From: redpointai
The future of search will be characterized by answers, with AI agents performing tasks for users [00:00:25]. It will evolve to feel like natural human communication [00:00:06], providing concise responses that don’t “ramble” [00:23:44].
Perplexity AI’s Approach to Next-Gen Search
Perplexity AI, a “next-gen search product,” aims to maximize “knowledge velocity” – the rate at which users gain knowledge [00:15:36]. This contrasts with traditional platforms like Wikipedia, which offer static, unpersonalized articles, and Quora, which relies on slow human responses [00:15:55]. Perplexity performs the human skill of research and summarization in seconds [00:17:08].
Core Dimensions of a Top-Tier Search Product
Building a smooth and seamless search experience requires focus on five key dimensions [00:01:16]:
- Accuracy [00:01:34]
- Reliability [00:01:37]
- Latency [00:01:38]
- Delightful User Experience (UX) [00:01:40]
- Iterative Improvement: Constantly improving for everyone or through personalization [00:01:42].
Achieving excellence in all five is crucial for market leadership [00:01:50].
Behind the Scenes: How Perplexity Delivers Answers
When a user queries Perplexity, the system undertakes several complex steps:
- Query Understanding and Reformulation [00:02:20]
- Page Selection: Identifying relevant web pages [00:02:24].
- Content Extraction: Determining which parts of selected pages to use [00:02:28].
- Answer Rendering: Deciding the format (summary, bullets) and ensuring citations for each sentence [00:02:30].
- Hallucination Minimization [00:02:47]
- Multimedia Integration: Including images or videos to enhance answers [00:02:54].
- Sharability: Implementing permalinks so answers can be easily shared and benefit others [00:03:09]. This innovation was later adopted by Bing, ChatGPT, and Bard [00:03:25].
- Follow-Up Questions: Suggesting further questions to help users articulate their curiosity, as people are often not precise in their initial queries [00:03:33].
- Co-pilot Feature: This feature assists users who may not be “great prompt engineers” or know how to ask the “right question” [00:04:25]. The philosophy is that “the user is never wrong” [00:05:27], a lesson learned from Larry Page’s early Google demos [00:04:51].
Building Models: Perplexity’s Evolution
Perplexity’s journey in model development serves as a case study for AI companies [00:07:40]:
- Starting with Off-the-Shelf Models: Initially, Perplexity used OpenAI models [00:07:27]. This approach is recommended for product-focused companies to “move fast” and prioritize getting a product to users [00:07:59]. The most important goal is to attract returning users and sustain usage [00:08:40].
- Fine-Tuning and Open-Source Adoption: They later fine-tuned smaller, faster models and transitioned to using open-source models, even releasing their own [00:07:34]. This strategic shift was driven by a desire to avoid over-reliance on competitors and to control their own destiny and costs [00:10:30]. They waited for significant open-source developments like Llama 2 to “take good advantage” of the “next wave” [00:10:49].
- Model Agnosticism: Perplexity aims to be model agnostic, willing to integrate any superior model to provide the best answer to users, as users prioritize results over the underlying technology [00:13:44].
Competition in Search: Lessons from the Past
Competing with established players like Google requires a different strategy than simply replicating their technology [00:20:12]. Previous attempts by former Googlers failed because they tried to build the “exact same technology” [00:20:20] that Google had already perfected over 20 years [00:22:27].
Successful challengers, like DuckDuckGo and Brave, focused on unique market positioning and “go-to-market” strategies (e.g., privacy, crypto incentives) rather than just technology [00:21:17].
The rise of AI presented a rare “big moment” and “rare failure in Google’s history” [00:22:46] where they were not the number one in the field, and others made powerful AI models accessible via APIs [00:23:07]. Perplexity capitalized on this opportunity [00:23:18].
The Role of Retrieval-Augmented Generation (RAG) and Hallucinations
Perplexity has been highly effective at using RAG to combat hallucinations in web search [00:27:42]. However, applying RAG to internal enterprise search is a “completely different” challenge [00:28:18]. The indexing, embeddings, and snippet generation for internal data differ significantly from web search [00:28:32]. It’s not as simple as merely training a large embedding model or “dumping garbage into the prompt” [00:29:08]; extensive work on the retrieval component, including indexing, embeddings, and ranking with diverse signals, is required [00:30:14].
Claims that a single company can solve RAG universally are misleading [00:29:31], as effective RAG solutions are highly dependent on the specific use case [00:29:39].
User Experience and Growth Challenges
A continuous challenge for Perplexity is balancing the needs of “power users” with making the product intuitive for new users [00:35:05]. While listening to existing users is important, prioritizing the experience for new visitors who have never seen the product is crucial for sustained growth [00:34:17]. Products that become too unintuitive or feature-heavy for new users are less likely to grow [00:36:34].
Broader Future Trends in AI
Search as Answers and Agents
In the next 10 years, search will primarily deliver direct answers, with AI agents capable of performing tasks for users [00:23:25]. This will transform interactions to be more like conversations with friends [00:23:34].
Customization and Personalization
The future of search will involve highly personalized content. Unlike static articles on Wikipedia, AI-powered search will adapt information depth and focus to individual user preferences (e.g., deep dive on black holes for a nerd, high-level celebrity stats for casual interest) [00:15:57]. This “fast, high bandwidth access to knowledge” will be a key differentiator [00:16:38].
AI Assistance APIs
The development of AI assistance APIs will unlock new possibilities, though the claim that a single API can lead to a trillion-dollar company by itself is likely an overstatement [00:29:14].
Voice AI and Communication
Voice AI, like ChatGPT’s voice feature, is becoming more prevalent [00:24:06]. However, for search, conciseness is key; users prefer to read answers quickly rather than listen to lengthy voice responses [00:23:51].
AI in Human Communication
A fascinating potential future application of AI is to preserve memories and enable communication with deceased loved ones through generated voices and conversations, based on extensive recordings [00:00:41]. This could involve AI generating personalized messages in the voice of a family member [00:56:03].
AI Regulation
Regulation of AI is currently “too premature” given that widespread economic benefits haven’t been fully realized [00:53:30]. While AI safety is important, slowing down development could hinder the discovery of potential issues and their solutions [00:54:17]. Over-regulation could also inadvertently centralize AI development among a few well-funded entities, which is more dangerous than broad participation [00:54:30].