How To Build The Future

Srinivas says research in deep learning brought him from India to Berkeley and then to OpenAI, where Ilya Sutskever pushed him to focus on unsupervised learning. He became interested in search because product use could generate data that improves the underlying AI, creating a cycle of better products and more usage.

An idea about using language models to reformulate queries helped motivate Srinivas to explore search. Early experiments included an enterprise product for querying structured data and a Twitter search tool built with SQL templates and OpenAI Codex models. The team later tried a more general approach: give search results to a language model and have it produce a sourced answer, a method that became more practical as instruction following improved.

The first answer product took about seven seconds to respond, and the team limited answers to keep them short. A mistaken biography helped attract attention, while users sharing searches about themselves gave the team signs of interest. Follow-up questions doubled engagement time and increased daily questions, persuading Srinivas that the product had enough promise to continue.

Srinivas says product teams should treat user requests as useful signals and make the product handle ambiguity, rather than expecting people to become better prompt writers. Perplexity tracks daily questions as its primary metric and reviews growth weekly. As the company grows, production bugs, deployment checks, and loss of codebase context make it harder to move quickly.

Srinivas wants Perplexity to help users move from a question to an action, such as finding a product and buying it, while recognizing that product links can look like ads. He describes a future search experience that routes queries among small models, knowledge graphs, widgets, and longer reasoning, and can clarify intent or complete tasks. He argues that product judgment and a new way to monetize search will matter as AI changes how people find information.