How To Build The Future: Aravind Srinivas (AI column)

Interesting. interview CEO and founding member Perplexity Y Cominator guys. I enjoy using Perplexity as my primary search engine and as a thought aid. That’s why I was interested to see the story of how the company was created, how it developed and where it is going now. As a result, the main ideas of this speech are about the following:

1) The Origins of Perplexity and the Way of the Founder Aravind began his journey in AI with training in India, after which he moved to the United States, where he trained at OpenAI under the leadership of Ilya Sutskever. It was Sutzkever who pointed out to him the importance of teacherless learning and reinforcement learning. After working at Berkeley on unsupervised and generative models and an internship at Google, Aravind came to the idea that starting an AI company and developing a product are inextricably linked. The idea of Perplexity was born in a post. Daniel Gross on creating the next Google and the concept of improving the wording of queries using suffixes. Initially, the team experimented with vertical search and databases using OpenAI Codex models to create SQL queries. It was a b2b search product for databases inside the pumping companies. 2) Product evolution The first prototype of Perplexity was created over the weekend and was inspired by WebGPT. Initially, a slow model was used. 175 AGP, the team took a heuristic approach, choosing the first k links and short descriptions from the cache. Aravind notes a paradoxical point: “The stupid approach worked when models began to follow instructions better.” This allowed to solve problems with the delay in the user interface, reducing the response time of the user. 7 seconds to an acceptable value. A key success factor was the ability to ask additional questions, which doubled the time users interacted with the product. 3) Competitive Advantage and Strategy Aravind didn't see direct competition with Google due to their interface congestion, though Microsoft's launch of Bing chat caused some concern. Unlike Google, Perplexity focuses on “the user is always right,” requiring refinements from AI. Consumer products, according to Aravind, should be simpler and more convenient for users. 4) Corporate culture and approach to development The main indicator of the effectiveness of the company is the number of requests per day. Weekly and monthly data help manage a team without a rigid hierarchy. The company actively uses Twitter to communicate with customers, appreciating direct criticism more than compliments. Aravind stresses the importance of "meticulous and obsessed people in the team" and constant "fighting with entropy" to maintain quality. 5) Future and vision Aravind sees Perplexity as “more intelligent search” than Google, aiming to create a place for a full user experience. He plans to build a system of small models and knowledge graphs, focusing on AI management and new monetization models. Unlike Google, the company describes itself as a “user and product-conscious” organization that can quickly implement and train the latest models. This gives Perplexity an edge in competition with tech giants who find it harder to integrate innovation because of their size and established business models.

#AI #Software #Engineering #Storytelling #ML #Architecture