Feb 21, 2025 · 34m · y-combinator
How To Build The Future: Aravind Srinivas · Y Combinator
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of Y Combinator's 'How To Build The Future', host David Lieb interviews Aravind Srinivas, co-founder and CEO of Perplexity, about the company's rapid rise, technical breakthroughs, and strategies for competing against search and AI giants.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The partners as informed peer 4.3 Guest teaching 3.2 Guest disagreement 1.2 The partners pushing back 0.7
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 33:09 Contrasting product obsession with benchmark chasing
Aravind dismisses frontier lab competitors like OpenAI and Anthropic, arguing they focus on datacenter builds and coding benchmarks rather than product design and user experience.
Hardest push from the partners ▶ 21:44 Pushing back on query volume as a success metricLieb directly counters Aravind's primary metric of daily query volume by citing Larry Page's philosophy that a great search engine minimizes time on site.
Biggest teaching moment ▶ 10:22 Explaining the shift to unstructured LLM inferenceAravind explains how scaling models invalidated traditional domain indexing in favor of fast inference-time extraction from search snippets.
The partners hold their own ▶ 19:02 Citing internal Google history and Paul BuchheitLieb brings his background working at Google to correct the origin story of Google's spellchecker, citing Paul Buchheit's role.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| Aravind Srinivas's Path to AI Research | 4 | 3 | 1 | 0 | David Lieb opens with an introductory question about Aravind's background. Aravind shares formative stories regarding Ilya Sutskever at OpenAI and reading 'In the Plex' at Google. Lieb demonstrates good context by synthesizing Aravind's core flywheel thesis. | |
| The Spark for Perplexity and Query Reformulation | 3 | 4 | 1 | 0 | Aravind details the spark behind Perplexity, citing Daniel Gross's query reformulation thesis and the realization that removing link clicks threatens the ad economy. The conversation is collaborative and explanatory. | |
| Early Demos and the Twitter Search Experiment | 4 | 2 | 1 | 0 | Lieb asks about early prototypes prior to Perplexity. Aravind describes pivoting from enterprise tabular search to a viral Twitter search demo built on OpenAI Codex SQL generation. | |
| Pivoting to Unstructured General Search | 3 | 5 | 1 | 0 | Aravind walks through the fundamental technical pivot from pre-indexed structured SQL querying to unstructured inference-time summarization with citations. He references Paul Graham's essay and OpenAI's early WebGPT research. | |
| Solving Latency and Achieving Initial Virality | 4 | 3 | 1 | 1 | Lieb characterizes their choice as the 'dumb approach' betting on model progression. Aravind agrees and recounts early virality driven by ego-searches, biography hallucinations, and follow-up query retention. | |
| Navigating Competition and Fundraising Pressures | 4 | 3 | 2 | 1 | Aravind discusses the panic when Bing Chat and Google Bard were announced right around their seed term sheet. He explains why big tech incumbents suffer from innovator dilemmas in consumer UX. | |
| Product Philosophy: 'The User Is Never Wrong' | 6 | 3 | 1 | 2 | Lieb leans on his decade of experience at Google to discuss product craft and Larry Page's design philosophy. When Lieb raises a potential issue with query count metrics, Aravind explains how Perplexity interprets follow-up queries. | |
| Building a Data-Driven and Feedback-Obsessed Culture | 4 | 2 | 1 | 0 | Lieb inquires about management and internal culture. Aravind explains their weekly review of growth numbers, his habit of filing 50 bugs daily, and gathering unfiltered feedback from X. | |
| Maintaining Agility and Quality While Scaling | 4 | 2 | 1 | 1 | Lieb questions how Perplexity avoids the slow-down of larger companies. Aravind candidly concedes they are already experiencing friction from regression testing, staging, and onboarding new engineers. | |
| The Future of Perplexity: From Answers to Action Fulfillment | 5 | 4 | 1 | 1 | Lieb asks where Perplexity goes over the next few years. Aravind lays out the vision of transitioning from answers to action fulfillment and building an intelligent orchestrator of models, widgets, and knowledge graphs. | |
| Competing with Giants and Long-Term Edge | 6 | 4 | 2 | 2 | Lieb pushes Aravind on why Perplexity can beat giants like Google, OpenAI, and Anthropic. Aravind breaks down Google's margin constraints in search vs cloud/YouTube, while arguing OpenAI and Anthropic prioritize benchmark chasing over user experience. |
Statements from this episode (19)
Ilya Sutskever defined AGI as unsupervised learning containing reinforcement learning
“And then he told me the only thing that matters is he drew two circles. One big circle, called it unsupervised learning, and then inside he said, reinforcement learning, another circle, and he said, this is AGI, every other research doesn't matter.”
Search and autonomous driving are the only AI product data flywheels
“There are probably only two problems where you can work on AI and also build product at the same time. One is like search, and the other is self-driving car.”
Perplexity's co-founders met after publishing the identical paper a day apart
“We had written the same paper a day apart, so we knew each other, and he was a visiting student in my lab”
Srinivas initially pitched seed investor Elad Gil on smart glasses, not search
“I was like, adoshes enough to go and pitch to the first seed investor of ours, Elad Gill, that like, hey, like you know, I want to disrupt Google but I kind of want to do it from pixels from a glass”
Perplexity's first demo queried Twitter tables using OpenAI Codex
“Twitter. The pre-Elon CEO moments academic access was allowed, legal, so we built a database of Twitter. We organized it in the form of tables. We tried to do it with the OpenAI codex models. This was even pre-GPT-III. We wrote a lot of templates.”
Inference-time search gives startups a structural advantage over Google's legacy index
“There's one way of doing these things where we go to each of these domains and like try to build an index of it and put it into specific formats like tables and then have the LLM, like read that. In a structured language of SQL, or you could do the other way w…”
OpenAI built a slow, internal web browsing prototype called Truthbot
“OpenAI even had a bot when I worked there called the Truthbot, which John, John built with this team. Where you could ask it a question and it'll go and search the web, and then it'll give you an answer with some sources. And it was very slow, and it was built…”
Releasing follow-up questions doubled user engagement time on Perplexity
“And then we released the ability to ask follow-up questions. That doubled the engagement time on the site, and also increased the number of questions every day, and the number of people, number of questions every day was increasing exponentially.”
Google cannot replicate Perplexity directly on its cluttered search homepage
“They cannot make this exact product on the Google homepage. It's so hard to know when a query is purely informational or not, and then the Google search page is already, like, so cluttered, that's the answer box, the knowledge panel,”
Microsoft messed up its AI search opportunity with Bing Chat
“Microsoft was never really good at consumer products for a long, long time. You can't suddenly change that. So they actually messed up the opportunity, in my opinion.”
Most Perplexity follow-up queries are completely irrelevant to the initial search
“Most of the follow-up queries actually we see are, like, completely irrelevant to the first query, because they just want to keep continuing the session.”
Perplexity is slowing its development speed to protect production reliability
“Well, it's beginning to happen already a little bit, right? We're not as fast as we used to be. I think some of it is not because of people. It's also because things breaking in production, people start losing trust in the product.”
Companies cannot expect their 250th engineering hire to be obsessively detail-oriented
“The obsessive detail-oriented people there are only that many people in the world, so obviously you cannot expect engineer number 250 to be like that.”
Whoever builds a billion-user AI query orchestrator will become the next Google
“That sort of router, that, that orchestrator, I think that's the hardest thing to build, and whoever builds that, and can operate that at a scale of billion users, and also knows how to monetize, like, some of those queries really well, is going to be the next…”
Google's current infrastructure is the closest system to an ideal AI orchestrator
“Whatever Google has already built is the closest system to something like this.”
Google has better consumer product taste than both OpenAI and Anthropic
“Out of the list you mentioned, Google is the only company that actually has the product taste to do this.”
YouTube will never achieve profit margins as high as Google Search
“YouTube is not never going to be a high margin business because number one, they don't serve ads on subscription like users. And number two, like they have to pay the creators, they have to pay the media partners. So it's never going to be as high margins as s…”