The Scene
Aravind Srinivas, co-founder and CEO of Perplexity, has been at the forefront of trying to shake up online search, with the help of advances in artificial intelligence.
Semafor has been tracking those developments in past stories. Now he’s talking to online platforms like Instacart about ways to incorporate a conversational search engine with commerce. If that effort is successful, the results could be profound.
Say you want to buy something: First, you start with a Google search. You end up at Wirecutter. Then you go to Amazon and read a bunch of reviews. Once you’ve decided what you want to buy, you go back to Google and try to find the best price. Then you go to another web site and order it. Every one of those sites has ads, the digital oil that powers a big chunk of our economy. Now imagine all of that happens in the form of a conversation. Maybe you don’t even look at a screen.
That’s what Perplexity and other companies are trying to build. Srinivas talks about those efforts and more below in an edited conversation.
The View From Aravind Srinivas
Q: How did you come up with the idea for Perplexity?
A: We were passionate about search and we thought one of the best ways language models could be useful is in summarization, where it takes the relevant aspects of the query and summarizes them. Then ChatGPT came out and people were comparing ChatGPT to Google and saying ChatGPT was better. That gave me the conviction that people are really looking for this service. We launched right away, at first on Discord.
Q: How does Perplexity work? How do you ensure the answers are accurate?
A: When a user asks a question, the search engine pulls up relevant web pages. The language model then reads through those pages and crafts an answer, providing citations to relevant sources. We focus on citing authoritative, high-page rank sources to ensure accuracy. The language model is also prompted to provide an “unbiased and journalistic tone” and only use information from the search results. If there are conflicting answers from different sources, Perplexity will provide separate answers for each.
Q: So it’s like a very elaborate ChatGPT prompt?
A: You can just prompt it. We didn’t train anything on that. That’s the magic about this.
Q: So is the prompt your secret sauce?
A: No, everyone’s prompts leak. One of our old prompts leaked on Hacker News.
Q: So you’re relying more on network effects and that kind of thing to keep people using the product?
A: Yeah, we’re adding new features and products all the time.
Q: Do you worry about being somewhat reliant on OpenAI?
A: I think we should be concerned about it, but there is a good amount of competition in the market.
Q: Tell me about this new co-pilot feature?
A: We developed the co-pilot feature during a company hackathon in Denver. We looked at how to use GPT-4 to provide answers to harder questions that require research and clarification. The co-pilot follows up the user’s initial question with clarifying questions to better determine what they need. It then provides a comprehensive, tailored answer based on their responses. The co-pilot is designed to work with the user rather than autonomously provide an answer.
Q: How do you see AI assistants and search evolving in the future? Are subscriptions the future?
A: AI assistants and search will fundamentally change how people interact with computers. Assistants will increasingly work together with humans to provide answers and complete tasks. However, human agency and oversight will remain important.
A subscription model for search is possible as search evolves to provide more value. As search engines gain the ability to understand complex questions and provide time-saving answers, people may be willing to pay for that additional value and experience. A subscription model also benefits search engines by allowing them to focus on user needs rather than advertising. Some people may continue using free, ad-based search options, but subscriptions could appeal to certain segments.
Q: Do you see advertising being a part of search in the future?
A: When a person is talking to a chatbot or a conversational search engine and trying to directly seek answers to precisely targeted queries, companies probably don’t even need to advertise, because we would already pull up the more relevant data and surface it to the user. A platform like Shopify or Instacart could let merchants reach the end user in an even more targeted way.
In this new world, where ads are just going to be citations, people should just put more effort into having well-documented information about them on the internet and an LLM can parse it on the fly. I feel like this might make the world better in some sense. Why do you need to advertise if an LLM can just go read about you and what you’re saying? You’re just part of the index, so make sure you build a good product and people will automatically know. I actually don’t know what’s going to happen, but it’s fun to think about.
Q: Are we headed to a place where AI search engines will start taking actions for us? For instance, do you think that eventually when people are doing research for their trip to Paris or whatever, they’ll want Perplexity to start actually booking the travel for them?
A: We would love to do that. The part of helping the human to arrive at the price and the time and like, like the activities to do in Paris, that’s the hard part. The actual booking of a ticket is so easy. People would have to approve each purchase once all the details are there. So that’s basically what we want to enable over time.
Q: Aren’t there roadblocks there that are not so much about technology. For instance, incumbent booking services are going to want people to go through their platforms, right?
A: We’ve been talking to Klarna and Instacart, so I think people are pretty interested in partnering with us.
Q: So Perplexity would have an Instacart API or something that would allow the order to occur and Perplexity would get a commission?
A: We haven’t talked through details like this. The first step is to make the experience really good. It’s mostly an early prototype. The point I’m trying to make is the traditional ideas of ‘hey, I’m only going to work with you if I get the traffic,’ those days are gone. It’s mostly like, ‘hey, let’s try to give our users the best possible experience.’ If you can enable that, it’s a win-win. That’s how people are thinking about it.