Aravind Srinivas - Building An Answer Engine - [Invest Like the Best, EP.363]

At a Glance

Perplexity founder and CEO Aravind Srinivas explains how his company is building an AI-driven "answer engine" to challenge Google's dominance in search. Aravind's core thesis is that the future of search lies in moving from "10 blue links" to delivering direct answers, and the truly disruptive experience ("insanely great") is having AI understand and actively guide users to ask questions, democratizing access to knowledge.

The Ultimate Form of Search: From Answer Engine to Question Engine

Aravind Srinivas believes that the traditional "10 blue links" were a "hack" of the search era, and the current "answer engine" is only the first step, with the ultimate goal being a "question engine."

• Current Stage: Answer Engine. Perplexity's core mechanism: after a user enters a question, the system first rewrites and expands it, then retrieves relevant web snippets from its own index, and finally uses a large language model (LLM) to generate a concise answer with precise citations. Aravind notes that 90% of users do not click on source links, indicating the immense value of delivering direct answers. Its ideal state is a hybrid of "research assistant" and "navigational search": the user first completes research and decision-making via AI, then goes to a specific website (e.g., Expedia) to execute the action.
• Ultimate Vision: Question Engine. Citing Jobs's philosophy of "bringing joy to the common man," Aravind believes that the future AI should not force users to become "prompt engineers." Instead, AI should be able to infer the user's true intent from their "clumsy questions" and even proactively ask interesting questions, making users smarter every day. This represents a fundamental shift from "passive answering" to "active guidance of knowledge exploration."

Competitive Strategy: Building Moat in the Cracks of Giants

Aravind clearly states that Perplexity's primary risk comes from OpenAI, not Google, because the latter is constrained by its advertising business model (counter-positioning).

• "Counter-Positioning" Advantage over Google. Aravind argues that Google's business model is rooted in the advertising revenue generated by "10 blue links," which gives it no incentive to disrupt itself and offer a more direct answer experience. Therefore, Perplexity aims to disrupt high-commercial-intent search categories such as shopping, travel, insurance, and law. These are core to Google's advertising revenue, but its AI product Gemini has no motivation to optimize them.
• "Execution Speed" Race against OpenAI. Aravind admits that the biggest concern is ChatGPT evolving toward search. Perplexity's response is to continuously focus on speed and accuracy, and build faster iteration capabilities. He quotes Bruce Lee ("I fear not the man who has practiced 10,000 kicks once, but I fear the man who has practiced one kick 10,000 times") to emphasize Perplexity's "laser-focused" strategy: it has rejected all features unrelated to its core search product, such as image generation, free-form chat, and GPT stores.

Technical Path: From "Wrapper" to "Fusion Architecture"

Perplexity's technical evolution has gone through a process from "validating the product" to "building a moat," with the core being the construction of a system that deeply integrates its own index, retrieval model, and reasoning model.

• Phase One: A "Wrapper" for Rapid Validation. With only $2 million in seed funding, Perplexity chose the fastest route: connecting to the Bing API and GPT-3.5 API to validate product demand. Any team could do this, but the real challenge began with the sustained growth phase after the product gained users.
• Phase Two: Building Proprietary "Infrastructure." When user volume surged, simply being a "wrapper" could not handle high concurrency, API rate limits, and third-party service outages. Perplexity began building its own index (web crawler, parser, ranking algorithm), inference infrastructure (running and fine-tuning open-source models like Llama and Mistral on its own GPUs), and an end-to-end orchestration layer to ensure low latency.
• Core Philosophy: Be an "Open-Book Exam" Top Student, Not a "Memory Master." Aravind likens Perplexity's architecture to an "open-book exam": you don't need a massive, omniscient model; you need a small model that is smart enough and good at reasoning, combined with a high-quality, real-time updated index to "look up" and extract information when needed. This stands in stark contrast to OpenAI's philosophy of "one model solving all problems."

Position Moves

Position	Guest's Attitude	Key Data
OpenAI	Primary risk (competitor, may enter search)	Its API was once used as underlying infrastructure; GPT-4 is considered the benchmark for reasoning ability, but its plugin store failed due to inability to handle complex API calls
Google	Constrained by business model, lower competitive risk	Its advertising business model gives it no incentive to disrupt its own search experience
Microsoft	Mentioned as a competitor	Its Bing API was once used by Perplexity as an underlying search source
Meta	Mentioned as a competitor for talent	A senior Meta researcher declined to join Perplexity, saying "come back when you have 10,000 H100s"
Anthropic	Mentioned as a technology benchmark	Its CEO Dario Amodei's physics background is considered an advantage for its team in data science and experimental methodology
Mistral	Mentioned as a technology partner/open-source model provider	Its Mixtral model is used by Perplexity as one of the base models for the reasoning layer
XAI (Elon Musk)	Mentioned as a potential competitor	Considered "has potential, but has not yet achieved anything significant"
Rabbit	Mentioned as an API customer	Its device is using Perplexity's online LLM API
Arc (Browser)	Mentioned as an API customer	Is using Perplexity's online LLM API

Investment Implications

1. The ultimate form of search is the "question engine." (Aravind Srinivas) Support: Future AI should not force users to become "prompt engineers." Instead, it should actively guide users to ask questions, even inferring their true intent from "clumsy questions," thereby democratizing knowledge acquisition.

2. The true moat for an AI startup is not technology, but execution speed. (Aravind Srinivas) Support: In the AI field, any technological advantage can be quickly caught up. The key is whether you can iterate products, optimize user experience, and build a data flywheel faster than competitors (especially OpenAI).

3. A search product that is "externally regulated" is more valuable than one that is "self-disciplined." (Aravind Srinivas) Support: Perplexity's "retrieval-augmented generation" (RAG) architecture makes it a top student in an "open-book exam." It doesn't need an omniscient giant model, but rather a smart reasoning model combined with a real-time updated index, thereby avoiding hallucinations and enabling honest responses of "I don't know."

4. Don't be divided by the "verticalization" narrative of AI investing. (Aravind Srinivas) Support: Investors often think one should invest in vertical SaaS, but Aravind holds the opposite view, citing Marc Andreessen's warning: "Don't try to make Perplexity a vertical product, or you're dead." He believes users expect a universal, natural language interface that can handle all problems, not an "AI chatbot" limited to a specific domain.

5. Execution is strategy; you need "muscle" before you can talk about "strategy." (Aravind Srinivas) Support: Citing Snowflake CEO Frank Slootman's view, founders should first build "execution muscle" through rapid iteration and product-market fit, rather than obsessing over "strategic planning" in the early stages. Perplexity's early failure with "TexSQL" demonstrates the importance of trial and error and iteration.

6. The value of "information retrieval" is more important than that of "information generation." (Aravind Srinivas) Support: Perplexity treats "hallucination" as a bug and is committed to combating it by training the model to "say it doesn't know" when information is insufficient. This is completely opposite to the popular view of "building products with hallucination as a feature," reflecting its adherence to the core value of "accuracy and reliability."

7. Don't try to beat Google on its own track; instead, race on a track where it cannot compete. (Aravind Srinivas) Support: Google's business model (advertising) prevents it from disrupting the "10 blue links" model it dominates, creating a "counter-positioning" opportunity for Perplexity to focus on disrupting high-commercial-intent search categories (e.g., travel, shopping), which are precisely Google's core profit sources.