In the realm of search engines, there’s a holy grail. Think human intelligence, as if you were telling your question to a friend—but the friend exists through the computer screen and in the digital ether, a pansophical internet librarian who will rummage through virtual bookshelves to retrieve the answers you seek.
There’s no such librarian in today’s world. To use modern search engines, we often pretend the machine is dumber than it is smart: We digest queries into a basic mash of keywords and awkward half-phrases, then deploy more brainpower to foolproof the language, so search algorithms cannot wander astray. But George Sivulka, an artificial intelligence scholar, wants to end the era of what he calls “Googlese.” He’s the founder of Hebbia, a two-year-old startup that’s building the world’s first “neural search engine,” set to launch today.
Since its founding in August 2020, Hebbia has racked up an investor list that includes Peter Thiel (as one of just six pre-seed investments, following Facebook, OpenAI, and Deepmind), Stanley Druckenmiller, Yahoo founder Jerry Yang, board members of Google parent Alphabet, a former Federal Communications Commission chairman, and venture capital firm Index Ventures. It recently closed a $30 million Series A funding round.
Tools for thought, not keywords
The company’s raison d’être stems from the fact that in today’s market, dominant commercial search engines including Google, Yahoo, Bing, Baidu, and DuckDuckGo, are all powered by keyword matching. They must be able to identify words from queries within any webpages they spin up; then they rank the pages through an abstruse synthesis of consumer statistics, word counts, and semantic mapping (for example, past data may have drawn links between thematically similar words like “solar” and “sun”). That model has its limits—in a way, you must already know something about the results you seek.
But Hebbia hopes to transcend the limits by offering answers keyword agnostically, using a set of machine learning-trained neural networks. If you were to search “What is the meaning of life?” Google might spin up a blog post with that very title—but Hebbia might go deeper, offering you a trove of scholarly literature from famous existential philosophers. According to the company, it performs 57% better than other state-of-the-art search algorithms in finding the most optimal results.
Make no mistake: The technology to understand like a human, called natural language processing (NLP), has existed for years. The mind-boggling brilliance of AI minds like GPT-3 have already wowed crowds. But when it comes to search, NLP is sorely underutilized by companies that thought it was interesting from an academic perspective, but perhaps not commercially. In recent years, Google has tinkered with “transformer” NLPs called BERT and MUM, but stopped short of integrating them into the core of its search engine; its current transformer architecture is minimal. For Sivulka, the question was glaring: “How do we apply and productionize this new technology to build tools for thought?”
Peering into the mysterious ‘Deep Web’
Hebbia’s first product, launched two years ago, was a Ctrl-F function that allowed users to search more intelligently within text on-screen, now followed by a stand-alone search engine. But the company’s business model also leans into its academic roots: It’s carving a specific niche in the financial, legal, and medical fields, where analysts toil away digging for details buried within mountains of documents and transcripts (they might now hopefully type: “What are the latest sales figures,” or “When was the patient diagnosed,” and have answers on hand in seconds). To achieve this, Hebbia partners with institutions and governments to spin up a database of private internal papers and texts—a solution that Sivulka claims will also attack one of Google’s pain points.
The data on the web is vast—almost unfathomably so, with experts claiming that the “Deep Web,” which is invisible to most web crawlers, is roughly 500 times bigger than the web we know—but it’s estimated that Google has only indexed 4% of the world’s data. A large part of reaching the last 96% involves cooperation from the data’s suppliers who hesitate to make their valuable archives fully public—such as Hebbia’s partner institutions. For them, Hebbia might offer peace of mind. Because its algorithms do not require quick access to text keywords, it’s able to encrypt all the content stored in its index, meaning that data would be secure even in the event of a hack.
Hebbia’s not a Google killer yet, but Sivulka says he hopes to eventually branch out into Google’s turf of public search engines that trawl the World Wide Web, perhaps as a hybrid product—likely for a skilled “knowledge worker”—who will log into a personal database of highly specialized information, operating within a broader search engine. He notes that Hebbia is already indexing public documents including Congress’s inflation reduction act and bipartisan infrastructure bill, as well as filings from Johnny Depp’s defamation trial, and he imagines all this scaling up in the next six months to a year.
A powerful status quo
The idea of a better search engine might seem like a no-brainer. So why hasn’t it been done yet? Google, with its trillion-dollar war chest, may have all the ammo it needs to build a version of Hebbia. But according to Sivulka, it’s held back by an “innovator’s dilemma”: Its current business model is far too lucrative, as its dominance over the search market allows it to sell valuable s based on its traditional method of keyword searches and page rankings.
“If you talk to a founder who believes in neural search and computers that understand you,” says Sivulka, “I think it will be inevitable that Hebbia, or another neural search company, might take over Google.”
Its investor, Index Ventures’ Mike Volpi, has a vision for Hebbia that’s more tempered. “I’m really happy our founders have big ambitions, but I don’t think [Hebbia] is going to challenge Google in public search,” he says. “I don’t see a consumer business in the near future. It’s a tool that Index Ventures, or Fast Company, might buy to do intensive research.”
For Volpi, part of the draw is Sivulka himself, whom he first met through his daughter, a classmate of Sivulka’s at Stanford (“I told her, if any of them seem smart or have a cool business plan, make sure you send them my way,” he laughs). Building a company takes “a special type of person,” he says—not just sharp, but also charismatic, “a personable, commercial, regular person. At [Index Ventures], when we invest, that’s a big factor. Where you end up is not always where you were pointing at when you get started.”
Sivulka started his first job at age 16 at NASA, working on software for satellite landmine detection, then did physics research on a dark matter detector for the U.S. Energy Department at Stanford, and then became the fastest undergraduate to complete a mathematics degree at Stanford. He then dropped out of his Stanford PhD program in computational neuroscience to start Hebbia.
The name, he says, was inspired by his doctoral research, on a neural network that was mimicking the patterns of a coral reef. “You could make the argument that if the coral reef had a soul, the AI that was doing the exact same thing also had a soul,” he says. “It was this weird tie-together between the artificial and the biological, or the natural and the human.” The training for that algorithm came through a process called Hebbian learning. Now the next evolution, he asks: How do we train neural networks that are like humans?