Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet | Lex Fridman Podcast #434
Aravind Srinivas argues Perplexity wins by doing what Google’s ad margins won’t allow, not by beating Google at its own game.
- Google Cloud + YouTube combined are already at $100B ARR, so Google is not existentially threatened even if search ad revenue collapses.
- The AdWords bidding model was invented by Overture; Google only improved the auction math and mapped it onto a growing search platform.
- Srinivas invokes Bezos: ‘Your margin is my opportunity’ — Google won’t cannibalize its high-margin ad links to serve direct answers, which is Perplexity’s opening.
- Answer Engine Optimization (AEO) already exists as an industry, directly analogous to SEO, with people reverse-engineering Perplexity’s ranking signals.
- Perplexity’s own model, Sonar, is a post-trained Llama 3 70B fine-tuned for citations and summarization; the product is model-agnostic and can swap GPT-4o, Claude, or Sonar.
- Perplexity tracks P90 and P99 tail latencies at every system layer — search, LLM, time-to-first-token — inspired by Jeff Dean’s tail latency paper.
- Netflix runs 100,000+ server instances on AWS with no proprietary datacenter; Srinivas uses this as proof you can scale a major internet business entirely on cloud.
- Decoupling reasoning from stored facts is the path to capable AI without million-H100 clusters — smaller models that reason iteratively rather than encoding all knowledge in parameters.
Guests: Aravind Srinivas, CEO of Perplexity · 2024-06-19 · Watch on YouTube
