Tubegraph

Challenges and opportunities in AIpowered search engines

6 min read

From: redpointai

The future of search is evolving rapidly, moving beyond traditional link lists to provide direct answers, intelligent agents, and conversational experiences [00:00:01]. Perplexity AI, a prominent next-generation search product, aims to lead this transformation [00:00:17].

Core Principles and Development of Perplexity AI

Perplexity AI’s sophisticated, seamless user experience belies significant underlying complexity [00:01:06]. Its development focuses on five key dimensions:

Achieving excellence across all five dimensions is critical for market leadership [00:01:52].

Behind the Scenes: Delivering Answers

Building Perplexity involves a complex process for each query:

  • Query Understanding and Reformulation [00:02:20]
  • Page Selection and Content Extraction [00:02:24]
  • Answer Rendering: Determining the best format (summary, bullets, visuals like images or videos) [00:02:33].
  • Citation and Hallucination Minimization: Ensuring each sentence has supporting citations and reducing factual errors [00:02:41].
  • Shareability: Innovating with permalinks so answers can be easily shared and benefit others [00:03:09]. This feature was later adopted by competitors like Bing and ChatGPT [00:03:25].
  • Follow-up Questions and Co-pilot: Suggesting further questions helps users, as articulating precise questions is a common human limitation [00:03:33]. The “Co-pilot” feature aims to improve user query formulation, reflecting the philosophy that “the user is never wrong” [00:04:25].

Resource allocation for these development areas is a continuous discussion among design, product, and AI leads, emphasizing vertical integration and shared values like quality, truth, and velocity [00:06:01].

Model Strategy and Moats

Perplexity’s journey in model development serves as a case study for other AI companies.

  • Starting with Third-Party Models: It is advisable for product-focused companies to begin with existing models (e.g., OpenAI models) to achieve product-market fit and attract users [00:07:55].
  • Building Internal Capabilities: Over time, as usage sustains, companies can fine-tune smaller, faster models and eventually transition to open-source or proprietary models to control destiny and drive down costs [00:08:08]. This transition should align with market shifts, like the arrival of open-source waves (e.g., Llama 2) [00:10:49].
  • Addressing the “Rapper” Criticism: Companies starting with external models are often labeled “wrappers.” Perplexity embraces this initial approach, prioritizing user acquisition over immediate model building. “I would rather be a rapper with 100,000 users than having some model inside and like nobody even knows who I am” [00:12:42].
  • The Concept of a Moat: A moat only matters once a product has achieved significant user adoption [00:12:26]. User value is paramount; users don’t care about a company’s underlying models or profitability, only about getting the best answer [00:14:06].
  • Model Agnosticism: Perplexity aims to be model-agnostic, ready to integrate the best available model, whether proprietary or open-source, to provide optimal answers to users [00:13:55].
  • Talent Acquisition: Recruiting top-tier engineers and AI researchers is often harder than securing funding in the current generative AI landscape [00:08:59].

In 10 years, search will primarily consist of answers provided by AI agents that can perform tasks and converse like friends [00:00:01].

  • Knowledge Velocity: Perplexity aims to maximize “knowledge velocity” – accelerating the acquisition of knowledge by providing fast, personalized access [00:15:36]. Unlike static platforms like Wikipedia or slow Q&A like Quora, Perplexity delivers concise, sourced answers in seconds [00:17:08].
  • Personalization: Search results will be highly personalized to individual user needs, offering varying depths of information depending on their query and preferences [00:16:00].
  • Vertical vs. Horizontal: While many venture capitalists advocate for “vertical search engines,” Mark Andreessen advised against this, noting that past attempts to build “Google for X” largely failed. Instead, successful companies built end-to-end experiences (e.g., booking.com for travel) [00:18:23]. Perplexity’s goal is to cater to all human curiosity while starting with a focused vertical [00:17:20].

Competing with Google

Perplexity acknowledges fortunate timing in the market [00:20:01]. Previous challengers to Google often failed by trying to replicate Google’s technology and market strategy [00:20:17]. Success stories like DuckDuckGo and Brave, which focused on strong go-to-market strategies and niche positioning (privacy, crypto), highlight the need for differentiation [00:21:17]. Google’s “rare failure” in leading the initial AI boom created a significant opportunity for competitors [00:22:50].

Challenges in using AI for research advancements and Hallucinations

Perplexity has been effective in using Retrieval Augmented Generation (RAG) to reduce hallucinations in web search [00:27:45]. However, solving RAG for internal enterprise search (like Google Drive) is a distinct and complex challenge, requiring different indexing, embeddings, and ranking strategies [00:28:18]. The common claim that a single embedding API can solve all RAG problems is misleading; effective RAG requires substantial work on the retrieval component, not just the language model [00:29:16].

Product Evolution and User Acquisition

Perplexity aims to build user trust by developing its own infrastructure and models, countering the perception of being merely a “wrapper” [00:24:43]. The goal is for users to not notice or experience an improvement when internal models are deployed [00:25:25].

A key challenge in AI product development is balancing the needs of power users with the need to make the product intuitive for new users [00:35:05]. Overloading new users with features or sign-up prompts can hinder growth [00:36:34].

Origin Story and Viral Growth

Perplexity’s initial ideas included a vision-based search (glasses, microphone, audio answers) and search over databases (text-to-SQL for enterprises) [00:39:13]. The text-to-SQL idea was abandoned because existing visual tools (e.g., Tableau, Power BI) were more efficient for enterprise users [00:41:15].

The pivotal moment came from an internal Slack/Discord bot that provided summarized answers with links [00:45:01]. This tool, originally for internal team use for quick answers on coding or health insurance, proved highly useful [00:45:12]. When launched to the public as a consumer-facing search, it gained viral traction, notably amplified by Jack Dorsey’s tweet, which led to a surge in users searching their own Twitter handles [00:47:16]. The product’s ability to pull holistic summaries of individuals across platforms “spooked people out” and generated significant screenshot virality [00:48:18]. This demonstrated real user excitement and confirmed the viability of a consumer-facing search product [00:48:48].

Overhyped and Underhyped in AI

  • Overhyped: The constant focus on “moats” and being “vertical-focused” [00:50:18].
  • Underhyped: The need for more focus on building delightful user experiences [00:50:42].

AI Regulation and the Broader Ecosystem

AI regulation is currently “too premature,” as the widespread economic benefits have not yet been fully realized [00:53:30]. Regulating too early risks stifling innovation and centralizing power in the hands of a few large organizations, which could be more dangerous than allowing widespread development and public scrutiny [00:54:30].

Looking beyond search, areas like AI-generated movies, realistic voices, and tools that can preserve memories (e.g., recording parents’ voices and personalities to interact with later) are exciting applications of generative AI [00:55:49]. The reduction in marginal cost of creation through these tools will foster a future of amazing visualizations and explanations [00:56:57].

The recent events at OpenAI suggest internal disagreements between those advocating for faster development and those for more cautious approaches [00:57:33]. While OpenAI remains a cutting-edge organization, it may not be able to pursue every possible AI application simultaneously, potentially leading to a more focused strategy [00:58:04].

Graph View

Backlinks