Descriptions:
Aravind Srinivas, CEO and co-founder of Perplexity AI, joins Lex Fridman to explain in technical depth how Perplexity works and where AI-powered search is headed. Srinivas brings unusual credibility to both sides of the conversation: he holds a PhD from UC Berkeley and worked as a research scientist at DeepMind, Google Brain, and OpenAI before founding one of AI’s fastest-growing consumer products.
Perplexity is described as an “answer engine” — distinct from a chatbot or traditional search — that combines web crawling, retrieval-augmented generation (RAG), and large language models to produce answers where every factual claim carries a citation to a source on the open web. Srinivas traces this design principle directly to academic writing norms: every sentence must be backed by evidence, and anything that cannot be sourced should not be stated. He walks through the full technical pipeline: query decomposition, retrieval and re-ranking, context injection into the LLM, and the post-training techniques that teach models to attribute accurately rather than confabulate.
The conversation also explores a deeper architectural question: whether the dominant paradigm of expensive pretraining — forcing models to memorize the entire internet — can be disrupted by small, reasoning-focused models paired with strong retrieval. Srinivas and Fridman discuss Microsoft’s phi-series work and what it would mean to fully decouple factual retrieval from reasoning capability. Developers and researchers interested in RAG systems, enterprise search, and the economics of AI product development will find this episode substantive and practically grounded.
📺 Source: Lex Fridman
🏷️ Format: Interview
