There are a number of startups aiming to replace Google with AI-powered search (Perplexity is one of them), but a startup called Exa has a different idea: an AI version of Google.
Exa's founders believe it's not humans who are in desperate need of a new kind of search engine: rather, as AI becomes more and more prevalent in businesses and consumers' lives, the AI platforms themselves will need to periodically go to the internet to search for information and return real answers, not hallucinated ones, and they won't just let you type requests into a keyboard.
Exa is building tools that enable AI models to perform things like web searches, but with an AI-native twist.
The co-founders bought $1M worth of GPUs (which were readily available at the time) and started building a machine learning model trained to natively understand links, rather than words or sentences, using vector databases and embeddings (not traditional transformer-based LLMs).
“Transformers typically predict the next word. We train search engines to predict the next link,” Brueck said. “People share links on the web. We use that data as a dataset to train models. And we train models to predict the next link. This is a new search algorithm.”
So, just as a law master completes a sentence by providing the most likely next word, Exa’s system also completes a sentence with the most likely link (or ten), but presumably excluding the SEO spam and (ironically) the AI-generated rubbish that clogs up search engines in general these days.
Image credit: Exa
The company exclusively told TechCrunch on Monday that it had raised $17 million in new funding in a Series A led by Lightspeed's Guru Chahal, with participation from Nvidia's venture arm NVentures and Y Combinator. Exa has raised $22 million in total to date, including a previous $5 million seed round. (Exa was in the summer 2021 YC cohort.)
“It's a very ambitious vision,” Chahal said. “They're building for AI what Google is building for humans.”
The team was founded about a year before ChatGPT was launched by two best friends who met during their freshman year at Harvard University: CEO Will Brueck (now 27) and co-founder Jeff Wang (26).
“We launched before ChatGPT. Our original goal as a company was not to provide AI. It was how do we leverage AI to build better search,” Wang said.
After ChatGPT took the tech world by storm, AI companies started asking Exa for an API version of the search engine they could incorporate into their models. Exa is based in San Francisco and is part of the cozy Cerebral Valley AI startup group. In fact, as TechCrunch previously reported, Wang's tweet looking for companies to help order office sleeping pods went viral and the response was overwhelming (the work-sleep culture is alive and well in the tech industry).
Currently, AI companies are the primary customers for Exa's search engine, and use cases range from AI chatbots that answer customer questions while searching the internet for information, to companies that want to curate training data.
Databricks, for example, is a major Exa customer, using it to find large-scale training sets for its own model training initiatives, the founders said.
An API version of the product was released about a year ago. “Since then, we've seen incredible adoption,” Wang says. Today, Exa says it serves thousands of developers, though it's worth pointing out that the company has a free plan that allows anyone to try out the search engine on a limited basis. It also has several tiered pricing levels. The founders declined to disclose revenue figures beyond saying that the company has some revenue and that it's growing. (Interestingly, in addition to running its own GPU clusters, Exa hosts its product on AWS rather than the AI-focused Google Cloud.)
The team isn't particularly focused on becoming the search startup that will upend Google, but if AI becomes as all-powerful as the tech industry thinks it will, a search engine for AI bots could pose an unexpected threat to search dominance.