This conversation is part of our AI Revolution series, which features some of the most impactful builders in the field of AI discussing and debating where we are, where we’re going, and the big open questions in AI. Find more content from our AI Revolution series on www.a16z.com/AIRevolution.
CEO of Quora Adam D’Angelo discusses how building infrastructure for creators can democratize AI with a16z general partner David George.
- [01:07] Social networks as compliment to AI
- [03:59] Poe: bringing AI to the masses
- [05:51] The future of AI is multi-model and multimodal
- [08:11] Is the model the product?
- [11:31] Building AI infrastructure for creators
- [13:41] Navigating platform shifts
- [16:02] Sharing human- and computer-generated knowledge
- [17:43] Knowledge sharing on the internet
- [20:41] The benefits of scale for AI
- [21:59] Competing on scale or feature differentiation
- [25:01] Fault tolerance as a wedge for startups
Social networks as compliment to AI
Adam D’Angelo: I was actually very excited about AI early on in my career. I remember trying to build some AI products in college, and it was just very difficult. The technology just 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 meanwhile, I just watched social networking start to boom. You can actually look at a lot of social networking technology as if it’s almost an alternative to AI. So, instead of trying to get the computer to do everything, you could just connect people with other people over the internet who could do those things.
In the same way that globalization can be a substitute for automation, social networking is, you know, like letting people access everyone else in the world for entertainment, for fun, for communication, for whatever you want to do. I think it was just this incredibly powerful technology, and given that AI wasn’t quite there yet, [social media] was the main thing that there was to do, to apply all the technology to.
So I first got interested in social networking, and then, through my experience at Quora, we started out with a product that was entirely human-driven. People would come and ask questions, and they would put topics on them, and other people would sign up to answer questions, and they would tell us about what they knew about by tagging themselves with these topics, and we would try to route the questions to the people who knew about the particular topic. And it was all manual.
But we knew that at some point we were going to get to the point where software would be able to generate answers. We ran some experiments using GPT-3 to generate answers and compare them to the answers that humans had written on Quora. And a lot of the time, GPT-3 could not write as good of an answer as what the best human answer was that had been written, but it could write an answer instantly to any question. The constraint for Quora had always been the amount of time that high-quality answer writers had to answer questions. And so, the thing that was really new about LLMs was the ability to, at extremely low cost, generate an answer instantly to any question. We realized that a chat kind of experience, where you can write a question and then get an answer instantly from AI, was more likely to be the best paradigm for interacting with AI, as opposed to this publication paradigm that Quora had.
David George: Sure. Yeah, of course.
Adam: And so, based on all that, we landed on building Poe as a new chat-oriented AI product.
Poe: bringing AI to the masses
David: I think many people will be familiar with Poe, but explain, just for us, how does the product work? How do you find it in the first place? How do you interact with it?
Adam: In the same way that Quora aggregates knowledge from many different people who have knowledge and want to share their knowledge, we want Poe to be a way for people to access AI from many different companies and many different people who are building on top of AI. You can come to Poe 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. We’ve got an open API, where anyone can hook in. So, anyone who’s training their own model—any of these research teams, anyone who’s doing fine-tuning—they can take their model and put it on Poe. What we allow is for them to reach a big audience quickly.
We thought about, [as Quora] , what is the role that we’re going to play in this new world with AI? What are the strengths that we had, and what have we learned over the past 10 years building and operating Quora? There’s actually a lot of this consumer internet know-how that’s important in getting a product to mass market. This [includes] things like building applications across iOS and Android and Windows and Mac, localization of the interface, A/B testing, subscriptions, all these other kinds of small optimizations that you need to make a good consumer product. We want Poe to be a way for anyone who’s creating AI, whether it’s one of the big labs or an independent researcher, to get that model and make it available to mainstream users all around the world.
The future of AI is multi-model and multimodal
David: There’s a lot that you just said that I would love to go deeper on. You listed off all the models that you make available. There’s one theory that one model [or] one company is going to provide everybody the solution that they need for everything, right? There’s another theory that there’s going to be tons of different models for different use cases. The world’s going to be multi-model and multimodal. The theory behind Poe is that the future is going to be multi-model and multimodal. Why do you think that’s the case?
Adam: I think nobody knows how the future is going to unfold, but we think that there’s going to be a lot of diversity in the kind of products that people build on top of these models, 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 you are going to train on it, what kind of fine-tuning you are going to do. What kind of instructions is the model going to expect 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 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 AI.
So, early on in the internet, the web browser came along and made it so that anyone who was making an internet product, they didn’t need to build a special client to get distribution to people around the world. They could just build a website, and this one web browser could visit any website. And in the same way, we want Poe to be a single interface so that people can use that to talk to any model. We’re betting on diversity, just because there are so many talented people around the world who are going to be capable of tuning these models. You can tune the open-source models today. There’s also products from OpenAI and Anthropic, and I think Google’s close to having something where you’ll be able to fine-tune all these models. Everyone has their own data sets. Everyone has their own special technology that they can add to the models. I think, through the combination of all of this, we’re just going to see a very wide diversity of things you can do with AI.
Is the model the product?
David: There’s two things that I’d like to maybe go deeper on there. One is the idea of, like, what constitutes the product itself? What is it today, and then what is it going to have to become? Then second, the idea of the long tail, right? Like, bet on the long tail, incentivize them, give them a platform, abstract away a bunch of the infrastructure that they don’t know how to build, and harness really what they’re great at, right?
So, on the first idea, what is the product like? Today, many people would probably say [the AI model] is largely the product. What are the advances that you anticipate seeing that are going to sort of change the way people interact with these, [and] enable new kinds of products being built? You know, one way of thinking about that is: are the model providers themselves going to be the ones that build all the products?
Adam: If you’re a large model creator and you have tens of employees that you can allocate to building a consumer product, and you have the culture to do that, then you can go direct to consumer, and you can build a good product. I think that most of the people who are training these models are not in that position. If you want to take your model and bring it to consumers all around the world, you’ve got to think about [how] you need an iOS app, you need an Android app, you need desktop apps, you need a web interface. You need to do billing in all these different countries. You’ve got to think about taxes. There’s just a lot of work. You could either spend, you know, you raise some venture funding, you could either spend some of that funding on hiring out a whole team and developing all those competencies, or you can spend that on making your model even better. I think different startups will choose different paths here. But I think, for a lot of them, the right path is going to be to just set up an API, or plug into the Poe API, and use that to get to a lot of consumers very, very quickly.
David: Yeah. Talk about the role that the sort of long tail of creators then plays. How do you want to engage with them and what’s the incentive for them to want to build on top of Poe, as opposed to other places?
Adam: Yeah. So, we have a revenue-sharing program that allows people to get paid as a result of people using their bots on Poe. It costs a huge amount of money to provide inference for these models. And so almost no other platforms provide this kind of revenue share today. So if you have a model that requires a lot of GPUs to do inference on, then this is really your best place to come, and you can have a real business, you can cover your inference costs, and make more. We think a ton of innovation is going to come from these companies.
There are other companies that are building things on top of some of the big models, so, say, from OpenAI. And in that case, they have to pay the OpenAI inference cost, which is another sort of source of need for money. So the Poe revenue share model works in the same way, where it’ll let you afford your cost that you’re then paying on to any other inference provider.
Building AI infrastructure for creators
David: Yeah, absolutely. What are some of the really fun and interesting things that creators have already built on top of Poe?
Adam: A lot of people right now are excited about image models. There’s Stable Diffusion, SDXL, and then we let users go and do some prompting to customize it to provide art of a particular style. There are these, like, anime-style SDXL bots on Poe. Those are popular. There’s this company called Playground. They’re making a product for people to edit images. But in the process, they’ve created a pretty powerful model, and they have that model available on Poe. That’s gotten pretty popular recently.
David: It’s so cool that you can have a long tail of these creators make their own sort of opinionated style of these base models. But I think there’s something to that, where you provide, you know, the sort of infrastructure and support, and then let the users or creators do what they do best.
Adam: Yeah. I think it’s super early days right now, but I think what we’re going to see over just the next year or two is going to be incredible. This will go from being sort of useful to some people right now to being something that’s just critical to many different tasks that anyone is going to try to accomplish.
David: There’s a really good analogous company that you and I both know very well, which is Roblox, right? Early days, you know, creators were on there building games. They were pretty basic [in] the early days, and it was a lot of kids learning how to build games, and then it graduated eventually to people who were able to earn a living. So, I think the ideal for you would be you build enough scale, they can build large enough audiences to actually be professionals at what they’re doing.
Adam: Yeah. And we’re spending millions of dollars already on inference. It’s mostly going to the large model providers right now, but we want to let as much of that as possible go off to these independent creators.
Navigating platform shifts
David: Cool. I want to shift topics, and get maybe a little bit more, like, conceptual AI. You were CTO at Facebook at the time where social was emerging, and then right when the platform shift to mobile was taking place, right? I’d love your thoughts on what are the similarities to the shift to mobile in this AI wave, and what are some of the big differences?
Adam: You know, I think it’s very hard to say. With Quora, I think we were a little bit slow to adopt to mobile. You know, mobile was one of the things on our list of many priorities, and it needed to be the number one priority, and we needed to make tougher trade-offs to prioritize it. You know, we needed to do things like hire a set of different people who were going to focus on it, and have a period where we’ve released no new features, and we were just simplifying things because the mobile UI called for a different experience. When you have such a critical change in the platform structure, you need to rethink so much that it’s only going to happen if you have this very strong kind of top-down leadership.
David: And so you’ve done it differently this time around?
Adam: Yeah, yeah, yeah.
David: Right. Talk about some of the organizational changes and, you know, what you’ve done to actually refocus yourselves on the big thing that’s right in front of us here.
Adam: I think the first thing is just identifying this trend and then starting off doing some experimentation early on, just to learn. That didn’t require any kind of strong, decisive leadership as much as it just required paying attention to the market. But then from that experimentation, that got us enough conviction that too much of the Quora product has been built up around this publication model that is fundamentally premised on the idea that expert time is going to be scarce. The AI, the LLM time is not scarce in the same way. And so, we need to rethink that. This was in, I think, August of 2022, we got to this conclusion that chat is the right paradigm for this, and we need a new product. Just trying to retrofit everything into Quora… we thought we were going to move too slowly. So we had a small team just start working on Poe based on that.
Sharing human- and computer-generated knowledge
David: Talk about the relationship between Quora and Poe, and how you actually envision that changing in the future. And then maybe there’s even an extrapolation of Quora and Poe, and human experts and AI experts answering questions. Do they do it in the same place? Is it a different way of interacting?
Adam: We’d love to have all of this as integrated as possible. You know, I think if you think about the relationship between Facebook and Facebook Messenger, these are two products built by the same company, but they share a lot. I think that Poe and Quora might evolve into a similar kind of relationship. We’d love to get more of the human aspects of Quora into Poe. We’d also love to get the whole Quora data set into the Poe bots. And we’re also working… We’ve launched some of this already, to get some of the Poe AI to generate answers that are available on Quora. As these models continue to scale up, the quality is going to go higher and higher, to the point where it actually will be as good as human quality in a lot of cases. And so, the Quora paradigm actually becomes more appropriate for AI as the cost of inference gets higher.
David: Gets lower. Yeah. And model quality gets better. Yeah.
Adam: Yeah, yeah. So, we’ll see what the exact right relationship is, but we think of this as we’re building a network for people and AI to all share knowledge together. Sometimes the people will be getting knowledge from the AI, and sometimes the AI will need to get knowledge from humans, and we’d love to be as much of a conduit for that as possible.
Knowledge sharing on the internet
David: Yeah. And Quora—or Poe, depending on how they interact—is a place you get answers, and sometimes your answer’s going to come from an expert and sometimes it’s going to come from AI, right?
Adam: Yeah, yeah.
David: What do you think about just the internet? [Can] you extrapolate that out? Are people going to be engaging with this collection of bots that have different personalities and different expertise? And will those be interspersed with real humans? You know, will real humans be interspersing in the AI? What do you think actually happens?
Adam: I, personally, I think that humans are always going to play some role. There’s knowledge that people have in their heads that is not on the internet and is not in any book. And so, no LLM is going to have that knowledge.
David: Yeah, Andrej Karpathy called the LLMs a lossy compression algorithm of the internet.
Adam: Mm-hmm. Yeah, yeah.
David: And just as on the internet, there’s experts that know a lot of stuff that’s not on it, right?
Adam: Right, right. I think there’s a lot of potential in the kind of interplay between humans and the LLMs going forward. I think a lot of people, they… LLMs have a problem with hallucinations right now. And I think hallucinations are going to… the rate is going to go down as the models get better, but it’s never going to get to the point where it’s 100% perfect.
I think there will be a lot of value placed on the idea that you know the source of your information, you know which human said it, or which publication originally printed it. I expect that that is going to lead to some kind of product or some kind of user experience where the LLM is helping you sort through your sources, and quoting exact experts or exact sources, as opposed to just synthesizing it all and giving you something where you can’t exactly trust where it came from.
David: And is that a new technology that gets built outside of the models themselves, or do you think that that’s incorporated inside of the model?
Adam: I could see it going 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 quotes. And so it’ll have to be some augmentation of the model, but how tightly integrated into the model it’ll be, I think we don’t know yet.
David: Yeah, I agree. I think that’s going to be critical. It’s one thing [where] we’ve started out with these use cases of companionship and creativity, and hallucinations are a feature of that, right? That makes it more fun and exciting, especially when you get into business use cases, or more utility-type stuff. You know, it’s obviously needed. What are the other big advances that you’re excited about, just broadly in the AI space for language models?
The benefits of scale for AI
Adam: I’m personally the most excited just about scale. Just continuing on the current paradigm, if you just play this forward, there’s so much further that it can go.
David: And you think the scaling laws will hold, are holding?
Adam: So far they have held. My prediction would be there are some issues that need to be overcome, but there’s just this incredible industry, so many talented people right now, who are trying to make this technology advance. And there’s so much money behind it. The force that’s there to help overcome any road bumps that we hit is so massive. So I expect that it’s just going to continue. I think there will be road bumps, there’ll be issues that need to be worked around, and there’ll be breakthroughs that probably need incredible creativity, but we have many of the smartest people in the world, the most determined people in the world, the most talented people in the world, all focused on this problem. I think we’re going to just continue to see the kind of exponential growth progress that we’ve had so far. I think that will go on for many years.
David: We talked about the last shift, right? The mobile shift that you lived through, and some of the lessons that you had from it. What do you think the ultimate market structure looks like in the genAI space?
Competing on scale or feature differentiation
Adam: In order to train these frontier models, you need billions of dollars of capital, and you need many years of investment in infrastructure. There’s a very small set of people who can do that. So that’s leading to this world where there’s only a small number of players that can be on the frontier. Right now it’s OpenAI, Google, maybe Anthropic. Maybe Meta can be there. Those who can get there, I think it’s going to be good business. You’ll be able to make a lot of money. You can have good profit margins. You’ll have to work very hard to stay on the frontier, to keep up. But it’s not a commodity. I think when you go six months behind the frontier, or definitely one year, it’s brutal.
There’s just way too many people that are able to get the capital and the resources to train those models. It’s going to be either fully open source, or there will be too many different competitors for anyone to make a good business at that point, on pure technology. I do think there’ll be very good businesses at that level that are not using frontier models, but are combining some other kind of unique thing with the model. So, it might be that you’re providing some tool that the model can use, or that you have some unique data that you’re using for fine-tuning, or there might be some unique product you build around the model. Then that ends up being the source of competitive strength.
I think there’s going to be this choice, where you’re either competing on scale, by being on the frontier, or you’re competing on some kind of feature differentiation. And in that case, you don’t need a frontier model. In some cases, you’ll have both. So, you know, you might be able to use the OpenAI API and combine that with some unique tool that you’re providing, and that could be a good business as well.
David: Yeah. Once you get beyond the foundation models, you get to more traditional forms of business differentiation, competitive differentiation, competitive advantage, you know, sources of moats and things like that, which I think totally makes sense.
Adam: Yeah. And I think what’s interesting about this is that it’s evolving. So, things are moving so quickly, and so every six months, the frontier moves forward. The frontier players, they have to invest more capital, but then they have much more powerful models that open up even bigger markets. But then, you know, the open source, one-year-back frontier, that’s also moving forward.
David: That keeps advancing. Of course.
Adam: Yeah. And so the markets that that can address are getting bigger and bigger. I think every year that goes by, we’re going to have this much larger market that can be addressed by the technology and all the products that are built on top of it.
Fault tolerance as a wedge for startups
David: So, yeah, that brings me to another topic, which is related to market structure. You know, incumbents versus startups. And in the seat that we’re in, we hope the startups always win. But in the last cycle, and maybe just from a B2B lens here, in SaaS and cloud, there were a bunch of things that made it really difficult for the incumbents to actually innovate. There was a business model innovation, and new talent and technology required, which opened the door pretty widely to startups.
There’s a take out there now on AI, which is, this time it’s different, and the incumbents are the real winners, right? Because the technology is available by simple API, you can plug it right in. And they have distribution, so they should be the winners. And, you know, if you just sum up Microsoft and Google’s business apps and all these things, it’s probably somewhere between $10B and $20B of revenue over the next one to two years. I’m curious if you have a take on that, if that’s consistent with how you see it, or if you see it differently.
Adam: Yeah, I think it’s going to vary. Definitely the incumbents, they’re going to have access to the technology and they’re going to have distribution. That’s a big advantage that they have. I think the opportunities for new players in this wave are more in the cases where the kind of product you want to build around this technology is somehow fundamentally different than what was built before. And so, as an example, the hallucination problem, that’s in some ways a good thing for startups, because a lot of the existing products out there have zero tolerance for anything that’s going to have a risk of producing something wrong. You can see this with, I think, with Perplexity getting share from Google right now. Google can’t just go and put something on all their search results where it has a few percent chance of being wrong. That would be a huge problem for them. Perplexity, that can just be the expectation when you’re using that product, that it’s almost always right, even though there’s a small chance that it’s wrong. I think that same thing is actually going to play out in a lot of other cases, where the products you build around this, they need some kind of fault tolerance, and there needs to be a user expectation that everything is not perfect.
David: And the cost advantage can be so great for this, right? Like, if you take a highly-paid person, like a lawyer, and you run it through an LLM, it costs cents, versus $1,000 an hour. Maybe you just 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. That’s the new way of engaging, right?
Adam: Yeah. Yeah. So, you know, you have these entrenched companies that maybe have a very strong brand of never making a mistake, or never messing up, or always being reliable. And a startup can just come in and say, like, okay, well, this is going to cost a 10th or a 100th the price, but it’s going to have a small chance of getting things wrong. Actually, a lot of people would prefer that. But it’s a real problem for the incumbent, because they can’t compromise their brand.
David: Yeah. That’s a great point. I guess, just to close it out, I’m sure a big part of the audience here is founders who are building, and probably earlier stage than you. What advice do you have for people building in AI?
Adam: I think what I would do if I was starting a new company right now is just spend a ton of time playing with the models, and playing with integrating them with different things. You know, there’s so many different inputs you can give to the models. You can make scrapers that ingest data from anywhere. You can get data from the user’s local screen. You can get data from voice. And there’s just such a huge space of needs people have, and such a huge space of different, like, inputs you can combine to try to address those needs. I think it’s very hard to just kinda, like, think top-down about where there’s demand in the market. I think experimentation is really the way to go to generate ideas and to set up a startup that’s going to be able to build something really valuable.
David: Yeah, and have a place in the world, for sure. Awesome. Well, thanks for being here, Adam. I appreciate it.
Adam: Yeah. Thank you.
“Andreessen Horowitz is a private American venture capital firm, founded in 2009 by Marc Andreessen and Ben Horowitz. The company is headquartered in Menlo Park, California. As of April 2023, Andreessen Horowitz ranks first on the list of venture capital firms by AUM.”
Please visit the firm link to site
