About this Video

• Video Title: How To Build The Future: Aravind Srinivas
• Channel: Y Combinator
• Speakers: Aravind Srinivas, David (Host)
• Duration: 00:34:39

Introduction

This Y Combinator interview features Aravind Srinivas, co-founder and CEO of Perplexity AI. The conversation covers 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 challenges.

Key Takeaways

• Perplexity's Origin and Evolution: Initially focused on enterprise search, Perplexity pivoted to a consumer-focused, AI-powered search engine after realizing the potential of its ability to answer follow-up questions and provide source citations.
• The "User is Never Wrong" Philosophy: Aravind emphasizes the importance of prioritizing user experience and iterating based on user feedback, even if it means revisiting design choices. This approach is inspired by Google's early days.
• Competitive Advantage: Perplexity's unique approach—using a simpler, less resource-intensive method—gives it a potential edge against competitors like Google and Microsoft, particularly regarding speed and scalability. This approach was initially seen as less sophisticated but became advantageous with improvements in large language models.
• Future of Search: Aravind envisions a future where search engines provide not only answers but also directly facilitate actions, like purchasing products or booking travel. However, this requires solving significant challenges in user experience, monetization, and integration with various services.
• Team Management: Aravind emphasizes a data-driven approach to team management, prioritizing metrics like daily queries and fostering open communication and a culture of brutal honesty in identifying and fixing bugs.

This Y Combinator interview features Aravind Srinivas, co-founder and CEO of Perplexity AI, discussing the company's journey, challenges, and vision for the future of search. Initially focused on enterprise solutions, Perplexity pivoted to a consumer-focused AI-powered search engine after discovering the high engagement generated by its ability to answer follow-up questions and cite sources. Aravind emphasizes a user-centric design philosophy, inspired by Google's early days, where the "user is never wrong." This, combined with a simpler, more scalable technical approach, provides a potential competitive edge against larger players like Google and Microsoft. He envisions a future where Perplexity not only provides answers but also facilitates actions, a significant challenge requiring innovative monetization strategies. Team management is data-driven, with a focus on daily query volume and open communication, even encouraging "brutal honesty" to improve the product.

Traduzione in italiano:

Questa intervista di Y Combinator presenta Aravind Srinivas, co-fondatore e CEO di Perplexity AI, che discute il percorso dell'azienda, le sfide affrontate e la visione per il futuro della ricerca. Inizialmente focalizzata su soluzioni per le imprese, Perplexity ha virato verso un motore di ricerca AI incentrato sul consumatore dopo aver scoperto l'alto coinvolgimento generato dalla sua capacità di rispondere a domande di follow-up e citare le fonti. Aravind sottolinea una filosofia di progettazione incentrata sull'utente, ispirata ai primi giorni di Google, dove "l'utente non sbaglia mai". Questo, combinato con un approccio tecnico più semplice e scalabile, fornisce un potenziale vantaggio competitivo rispetto a player più grandi come Google e Microsoft. Prevede un futuro in cui Perplexity non solo fornisce risposte ma facilita anche le azioni, una sfida significativa che richiede strategie di monetizzazione innovative. La gestione del team è orientata ai dati, con un focus sul volume giornaliero di query e una comunicazione aperta, incoraggiando persino "l'onestà brutale" per migliorare il prodotto.

The transcript does not provide the exact content of the Polyram tweet Aravind mentions. It only states that he read a tweet from Polyram describing a situation where attempting to solve a harder version of a problem resulted in a simpler, more general, and scalable solution. This insight led Aravind and his team to realize that their approach of using unstructured data and relying on the LLM to handle most of the work at query time was superior to their earlier, more structured approach. This realization was a key factor in Perplexity's shift towards its current, more general and scalable design.

The transcript doesn't offer specific examples of "boring work" Aravind believes Perplexity needs to do. However, he implies that it involves the same type of work Google undertook to create Google Finance, Google Shopping, and Google Flights – integrating with various services and creating specific, verticalized features for different tasks and industries (like shopping, travel, etc.). He suggests Perplexity needs to solve many "boring" problems related to integrating with merchants, hotels, booking systems, handling cancellations, and other logistical challenges needed for a seamless user experience in those domains. The focus is on the less glamorous, but crucial, aspects of building a truly useful and widely adopted product beyond providing simple informational answers.

The transcript mentions Aravind's inspiration from a blog post about building the next Google, which suggested using "suffixes" or special strings to refine search queries.

Here are the examples given in the transcript:

• If someone is looking for reviews of a movie, suffixing the query with site:rottentomatoes.com.
• If someone is looking for reviews of a new gadget, suffixing the query with site:that_corresponding_subreddit.

Aravind notes that these suffixes and special strings were used to filter results and make Google better, even without considering the ads or more sophisticated ranking. He also mentions that LLMs could automatically figure out these suffixes.

Aravind encountered challenges in securing data from companies like Crunchbase because "they just don't want to give it to us."

To overcome this, he initially tried to hustle for calls with companies like Bitbook or Crunchbase to build a demo that would make sense to investors and allow them to raise capital and hire people. When direct data acquisition proved difficult, they pivoted to using publicly available data from Twitter through its academic access program. They organized this data into tables and used OpenAI's Codex models (pre-GPT 3.5) with templates and SQL generation to query the data.