How Perplexity CEO Aravind Srinivas is Building the Future
The evolution of generative AI and Large Language Models (LLMs) is fundamentally changing how we search for information. While Google still holds an overwhelming share of the search market, startups are emerging that redefine the user experience with new approaches. One of these is Perplexity. Co-founder and CEO Aravind Srinivas, having spent time at the forefront of AI research, realized the potential of combining search with product development and set his sights on building a "more intelligent Google." This article explains his research background, the company's founding story, and the vision of "next-generation search" that Perplexity is tackling, illustrated with specific episodes and quotes.
1. Aravind's AI Research and the Path to Entrepreneurship
1-1. From Berkeley to OpenAI
After completing his undergraduate studies in India, Aravind moved to the United States with a passion for research and enrolled in a doctoral program at the University of California, Berkeley. At the time, "Reinforcement Learning (RL)" was gaining attention following the shock of AlphaGo, but Aravind intuited that "research into generative models is essential if we are to aim for more general-purpose artificial intelligence," and shifted his focus to the field of generative AI.
He later interned at OpenAI, where he deepened his interactions with Ilya Sutskever and other leading researchers at the company. During their first meeting, although Ilya made stern comments like "anything other than reinforcement learning is not very meaningful," he ultimately shared the perspective that "the essence lies in how to combine large-scale pre-training (unsupervised/generative) with RL." Furthermore, his encounter with the "GPT prototypes" at OpenAI at the time significantly changed Aravind's research themes and vision for the future.
1-2. Interest in Search and the "Next Google"
During his internship at Google's headquarters, he read books about the founding story of Google, such as "In the Plex." His admiration for the scale and impact of the search business, and for a model that "advances AI while handling information and is used by the masses as a product," only grew stronger. Furthermore, inspired by a blog post titled "How to build the next Google" by then-Y Combinator (YC) partner Daniel Gross, he began to think, "If I can redefine search with generative AI (LLM), couldn't I build a business as big as Google?"
2. Perplexity's Prototype and Trial and Error
2-1. The First Demo: Twitter Search
Thinking he would explore new possibilities in "Search x Generative AI," Aravind first tried the idea of enterprise data search. However, dealing with actual data from sources like Crunchbase or Pitchbook faced high barriers regarding licenses and data provision, making progress difficult. That's when he turned his attention to "Twitter". At the time, API usage restrictions were not as strict as they are now, so he developed a "Twitter search demo" in just about a month where tweets were turned into a database, searched via SQL, and the results were summarized by a generative model in a chat UI.
Users could ask questions in natural language, such as "Which accounts did person X recently unfollow?", and the AI would create and execute SQL queries in the background to display a summarized result. This was groundbreaking, but he also keenly felt that it was inefficient to deploy generally because of the constraint that "data to be searched must be individually converted into tables."
2-2. The Victory of a "Simple Approach"
In the process of such trial and error, Aravind realized that "clever indexing and structuring are too difficult to develop, but thanks to the performance improvements of large language models, even 'simple methods' are reaching a level of accuracy that is sufficient for use." Specifically, this was the method of "passing the top K links from a Web search API directly to an LLM and returning a 'summary with references to the original articles'." Previously, the model's performance was insufficient, making it slow and unstable, but the arrival of GPT-3.5 and others brought significant improvements, and the feature of "attaching sources to answers" made a vivid impact.
3. Early Breakthroughs and the Sense of Crisis Regarding Competitors
3-1. Unexpected Viral Spread
When the initial version of Perplexity was tweeted and released, the first thing many users tried was "searching for their own names." The LLM would then instantly summarize their online activity from childhood to the present, leading many to be surprised, saying, "Even posts from that long ago?" Also, some answers contained misinformation, and complaints like "I was declared dead even though I'm still alive" went viral on social media.
While it might seem negative at first glance, this attracted a lot of attention and a flood of continuous feedback. Furthermore, when they implemented a "chat-type UI that allows for follow-up questions" in the same context, the time spent on the site and the number of questions per visitor doubled, which he says gave him the conviction that "the experience users are looking for is exactly here."
3-2. The Rise of Bing Chat and Bard
However, while fundraising with investors was underway, major competitors like Microsoft's Bing Chat and Google's Bard announced similar products one after another. Bing, in particular, adopted OpenAI's technology and made a high-impact move by embedding generative AI directly into its search browser. Aravind himself seemed to feel anxious, thinking, "My funding might be canceled," or "Won't I be unable to beat the big players at this rate?"
However, investors judged that "large companies have their own constraints. Companies like Perplexity that can prioritize 'truly rebuilding search' have a chance to win." As a result, Perplexity grew to a valuation of $1 billion in just a few years.
4. The belief that "the user is never wrong"
4-1. Inheriting Larry Page's philosophy
Aravind's product philosophy is deeply rooted in the famous words of Google founder Larry Page: "The user is never wrong." For example, he argues that even if a user makes an input error or their intent is ambiguous, the stance should not be to "prompt the user to provide the correct input," but rather for the system to "infer the intent and ask clarifying questions if necessary."
This culture of "user-centricity" permeates every aspect of engineering and design, and members of Perplexity all view even minor bug reports or complaints as valuable signals provided by the user.
4-2. Balancing "usability" and "accuracy"
On the other hand, "immediacy" is also extremely important for a search system. If the system performs overly complex reasoning or reads through massive amounts of documents, the response time slows down, causing users to leave. Perplexity actively works to "eliminate waste and increase speed even with the same methods" as a compromise, and is also considering optimizations such as selecting and switching models in stages.
5. The business model and future of next-generation search
5-1. The dilemma of the advertising model
The current search business is dominated by the "link clicks + ad revenue" model established by Google, but once "summarized answers that eliminate the need for users to click links" are realized, the advertising business will face a major turning point. As Perplexity pursues a design where users can get direct answers and complete their tasks, the challenge of "where to generate revenue" inevitably arises.
Aravind suggests that "the nature of advertising also needs to be redesigned," and is exploring new revenue opportunities, such as creating pathways that allow users to complete purchases, reservations, and service usage directly within the app.
5-2. The integrated platform Perplexity aims for
Ultimately, Aravind emphasizes that the ideal is "orchestration," where search, reservations, purchases, and various actions can be completed end-to-end. Even if a user asks an ambiguous question, an agent-based AI understands "what they really want to know," and if necessary, "asks follow-up questions and guides them to the appropriate service." This is a business model where revenue is generated naturally as a result of supporting the actions the user desires.
While major players like Google and Microsoft are moving to target this area, it is a fact that they face dilemmas due to their massive existing businesses and internal structures. That is precisely why Aravind expresses confidence that "startups that are thoroughly focused on user experience can release innovative services more flexibly and quickly."
Perplexity, led by Aravind Srinivas, is taking on the established "search engine" market head-on and attempting to pioneer a new information experience utilizing large language models. Behind this is the motivation to simultaneously conduct "advanced AI research and integrated product development," as well as a thorough commitment to "user-centricity," which aligns with the philosophy of Google's founders.
It is still in the development stage, and there are many hurdles to clear, such as search speed, accuracy, and the advertising revenue model. However, the vision of building an "end-to-end information platform" by leveraging the vast amount of data and feedback gathered through user interactions has the potential to not only "replace the existing Google" but to fundamentally change our style of information seeking.
Related Articles