In its early weeks, ChatGPT, the wildly popular artificial intelligence tool from OpenAI, has offered up a potential new model for online search. The chatbot responds to questions about topics such as political science and computer programming with detailed explanations, and its question-and-answer format means users can drill down until they fully understand. Users doing similar research on Google must typically scan search results and peruse various websites until they arrive at their own conclusions. ChatGPT, by contrast, delivers a decisive (or at least decisive-sounding) answer in seconds.
Alphabet Inc.’s Google has been essentially untouchable in search, but a handful of companies, some founded by former Googlers, think that’s about to change. The entrepreneurs say a shift is under way from the prevailing model of keyword search, in which search engines comb the web for specific terms, to searches powered by large language models, which analyze enormous text databases to develop the ability to understand user questions and produce direct answers. This is the technology that ChatGPT uses to compose its rapid-fire responses.
Some of the technological breakthroughs underpinning large language models were forged in Google’s own research labs. But entrepreneurs who have left in recent years say the company may struggle to fully capitalize on the technology’s potential, in large part because its business model, in which ads are displayed alongside search results, is too lucrative to disrupt. Google raked in $54.48 billion in advertising revenue in the most recent quarter, representing 78.9% of its gross sales. Search ads were the biggest driver by far.
“Google is just a victim of its own success,” says Sridhar Ramaswamy, once Google’s top ad executive, who’s now chief executive officer of Neeva Inc., a privacy-sensitive search engine. “They’re trapped to a certain extent in how that page looks and behaves.”
Google may have to hold itself to a higher bar when using emerging technology than startups do. ChatGPT and other large language models often deliver incorrect information in a convincing fashion, a phenomenon computer scientists sometimes refer to as “hallucinating.” Unlike a Google search, ChatGPT currently offers no clue about where it sourced the information it’s providing to users, and OpenAI has acknowledged it produces incorrect answers at times. Google may direct users to sites that promote misinformation, but delivering falsehoods in its own voice is a fundamentally different—and riskier—proposition.
Google declined to comment. OpenAI said in a statement that ChatGPT doesn’t incorporate data from the internet into its answers, adding that the current version is a research preview to help develop models that are “safer, more reliable, aligned and more useful,” and is not intended to be used for advice.
Google introduced its search engine in 1998, powered by its signature PageRank algorithm, which measured each website’s importance by the ways other sites linked to it. It quickly became the dominant search tool. Google has spent decades indexing the web, and the breadth of queries it can field is unmatched. ChatGPT, by contrast, is trained on a dataset that contains only limited information after 2021, freezing its knowledge in time.
In recent years, Google has taken steps to help users conduct searches in new ways, including through the lens of their smartphone camera and with image and text combined. It uses large language models to understand users’ queries and has also incorporated the technology into its “featured snippets,” which spotlight key information on search results pages.
A company such as Google, with access to vast data and computing resources, would seem to be an ideal place to do advanced AI work. Yet engineers often want to move faster than they can at such a big company, and Google has seen a lot of departures, including some of its AI researchers. One of the tech giant’s most influential contributions to the field was the 2017 paper “Attention Is All You Need,” which introduced the concept of transformers, systems that help AI models zero in on the most important pieces of information in the data they are analyzing. Of the paper’s eight authors, all but one have jumped to startups in recent years, a review of LinkedIn profiles shows; at least five have founded their own AI ventures. Asked why he embraced the chance for entrepreneurship, one author, Character.AI founder Noam Shazeer, puts it bluntly: “Startups can move faster and launch things.”
Industry watchers describe the threat to Google as anywhere from worrisome to existential. In a note to investors in December, Morgan Stanley acknowledged the possibility that users could use AI programs for queries such as product reviews and travel. Paul Buchheit, the former Google employee who created Gmail, wrote in a series of posts on Twitter that the company may be “only a year or two away from total disruption.”
Yet about two-thirds of Google searches end without users clicking on another site, according to market research firm SparkToro and analytics firm Similarweb. That suggests Google may simply need to tweak its user interface to highlight all that it’s already doing to head off users’ queries, says Mandeep Singh, an analyst with Bloomberg Intelligence. “Google has a very strong moat, and it’s unlikely to be disrupted in search,” he says.
That hasn’t stopped startups from trying. With many more open source and commercial AI tools available, the barriers to entry have fallen, says Edwin Chen, founder of Surge AI, a data labeling platform that works with search startups. His company has conducted consumer research that shows users often prefer search results from newer entrants such as Neeva, You.com and Kagi, particularly for queries such as recipes. “You don’t need hundreds of millions of dollars to beat Google now,” he says.
Search startups have adopted a variety of business models. Vectara, founded by former Google employees, is selling its software to businesses, enabling them to offer search powered by large language models on their websites. Such models require massive computing power, giving the advantage to large incumbents. But Vectara says it’s found a way to make the economics work, in part because using large language models to process text is much cheaper than generating responses.
Kagi, Neeva and other consumer search engines are subscription-based, offering users unlimited searches for a monthly fee. Strategies such as distillation, in which the output from a large language model is run through a smaller model, have enabled Neeva to use its computing resources efficiently, says Ramaswamy, who spent 15 years at Google before co-founding the company in 2019. “I expect that usage of these models will essentially become table stakes,” he says, “and power and value will again shift back to the product creators.”
Ramaswamy says he still sees great value in the legacy model of search, noting that signals such as how websites link to one another can reveal much about the authority of a source. Neeva has already woven large language models into its search engine and is planning to release products in which the tools are more front and center soon. It also plans to use the models to draft “one-pagers” that disclose sources, so users can evaluate the credibility for themselves. “We are thinking a lot about how we can have the best of both worlds,” Ramaswamy says.
Shazeer’s Character.AI may offer another model of search. He founded the company with Daniel De Freitas, who led Google’s work on the pioneering LaMDA chatbot, which Google first announced in May 2021. In September, Character.AI introduced a site with a diverse cast of chatbot characters that users can converse with in real time, all powered by large language models. The site includes disclaimers—“Remember: Everything Characters say is made up!”—that offer some cover if the chatbots utter falsehoods. “Even now, we can do something extremely, extremely valuable, which is to bring people joy and fun and help people feel better,” Shazeer says. “But that’s really just the tip of the iceberg.”
Google continues to polish LaMDA and other large language models internally, and its dominant position gives it some luxury to be patient. But there is at least one powerful argument for a sooner introduction: the data gleaned along the way.
“One should never underestimate Google,” says Oren Etzioni, an adviser and board member at the Allen Institute for Artificial Intelligence. “But the challenge they have is GPT is learning as it goes.”Read next: What Are the New AI Chatbots For? Nothing Good
©2022 Bloomberg L.P.
