
An answer engine, not a link list
Perplexity co-founder and CEO Aravind Srinivas describes the product with a phrase of his own: an “answer engine.” The distinction matters to him. Where a conventional search engine returns an exhaustive list of links, Perplexity synthesizes relevant material from across the web, and — as the company's own description has it — ties natural language processing, web indexing and LLMs to transparent sourcing so that answers stay current. The goal is a knowledge app that delivers information at the right altitude. 
Rabbit holes and the courage to launch
Srinivas traces the idea to his childhood in Wikipedia rabbit holes, clicking from link to link. Perplexity, by his account, is built for those “Wiki nerds”: the encyclopedia has moved from bound volumes to CDs, DVDs and pen drives, and now to an answer engine.
The path from that idea to a released product ran through doubt. Srinivas and his co-founders were first-time entrepreneurs, and they built a Slackbot to answer their own questions — how fundraising mechanisms work, how to set up health insurance for employees. They found it useful but never had the courage to launch, worried about how people would react. Would they think we are that audacious to compete with Google?
Investor Nat Friedman reframed it: "Why are you taking yourselves so seriously? You have literally nothing to lose." The worst case, he argued, is still progress; the best case is the next big thing. Srinivas describes that as an asymmetric bet, and the team launched a week after ChatGPT.
Grounding answers in sources
That timing shaped the product decisions. ChatGPT had established in users' minds that a knowledge cut-off existed and that models hallucinate. Perplexity took what Srinivas calls a different twist: give the answer, but also give the sources, making it as accurate and as grounded in facts as possible — with a UI that speaks directly to the user. One user, he recalls, described it as Wikipedia and chat coming together and having a baby, except the underlying data comes from the whole internet.
On how Perplexity separates itself from other AI tools, Srinivas splits the work in two. One part concerns facts — pulling relevant documents from the web in response to a query. The other concerns reasoning — synthesizing those documents concisely, with appropriate references. What sets Perplexity apart from other generative AI companies, in his telling, is that they focus either on one bot that does everything or on a single vertical. Perplexity is horizontal in how many professions can use the product, yet vertical in the use case it targets.
Surfacing the question you couldn't ask
Srinivas wants the product to serve curiosity directly. “Catering to humanity’s curiosity is a big overarching goal for us,” he says, and the first step is letting people ask whatever they want without fear of judgment. Nobody needs dancing videos to make an app engaging, he argues — curiosity can do it.
The interface supports this through related questions: after each answer, the display shows three of them. Srinivas' hypothesis is that everyone is extremely curious, but most of us lack the inclination to channel it — and even when we do, we cannot always articulate what we are curious about. Good questions are not frequent, as he puts it, which is why “Let Me Google That for You” exists.
A good question depends on the asker's level of understanding; someone learning about black holes and a PhD physicist need very different answers, and the AI must sense the right granularity. Perplexity's design philosophy places the burden on the AI for this reason: “the human is never wrong,” Srinivas says. That, in his view, is what makes a true copilot.
The 80/20 idea behind it comes from a conversation with Sam Altman, who pointed out that reading the intro and first chapter of a book delivers most of its value. Someone trying to understand LLMs, Srinivas reasons, needs roughly 20% of the material for 80% of the understanding — and Perplexity can gather that by sourcing many web pages, picking the most important parts and forming a concise answer.
What still breaks
Asked for a current bug, Srinivas points to insufficient data. Rather than saying it can't reach a conclusion, the model may infer from an older snippet — a question like “Who won the Olympics in 2024?” can return an answer reflecting a previous U.S. win. Building those reasoning skills, he says, is ongoing work.
From the restaurant booking to universal superpowers
The gap between AI in theory and AI in practice shows up in tasks that sound trivial. An agent booking a restaurant would have to recognize the intent, check your calendar for free slots, compare them with availability, and — if nothing overlaps — call to ask about wait times or look for an alternative. That kind of iterative, multistep reasoning is beyond today's models.
The payoff, if it works, is a simpler life: personalized AI lifting burdensome tasks, affordable and available to every person on the planet. “AI is an equalizing force,” Srinivas says. “It gives everyone superpowers.”




