#434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet

June 19, 20243hr 11min

#434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet

Lex Fridman Podcast

Arvind Srinivas is the CEO and co-founder of Perplexity, a startup aiming to revolutionize how humans find answers and discover knowledge on the internet. In this wide-ranging conversation, Arvind discusses the technical details behind Perplexity's "answer engine", his vision for the future of search and AI, lessons learned as a startup founder, and perspectives on the broader AI landscape. The discussion covers Perplexity's origins, its core technologies like retrieval augmented generation, the challenges of web indexing and ranking, and Arvind's thoughts on topics like open source AI development, AGI timelines, and maintaining a human-centric approach to AI progress. Arvind also shares advice for aspiring entrepreneurs and his views on work-life balance for founders.
#434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet
#434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet
0:00
-0:00

Key Takeaways

  • Perplexity is an "answer engine" that combines search and large language models to provide answers with citations to human-created web sources
  • The company aims to be "the world's most knowledge-centric company" and cater to human curiosity
  • Key technical components include retrieval augmented generation (RAG), custom-trained language models, and a focus on reducing hallucinations
  • Arvind believes AI will augment human intelligence and curiosity rather than replace it
  • The future of search and knowledge discovery will likely involve more personalized, conversational interfaces
  • Building a startup requires genuine passion, relentless determination, and a willingness to work extremely hard, especially when young
  • Open source AI development may be key to ensuring safety and transparency as capabilities advance

Introduction

Arvind Srinivas is the CEO and co-founder of Perplexity, a startup aiming to revolutionize how humans find answers and discover knowledge on the internet. In this wide-ranging conversation, Arvind discusses the technical details behind Perplexity's "answer engine", his vision for the future of search and AI, lessons learned as a startup founder, and perspectives on the broader AI landscape.

The discussion covers Perplexity's origins, its core technologies like retrieval augmented generation, the challenges of web indexing and ranking, and Arvind's thoughts on topics like open source AI development, AGI timelines, and maintaining a human-centric approach to AI progress. Arvind also shares advice for aspiring entrepreneurs and his views on work-life balance for founders.

Topics Discussed

How Perplexity Works (10:52)

Arvind explains that Perplexity is best described as an "answer engine" rather than a traditional search engine. The key components are:

  • Search: Extracts relevant results and paragraphs from indexed web pages
  • Large Language Model: Takes the relevant paragraphs and query, generates a well-formatted answer with citations
  • Citations: Every sentence in the answer is backed by sources, similar to academic writing

He notes: "The magic is all of this working together in one single orchestrated product. And that's what we built Perplexity for."

Perplexity's Origins and Vision (50:13)

Arvind shares the origin story of Perplexity:

  • Initially explored building search over Twitter data and relational databases
  • Pivoted to general web search and question answering after seeing strong user engagement
  • Developed the vision of being "the world's most knowledge-centric company"

He explains: "We want to obsess about knowledge and curiosity. And we felt like that is a mission that's bigger than competing with Google."

Technical Details of Perplexity (2:05:25)

Arvind dives into the technical components powering Perplexity:

  • Retrieval Augmented Generation (RAG): Retrieves relevant documents for a query, uses them to generate an answer
  • Web Crawling: "Perplexity bot" crawls the web, makes decisions on what/how to crawl
  • Indexing: Processes crawled content into a format for efficient retrieval
  • Ranking: Uses both traditional information retrieval (e.g. BM25) and neural approaches
  • Custom Language Models: Fine-tuned models like "Sonar" optimized for Perplexity's use case

On reducing hallucinations, Arvind notes: "The principle in Perplexity is you're not supposed to say anything that you don't retrieve, which is even more powerful than RAG."

The Future of Search and Knowledge Discovery (2:42:52)

Arvind shares his vision for the evolution of search and knowledge discovery:

  • Moving beyond just search to broader knowledge discovery and curiosity-driven exploration
  • More conversational, personalized interfaces
  • Ability to ask more complex questions not possible with traditional search
  • AI augmenting human intelligence and curiosity rather than replacing it

He states: "I just believe that we're working towards neither search or answer engine, but just discovery, knowledge discovery, that's a bigger mission."

Advice for Startup Founders (2:30:15)

Arvind offers advice for aspiring entrepreneurs:

  • Work on things you're genuinely passionate about, not just what you think will get funding
  • Be prepared for extreme hard work and dedication, especially when young
  • Surround yourself with driven, passionate people
  • Have a strong support system to cope with the challenges

He emphasizes: "If there is a certain idea that really just occupies your mind all the time, it's worth making your life about that idea and living for it, at least in your late teens and early twenties."

Perspectives on Open Source AI (1:05:11)

Arvind discusses the importance of open source AI development:

  • Enables more players in the field beyond a few large companies
  • Allows for faster discovery of potential misuse and development of safeguards
  • Academic researchers can make breakthroughs that benefit frontier models

He notes: "I think there's lot he's controversial for like, whatever's happened in social media in general, but I think his positioning of Meta and like himself leading from the front in AI, open sourcing great models, not just random models really. Like LLaMA 70B is a pretty good model."

Thoughts on AGI and the Future of AI (3:00:29)

Arvind shares his perspective on artificial general intelligence (AGI) and AI progress:

  • Believes AGI development will be more gradual rather than a sudden breakthrough
  • Sees AI augmenting human intelligence rather than fully replacing humans
  • Hopes AI will make humans "even more empowered, even more curious, even more knowledgeable and truth seeking"
  • Cautions against AI systems that provide artificial emotional connections

On the positive potential, he states: "We'll all have a much more fulfilling life, and that way, like, have a lot more time for other things and channelize that energy into, like, building true connections."

Conclusion

This wide-ranging conversation covered the technical details behind Perplexity's novel approach to search and knowledge discovery, as well as Arvind's broader perspectives on the future of AI and advice for entrepreneurs. Key themes included:

  • The potential for AI to augment human intelligence and curiosity rather than replace it
  • The importance of maintaining a human-centric approach as AI capabilities advance
  • The value of open source development in ensuring safe and beneficial AI progress
  • The need for genuine passion and extreme dedication when building a startup

Arvind's vision of AI empowering humans to better satisfy their curiosity and expand their knowledge offers an optimistic view of how these technologies could shape the future. At the same time, he acknowledges the potential risks and emphasizes the importance of thoughtful development focused on long-term human flourishing.