Lex Fridman | Aravind Srinivas on Perplexity, Google & the Next Web
[HPP] Aravind SrinivasJanuary 2, 20262h 59min
41 connectionsΒ·40 entities in this videoβPerplexity's Answer Engine Approach
- π‘ Perplexity functions as an answer engine, combining traditional search with Large Language Models (LLMs) to provide direct, sourced answers.
- π― The core principle is to back every statement with citations, similar to academic papers, ensuring factual grounding and accuracy.
- π This approach was developed to combat chatbot hallucinations by forcing the model to only use information it retrieves from multiple internet sources.
Differentiating from Traditional Search
- π Perplexity aims to be a knowledge discovery engine, focusing on direct answers and guiding users to deeper insights, rather than just providing a list of links like Google.
- π Google's highly profitable AdWords business model disincentivizes it from adopting a direct-answer UI, as it would reduce ad clicks.
- β Perplexity's strategy is to rethink the entire search UI, betting on the exponential improvement of AI technology to reduce hallucinations and enhance user experience.
Key AI & LLM Breakthroughs
- π§ The evolution of LLMs involved crucial insights like attention mechanisms, parallel computation in Transformers, and scaling laws for pre-training.
- π¬ Retrieval Augmented Generation (RAG) is central to Perplexity, ensuring models only "say" what they retrieve, improving factual accuracy.
- π Post-training phases, including Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), are vital for making models controllable and product-ready.
Inspiration from Tech Visionaries
- π‘ Founders were inspired by Larry Page and Sergey Brin's PageRank insight (link structure for ranking) and obsession with low latency.
- π₯ Jeff Bezos's principles of clarity of thought and "your margin is my opportunity" influenced their business strategy.
- β¨ Yann LeCun's early advocacy for self-supervised learning as the "cake" of intelligence, rather than just RL, proved prescient for modern LLMs.
The Future of Knowledge Discovery
- π The internet is evolving beyond simple search to knowledge discovery, where tools guide users to understand complex topics and foster curiosity.
- β Perplexity aims to make people smarter and more truth-seeking by providing accessible, unbiased information, potentially breaking echo chambers.
- π The ultimate goal is to empower humans with AI tools that facilitate lifelong learning and deeper understanding, leading to a more fulfilling and knowledgeable society.
Knowledge graph40 entities Β· 41 connections
How they connect
An interactive map of every person, idea, and reference from this conversation. Hover to trace connections, click to explore.
Hover Β· drag to explore
40 entities
Chapters14 moments
Key Moments
Transcript665 segments
Full Transcript
Topics15 themes
Whatβs Discussed
PerplexityAnswer EnginesLarge Language Models (LLMs)Retrieval Augmented Generation (RAG)Factual GroundingChatbot HallucinationsGoogle AdWordsSearch Engine Optimization (SEO)PageRankTransformer ArchitectureSelf-supervised LearningReinforcement Learning from Human Feedback (RLHF)Inference ComputeTail LatencyKnowledge Discovery
Smart Objects40 Β· 41 links
ProductsΒ· 5
PeopleΒ· 13
CompaniesΒ· 7
MediasΒ· 2
ConceptsΒ· 13