#434 – Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet

Here is the English translation of the provided Chinese investment research notes.

At a Glance

Guest: Aravind Srinivas, CEO of Perplexity, former researcher at OpenAI, DeepMind, and Google AI, PhD from UC Berkeley.

Main Theme: An in-depth exploration of how Perplexity, through its "Retrieval-Augmented Generation (RAG)" architecture, combines search engines with Large Language Models (LLMs) to redefine knowledge discovery, along with an analysis of its competitive strategy against Google, business model, and the future of the AI industry.

Core Thesis: Aravind Srinivas argues that Perplexity's disruptive potential lies not in competing with Google on "10 blue links," but in completely rethinking the search UI—making the answer, rather than links, the primary element, and using a "citation" mechanism to fundamentally reduce hallucinations. This positions Perplexity as a "knowledge discovery engine" rather than a search engine.

Thematic Sections

1. The Path to Disrupting Google: Don't Play the Opponent's Game, Change the Rules

Aravind Srinivas explicitly states that Perplexity never attempted to beat Google on its own turf (i.e., providing 10 blue links). He believes Google has perfected link-based search, making any incremental improvement unlikely to shift its position. Perplexity's disruption lies in rethinking the entire user interface: placing the "answer" rather than "links" at the center of search results. This decision was strategic, aiming to leverage the exponential progress of AI technology (smarter, cheaper models; fresher indexes) to continuously reduce hallucinations, rather than trying to catch up on Google's track.

• Historical Context: Aravind points out that Google's business model (AdWords) is both its greatest strength and its greatest weakness. Any ad unit with a lower profit margin than a "link click" lacks the incentive for Google to actively develop. He cites the Amazon case: "Your margin is my opportunity." Google's high-margin advertising model makes it difficult to pivot towards a potentially lower-margin "answer engine" model.
• Mechanism Breakdown: Perplexity's "answer engine" model decouples advertising from "link clicks." This allows Perplexity to explore entirely new advertising formats, such as subscription-based (Pro) or ads that are highly integrated with the user experience without disrupting the "truth-seeking" process. Aravind believes it's sufficient to build a "good business" without necessarily striving for "the greatest business model in human history."
• Data Chain: Aravind mentions that Google's advertising system is based on a dynamic auction system with extremely high margins. However, he also notes that Alphabet's (Google's parent company) cloud services and YouTube business have an annualized revenue of $100 billion, so the company itself is not at risk even if search ad revenue declines.

2. Technical Core: RAG Architecture and "Academic" Rigor

The technical core of Perplexity is Retrieval-Augmented Generation (RAG), but Aravind emphasizes that its execution is stricter than standard RAG. Standard RAG simply uses retrieved documents as context to generate an answer, whereas Perplexity's principle is "don't say what you didn't retrieve." This forces the model to strictly base its answers on human-created text retrieved from the internet, attaching citations sentence by sentence. This philosophy is directly inspired by the norms of academic paper writing.

• Historical Context: Aravind recalls that in the early days of the startup, they used a Slack bot connected to GPT-3.5 to answer internal company questions (e.g., about health insurance), but found the model frequently gave incorrect answers. This prompted them to return to their academic roots: "Every sentence you write should be supported by a citation." They realized this is exactly how Wikipedia works.
• Mechanism Breakdown: The system workflow is as follows: User asks a question → Traditional search engine retrieves relevant links → Relevant paragraphs are extracted → Paragraphs and the question are fed into an LLM → The LLM is instructed to generate an answer with footnotes based on these paragraphs. Aravind points out that hallucinations stem from four sources: 1) Insufficient model capability to understand semantics; 2) Poor index quality (outdated or insufficiently detailed information); 3) Too much context provided to the model, causing confusion; 4) Retrieval of completely irrelevant documents. Perplexity's optimization focuses on continuous improvement across these four dimensions.
• Unique Insight: Aravind believes that Vector Embeddings are not a panacea for retrieval. He reveals that the traditional BM25 algorithm (a term-frequency-based retrieval algorithm) still outperforms the latest embedding models released by companies like OpenAI on many benchmarks. Therefore, Perplexity's retrieval system is hybrid, combining traditional lexical retrieval with semantic retrieval, supplemented by PageRank-style domain authority and timeliness signals.

3. Product Philosophy: The User is Always Right, Curiosity is Core

Aravind positions Perplexity as a "knowledge discovery engine," with a mission to serve human curiosity. He believes most people are naturally curious but are not skilled at translating that curiosity into precise questions. Therefore, the core task of the product is not to wait for users to ask perfect questions, but to help users "ask the next question." This is reflected in features like "related questions" recommendations, autocomplete, and tolerance for poor input.

• Mechanism Breakdown: Aravind shares several key design philosophies:
• The user is always right: A philosophy learned from Larry Page. If the user input is poor and the answer is irrelevant, it's the product's fault, not the user's. The product should understand the user's intent, even with typos, grammatical errors, or unclear phrasing.
• Help users ask questions: The biggest enemy is not Google, but users who are not good at asking questions. Perplexity guides the user's exploration journey by automatically generating "related questions," transforming a one-off Q&A into a continuous knowledge discovery process.
• Pursue extreme latency: A lesson learned from Larry Page and Spotify. Aravind tests Perplexity's latency on airplanes (the worst network environment) and tracks P90 and P99 tail latencies to ensure the system remains fast under pressure.
• Deduction and Validation: Aravind believes a good product should allow users to be "lazier," not require them to work harder. He predicts that Prompt Engineering will not be a long-term skill, as products will eventually become intelligent enough to understand users without requiring carefully crafted prompts.

4. Future Outlook: From Search to Knowledge, From Facts to Truth

Aravind believes the ultimate form of search is not an "answer engine," but "knowledge discovery." In the future, the entry point for user interaction with information will not be limited to a search box; it could come from reading an article, listening to a podcast, or even a page being read aloud. The Perplexity Pages feature embodies this vision—transforming a personal exploration journey into a shareable, Wikipedia-style knowledge page.

• Mechanism Breakdown: Aravind proposes that the next major breakthrough in AI lies in decoupling reasoning from facts. He argues that current large models spend significant compute during pre-training to memorize facts, which is "brute force" and inefficient. The future direction is to train a "small model" that excels at reasoning, which can acquire facts through retrieval (RAG) or tool use, thereby drastically reducing compute requirements.
• Unique Insight: Aravind believes that the key bottleneck for AGI is not pre-training, but "Inference Compute." Once we crack how to make models improve answer quality through iterative thinking (like Chain of Thought), whoever has the most inference compute (GPUs) will have a decisive advantage. He cites Jensen Huang's philosophy of "paranoia," emphasizing that a single generational misstep in hardware can be fatal.
• Falsification Condition: Aravind points out that the true sign of AGI is not passing exams, but being able to create new knowledge—for example, proposing a counterintuitive but correct, engineer-verifiable new idea like PageRank or FFT. The real milestone will be when an AI can offer deeper, more truthful insights on a controversial issue (like the origins of COVID-19) than human experts, much like an excellent PhD student.

Position Moves

Position	Guest Sentiment	Key Data
Google (Alphabet)	Risk Alert / Neutral	Its ad model is "the greatest business model of the last 50 years," but also its weakness; Cloud + YouTube annualized revenue reaches $100 billion.
OpenAI	Neutral	Its models (GPT-4) are among the options Perplexity can use; still leads Perplexity's in-house models in reasoning capability.
Meta	Bullish	Open-sourcing Llama 3 70B, which is "close to GPT-4"; its open-source strategy benefits the entire AI ecosystem.
NVIDIA	Bullish	An investor in Perplexity; its GPUs are key for future inference compute; B100 offers 30x improvement in inference efficiency over H100.
Amazon (AWS)	Neutral	Perplexity runs on AWS; its "customer obsession" and "your margin is my opportunity" philosophy inspired Aravind.
Tesla / SpaceX (Elon Musk)	Bullish	Admires the "sheer grit" and "first principles" thinking, as well as the direct-to-consumer strategy.
GitHub Copilot	Bullish	Served as inspiration for starting Perplexity, proving that AI itself can be a product.
Wikipedia	Bullish	Served as inspiration for Perplexity's product philosophy (citation-driven), but acknowledges it has human biases.
Character.AI / Replika	Risk Alert	Views "AI companions" as using hallucinations as a "feature," a "dangerous" and "easy" path; Perplexity chooses the harder path of "truth-seeking."

Judgments Worth Remembering

1. "Your margin is my opportunity." (Aravind Srinivas quoting Jeff Bezos) — Google's high-margin advertising model makes it difficult to pivot to the lower-margin "answer engine" model, creating a disruptive opportunity for Perplexity.

2. "We never tried to beat Google at its own game; we redefined the game." — Perplexity's disruption lies not in providing better links, but in fundamentally transforming the UI from a "list of links" to "answers with citations," which is its strategic core.

3. "BM25 still beats the latest vector embeddings on most retrieval benchmarks." — This counterintuitive insight reveals the complexity of search technology: pure semantic search is not a panacea, and traditional term matching (an advanced version of TF-IDF) remains crucial.

4. "The bottleneck for AGI is not pre-training, but Inference Compute." — Once models learn how to improve answer quality through iterative thinking, whoever has the most GPUs to run these "thought processes" holds the key to AGI.

5. "The true sign of AGI is the ability to create new knowledge, like proposing an algorithm such as PageRank or FFT." — This defines the ultimate goal of AGI: not answering questions, but, like the best scientists, proposing counterintuitive yet correct new theories.

6. "Prompt Engineering will not be a long-term skill." — A good product should be intelligent enough to understand poor user input, rather than requiring users to become "prompt engineers." Products should allow users to be "lazier."

7. "The biggest enemy is not Google, but users who are not good at asking questions." — Perplexity's core value lies in helping users translate vague curiosity into precise questions and guiding them to the next step of exploration, thus initiating a journey of knowledge discovery.

8. "AI companions use hallucinations as a 'feature'; we choose the harder path of 'truth-seeking'." — Aravind explicitly rejects the "easy" path of using AI hallucinations to build emotional companion products, insisting on "reducing hallucinations" and "pursuing truth" as the company's core mission.