What would you ask if no one could judge you?
Co-founder and CEO Aravind Srinivas wants Perplexity to be not only a resource, but an engine for curiosity.
Share What would you ask if no one could judge you?
Hero illustration by Kyle Platts
If you ask the search platform Perplexity how Perplexity works, it says: “Perplexity AI combines natural language processing, web indexing, LLMs, and transparent sourcing to deliver accurate, up-to-date answers by seamlessly integrating web search and conversational AI capabilities.” If you ask Perplexity co-founder and CEO Aravind Srinivas, it’s an “answer engine.” This distillation is a nod to Aravind’s vision for a knowledge app that synthesizes relevant information from around the web instead of returning an exhaustive list of links. We sat down with him to learn how Perplexity delivers information at the right altitude, the relationship between footnotes and sources, and why catering to our collective curiosity is a personal endeavor.
What led you to build Perplexity?
As a kid, I would spend a lot of time in Wikipedia rabbit holes clicking on links. Perplexity is kind of built for those Wiki nerds. Back in the day, encyclopedias were big volumes of books; then they became CDs, DVDs, pen drives, Wikipedia, and now Perplexity.
We want to be the reference for others by building an app that delivers accurate information and knowledge. We want to make learning fun and interesting. Everybody talks about making mobile apps engaging, but that doesn’t need to be done through dancing videos. It can be done through making people curious. So catering to humanity’s curiosity is a big overarching goal for us, and that starts with allowing people to ask whatever they want without any fear of being judged.
How did building Perplexity ease your own fear of judgment?
My co-founders and I were first-time entrepreneurs and didn’t have any company or product building experience, so we built a Slackbot to answer our own questions—like understanding the different mechanisms to do fundraising, or how to set up health insurance for our employees. It was so useful, but we never had the courage to launch it. We were always worried about what people would think about us. Would they think we are that audacious to compete with Google?
And then one of our investors, Nat Friedman, said, “Why are you taking yourselves so seriously? You have literally nothing to lose. The worst case outcome is actually progress from where you are, and the best case is that it takes off and becomes the next big thing.” And I think that appealed to me—that sort of asymmetric bet where loss is fine and a win is amazing. We launched a week after ChatGPT.
Why are you taking yourselves so seriously? You have literally nothing to lose.
How did that timing inform the big decisions you made around how Perplexity works?
ChatGPT clearly established in everybody’s mind that there was a knowledge cut-off; it would hallucinate and answer the wrong things. So we came up with a different twist: What if we literally gave you the answer, but also gave you the sources so that it’s as accurate and as grounded in facts as possible? And what if we built a UI that truly speaks to the user? In fact, I’ve heard one user describe Perplexity as Wikipedia and chat coming together and having a baby. The only difference is that the data for it comes from the whole internet.
How do you think about making information approachable without being reductive?
I was once at [OpenAI CEO] Sam Altman’s house and noticed that he had a lot of books. I said, “You have no time to read all these books. Why do you even have them?” He told me, “You don’t need to read more than 20 pages. You read the intro, the first chapter, you get the 80/20.” The idea is that you get most of the value of the book very quickly. That’s what Perplexity wants to establish for you.
Let’s say I want to understand LLMs. I just need to understand 20% of the bits to get 80% of the understanding, and that can be achieved through a tool like Perplexity that sources content from many different web pages, picks the most important things, and forms a concise answer.
How does Perplexity balance answering someone’s question with encouraging further curiosity?
After an answer to a query, you see three related questions. My hypothesis is that every person is extremely curious, but most of us don’t have the inclination to channel it. Even if we do, we don’t have the ability to articulate what we’re curious about. Good questions are not that frequent. That’s why “Let Me Google That for You” exists. That’s where we can really help people.
A good question is very personal to your understanding of the topic. Maybe you’re trying to understand black holes. Some physicists might be at the PhD level. The AI has to know that a user wants something at this granularity. We don’t blame the human for not being able to come up with a good, precise question. We always work with the product design philosophy that the human is never wrong. The burden is on the AI, not on us. That’s a true copilot.
We always work with the product design philosophy that the human is never wrong. The burden is on the AI, not on us. That’s a true copilot.
What’s a bug that you’re currently working through?
When there isn’t enough information, the model doesn’t say, “Hey, we don’t have enough data, and I can’t make a conclusion.” If you ask a question like, “Who won the Olympics in 2024?” it might give you an answer based on a snippet from a previous time when the U.S. won. Those sorts of reasoning skills are what we need to keep working on.
What differentiates Perplexity from other AI tools?
With Perplexity, one part is about facts—responding to a query by pulling in relevant documents from the web—and the other part is about reasoning—synthesizing those documents in a concise way, with appropriate references. We train Perplexity for the skill of taking in a bunch of information, a bunch of snippets, and then synthesizing a concise answer.
That’s what’s different from other generative AI companies that are either focused on building one bot that does everything, or on one particular vertical. Perplexity is horizontal in terms of how many different professions can use the product, but vertical in terms of the use case we focus on.
How big is the gap between how AI works in theory and how we’re able to use it in practice today?
We are all talking about scenarios where AI agents will take over. But can I have AI reliably find a restaurant for me? It’s not a simple case of finding where there are spots available on Google Maps. The next scenario is an AI agent understands you’re looking for a dinner reservation, goes to your calendar, finds the free slots in the restaurant, and sees what overlaps. If there are no open slots, it calls the restaurant and gets the details about the wait time or looks for an alternate place. This is the iterative, multistep reasoning that today’s models aren’t able to do.
What is the promise of AI?
If AI really works, life will be simple for a lot of us. AI will be personalized to us. You won’t have to do all of these burdensome tasks. And if that’s available to every person on the planet in an affordable manner, our world will be a much better place. AI is an equalizing force: It gives everyone superpowers.
Explore the rest of The Prompt, a magazine available online and in the Figma Store as a limited print edition.
Alia Fite is a writer and editor on Figma's Content & Editorial team. She has previous experience at Stripe and Dropbox.
