Aravind Srinivas (Perplexity) and David Singleton (Stripe) fireside chat
TLDRIn a fireside chat, Aravind Srinivas, CEO of Perplexity AI, discusses the journey and evolution of his AI-powered search engine company. Starting with a focus on natural language to SQL, Perplexity has shifted towards a search tool that leverages large language models (LLMs) to process and structure data, allowing for faster and more efficient searches. The company gained traction by creating a demo using Twitter data, which attracted investors like Jeff Dean. Perplexity's strategy has been influenced by the growth and capabilities of LLMs, leading to a more conversational search experience that retains context for follow-up queries. The company has seen significant growth through word of mouth and aims to continue this trajectory by focusing on product market fit and engineering excellence. Srinivas also touches on the challenges of data collection, the potential for changing content creation, and the future of advertising in the context of AI-driven search.
Takeaways
- 🚀 **Innovation Journey**: Aravind Srinivas founded Perplexity AI to solve a specific problem—translating natural language to SQL, inspired by the success of Google's search engine model.
- 🌐 **Data Scraping Strategy**: Initially, Perplexity gained traction by scraping Twitter data to build a demo, which attracted investors and showcased the potential of their AI technology.
- 📈 **Product Market Fit**: Perplexity's product spread through word of mouth, indicating strong product market fit, with users finding the experience fast and engaging.
- 🔍 **Search Engine Evolution**: The company transitioned from a simple wrapper to building its own index and models, focusing on reducing latency and improving the user experience.
- 🤖 **Conversational AI**: Perplexity introduced a conversational element, allowing users to ask follow-up questions while retaining context from past queries.
- 📊 **User Insights**: User feedback led to the development of features like 'collections', showing the importance of direct interaction with users in product development.
- 🤝 **Partnerships**: A strategic partnership with Arc browser as the default search engine expanded Perplexity's reach, driven by user demand and shared investor interests.
- 💡 **Future of Search**: Aravind envisions a future where the value of traditional search engines diminishes as users prefer quick answers, potentially shifting the search paradigm.
- 📉 **Content Creation Impact**: There's hope that Perplexity's model of citing sources will encourage higher quality content creation, as producers strive to be recognized by AI.
- 🛡️ **Bias and Attribution**: Perplexity always provides citations, aiming to avoid biases and ensure fair use, though challenges with data scraping from certain sources are anticipated.
- ⏳ **Growth Aspirations**: The goal for the upcoming year is to increase the user base and query volume tenfold, continuing the rapid growth trajectory of Perplexity AI.
Q & A
What motivated Aravind Srinivas to start Perplexity AI?
-Aravind Srinivas started Perplexity AI to focus on a specific problem: building a great natural language to SQL interface. They were inspired by search engines and the journey of Google, aiming to approach the SQL problem as a tool for searching over databases rather than a coding copilot.
How did Perplexity AI initially gain traction and investors?
-Perplexity AI gained traction by scraping all of Twitter, organizing it into tables, and powering search over that data. This demo impressed investors, including Jeff Dean, and helped secure funding.
What was the strategic shift in Perplexity's approach as language models evolved?
-As language models like GPT-3.5 and later versions became more capable, Perplexity shifted its strategy to do less offline pre-processing and more post-processing at inference time, leveraging the increasing capabilities of LLMs.
How does Perplexity AI ensure a fast and snappy user experience?
-Perplexity AI ensures speed by building its own index, serving its own models, and orchestrating these elements together efficiently. It also focuses on minimizing tail latencies and improving perceived latency through UX innovations.
What is the hiring process like at Perplexity AI?
-Perplexity AI's hiring process involves a trial phase where candidates work on real tasks for several days. This approach helps assess their capabilities and cultural fit, and it has proven effective for the company's initial hires.
How does Perplexity AI's product development process work?
-The company has moved from a phase of experimentation to one of exploitation, with a clear roadmap. It organizes work into small projects with specific timelines, and team members may work on multiple projects simultaneously.
What is the Collections feature in Perplexity AI?
-The Collections feature allows users to organize their search threads into folders, making it easier to manage and revisit their work. This feature was inspired by user feedback from pro users.
How did the partnership with the Arc browser come about?
-The partnership was initiated by user demand. Users of both Perplexity and Arc browser reached out to the respective companies expressing interest in integration. The CEOs of both companies, who shared investors, decided to collaborate for mutual user benefit.
Does Perplexity AI have plans to replace traditional search engines?
-Perplexity AI does not aim to completely replace traditional search engines. Instead, it sees itself as offering a different approach on a spectrum of search functionalities, focusing more on providing direct answers rather than just navigating to links.
How does Perplexity AI handle user link clicks to refine its search rankings?
-Perplexity AI collects data on user link clicks and uses these signals to train its ranking models. The company believes that it doesn't need billions of data points to create effective ranking models, especially with the advancements in unsupervised generative pre-training.
What are Aravind Srinivas's thoughts on the future of advertising in the context of AI-driven search?
-Aravind Srinivas believes that there will be a new, more effective way to integrate advertising in AI-driven search interfaces. He suggests that ads could become more targeted and personalized, potentially blending in more naturally with the search results, but it's an area that still needs to be explored and developed.
Outlines
🎉 Introduction and Perplexity's Beginnings
The video begins with an enthusiastic introduction to Aravind Srinivas, CEO of Perplexity AI, and an overview of the company's journey. Aravind explains that Perplexity was not initially intended to be a new search engine but started with a focus on solving the natural language to SQL problem. The conversation also touches on the company's early inspirations from search engines and their academic roots, leading to the creation of a prototype tool for Stripe analytics. Despite initial excitement, they struggled to gain traction with real-world usage, leading to a pivot towards scraping Twitter data to build a compelling demo.
🚀 Perplexity's Evolution and Product Market Fit
Aravind discusses the evolution of Perplexity, moving from a SQL Pro solution to a search over databases tool. The company's strategy shifted with advancements in Large Language Models (LLMs), allowing for less pre-processing and more reliance on the models for post-processing. This change in approach led to a generic search feature that summarized content from links, attracting the attention of notable investors like Jeff Dean. Perplexity's usage sustained over time, prompting the team to further develop the product, including adding a conversational feature to retain context from past queries.
🤖 Hiring Process and Internal Operations
The conversation shifts to Perplexity's internal operations, including its hiring process. Aravind explains that the first 10 to 20 hires went through a trial process to ensure they were a good fit for the company culture and work style. This approach proved successful in identifying strong candidates quickly. The company's weekly operations are structured around small projects with clear timelines, and communication is maintained through regular meetings and stand-ups, inspired by Stripe's operational model.
🔍 User Insights and Perplexity's Features
Aravind shares insights gained from user feedback, such as the introduction of the 'collections' feature, which allows users to organize their search threads into folders. This feature was not directly related to search quality but was valuable for user organization. He also discusses a partnership with the Arc browser, making Perplexity the default search engine, which was a result of user demand and a shared investor connection.
🌐 Content Generation and the Future of Search
The discussion explores the impact of search engines on content generation and how Perplexity might influence this dynamic. Aravind hopes that Perplexity will encourage the creation of better content by prioritizing quality sources in its citations. He also addresses the challenges of data collection and the need for attribution, anticipating difficulties as the company grows but emphasizing the importance of fair use and citation.
📈 Monetization Strategies and AI Industry Insights
Aravind talks about the decision to monetize Perplexity early in its lifecycle, driven by the need to validate product-market fit and ensure the product's value beyond being a free service. He believes that monetizing early provides leverage and sustainability, allowing for future growth and investment. He also provides feedback to Stripe on improving fraud detection and customization options for referral programs and discusses the potential for more integrated and relevant advertising in the future.
📊 Addressing Bias and Advertising Challenges
The conversation concludes with questions from the audience about avoiding biases in Perplexity's answers and the potential for content creators to manipulate search results through 'prompt injection.' Aravind acknowledges these challenges and emphasizes the importance of prioritizing domains with established content review processes. He also addresses the issue of advertisements blending in with search results, suggesting transparency as a key to maintaining user trust.
🎯 Future Goals and Closing Remarks
Looking ahead, Aravind sets an ambitious goal to increase Perplexity's monthly active users and queries tenfold. The discussion ends on an optimistic note, with the host expressing support and enthusiasm for Perplexity's future growth.
Mindmap
Keywords
💡Perplexity AI
💡Natural Language to SQL
💡Bird-SQL
💡LLMs (Large Language Models)
💡Product-Market Fit
💡GPT-3.5 and Bing
💡Conversational Search
💡Word of Mouth
💡Enterprise Search
💡Tail Latencies
💡Collections Feature
Highlights
Aravind Srinivas, CEO of Perplexity AI, discusses the company's journey and its AI-powered search engine.
Perplexity was inspired by Google's success and focuses on solving the natural language to SQL problem.
The company initially built a tool for analytics over Stripe data, which led to insights on user behavior and data handling.
Perplexity gained traction by scraping Twitter data to create a demo that attracted investors like Jeff Dean.
The strategy evolved to leverage large language models (LLMs) for more work during the post-processing at inference time.
Perplexity's search tool provides summaries in the form of citations, differentiating it from traditional search engines.
The company has experienced strong product-market fit, growing largely through word of mouth.
Perplexity's fast response times are achieved through parallel processing and minimizing tail latencies.
The company culture values engineering excellence and focuses on improving latency and user experience.
Perplexity's hiring process involves a trial phase to assess candidates' real-world work capabilities.
The company has shifted from experimentation to a more exploitation-focused approach with a clear roadmap.
User feedback has driven the development of features like 'collections' for better organization of search threads.
Perplexity partnered with the Arc browser to become the default search engine, a decision driven by user demand.
Aravind envisions a future where the value of traditional search engines decreases as users seek quicker answers.
Perplexity uses link clicks to train ranking models, relying on less data due to advancements in unsupervised generative pre-training.
The company operates on a subscription model, which provides leverage and sustainability in business growth.
Aravind suggests that the next wave of consumer application startups may utilize open-source models fine-tuned for specific needs.
Perplexity aims to avoid biases by summarizing from multiple sources and maintaining a focus on truth and helpfulness.
The company is transparent about its monetization strategies and is open to exploring new advertising models that blend naturally with search results.
Aravind predicts that content creation may evolve with the influence of AI-driven search, potentially leading to higher quality content.
Perplexity plans to grow its user base and query volume by tenfold in the coming year.