more to the point where it actually will. be as good as human quality in many cases, so the core of Paradigm actually I think becomes more appropriate for AI, because as the cost of inference increases, it decreases, yes, and the quality of the model improves, yes, yes.

Yes, we will see what exactly the right relationship is, but we think about this as we are building a network for people and AI to share knowledge together and sometimes people will get knowledge from AI and sometimes AI will need to get knowledge of humans and we would love to be a conduit for that as much as possible, yes, and cor o po, depending on how they interact, is a place where you get answers and sometimes your answer goes. it comes from an expert and sometimes it will come from AI, yes, sure, yes, yes, what do you think about the Internet?

If you extrapolate that, people will engage with this collection of robots that have different personalities and different knowledge. And those will be interspersed with real humans? You know, they'll like it if real humans are interspersed with the AI. What do you think is really happening? I personally believe that humans are always going to play some role. There is knowledge that people have in their heads. it's not on the internet and it's not in any book so no movie will have that knowledge so yeah Andre Kathy called movies an internet lossy compression algorithm and it's like it's just from the internet like there's experts.

That knows a lot of things that are not so correct, so you know. I think there's a lot of potential in the kind of interaction between humans and movies in the future. I think a lot of people, LM, have this problem. with hallucinations at the moment and I think the rate of hallucinations will decrease as the models improve, but it will never get to the point where it is 100% perfect, so I think there will be a lot of value. It's based on the idea that you know the source of your information, you know which human said it or which publication originally printed it, and I hope that leads to some kind of product or some kind of user experience. the movie helps you sort through your sources and cite exact experts or exact sources instead of just synthesizing everything and giving you something that you can't trust exactly where it's coming from, yeah, and it's a new technology that's being built outside. of the models themselves or do you think that is built into the model.

I could see it either way. I mean, if you just look at a model, the raw model doesn't have access to these other databases where it can get exact, so there will have to be some model augmentation, but how tightly integrated will it be into the model. I think we don't know yet. Yes I agree. I think that will be critical. I know that's one thing that we've started with these use cases of camaraderie and creativity and that hallucinations are a feature of that, that's how it makes it more fun and exciting, especially when you get into commercial use. cases is or more utility type things, you know, it's obviously necessary, what are the other big advances that you're excited about generally in AI in the AI ​​space for language models?

Personally, the one I'm most excited about is the scale. Continuing with the current paradigm, if you simply play forward, there is much further you can go and you believe that the scaling laws will hold until now. They have remained. My prediction would be that there are some issues that need to be overcome, but. There is just this incredible industry, so many talented people right now that are trying to move this technology forward and there is so much money behind it, the force that is there to help overcome any obstacles that we encounter along the way is so enormous that I hope let it just go Continue I think there will be bumps in the road and there will be problems that need to be solved and there will be breakthroughs that probably need creativity, but we have a lot of the smartest people in the world, the most determined people.

In the world, the most talented people in the world focused on this problem and I think we're going to continue to see the kind of exponential growth progress that we've had so far. I think we'll continue for many years, we talk about the latest change, like the mobile change that you experienced and some of the lessons you learned from it. What do you think the final market structure looks like in the Gen space for training these? Frontier models need billions of dollars of capital and many years of investment in infrastructure because there is a very small group of people who can do it and that is leading to this world where there are only a small number of players who can be on the frontier, so, you know, right now it's open AI, Google, maybe anthropic, maybe meta, may be there, those who can get there, I think it will be a good business, you will be able to make a lot of money, um , you can have. good profit margins, um, you're going to have to work really hard to stay on the frontier and keep up, but it's not a commodity, mhm, I think when you're six months behind the frontier or even definitely a year, it's brutal, There are just too many people who can get the capital and resources to train those models, so it will either be completely open source or there will be too many different competitors so that anyone can do a good business on that at that point in pure technology, I think there will be very good deals at that level, um, that you know, they don't use Frontier models but they are combining some other kind of unique thing with the model, so it could be that you are providing some tool that the model can use or you have some data unique ones that you are using to make adjustments or there may be some other unique product that you build around the model and then that ends up being the source of competitive strength, so I think there will be this type of choice where you will be competing at scale being on the frontier or competing on some kind of differentiation of similar features and in that case you don't need a Border Model and something in some cases, you will have both, so you know you could use the Open ey API and combine it with some unique tool that you are providing and that could also be a good business.

Yeah, you kind of get it once you get through it. The basic models that you come to are more traditional forms of business differentiation, competitive differentiation as competitive advantage, you know, mode sources and things like that, which I think makes a lot of sense, yeah, and I think the interesting thing about this is that It is evolving. So things are moving very fast and every 6 months the frontier advances and the frontier players know they have to invest more capital, but then they have much more powerful models opening up. Even bigger markets, um, but then, you know, the open source from a year ago, Frontier, is also moving forward, so the markets that can be addressed are getting bigger and bigger, so you know, I think every year.

Over time, we are going to have a much larger market that can be addressed by technology and all the products that are built on it, so yes, that brings me to another topic that is related to Market Structure, which is, you know, established companies versus startups, and we, in the seat that we're in, we expect startups to always win, but in the last cycle and maybe just from a B2B lens here, like in the last cycle, um . You know, SAS and the cloud, there were a lot of things that made it really difficult for incumbents to innovate, there was a business model.

It required innovation and sort of new talent and technology, which opened the doors pretty widely for startups. Now there is artificial intelligence, which this time is different and the incumbents are the real winners, because the technology is available through a simple API, you can plug it directly and they have distribution, so they should be the winners, and you know. If you just sum up Microsoft and Google's business applications and all these things, it's probably $10 to $20 billion in revenue over the next year or two. I'm curious if you have an opinion on that, if that's consistent. with how you look at it or if you look at it differently, yeah, I think it's going to vary, so definitely the incumbents that have it will have access to the technology and they will have distribution, and that's it. a great advantage they have.

I think the opportunities for new players in this wave are more in cases where the type of product you want to create around this technology is somehow fundamentally different from what was created before and, for example, the hallucination. problem which in some ways is a good thing for startups because many of the existing products have zero tolerance for anything that might have the risk of producing the wrong thing, so you might look at this with uh, I think with bemusement. get share from Google right now, yeah, Google can't just go and put something in all of their search results where it has, you know, some chance of being wrong, yeah, that would be a big problem for them, the puzzlement can be just an expectation.

When you use that product, you know it's almost always right, even though there's a small chance it's wrong. I think the same thing will happen in many other cases where the products that you build around this need, need some kind of fault tolerance, um, and there has to be a user expectation that not everything is perfect and the cost. The advantage can be as great for this right, as if you hire a well-paid person, such as a lawyer, and you know yourself. running it through a movie that costs pennies instead of dollars per hour, like maybe you should have a really high fault tolerance and you just have to double check a lot of the work and that's just a different workflow which is a different way of engaging, yes, yes, no, and then you know you have these entrenched companies that maybe have a very strong brand of never making a mistake or never being wrong, always being reliable and a startup can just come in and say, " "Okay, this goes." It costs a tenth or a hundredth of the price, but there will be a small chance of doing things wrong and that's actually a lot of people would prefer that, but it's a real problem for the incumbent because they can't compromise their brand.

Yeah, that's a great point, I guess just to close it, I'm sureSince a large portion of the audience here are Founders who are probably building at an earlier stage than you, what advice do you have for people building in AI? I think what I would do if I were starting a new company right now is just spend a ton of time playing with the models and playing with their integration. with different things, you know, there are so many different inputs you can give to models, you can make scrapers that ingest data from anywhere, you can get data from the user's local screen, you can get voice data and there's a huge space. of needs that people have and such a large space of different inputs that you can combine to try to address those needs.

I think it's very difficult to think from the top down about where there is demand in the market. I think experimentation is really the way to go to generate ideas and establish a startup that will be able to build something really valuable, yes, and have a place in the world, safe, amazing, thank you for being here. I appreciate this is fun thanks for sharing the time thanks