Search Has Its Goliath. Could Richard Socher Be Its David?

Richard Socher, CEO ready for his next paramotor adventure
Richard Socher, CEO ready for his next paramotor adventure

Disclaimer: This blog post was published prior to’s latest AI advancements and may not reflect our current capabilities. has transformed from a search engine to an AI assistant. With a foundation in search and the team’s AI expertise, was perfectly positioned to enhance LLMs with live access to the Internet to address issues around hallucinations and transparency. As such, is capable of tasks ranging from searching online to writing an essay, debugging code, creating digital art, solving complex problems, and more. Learn more about getting the most out of also offers its core technology through a suite of self-serve APIs. Get complete details about the YOU API.

*This story was originally published in The Information on April 3, 2023.

by Arielle Pardes

On a rare rain-free afternoon this spring, Richard Socher stood atop a hill, surveying the direction of the wind. The wind often interests Socher, the 38-year-old co-founder and CEO of search engine, because in his spare time he likes to fly around on a paramotor, a parachute-like contraption supported by a backpack propeller. We were standing at one of his launch spots, a short hike from the 40-acre ranch where he lives, about half an hour west of Palo Alto, Calif.

Before making the jump offa hill such as this, Socher explained, you have to know the wind’s direction, the speed, the gust factor and the weather forecast. You cannot simply Google this information. You could ask ChatGPT, but you’d be let down. (“As an AI language model, I don’t have access to real-time weather data,” the chatbot will admit.) So Socher uses an app called Windy, which offers precise forecasts for his airborne adventures.

Soon, he said, apps like Windy will be integrated into his artificial intelligence–powered search engine,, so you can simply search for current wind conditions, with the AI model surfacing the app. “The beauty of an open-platform search engine like is that you can select which weather app you want, versus on Google you cannot,” he said. “My hope is that we put users somewhat more in control.”

Socher started working on in 2020, convinced that he could build a better search engine with AI. Still, the idea of out-Googling Google seemed like a long shot. “I’ve had very smart people and investors say, ‘Richard, why would you build a search company? The space is dead. It’s never going to work. It’s a monopoly. It’s a fool’s errand,’” he told me recently.

But three years later, the future that Socher had imagined, anticipated and prototyped has suddenly and dramatically arrived. What once seemed like a quixotic obsession is now the hottest space in tech, and the real fools are those who didn’t see the promise of AI search coming.

For Socher, who has spent a decade working on natural-language processing, all of this felt inevitable. “Somehow, while the technology to understand language has gotten better, the search experience has gotten worse,” he said. Ask Google a question, and it sends back a scavenger hunt of hyperlinks. Ads are search-optimized spam; privacy is a joke. And how come you have to use computer-speak (“showtimes avatar AMC”) instead of talking like a human? Socher thought there had to be a better way.

On our hike, he shared his vision for, a small but vibrant challenger in the emerging chatbot wars. He was building a search engine that gave reliable answers, not just hyperlinks, and that made a customized experience possible with AI tools and apps. It also protected privacy, with limited ads and an option to browse in a sort of “paranoid mode,” without cookie tracking. This vision had already earned the platform a small fan base — which was feeding it “millions of queries a day,” according to Socher — and $45 million in funding from investors including Time Ventures, Breyer Capital, Sound Ventures, Day One Ventures and Salesforce.

Now Socher must prove he can take on Google, Microsoft and OpenAI. To some, that still seemed like an impossible quest. But from where Socher stood, the winds of change had already started blowing.

In 1996, Marc Benioff was relaxing on the beach in Hawaii when he had an epiphany about the future of the internet. Benioff, who was on sabbatical from his job as a senior vice president of sales at Oracle, realized something big was about to happen on the web.

The online world would one day soon be a simulacrum of real life, offering an ersatz copy of the sun and wind and sand beneath his toes. He thought about the concept of the digital twin, a virtual model that represents a physical one, and wondered what kinds of things would wind up on the internet. People were starting to buy domain names, snapping up little plots of digital land on which to build digital futures. Inspired, Benioff bought the domain — a representation of you, online. “Ultimately, we all want to be able to express ourselves in the system,” he told me, “so that when we look at the internet, we’d say, ‘There we are. That’s you.’”

Benioff held on to for the next 24 years. It gathered digital dust through the creation of Google and Facebook and Uber and Benioff’s own massive internet company, Salesforce, until he found an inheritor in the form of a German-born entrepreneur with rumpled red hair: Richard Socher.

When the two met, in 2014, Socher had just finished his doctorate at Stanford University, where he’d pioneered a new method for teaching computers to understand natural language. Neural networks had gained traction for things like recognizing faces in Facebook photos and translating phrases from English to Chinese, but they weren’t very good at grokking the meaning of language. Even tasks as simple as labeling a sentence as “positive” or “negative” could get complicated by basic grammar, like double negatives. Socher’s research demonstrated how a neural net could solve these problems by modeling the context and relationship between words.

It also won him a job offer from Princeton University. But academia seemed a black hole for bleeding-edge research, and Socher decided to start a company instead, bringing his research into industry. His company—MetaMind—used neural networks to perform a wide range of tasks for other companies, from analyzing the sentiment in tweets to detecting brain bleeds in medical scans. The technology was designed to be dead simple, so you didn’t need to be a developer to use it; if you wanted to perform a complex machine-learning task, all you had to do was formulate the question. “This is similar to web search,” Socher told Wired in 2015, “except you give the actual answer rather than just a bunch of links.”

The technology attracted the interest of Khosla Ventures and Benioff, who together invested $8 million in seed funding. Khosla also installed Sven Strohb and, the firm’s chief technology officer, to serve as CEO; Socher became CTO. Salesforce acquired the company for an undisclosed sum in 2016, bringing Socher on to build Einstein, an AI platform that lets Salesforce clients better understand their customer data.

Besides the “dramatic impact” Socher had on Einstein, Benioff credited him for attracting a high-quality team of machine-learning engineers to Salesforce, including its current head of research, Silvio Savarese. “It’s quite extraordinary what Richard’s been able to do intellectually,” Benioff told me. “He has an amazing body of work to date in what is clear to everybody will be one of the most important fields of our time.” Chien-Sheng Jason Wu, who worked with Socher at Salesforce AI Research, characterized him as an ambitious researcher with a penchant for moonshot projects: “He’s always willing to take the risk.” CEO and co-founder Richard Socher at his San Mateo County hills home
Richard Socher at his home in the San Mateo County hills. Photography by Katie Thompson.

Flush with cash from the Salesforce acquisition, Socher — still living in his grad-student apartment in Palo Alto — bought a $3 million ranch in the San Mateo County hills. It had sweeping views and access to a network of private hiking trails. He started paramotoring more often, flying above Egyptian pyramids, Slovenian castles and verdant Icelandic valleys.

He also started angel investing, including in some students who had taken CS224, the deep-learning course at Stanford that Socher created in 2015. It was the first class to teach neural networks for natural language processing, and quickly became one of Stanford’s most popular courses, with more than 600 students enrolled at a time and even more auditing the course with the videos Socher posted to YouTube. Julien Chaumond, who formed a study group for people following along online, told me the class was “pivotal” for him. He and a friend, Thomas Wolf, used its techniques to build out their startup, called Hugging Face. Wolf later pitched the startup to Socher, and Socher offered both seed funding and “enough science credibility to help with our early fundraising in the Valley,” Chaumond said. Hugging Face has since raised $160 million; last year, it was valued at $2 billion.

Between his Salesforce job, his new investments and his paramotoring hobby, Socher’s life might have felt complete. But by 2020, he realized there was one idea that still nagged at him. He had worked on natural-language processing for a decade, and still no one had applied it to the biggest business model of all time: online search.

“Search had always been something Richard was interested in, because it’s the gateway to the internet, to information,” said Bryan McCann, who worked with Socher at Salesforce and joined him as the other co-founder of “He’s always looking for the biggest impact problem he can work on.”

Socher offered Benioff his resignation, explaining that he was going to start another company. Benioff felt “disappointed that he was ready to move on,” but understood, perhaps better than anyone, what motivates brilliant young talent to fly the coop. He also felt certain that he would invest in whatever Socher would do next, even without knowing what it was. “I’m just more of a fan of his,” he explained. As a parting gift, Benioff offered Socher something for his new venture: the domain name

Earlier this year, Socher met me at a WeWork in downtown San Francisco to show me what could really do. It was one of his first in-person interviews for the company, which had mostly been built in Covid-19 isolation, and he showed up with his boyish flop of red hair a little messy. Under a black hoodie, he wore a T-shirt that said “Predict I Will.”

Socher opened his laptop to, the company name appearing in technicolor. It had taken a year and a half for about 50 employees to build a feature set equivalent to Google’s — things like Maps and News — which now has for everything but sports. (“I sort of like sports to some degree,” Socher said, “but I never watch them.”)

The main thing differentiating from Google: The AI tools are built into the main search page. “You come to us with any question, and we give you useful answers using natural-language processing,” Socher said. Searching for the Fibonacci function in the Python programming language? surfaces the Stack Overflow app — you just copy and paste the function, then move on — rather than a list of links to click through. Besides using the search bar, you can also ask questions via YouChat, which works much like ChatGPT. There’s also YouCode (for AI-generated code), YouWrite (for writing), and YouImagine (an AI image generator that Socher demonstrated by creating a photo of a skydiving baby). “We are the first to bring those into the search engine world,” Socher told me at the time. Of course, this was only weeks before some of YouChat’s features would be replicated in other products, like Microsoft’s BingChat.

In addition, Socher showed me, had added over 250 apps to its platform, allowing its AI search to call on content from sites such as Reddit, Yelp or Wikipedia. Socher called this paradigm the “CAL future,” an acronym that encompasses chat, apps and links. He suggested this combination of information could help circumvent one of the main problems with AI search bots, which is their tendency to make things up.

Of course, YouChat — like its main AI search rivals — does sometimes make things up. After our demo, I asked the bot to identify a “leading expert in the field of search technology” to whom I could speak for this article. Immediately, it responded with a name: Alan Lu, a professor at University of California, Berkeley, who has studied search engines for over 15 years and published a book on the subject, “Search Engines and Information Retrieval Algorithms.” He sounded perfect. The only problem: When I went to look up Lu’s contact information, there was no such professor at UC Berkeley. Nor was there an Alan Lu who specialized in search engines at another university. When I tried looking up the book, I found that it, too, did not actually exist.

I brought that example to Socher this week, who told me to try again, since YouChat had recently been updated. (Indeed, YouChat provided a more cogent answer this time.) Chirag Shah, a real researcher who studies search and information technologies at the University ofWashington, told me that even with the occasional hallucination, YouChat’s technology seemed “on par” with Bing and Google. “But we all know that it’s not just the tech that makes or breaks things,” Shah added. “Google and Microsoft have deep pockets and a dire need to take or defend market shares. Many search startups have come and gone because they couldn’t match the might of the big players in terms of spending and investing. You also have to ask — who’s going to pay for all this ultimately?”

Socher plans to monetize by charging apps to build on its platform, and charging users for some of its AI products. For now, most of those costs are subsidized: Generating 100,000 words on YouWrite costs about $2, compared to $82 on JasperAI. That may still make it difficult to compete with OpenAI, which seems determined to give away many of its products practically for free. “They’re very clearly trying to create a Google-like presence, where competing with them is something you don’t even conceive of,” said Jeremy Howard, an AI researcher who wrote the first paper describing the kind of fine-tuned language models used by ChatGPT and YouChat. “You’d be brave — or stupid — to try to compete with them.”

Yet Socher seemed unphased. On our hike near his ranch, I asked him if he felt worried about competition from OpenAI. Socher conceded that the company was “doing their thing to build a monopoly,” which came with both upside and downside: OpenAI had been dropping its prices, which made things cheaper for everyone else, but it also meant that dozens of other search startups were now trying to do what was doing, cribbing its features and fighting for the same users. Some, like Perplexity, had decided to focus on a smaller subset of users, hoping a narrow scope would make them more competitive. But with a name like, Socher was making a big statement that his search engine was for everything and everyone.

Could a tiny search engine stand up to the goliaths? Maybe. Socher suggested that his company’s foresight was its edge. “Two years ago nobody thought we could compete with [Google] and innovate on anything they could not,” Socher told me. Even with a fraction of its resources, had been early on a number of features now considered standard in AI search, and Socher insisted that they would be first on more to come. “Maybe it’s a personality thing,” he said, half-smiling, “but once I set my mind on something, I’m not going to give up.”

Arielle Pardes covers tech culture for The Information’s Weekend section. Previously, she was a senior writer at WIRED in San Francisco.

Yay, you’ve read the whole article. If you enjoyed this story, please recommend 👏 and share to help others find it!