About this Video

• Video Title: How To Build The Future: Aravind Srinivas
• Channel: Y Combinator
• Speakers: Aravind Srinivas, David (Interviewer)
• Duration: 34:39

Introduction

This Y Combinator interview features Aravind Srinivas, co-founder and CEO of Perplexity AI. The conversation explores Aravind's journey into AI, the development of Perplexity, its competitive advantages, and the future of search. The video aims to provide insights into building a successful AI company and its potential impact on the search landscape.

Key Takeaways

• Perplexity's Evolution: Initially focused on enterprise search, Perplexity pivoted to a consumer-facing product after realizing the potential of its approach using LLMs for improved search and follow-up questions.
• The "User is Never Wrong" Philosophy: Aravind emphasizes a user-centric design philosophy, prioritizing user experience and addressing user queries directly, even if it requires clarifying ambiguities. This contrasts with approaches that place the onus of effective query formulation on the user.
• Competitive Advantage: Perplexity's focus on user experience, the efficient use of LLMs, and its potential for end-to-end user experiences (answer plus action) position it as a potential competitor to Google in the future of search.
• Future of Search: Aravind envisions a future where search is not just about providing answers but also facilitating actions, such as direct purchasing or booking. However, this requires solving complex challenges related to monetization and user experience.
• Importance of Iteration: The video highlights Perplexity's iterative development process, starting with a Twitter-based search prototype and evolving into a more general-purpose search engine through continuous experimentation and learning.

1. What specific challenges did Perplexity face in its initial attempts to focus on enterprise search, and how did these challenges contribute to the company's pivot?

Perplexity initially struggled to gain traction with enterprise clients who were unwilling to share their data. This difficulty in securing necessary data for their enterprise-focused demo hindered progress. The inability to obtain data from companies like Crunchbase and Bitbook led them to use publicly available Twitter data instead, which unexpectedly proved to be a more effective path to product development and user engagement. This experience highlighted the potential of a consumer-facing product built on readily available data.

1. What specific examples from the transcript illustrate Aravind's "user is never wrong" philosophy in Perplexity's product design?

Aravind recounts an instance where a user received an inaccurate biography because the LLM confused her with a deceased person of the same name. Instead of blaming the user for an unclear query, Aravind viewed the error as an opportunity for improvement and highlighted the need for clarification and robust error handling to ensure user satisfaction and enhance the accuracy of responses. His emphasis on directly addressing user queries even when there is ambiguity demonstrates the core principle of the "user is never wrong" philosophy. He contrasts this with approaches that make the user responsible for perfect prompt engineering, emphasizing Perplexity's focus on making the product work seamlessly for the user.

1. What were the key technical innovations or design choices that enabled Perplexity to achieve its current efficiency in using LLMs for search, compared to earlier attempts like WebGPT?

Unlike the slow and inefficient WebGPT which used a 175B parameter model, Perplexity adopted a "dumb" heuristic approach. They used a much faster method: always selecting top-k links from a search API, utilizing cached summary snippets to avoid slow browsing and scrolling, and feeding all links into the prompt to eliminate the need for selection. This streamlined process, coupled with improvements in instruction following capabilities of newer LLMs (around 3.5 turbo models), significantly improved efficiency and speed. The initial version's 7-second response time, due to a lack of streaming answers and inability to control verbosity, highlighted the problem, emphasizing the importance of this faster, simpler design.

1. What are the primary monetization strategies Aravind considers for Perplexity's future, given the challenges of competing with established players like Google?

Aravind acknowledges the difficulty of monetizing a search engine directly, especially given the potential for competitors to undercut pricing. His vision focuses on providing an end-to-end user experience that goes beyond answering questions to include facilitating actions. He cites examples like integrating product cards with buy buttons for shopping queries. Although this approach presents challenges regarding user perception of ads versus integrated actions, he believes that this complete experience, coupled with efficient model orchestration, could be the key to establishing a new standard for search and monetization. The discussion implicitly suggests that this complete user journey and experience will be the basis for new revenue models, rather than relying solely on advertising.