Half Baked Newsletter October 14th.jpg

The Idea:

In 2020, Will Bryk was an early engineer at customer service AI startup Cresta when OpenAI released GPT-3. He later described it as a magical tool that understood text at near-human level. Bryk saw something else in it: if a machine could understand a page, why was the world still finding things by keyword?

He knew who to call. Jeff Wang had been his roommate at Harvard, where he ran a GPU cluster in his dorm room, and had since spent three years building data and web infrastructure at Plaid. The pair had already tried building a crowdsourced search engine to take on Google, before realising language models offered a far more powerful approach.

In 2021, Bryk left Cresta and the old roommates got into Y Combinator with a new plan. Instead of matching keywords, they would train a search engine to predict the next link, the way GPT predicts the next word.

They thought they were building a better Google for people. They were actually building one for machines.

This was the beginning of Exa.

The Execution:

The lesson? Exa launched for curious humans in November 2022, and within months the loudest knocking came from AI companies wanting an API. Bryk and Wang kept the engine and swapped the customer, admitting they had been building for LLMs “this entire time”. Three years of 10x growth later, that switch is worth $100M a year. Google built search for the person typing. Exa built it for the thing doing the typing.