Berkeley alum wants to ‘make the planet smarter’

Since ChatGPT’s launch two years ago, people have been exploring new ways to use generative AI chatbots to automate tasks and answer questions. But while AI chatbots are useful tools, they are not always trustworthy sources of information.

Now imagine an AI chatbot that operates like a virtual Ph.D. student, citing sources for all the information that it provides. That’s what Aravind Srinivas (Ph.D.’21 CS) is trying to do through his AI-driven search company, Perplexity, which he co-founded in 2022, along with Andrew Konwinski (M.S.’09, Ph.D.’12 CS), among others, shortly after completing his doctoral studies at UC Berkeley.

“Ph.D. students write papers, and for every sentence they write, they reference another paper. That way, the paper is [based on] scientific facts,” said Srinivas. “We baked that principle into an AI chatbot, and that ended up becoming Perplexity.”

A new kind of search experience

AI technologies can sometimes hallucinate and make up answers. Srinivas aimed to reduce this problem by building a search engine that would only provide results that it could cite or reference.

“From the beginning, Perplexity has been this marriage of Wikipedia and ChatGPT having a baby together. Except the data for that marriage is coming from the entire internet, not just Wikipedia,” said Srinivas. “You can still converse and chat like in ChatGPT, but the response would be like a Wikipedia article, with subsections, citations and sources. It’s like ChatGPT’s educated uncle.”

Perplexity is also part of Srinivas’s quest to help humans satisfy their need for entertainment and their unending curiosity about the world.

“Our broader vision is to make the planet smarter. After just 30 seconds on Perplexity, you should learn something,” said Srinivas. “Catering to curiosity at the scale of humanity is the grand mission we’re working toward.”

From Chennai to Berkeley

Srinivas first became interested in AI while at IIT Madras in India. There he focused on machine learning and was invited to join a research project on teaching AI to play the 1970s Atari games Pong and Breakout.

Srinivas then ran similar experiments on transfer learning, a machine learning technique that uses knowledge gained from one task to improve performance on a related task, and hierarchical reinforcement learning, in which AI tackles more complex decision-making problems by breaking them down into smaller sub-tasks. After publishing a few papers about his findings, he undertook an internship at Montreal Institute of Learning Algorithms, where Turing Award winner Yoshua Bengio introduced him to deep learning concepts. Srinivas then set his sights on Berkeley.

“When it came to AI, I think Berkeley was the most happening university at the time,” he said. “Two [EECS] professors, Pieter Abbeel and Alexei Efros, were cranking out these amazing papers. And a Berkeley student named John Schulman [Ph.D.’16 EECS] was publishing a lot of cool papers and open-source code. I knew this was the place to be.”

Studying at Berkeley was both challenging and inspiring for Srinivas. He appreciated that Ph.D. students at Berkeley were expected to drive every step in the research and development process, from conceiving the idea to running the initial experiments and performing all the engineering.

“By watching other students in Pieter Abbeel’s lab, like Jonathan Ho [B.S.’14 EECS, Ph.D.’20 CS] and Peter Chen [B.A.’14 CS, B.A.’14 Statistics], I saw the amount of work they did all on their own,” said Srinivas. “From them I learned that software engineering is critical to doing great work in AI. It’s not just about mathematical equations on a whiteboard, but rather making them work in practice. So you need to be a good engineer.”

Following his first year at Berkeley, Srinivas was recruited by Schulman to work as a summer intern at OpenAI. It proved to be a transformative experience.

“The core idea of GPT was being conceived there at the time, and that changed my world view a lot,” said Srinivas. “It also increased my level of ambition by about ten thousand times.”

From engineer to entrepreneur

Srinivas wanted to start a company the moment he came to Berkeley. Although he spent time in the Berkeley Artificial Intelligence Research (BAIR) Lab under Abbeel and considered himself a researcher, he also wanted to be an entrepreneur. He just needed to figure out his angle.

Many of the startups that he had heard about, like those that came out of Y Combinator and similar incubators, were usually founded by people who had dropped out of college with the goal of changing the world. But Srinivas was looking for a different, more practical startup model. Specifically, he wanted to leverage his academic roots to build a solution that, once in the hands of users, could drive momentum and sustain growth.

“The immediate example I could think of was Google,” said Srinivas. “It was started by academics. The core idea is you can have a great academic idea or insight, convert that into a product and get it into the hands of a lot of people. As they use it, you gather a lot of data, and that data should fuel the product to be even better and help you garner more users.”

He also drew inspiration from the comedy TV show “Silicon Valley.” “The plot revolved around this company developing the fictional Pied Piper compression algorithm,” said Srinivas. “At the time, some people in my lab and I were very interested in generative models and compression. And so that got us all into generative AI without us even realizing it.”

Srinivas soon realized that launching a startup wouldn’t be easy. Many things had to come together: the idea, the people and the funding. After completing his Ph.D., he worked for a year at OpenAI and saw “how AI went from a research thing to driving real usage.” He then crystallized his idea for Perplexity, reached out to investors and went to work building his team.

For Abbeel, his Ph.D. adviser, Srinivas’s talent and drive were evident early on. “When it came to research, Aravind was the kind of student who not only brought great execution, but who was also capable of bringing great vision,” he said. “And now we are seeing him do the same thing with Perplexity, combining deep AI expertise with unique vision and exceptional execution, building one of the most useful AI products and one of the biggest AI companies to date.”

Advice for future Berkeley engineers

Srinivas’s advice to other Berkeley engineers who want to make the leap from researcher to entrepreneur is simple: start with a vision.

“With Perplexity, for example, we had a vision for creating a completely new search experience, where you’re now able to ask questions that you otherwise would not have been able to ask before,” he said. “Remember that a true vision, a true obsession about something, can be seen by other people, too. And that is what’s going to help you recruit people and attract funding.”

Srinivas added, “But the most important part — and the hardest — is to persist through the phase where you’re not sure. Never give up, just keep at it.”