AI summary by TwinMind

How To Build The Future: Aravind Srinivas

Y Combinator

34m 39s

Background and AI Journey

  • Aravind Srinivas, co-founder/CEO of Perplexity (9 billion valuation in <3 years)
  • Started with AI/deep learning research interest, came to US from India for Berkeley PhD
  • Life-changing OpenAI internship with Ilya Sutskever
    • Ilya dismissed fancy ideas in 5 minutes, said only unsupervised learning + reinforcement learning matter for AGI
    • This was during GPT-1 development (not called GPT-1 yet)
  • Shifted focus from RL (popular due to AlphaGo/DeepMind) to unsupervised/generative models
  • Google internship: read "In the Plex" book, inspired by Google's research+product combination
  • Identified only two AI+product problems: search and self-driving cars (data flywheel effect)

Company Formation and Early Pivots

  • Inspired by Daniel Gross YC blog "How to Build the Next Google" about query reformulation
  • Co-founder Dennis (CTO) met through shared research paper written day apart
  • Initial pitch to first investor Elad Gil: disrupt Google through visual/glass interface
  • Investor advised against consumer approach initially
  • Enterprise focus: searching specific verticals/databases/tables
    • Nobody wanted to share their data (tried Pitchbook, Crunchbase)
    • Built Twitter search using academic API access (pre-Elon)
    • Created chat UI for querying Twitter database with SQL templates
    • Used OpenAI Codex models (pre-GPT-3.5) with extensive error handling
  • Product took 1 month to build with 3 people, generated significant interest

Transition to General Search

  • Attempted similar approach for GitHub, LinkedIn but data access issues persisted
  • Realized two approaches: structured (tables/SQL) vs unstructured (LLM inference-time processing)
  • Chose unstructured approach betting on model improvements
  • Inspired by OpenAI's WebGPT (John Schulman's team) and internal "Truth Bot"
    • WebGPT was slow, agentic with RL-based browsing
  • Built simpler heuristic: always take top-k search results, use snippets only, feed all to prompt
  • GPT-3.5 Turbo timing made "dumb approach" work due to better instruction following
  • First version took 7 seconds, no streaming, hard-coded length limits

Launch and Early Traction

  • Viral moment: academic searched herself, got past-tense biography (she was alive)
    • Actually confused with deceased person of same name
  • Pattern: people searching for themselves became common behavior
  • Added follow-up questions feature: doubled engagement time and daily query volume
  • Exponential growth in daily questions convinced team not to pivot to enterprise

Competition Concerns and Funding

  • Never seriously considered Google competition initially
    • Google homepage too cluttered for this product
    • Difficult to determine purely informational queries
  • More worried about Microsoft Bing Chat launch
    • Happened same day as term sheet handshake with NEA
    • Other investor extended due diligence from 30 to 45 days
    • Original investor honored commitment despite competitive concerns
  • Google announced Bard week later with just screenshots
  • Microsoft historically poor at consumer products, Google had own challenges

Product Philosophy and Culture

  • "User is never wrong" principle (Larry Page influence)
    • Product should clarify ambiguous queries rather than blame user
    • Opposite of enterprise software approach
  • Google-inspired design details: cursor ready in search bar, auto-suggest, typo handling
  • Primary metric: number of queries per day (like early Google)
    • Shared company-wide at all-hands meetings
  • No hierarchy for bug reporting - direct communication encouraged
  • High standards: "if it doesn't feel good for ourselves, user won't feel it either"

Team Management and Scaling Challenges

  • Hire people who care about good work, not just treating it as job
  • Obsessive detail orientation naturally filters hiring
  • Slowing down due to production stability concerns, not just team size
  • Engineer #250 can't have same context as early engineers
  • Fighting entropy through continued founder involvement in details
  • Co-founders push quality principles in their teams

Future Vision and Strategy

  • Current view: "more intelligent Google search useful in certain scenarios"
  • 3-4 year vision: end-to-end experience from problem to action fulfillment
    • Example: research best sweater → buy directly through platform
    • Challenge: early adopters want ad-free vs mass market monetization needs
  • Building orchestration system combining:
    • Small models, knowledge graphs, widgets
    • LLM streaming answers, multi-step reasoning
    • Instant answers for simple queries (weather, scores, ages)
  • User shouldn't know which system component handles their query

Competitive Advantages

  • Against Google: obsession about user + good product taste
  • Google's dilemma: search company that's also ads company
    • Search exists to serve ads (~$200B annual revenue)
    • Other revenue (YouTube, Cloud) lower margin
    • Stock price tied to search revenue performance
  • Hypothesis: Google can't easily put AI chat on core search with billion users
  • New monetization model needed for AI-first search
  • Perplexity advantages over OpenAI/Anthropic:
    • Product-focused DNA vs pure AI research focus
    • Capable of using open source models, fine-tuning, serving
    • Won't spend bandwidth on data centers/chips
    • Focus on user experience over benchmarks

Technical Approach

  • Capable of taking latest open source models and self-serving
  • Post-training, fine-tuning, evaluations expertise
  • Not "AI illiterate" but not focused on infrastructure
  • Building next-generation information experience vs pure reasoning capabilities

TwinMind

Get summaries like this for your own meetings

TwinMind takes notes in the background during meetings, lectures, and calls, then writes the summary for you.

Free plan to get started

No bot joins your calls

Phone, desktop & Chrome apps

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome

Your next meeting, summarized like this

TwinMind listens in the background and turns meetings, lectures, and calls into searchable notes you can chat with. Start free.

Available on phone, desktop & Chrome