While there’s no shortage of startups aiming to replace Google with AI-powered search (we’re looking at you, Perplexity), a startup called Exa has a different idea: a Google for AI.
Humans aren’t the ones who desperately need a new kind of search engine, Exa’s founders believe. Rather, as AI increasingly takes hold in corporate and consumer life, it is the AI platforms themselves that must regularly venture out onto the internet to search for information and return bona fide answers, not hallucinations. And they can’t just type their requests on their keyboards.
Exa is building a tool that allows AI models to perform something like a web search, but with an AI-native twist.
The co-founders bought a million dollars’ worth of GPUs (which were easier to get in those days) and, using a vector database and embeddings (not a classic transformer-based LLM), they began to build a machine learning model trained to natively understand links rather than words and sentences.
“Transformers normally predict the next word. We train our search engine to predict the next link,” CEO Will Bryk says. “So people share links on the web; we use that data as a dataset for our model that we train. And we train the model to predict the next link. So it’s a novel search algorithm.”
So just as an LLM would complete a sentence by furnishing the most probable next word, Exa’s system does so with the most probable link (or 10), but presumably minus the SEO spam and (ironically) AI-generated chum clogging every ordinary search engine these days.
On Monday the startup announced it raised a new $17 million Series A led by Lightspeed’s Guru Chahal, with participation from Nvidia’s venture arm NVentures and Y Combinator, it exclusively told TechCrunch. Exa has now raised a total of $22 million, including its previous $5 million seed. (Exa was in the summer 2021 YC cohort.)
“This is a very ambitious vision,” says Chahal. “What Google is to humans, they are building for AI.”
The team was founded about a year before ChatGPT was launched, by two best friends who met their freshman year at Harvard: CEO Will Bryk (now 27) and co-founder Jeff Wang (26).
“We launched before ChatGPT. Our initial goal as a company was not to serve AIs at all. It was: how do you use AI to build better search?” Wang said.
After ChatGPT stormed the tech world, AI companies began asking Exa for an API version of their search engine that they could plug into their models. Exa is located in San Francisco, part of the cozy Cerebral Valley AI startup set. In fact, as TechCrunch previously reported, a tweet by Wang went viral when he was looking for other companies that wanted to go in on an order of office nap pods and the response was overwhelming. (The work-nap-repeat culture is alive and well in this part of the tech industry.)
With AI companies now as its primary customers, use cases for Exa’s search engine span everything from an AI chatbot looking up info on the internet while answering customers’ questions to companies looking to curate training data.
Databricks, for instance, is Exa’s marquee customer, using it to find large training sets for its own model training initiatives, the founders say.
The API version of the product was launched around a year ago. “Since then it’s gotten a crazy amount of traction,” Wang says. Today Exa says it’s serving thousands of developers — although it’s worth pointing out that Exa has a free tier that allows anyone to try its search engine in a limited way. It also has multiple tiered-fee levels. The founders wouldn’t reveal revenue except to say that they have some and the number is going up. (Interestingly, apart from running its own GPU-cluster, Exa hosts its product on AWS, not the AI-forward Google Cloud.)
The team isn’t particularly focused on being the search startup that upends Google. Although if AI becomes the be-all/end-all that the tech industry thinks it will be, search engines for AI bots could be the unexpected threat to the search hegemony.
Kommentar