Aravind Srinivas

Aravind Srinivas co-founded Perplexity in 2022, helping build an AI answer engine that combines web search with language models and citations. His consequential design choice was to put a synthesized answer at the center of the experience. He connected that approach to academic writing: claims should be backed by sources or experimental results, rather than simply sound convincing.[6][8]

In his June 2024 interview with Lex Fridman, Srinivas recalled an employee asking the founding team about health insurance. The team tried a Slack bot connected to GPT-3.5, but it gave answers they could not trust. Srinivas and his co-founder turned to their academic practice of citing what they wrote. The resulting goal was a chatbot that would answer from information it could retrieve and reference.[8]

Srinivas described the pipeline as retrieving relevant documents, extracting useful paragraphs and feeding them to a language model instructed to produce a concise, cited answer. The additional constraint was to stay within the retrieved information and acknowledge when it was insufficient. He attached an explicit qualification: “I’m not saying it’s foolproof.” Errors could still arise from misunderstanding the question, stale or incomplete snippets, irrelevant documents, or too much distracting context.[8]

His technical argument extended beyond choosing a language model. Srinivas argued that web retrieval needs a hybrid approach: semantic representations, traditional term-based retrieval such as BM25, and ranking signals including domain authority and recency. He emphasized that packing a page into a vector representation does not settle which aspect of that page matters to a particular question.[8]

Srinivas favors competing on the complete product rather than relying on lasting model superiority. In a Perplexity AMA, he argued, “The product is the ultimate place to compete on.” In June 2024, he described choosing whichever customized model gave the best answer, whether adapted from another organization’s base model or hosted by Perplexity. He also described the team’s Llama-3-based Sonar work as post-training for summarization, citations and context handling.[3][8]

That product focus includes understanding imperfect queries and reducing latency. Srinivas recalled initially blaming his mother’s poorly formed English query when an answer was irrelevant, then recognizing that the product should infer her intent. He also described testing Perplexity on unreliable flight Wi-Fi and tracking latency across both search and model components, including the slowest portions of the response-time distribution.[8]

In the Perplexity Labs AMA, Srinivas framed the next opportunity as moving from “ask a question” to “do this for me.” He wanted products to complete work rather than leave users copying answers between tools. He described browser-using agents as the first bet, with broader computer use through desktop apps as a possible direction.[4]

Before Perplexity’s cited web answers became the central product, Srinivas and his co-founders experimented with natural-language access to relational data. He recalled using Codex to turn questions about Twitter relationships into SQL queries, execute them and return the records. The team wrote query templates, retrieved similar examples for dynamic few-shot prompts, and added error handling and retries. They initially chose SQL because its restricted, templated output seemed easier than general Python; Srinivas later said that rationale had turned out to be wrong.[8]

Srinivas recalled releasing the web answer product on December 7, 2022. Continued use during the Christmas holiday convinced him there was more than a passing novelty. The initial version answered one query at a time; he said the team added conversation and suggested questions a week after New Year. He and his co-founders subsequently framed their mission around knowledge discovery, with follow-up questions helping users explore beyond the first answer.[8]

In Latent Space’s July 2023 Code Interpreter discussion, Srinivas described related early Perplexity experiments that let users plot follower distributions or likes over time using Codex. He said he had not yet tried Code Interpreter and expressed skepticism about its value for experienced programmers. Simon Willison responded with examples of coding projects it had enabled him to undertake.[1]

Before founding Perplexity, Srinivas worked on learning representations for reinforcement learning. He co-authored CURL: Contrastive Unsupervised Representations for Reinforcement Learning, first submitted in April 2020, with Michael Laskin and Pieter Abbeel. Srinivas and Laskin contributed equally. CURL extracts high-level features from raw pixels through contrastive learning, then performs off-policy control using those features. The authors reported that, on the DeepMind Control Suite, it nearly matched the sample efficiency of methods using state-based features.[5]

Srinivas completed his computer-science Ph.D. at UC Berkeley in 2021, then worked at OpenAI for a year before co-founding Perplexity; Andrew Konwinski was among his fellow founders. Reflecting on Berkeley’s expectation that researchers handle ideas, experiments and implementation, he said, “I learned that software engineering is critical to doing great work in AI.” That training provides context for his later emphasis on making models work within a reliable product.[6]