The evolution of AI startup strategies

From: redpointai

Perplexity AI, led by CEO Arvin Seros, offers a compelling case study in the evolving strategies for building and maintaining a high-growth AI startup [00:00:15]. The company, valued at $500 million, provides a “next-gen search product” that aims to deliver answers rather than links [00:00:17], [00:00:01].

Initial Strategy: Product Focus over Model Building

A foundational aspect of Perplexity’s strategy was to prioritize product-market fit and user adoption over immediately developing proprietary models [00:08:04]. Seros unequivocally states that if he were to restart, he would follow the exact same path: begin with existing models, such as those from OpenAI [00:07:28], [00:07:57].

The rationale behind this approach aligns with Lean Startup principles:

• Product Focus: For a product-focused company, the initial emphasis should be on getting a product out, attracting users, ensuring returning users, and sustaining usage [00:08:36], [00:08:42]. This first barrier, if overcome, solves “almost half of your problems” [00:08:51].
• Talent Attraction: Having a product with users is crucial for attracting top engineering talent, which is considered harder than securing funding in the current AI landscape [00:08:57].
• “Rapper” Analogy: Perplexity embraced the “rapper” moniker (companies building on top of existing LLMs) by focusing on distribution and user value [00:59:38]. Seros preferred being a “rapper with 100,000 users” over having an internal model nobody knows [00:12:42]. The concept of “moats” (defensible competitive advantages) is deemed “overhyped” for early-stage AI companies, as one must first “have something to protect” [00:12:28], [00:50:18].

Core Product Development Philosophy

Perplexity’s product development and prioritization focuses on five key dimensions for a top-of-market product:

1. Accuracy [00:01:34]
2. Reliability [00:01:37]
3. Latency [00:01:38]
4. Delightful User Experience (UX) [00:01:40]
5. Iterative Improvement (overall or through personalization) [00:01:42]

Behind the scenes, the search process involves:

• Understanding and reformulating queries [00:02:20].
• Selecting relevant pages and parts of pages [00:02:24].
• Rendering answers (paragraphs, bullets) with supporting citations to minimize error and hallucinations [00:02:33].
• Incorporating multimedia (images, videos) when text is insufficient [00:02:52].
• Innovations like permalinks for shareable answers [00:03:09] and AI-suggested follow-up questions [00:03:30].
• The “Co-pilot” feature helps users formulate better queries, based on the philosophy that “the user is never wrong” [00:04:25], [00:05:27].

Resource allocation is managed through vertical integration and constant communication between design, product, and AI leads, ensuring everyone appreciates the importance of different aspects of the product [00:06:01]. The company’s core values—quality, truth, and velocity—directly reflect its product goals [00:07:05].

Strategic Shift: Towards Model Agnosticism and Own Models

Perplexity initially used OpenAI models, then fine-tuned smaller, faster models, and now increasingly uses and releases its own open-source models [00:07:27], [00:07:38]. This shift was driven by:

• Reducing Dependency: A desire not to be “overly depend[ent] on somebody,” especially if they are building competing products [00:10:30].
• Cost Efficiency: Driving down inference costs [00:10:42].
• Market Waves: Strategically waiting for significant shifts in AI models and infrastructure, such as the release of Llama and the anticipation of Nvidia’s TensorRT library, which promised faster inference [00:10:49].

Perplexity aims to be “model agnostic,” meaning they want the flexibility to use the best available model, whether internal or external, to provide users with the best possible answers [00:13:44]. Building their own models addresses user concerns about Perplexity being merely a “rapper” without underlying infrastructure [00:24:51]. The goal is to build the “muscle” to serve everything internally, aspiring to rival GPT-3.5 turbo and eventually approach GPT-4’s level with their own fine-tuned models [00:25:08]. This self-serving capability is crucial for maintaining high margins [00:26:48].

Competitive Strategy in Search

Competing with tech giants like Google requires a differentiated approach. Perplexity acknowledges its fortunate timing in the AI revolution [00:20:01]. Previous attempts to challenge Google often failed by trying to replicate Google’s technology stack directly [00:20:20].

Perplexity’s strategy contrasts by focusing on:

• Market Positioning and Branding: Unlike technology-focused competitors, successful challengers like DuckDuckGo and Brave invested in go-to-market strategies and privacy/crypto-focused branding [00:21:17].
• Leveraging Google’s “Rare Failure”: The emergence of advanced AI models like OpenAI’s (GPT-3.5) before Google could match them, and their accessibility via APIs, created a unique opportunity that Perplexity capitalized on [00:22:46].

Future of Search and AI Applications

Seros envisions the future of search in 10 years as primarily answers delivered by agents that perform tasks, behaving like natural conversations with friends [00:23:25].

Perplexity focuses on “knowledge velocity” for deep knowledge work, going beyond Quora or Wikipedia’s goal of merely maximizing knowledge [00:15:21]. This means accelerating access to personalized information (e.g., different levels of detail for black holes vs. celebrity gossip) [00:16:02]. Perplexity’s AI streamlines the human research process, providing concise, sourced answers in seconds [00:16:56].

The company aims to cater to “all of human curiosity” [00:17:22]. While starting with a vertical focus on knowledge and research, Perplexity intends to expand, heeding advice not to become a narrow “vertical search engine” but rather build end-to-end experiences like Booking.com or Expedia [00:18:26].

A constant challenge is balancing features for “power users” (like those on Twitter) with an intuitive experience for new users, following the “Mom’s test” [00:34:55]. The focus is on frictionless growth for new visitors [00:34:19].

The “collections” feature, though widely used, hasn’t yet become a seamless collaborative research tool because it still requires too much manual input [00:51:04].

Addressing Hallucinations and RAG

Perplexity is recognized for its effective use of Retrieval Augmented Generation (RAG) to reduce hallucinations [00:27:45]. However, Seros emphasizes that solving RAG for web search doesn’t automatically translate to internal enterprise search, as indexing, embeddings, and ranking are vastly different [00:28:11]. This highlights the challenges and strategies in enterprise AI deployment and the problem-specific nature of RAG solutions, counteracting the notion that a single API can solve all RAG issues [00:29:39].

Evolution of Perplexity’s Product Line

The company’s journey to its current search product involved several pivots:

• Initial Pitches: Early ideas included a vision-based search using smart glasses and audio [00:38:59].
• Search over Databases (Text-to-SQL): Advised to narrow focus, they explored search over internal databases [00:39:24]. However, feedback revealed that most SQL is already written, and new queries are often done visually, not via text, showing that “not everything’s going to be revolutionized by text as the universal interface” [00:41:48].
• Twitter Search: The team built a tool to search Twitter data, which gained significant traction, especially for users searching their own handles [00:42:46]. This product was initially more about “Graph Search” to connect users or find specific tweets [00:42:47].
• Summarization-Based Search: Developed internally as a Slack bot for the team’s own questions (e.g., coding, health insurance terms), its utility for condensing news and complex topics led to its public launch [00:45:12]. This, combined with the Twitter search’s virality (fueled by Jack Dorsey’s tweet and user self-searching), affirmed their consumer-facing search strategy [00:46:09].

AI Industry Insights

• Overhyped/Underhyped: Seros believes AI “moats” and overly vertical-focused strategies are overhyped [00:50:18]. Underhyped is the focus on building delightful user experiences [00:50:42]. This perspective offers a valuable lens for venture capital perspectives on AI applications and startups and investment strategies in the AI landscape.
• Open Source vs. Proprietary Models: Seros predicts that capabilities currently exclusive to state-of-the-art models like GPT-4 will eventually become possible with open-source models, at a cheaper price and faster latency [00:51:59]. This will, in turn, enable entirely new applications not yet imagined [00:52:09].
• AI Regulation: Regulation is considered “too premature” as widespread economic benefits haven’t yet materialized [00:53:30]. Restricting development through regulation can centralize power and hinder the discovery of potential issues and solutions [00:54:26].

Arvin Seros also shared a personal vision for AI: recording his parents’ voices and life stories to create an AI model he could “talk to them again” if they were no longer present [00:56:03]. He also envisions a future with AI-generated movies that dramatically lower creation costs, democratizing storytelling and explanation through visualization [00:56:49].