Adam D’Angelo left Facebook in 2008 to found the question-and-answer service Quora. Nearly 15 years later, he’s back in startup mode as he reshapes Quora at breakneck speed for an artificial intelligence future.
He believes Quora’s new AI chat product, Poe — “Platform for Open Exploration” — will become a kind of web browser for AI, providing a one-stop shop to try an increasingly diverse set of services from companies like OpenAI and Anthropic.
When D’Angelo was a student at the California Institute of Technology, it was so hard to get artificial intelligence to work that he left it behind, deciding instead to join Facebook, where he became chief technology officer.
A board member of OpenAI, D’Angelo never stopped thinking about the technology’s potential. Now at 38, he’s finally making AI the focus of his life’s work. Below is an edited version of our conversation.
The View From Adam D'Angelo
Q: You became CTO of Facebook pretty much right out of college. Was AI a part of Facebook at that time?
A: Facebook basically didn’t use AI at all when I was there. I left in 2008 and then started Quora in 2009. And the whole first year of Quora, we didn’t really use machine learning.
We started to use AI in 2011 to rank the feed, very similar to what Lars [Backstrom] was doing at Facebook around that time. It’s in the category of recommender systems but you’re not generating the content.
That’s what’s really exciting about this new wave of AI. You can call it all AI but it’s fundamentally a very different thing today from what that was back then.
Q: Were you keeping tabs on the development of this?
A: Machine learning, specifically for recommender systems, became very important to Quora. We knew that someday, AI was going to be able to generate text and that that was going to be an important thing for us.
In 2018, I was asked to join the board of OpenAI. One of the reasons I was excited to join was that it would let me keep tabs on the world of AI. It’s been great to see the organization evolve.
Early last year as we were monitoring this, we said, ‘hey, this is the point where this is now crossing the barrier where this matters, and it’s going to be very important for Quora to adapt and take advantage of the big opportunity.’
Our first instinct was to use it to generate answers on Quora. But we learned the quality still wasn’t there. Quora is often able to find the best expert in the world to answer a question, and AI can write an answer maybe at the level of an average human.
The real advantage of AI was that it could instantly answer almost any question. So you want it for real-time use cases where someone has a new question that no one else has ever had before, and they want to get an immediate answer.
That’s where AI has this relative strength because you don’t want to wait two weeks for a human to write an answer. Because the AI can answer immediately, you can go back and forth and ask it to clarify its answers. We concluded that the right user experience would be a chat interface as opposed to this structured question-answer format that Quora is built around.
There were a lot of fundamental changes we needed to make very quickly and we needed to operate a little more like a startup. We decided this should be a new app so we started building it in August, and that’s what turned into Poe.
Q: You’ve worked in ad-supported models and now you are looking at the subscription model. Do people seem to be willing to pay for this?
A: It’s a little too early to know what percentage of people are going to be willing to pay. But it definitely seems like this business model will work.
It’s difficult to operate an ad-supported business right now. Right now it’s a recession. You need scale and be able to reach a big audience to provide a good experience to advertisers. There are cases where you can become misaligned with your user’s interests.
With this subscription product on Poe, we’re now strongly incentivized to keep making it better, and all of our energy can go into things that are going to benefit the user.
Q: Is Poe financially successful? Are you making money on the subscriptions?
A: We don’t have any numbers that we want to share yet, but it’s going well. These large language models are very expensive to operate and we think we’re going to be able to make the economics work.
Q: Poe gives you this menu of options to try out different models. What were the conversations like with each of the companies?
A: We’re customers of OpenAI and we have been for a while. We’ve been talking to Anthropic for a while as well. I’ve talked to a lot of other AI companies and one thing I noticed was they’re really good at machine learning and AI research. But they often didn’t really have the consumer DNA to build a good internet product.
From using their models, we had this insight. Hey, this is at the point where it’d be good if you could build a nice user interface. ChatGPT came out while we were developing Poe and the whole world learned there’s a lot of potential.
Ultimately, OpenAI is going to be able to build this consumer DNA, but that’s not something every AI research company or hobbyist developer can build.
We want to make it really easy for anyone who is building a chat-based bot to not have to have any consumer DNA. We’ll handle all of those nice interfaces across all platforms and social features.
Q: You’ve been able to roll up all of these apps into one. These companies are for profit, but they seem to act differently than your typical Silicon Valley company with the walled gardens.
A: It’s a function of the economics. In the past, they were generally aggregating free products. So there was a lot of fighting over who would get to show the ads and where they would be.
We’re more like a game console or an operating system, or a web browser. If you are building an internet product, you want it to work on all the web browsers. If you’re building a video game, you want it to work on all consoles.
The economics flow very clearly. Customers pay us and then we pay the AI labs. And if you’re an AI lab, you can get more customers if you integrate with us.
Q: Do you think eventually one might win?
A: I don’t believe there’s going to be one winner. GPUs are getting more efficient every year, there’s more funding available to train these models, there are more open-source models. It’s going in the direction of a greater number of models as opposed to concentrating on one winner.
The models available on Poe right now, they’re all in this assistant framework. But there’s room for a really wide diversity of applications.
You could have a language tutor or a tutor for any subject. Building a good tutor is going to involve more than just the AI research to train the language models. It’s going to involve things like making it have some memory: What you did yesterday and where you are today, and provide some structure and encouragement.
And with many of those applications, you’ll want a chat UI. We hope Poe can be this general interface to many of these different applications, similar to what the web browser did for the internet.
Q: Will Poe merge with Quora or are they on their own tracks?
A: We want to integrate them. I think the right long-term relationship is going to look something like Facebook and Messenger. It’s two apps that are standalone, but there are links between them.
So what makes sense to be shared between Poe and Quora? We may have high-quality answers generated from Poe available on Quora. And we might have the best Quora answers get cited by the bots on Poe. We want to get them feeding off each other as much as we can.
Q: We’re in the midst of a big discussion on AI safety. What are you worried about, if anything?
A: I’m representing Quora here, not OpenAI. There’s this legitimate fear that once this technology gets powerful enough, it could be more intelligent than humans. Then what does that mean for society and how can we make sure that humans still stay in control of society?
We want to ensure that whatever is in control values humans or shares human values. It’s a solvable problem and it’s important that this goes well.
My role on the OpenAI board is to hopefully help make sure that we steer things in that direction. I don’t know when we get to this point where this is a risk, but it’s useful that people are thinking about it now.
Q: You also have some alarmist stuff out there. Is it helpful?
A: I don’t think it’s helpful. There’s a thought that we’re basically doomed and because of that, there’s some advocacy for very drastic actions to stop it, like airstrikes on Chinese or Russian data centers if they don’t shut down. There are some people who hear these things and might commit some kind of violence. I don’t think people should be advocating these things.
When people theorize about the future of AI, they often make some wrong assumptions. There’s often this idea that there’s going to be one AI system that’s incredibly powerful and no one else will have any comparable technology. Then you can tell the story that this thing can just run wild and take over the world.
At the point at which there is this super powerful AI, there will also be super powerful AI that is available to everyone else to defend themselves or rein it in or enforce the laws.
Q: Is there anything that has surprised you about it?
A: Because Poe is a general interface into many models, it puts us in this position where a lot of people who are making these models want to get theirs into Poe.
We’re about to open up an API that will let them just self-serve, add their own models. But we haven’t had that open so people need to talk to us to get access right now. And it has been very surprising how many people are building these models and also the quality of some of them.
There’s this huge wave of models coming. And the pace at which they’re advancing is really unbelievable. Hopefully when we have this API out, we’ll be able to let everyone participate without being pumped on our time.
Q: You’re not using Quora data yet to train any models, right? Is that on the roadmap?
A: We’re making a roadmap for the next month and it’s not on there. That’s about as far out as we can plan because this field moves so fast.
There are so many other things to do that are even more important. We implemented subscriptions, but you can only subscribe on iOS.
We want to add voice support, support for images. Meanwhile, there are the AI companies like OpenAI and Anthropic, and there are a few others that we’re in the process of adding.
There’s this really intense competition in that market, so they’re making their models better every week. We’re happy to be in this position of providing this network, as opposed to feeling like we need to invest in building our own.
Long term, it may make sense for us to use the Quora data set. But there are a ton of other things that we want to do first.
Q: When you say that you are planning one month at a time, is that a new thing?
A: This is just Poe. Quora is on a quarterly cadence. Poe has been operating on a weekly, or even daily cadence at some points. The world changes so much every week in AI.
It’s definitely moving faster than any field has ever moved. Every week, there are releases of new models. I have to spend hours every day just learning about what just changed and making sure what we’re doing still still makes sense.