Interview, Fireside Chat
How To Build The Future: Aravind Srinivas
- Perplexity aims to double engagement and daily question volume through a follow-up questioning feature that automatically determines the optimal use of components like knowledge graphs versus LLM streaming.
- The company projects that within three to four years, the platform will become essential for specific commerce scenarios such as shopping for clothing or booking hotels, though it currently lacks direct revenue from off-platform transactions unless users subscribe to Pro.
- To achieve mass-market utility, the firm plans to build an orchestration system integrating small models, knowledge graphs, widgets, and multi-step reasoning over a decade-long timeline, while resisting the "entropy" of becoming a slow organization as the team grows beyond 250 people.
- Strategic plans include serving open-source models via fine-tuning and evaluation rather than focusing on hardware, improving query reformulation with LLMs, and handling user ambiguities by clarifying queries instead of blaming the user.
- The business faces significant risks including competitors undercutting prices with free services or larger cash reserves, potential declines in search revenue due to AI agents fulfilling tasks directly, and the threat of Wall Street pressure if ad-based models collapse.
- A fundamental tension exists between maintaining an ad-free experience for early adopters and introducing monetization features like buy buttons, which early adopters may perceive as ads rather than helpful tools.
- The company anticipates a challenging competitive landscape where it must compete with Google's near $200 billion annual search revenue and Microsoft's potential consumer product failures, aiming to become the "next Google" by operating a router and orchestrator at a billion-user scale.
- Long-term success depends on navigating the "AI completeness path" where product quality improves as AI improves, requiring the solution of complex logistics problems for merchants and hotels to enable end-to-end fulfillment without users leaving the ecosystem.
- Technical execution must balance speed against the increasing complexity of models, manage the shift from indexed data to unstructured inference at query time, and ensure streaming answers appear instantly to satisfy user expectations for latency.
- The organization must maintain a culture of obsessive detail-oriented work and direct communication hierarchies to prevent the loss of engineering velocity and trust that often accompanies growth and the introduction of production bugs.
- Future capabilities will include handling agent-like behaviors where the system decides to click, scroll, or browse, and ensuring the product works reliably even when users do not know they want to ask a follow-up question or when models are pre-GPT 3.5.
- The company must convince users and investors that a new monetization model is viable, overcoming the risk that people stop clicking links, while ensuring the product feels "magical" and provides a "healthy meal" compared to the "fast food" of legacy search results.