Written by: Techub News Compilation

Introduction

In the latest episode of the popular Y Combinator interview series "How To Build The Future", host David spoke with Aravind Srinivas, co-founder and CEO of the AI search newcomer Perplexity. In less than three years, Perplexity has grown from zero to a star company valued at over $9 billion, and its "conversational answer engine" model is seen as one of the most promising challenges to traditional search engines. This conversation delved deeply into Aravind Srinivas's journey from AI researcher to entrepreneur, revealing the key decisions behind the iteration of Perplexity’s products, and looking ahead at how a startup can find its path of survival and growth in the AI search arena surrounded by giants, thanks to its unique product philosophy.

Summary

• The founding of Perplexity stemmed from an insight into "AI completeness": only in fields like search and autonomous driving, where product iterations can directly feed back improvements in underlying AI capabilities, is it suitable to conduct AI research and product building simultaneously.
• The key turning point in product success came from abandoning the "smart" approach of building complex indexes for verticals, opting instead for a "dumb" method relying on the capabilities of large models for real-time processing, which unexpectedly solved delays and generality issues.
• The core advantage in competing with Google is not technology, but an extreme focus on user experience and a product design philosophy of "the user is never wrong", which is difficult to implement in large companies due to organizational structure and business models.
• Perplexity's long-term vision is to build an end-to-end intelligent information experience that understands user intent, integrates answers and actions (such as shopping and booking), which requires complex AI routing and orchestration capabilities.
• The company's agile culture faces challenges, but by focusing on details, data-driven metric management (like daily query volume), and flat communication, the founders combat entropy.

From Researcher to Entrepreneur: The Inspiration of AI Completeness

Aravind Srinivas's journey in AI began with an interest in deep learning during his undergraduate studies in India, prompting him to pursue a PhD at the University of California, Berkeley. A turning point in his life occurred during his internship at OpenAI when he met co-founder and chief scientist Ilya Sutskever. He prepared various "fancy" research ideas, but after listening for five minutes, Ilya Sutskever bluntly stated that "none of this matters," and drew two circles: a large circle representing unsupervised learning and a smaller one inside representing reinforcement learning. Ilya Sutskever told him, "This is AGI (Artificial General Intelligence), other research doesn't matter." At the time, OpenAI was building the predecessor to GPT-1. This conversation made Aravind Srinivas realize that his focus on reinforcement learning (popularized by AlphaGo) might have been following a trend, and he quickly shifted to researching unsupervised and generative models.

During his internship at Google, he was profoundly impacted by a book detailing Google's early history, "In the Plex." He realized that combining cutting-edge AI research with solid product building is extremely challenging. After discussing with Ilya Sutskever, he concluded that there may only be two types of problems in the world that can simultaneously satisfy AI research and product building—search and autonomous vehicles. Because in these two fields, every deployment of a product and user interaction becomes a data point for improving underlying AI models, forming a "better product -> more users -> more data -> better AI -> better product" flywheel. More importantly, these problems are on the path of "AI completeness", meaning that advancements in AI capabilities would directly make the company's products stronger, rather than being undermined by more advanced AI.

Starting Point: Born from a Vague Ambition to "Disrupt Google"

With the idea of starting a business, how to gain the initial energy? Aravind Srinivas mentioned that he was inspired by a blog titled "How to Build the Next Google" by former YC partner Daniel Gross. The core point of the blog was that through better query reconstruction (such as automatically adding the "site:rottentomatoes.com" suffix to movie review queries), even utilizing existing Google rankings can significantly enhance the search experience. He realized that large language models might automatically accomplish this complex query reconstruction, serving as the foundation for building a new generation of search engines.

At the same time, he met future co-founder and CTO Denis Yarats, who coincidentally published the same paper on a similar day. The two often discussed how to build AI agents that could control Android environments. However, when he mentioned the idea of "disrupting Google" to potential investors, the first feedback he often received was: "Isn't Google going to do this themselves?" After all, search was Google's "crown jewel," unlike document processing, which was a secondary business. Aravind Srinivas acknowledged that at that time, they were more drawn to a grand idea with scale and ambition, rather than purely motivated by a desire to "kill Google." They also realized that on the technological trajectory of the emerging multimodal models, there was an opportunity to build something remarkable.

On the advice of investors, they initially did not directly challenge general search but focused on vertical and enterprise search, such as querying specific databases or data from Crunchbase. However, they found that no companies were willing to easily hand over data. So, they turned to Twitter (now X), which still allowed academic access at the time. They utilized OpenAI’s Codex model (pre-dating GPT-3.5) to build a chat interface that could convert natural language queries into SQL and retrieve and visualize information from organized Twitter datasets. This demo was completed by a small team of three within a month and gained early attention due to its unprecedented interactive approach and the fun of "social search" (such as viewing whom someone followed or unfollowed that week).

Key Turning Point: Embracing the "Dumb" Approach and Betting on Model Progress

As they attempted to build similar vertical search tools for more data sources like GitHub and LinkedIn, the team encountered significant resistance to data acquisition and productization. At this point, they realized a more fundamental solution: rather than laboriously building structured indexes (like tables) for each domain, it was better to keep the data unstructured and delegate all the heavy lifting to the large model during queries (reasoning). This seemed like a more "dumb" and model-dependent approach, but it was more universal, scalable, and better at countering Google's established massive traditional indexing system.

They were inspired by an early project called "WebGPT" from OpenAI researcher John Schulman's team, but WebGPT used a 175 billion parameter model and operated in an agent-like manner to click and scroll through web pages, which was very slow. The Perplexity team took the opposite approach, using a simpler heuristic method: always retrieving top-ranked links from the search API and only using cached summary snippets from the search engine (without clicking or scrolling), then feeding all this information to the large model at once, asking it to generate summaries with academic-style citations.

The success of this strategy heavily relied on the enhanced instruction-following capabilities of the model. Aravind Srinivas admitted that if they had attempted this “dumb” method a year earlier, it would not have worked due to insufficient model capabilities, and people would have concluded that a more "intelligent" architecture was needed. But with the emergence of models like GPT-3.5 Turbo, instruction-following capabilities were greatly enhanced, making the simple approach feasible and addressing the core user experience issue of delays. Even so, the initial version of the answer generation still took about 7 seconds, and they could not control the verbosity of the answers, leading them to hard-limit response lengths through prompting to speed things up.

Product Validation and Competitive Awakening: Finding Space in the Shadow of Giants

Perplexity's first "breakout" moment was quite dramatic. After the product was launched, a scholar searched for their name, resulting in the model generating a biographical summary in the past tense, due to the existence of a deceased person with the same name. This scholar tweeted a "complaint" about still being alive, unexpectedly bringing a lot of attention to Perplexity. People began to enthusiastically search for themselves, with the model compiling their online historical activities and generating interesting summaries that users were eager to screenshot and share. Subsequently, the team launched a follow-up question feature, which exponentially increased user engagement time and daily question volume on the site, leading Aravind Srinivas to be confident that the product had found a market fit and should not pivot to enterprise services.

The realization of the severe competition with giants truly hit during a critical funding moment. Just a few days after reaching an investment agreement with a venture capital firm, a screenshot of Microsoft Bing Chat's interface was leaked. Shortly afterwards, another interested investor subtly extended their due diligence period from 30 days to 45 days. A week later, Google CEO Sundar Pichai announced the launch of Bard. Suddenly, a storm was brewing. Aravind Srinivas even considered selling the company. However, the initial investor who shook hands with them stood by their commitment and did not require them to pivot, which gave the team confidence.

They calmly analyzed the competitive landscape: Microsoft has long been inept in consumer products, and Google faced its own challenges—unable to easily integrate an experience similar to Perplexity directly onto its existing homepage filled with ads and various informational modules for billions of users. This left room for independent players to survive. Aravind Srinivas believes that the difference in experience between using Perplexity and Google for pure information queries is akin to the difference between "healthy meals" and "fast food."

Product Philosophy: The User is Never Wrong, Details Determine Success

Aravind Srinivas was deeply influenced by Google's early culture, particularly the philosophy of co-founder Larry Page. He repeatedly emphasized the principle that "the user is never wrong" to the team. He gave an example that when testing new features led to failures due to vague queries, the correct product mentality is not to blame users for unclear questions, but to have the AI proactively clarify and inquire about the user's true intent. This is in stark contrast to the approach of many enterprise software that "educates users to be better prompt engineers." Magical consumer products belong to the former category, as they take on the complexity of understanding users.

This obsession with detail permeates every aspect of the product: ensuring the search box cursor is always ready (without needing a mouse click), pursuing extreme loading speeds. The company's core north star metric is "daily query volume," which aligns with the early focus of Google. In weekly all-hands meetings, the first metric they look at is the weekly and monthly growth trends of this number. If there’s a decline, the team will analyze the reasons deeply. Aravind Srinivas believes that users asking repeated questions is not due to the product failing to meet their needs (which would lead to users leaving), but rather because they discovered more questions they wanted to explore within a session, which reflects the product's attraction.

In terms of company culture, he strives to create a flat, work-focused atmosphere. He personally raises numerous product bugs daily and communicates directly with the responsible engineers without hierarchical barriers. He believes that if the founding team does not love and scrutinize their own product, users will not be satisfied either. He mainly communicates with users through X (formerly Twitter), as feedback there is "brutal and honest," revealing the most serious problems.

Growth Challenges and Future Vision: Building an End-to-End Intelligent Information Body

As the team expanded, Perplexity faced all the typical challenges of company growth: slowing down. Reasons include trust issues from production environment problems, new engineers needing time to familiarize themselves with the codebase, and the necessity to implement stricter testing and A/B testing processes. Aravind Srinivas admits that maintaining the early culture of detail obsession has become difficult as the workforce expands, but the founding team still strives to combat "entropy."

When discussing the future, he sketches a long-term vision for Perplexity. Currently, Perplexity is a smarter version of Google search in specific scenarios. But in three to four years, he hopes it will become an end-to-end platform capable of understanding user questions, providing the best answers, and helping users complete subsequent actions. For example, searching for "the best sweater" would not only yield recommendations but also allow users to purchase directly; searching for hotels would not only compare options but also facilitate bookings.

The immense challenge within this is the business model and user perception. If it only provides answers while directing the transaction link to Google, then monetization credit would still belong to Google. Perplexity might only gain a small amount of subscription revenue and could easily be squeezed by better-funded competitors offering free services. To close the loop, it would require building an incredibly complex AI routing and orchestration system capable of intelligently determining when to use small models, knowledge graphs, plugins, and when to generate streaming answers or multi-step reasoning results. Users do not care about the underlying tech stack; they only want accurate, fast, and complete solutions. He believes that Google's existing system is the closest to this vision in the world, and that the next generation of systems must be built, but it requires a commitment of ten years or even longer.

In analyzing competitors for the next decade, Aravind Srinivas believes that Perplexity's core advantage lies in its obsession with users and its excellent product taste. While Google has product sensibility and all the resources, its fundamental dilemma lies in being both a search company and an advertising company, where search somewhat serves its advertising business. The search business, with quarterly revenues soaring into the tens of billions, is the pillar of Google's high profits and stock prices. Any changes that could destabilize this income source (for example, providing users with direct answers without requiring clicks on ad links) would face substantial internal resistance. In contrast, companies like OpenAI and Anthropic focus more on breakthroughs in the capabilities of the models themselves, rather than constructing a refined end-to-end user experience. Perplexity possesses both an understanding of AI technology (being able to utilize and fine-tune the latest open-source models) and a user-centric product-building DNA, which could be key to securing a place in the long competitive journey.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。