How Exa is Building the Perfect Search Engine | Deep Dives with a16z

影片總結

Exa CEO Will Bryk says agents need comprehensive search, and retrieval can cut customer token costs by 20x.

Exa cofounder and CEO Will Bryk tells a16z’s Sarah Wang that transformers made a Google alternative feasible, and Exa launched its first search engine in November 2022—two weeks before ChatGPT. He argues agents need controllable, complex-query search that can return 1,000 or 10,000 results, not just the small result sets designed for people. Bryk says retrieval lets customers use smaller models and can reduce token costs by 20x; he expects agentic search to become larger than Google Ads around 2030. He describes reinforcement-learning tests in which Exa required fewer search calls and performed better, while warning that agentic search faces infrastructure and data bottlenecks. Bryk also connects better information access to business research, coding, and social problems, and says Exa’s culture emphasizes fun, ownership, and hiring people with passion.

章節

1. 0:00Exa Origin Story: Bryk’s Lifelong Search Mission and the 2021 Transformer Breakthrough
2. 2:35Beating Google With LLMs: Exa’s GPT-3 Search Launch and Early AI API Demand
3. 9:00Search Built for Agents (Part 1): Complex Queries, Search Controls, and Biotech Coverage
4. 12:11Search Built for Agents (Part 2): Large Result Sets, LLM Reranking, and Rising Quality Demands
5. 15:47Search Built for Agents (Part 3): Higher Reliability and the Persistent Company-Search Problem
6. 19:03Search Built for Agents (Part 4): Go-to-Market Intelligence, Social Discovery, and Retrieval
7. 22:30Search Built For Agents: Exa Powers Devin and Bryk Envisions a Trillion-Dollar Agent Economy
8. 26:01Tokenpocalypse and Retrieval: Exa Says Focused Retrieval Can Cut Costs 20x
9. 29:21RL for Better Search: Exa’s RL Tests Use Fewer Calls and Outperform Traditional Search
10. 32:23Benchmarks and Agentic TAM: Customer A/B Tests and a Potentially Larger-Than-Google-Ads Market
11. 36:52Benchmarks and Agentic TAM: Scaling Search Beyond Google and Unearthing Unrecorded Data
12. 40:19Benchmarks and Agentic TAM: Exa’s Detail-Driven Culture, Office Goat, and 10¹⁸ Name
13. 44:20Culture Hiring and Wrap: Exa’s project autonomy and Bryk’s passion-first hiring