Sundar Pichai, the chief executive of Google's parent company Alphabet, has a packed audience in front of him at the Shoreline Amphitheater in Mountain View, Calif. He's doing his best to play the role of the new-school tech company head pioneered by Steve Jobs and Bill Gates: a bit of a pop icon, a bit of a church preacher communicating God's will, only no longer through songs or sermons, but through software and silicon. Unfortunately, the soft-spoken, introverted Pichai is not naturally well suited to this role. Somehow the atmosphere of his speech was more like a high school musical, without the warmth of a Hollywood theater.
As early as 2016, Pichai announced Google's "artificial intelligence first" strategy. Now AI development is at an important moment, but Google's competitors are attracting all the attention. Google was caught off guard by the unveiling of ChatGPT last November, and has been busy launching products over the past six months, competing with OpenAI, the developer of ChatGPT, as well as Microsoft, its partner and supporter.
At the company's annual I/O developer conference in May, Pichai wanted to showcase Google's efforts over the past six months. He introduced a new Gmail feature called "Write for Me," which automatically drafts entire emails based on text prompts; an AI-powered immersive view in Google Maps that builds realistic 3D previews of a user's route; a generative AI photo editing tool; and more. He talked about the powerful PaLM 2 Large Language Model (LLM), which supports a variety of new features, including Bard, which Google is using to benchmark ChatGPT, and a powerful family of AI models under development called Gemini, which could dramatically expand the impact of AI and amplify the risks involved.
But the topic that the audience on the site and watching the live stream was most interested in hearing about, Pichai avoided - what exactly does Google have planned? After all, search is Google's core product, generating more than $160 billion in revenue last year, or about 60 percent of Alphabet's total revenue. Now that AI chatbots can gather information from across the web to answer users, no longer offered as a list of links but as a conversation, what are the implications for Google's profit machine?
Pichai blinks at this. "We are reimagining all of our core products, including search, in a bold and responsible way," Pichai said. It was an oddly understated way to introduce a product that has a stake in the company's fate and in his own future. Later in Pichai's speech, the audience's impatience was felt in each round of lukewarm, polite applause.
Yet Pichai never touched on the key topic again. He chose to let Google's vice president of search, Cathy Edwards, introduce the somewhat awkwardly named Search Generative Experience (SGE). This feature combines search and generative AI to provide users with a single, summarized "snapshot" of the answer when they search, as well as a link to a website that confirms the answer. Users can continue to ask questions as if they were using a chatbot.
The new product has the potential to be a powerful answer generator. But will it generate revenue? That's the heart of the innovator's dilemma facing Google.
Alphabet says SGE is an "experiment." But Pichai has made it clear that SGE or similar products will play a key role in the future of search, telling Bloomberg in June that "similar products will become part of the mainstream search experience." The new technology has clearly not lived up to expectations; SGE is relatively slow and, like other generative AI, is prone to what computer scientists call "hallucinations," where it confidently provides false information. Pichai also acknowledged that the problem could be dangerous in search engines. He told Bloomberg that if a parent searches Google for the dose of Tylenol a child should take, "there's absolutely no room for error."
The launch of SGE gives an indication of Google's responsiveness in the AI arms race. The technology draws on Google's decades of experience in AI and search, showcasing Alphabet's firepower. But it also exposes Alphabet's vulnerability at a time of big change. Chatbots' information gathering could cannibalize Google's traditional search business and its lucrative ad-driven business model. Unfortunately, many people prefer ChatGPT's answers to the familiar list of Google links. "The search line as it's now known is going away," so predicts Jay Patisar, an analyst at research firm Forrester.
So it's not just the Tylenol dose that Pichai and Alphabet can't get wrong. Google has powerful AI tools but has failed to develop a strategy that matches Alphabet's advertising revenue as the 17th largest company in the world. How it responds to the shift will determine whether Google, whether as a verb or a company, can continue to survive in the next decade.
When ChatGPT was launched, some critics saw its significance as close to the introduction of the iPhone or the personal computer; others were more radical, arguing that chatbots were comparable to electric motors or printing presses. But to many executives, money managers, and techies, one thing was clear from the start: ChatGPT was a dagger straight to the heart of Alphabet, and within hours of its debut, users of the chatbot were calling it a "Google killer.
Despite ChatGPT's inability to access the Internet, though, observers accurately guessed that it would be relatively easy to give the AI-powered chatbot access to a search engine to provide answers. When dealing with multiple queries, it would seem that it would be easier for ChatGPT to respond in unison than to have to piece together information through multiple links. In addition, chatbots can write code, compose haiku, write high school history essays, develop marketing plans and provide life coaching. None of these Google searches can do.
So far, Microsoft has invested $13 billion in OpenAI and has been quick to announce the integration of OpenAI's technology into the search engine Bing, which has never had more than 3% market share before. Critics believe that this integration could be Bing's best chance to take down Google Search once and for all. Microsoft CEO Satya Nadella joked that Google is the "invincible giant" of search, before adding, "I want people to know that we can make the giant dance."
Nadella is more confident in Google's "dancing skills" than some critics who think Google is too bureaucratic and slow to dance. Google's world-class AI team has long been the envy of the tech world, and in 2017, Google researchers invented the basic algorithm underpinning the generative AI boom, an artificial neural network called transformer. (The T in ChatGPT is "transformer.") However, Alphabet can't seem to figure out how to turn its research into products that will spark the public's imagination. 2021 also saw the creation of a chatbot called LaMDA, which is quite powerful, and LaMDA's conversational skills are top-notch. But just like other big language models, this bot's responses can be inaccurate, biased, and sometimes just weird and disturbing. Since the problem has not been solved and it's really hard for the AI community to try to fix it, Google is worried that rushing to release it is irresponsible and could pose a risk to its reputation.
Perhaps just as importantly, chatbots don't fit well with Google's main business model - advertising. Summary answers or conversation streams seem to offer far fewer opportunities for ad placements or sponsored links than Google search.
To many, the clash reveals deeper cultural barriers. Some former employees say Google is too comfortable with market dominance and too complacent and bureaucratic to cope with rapid shifts, as it did in 2020 when Google acquired generative artificial intelligence entrepreneur Pravin Seshadri's startup AppSheet. Earlier this year, shortly after his departure, he wrote a blog post claiming that Google had four core problems: no mission, no sense of urgency, delusions of exceptionalism, and mismanagement. All the problems, he said, stem from "having a money-printing machine called 'advertising' that keeps growing year after year while covering up other mistakes."
Four other former employees who left Google in the past two years gave similar descriptions. (They requested anonymity in their interviews with Fortune magazine for fear of violating their severance agreements or jeopardizing their careers.) "The amount of red tape you had to endure just trying to improve existing features, let alone develop new products, was unbelievable," one of them said. Another former employee said Google often used its large user base and revenue as an excuse not to embrace new ideas. "They set the bar for impact so high that there's little way to get around it," said another.
Similar internal dissatisfaction has only fueled a broader narrative: that Google is in big trouble. In the five weeks between the release of ChatGPT and New Year's Day, Alphabet Inc.'s stock fell 12 percent.
By mid-December of last year, there were signs of panic within Google. The New York Times reported that Alphabet was on "red alert" to catch up with OpenAI and Microsoft.2019 Google co-founders Larry Page and Sergey Brin had stepped down from their day-to-day responsibilities, controlling the company only through super-voting shares. Now the two are suddenly returning, with reports that Brin has rolled up his sleeves to help write code.
The sudden return of the co-founders is hardly a show of support for Pichai's leadership. But Google executives see enthusiasm, not panic, behind the return of Page and Brin, and the recent scramble. "It's important to remember that Larry and Sergey are both computer scientists," said Kent Volker, president of global affairs at Alphabet. "Both are excited about the possibilities of the future." Later in an interview with the New York Times podcast, Pichai said a "red alert" was never issued. But he acknowledged "asking the team to act urgently" to find ways to turn generative AI into "deep, meaningful experiences."
The external stimulus has clearly had an effect. In February, Google announced the launch of Bard, which competes head-to-head with ChatGPT, and by March, it introduced its Workspace writing assistant feature, as well as the Vertex AI environment, which helps cloud customers train and run generative AI applications using their own data. At the I/O developer conference in May, almost every Google product shone with next-generation AI light. Some investors were impressed, and immediately after I/O, analysts at Morgan Stanley wrote that the company's "pace of innovation and go-to-market tempo has increased." Google's stock fell to as low as $88 a share after the launch of ChatGPT, and by the time Pichai took the stage in Mountain View, it was trading at more than $122. shares were trading at more than $122.
The cloud of suspicion still hasn't dissipated. "Google has a lot of inherent strengths," said Richard Kramer, founder of Arete research, an equity research firm; after all, Google has unmatched artificial intelligence research and access to some of the world's most advanced data centers. "Google just isn't pushing as hard as it could commercially," he added, noting that Google's divisions and product teams are too siloed to collaborate at the corporate level. (By far the most obvious organizational tweak Google has made in the wake of the AI onslaught is the merger of two leading AI research programs, Mountain View-based Google Brain and London-based DeepMind, into an entity called GoogleDeepMind.)
In addition to the Arete analysts, there are others who believe Google is not realizing its potential. Morgan Stanley noted that despite the recent recovery of Alphabet, there is still a "valuation gap". Its stock price has previously been higher than other tech giants such as Apple, Meta and Microsoft, but as of July Google's price-to-earnings ratio was about 23% lower than its competitors. To many, this suggests that the market sees Google as incapable of getting out of its artificial intelligence rut.
Jack Krawczyk, 38, is a "second-time" Google employee. He joined Google in his 20s, left in 2011 to go to a startup, and later in the streaming media broadcasting service Pandora and WeWork. 2020, he returned to the company to develop the Google Assistant, used to deal with Apple's Siri and Amazon's Alexa.
Google's LaMDA chatbot intrigued Krawczyk, who wondered if chatbots could improve the smart assistant. "I've been talking about this for most of 2022 and probably 2021," Krawczyk said. The obstacle is reliability, a persistent "illusion" problem. Can users accept being confident that the answer is wrong?
"We've been looking forward to the moment when we announce that we're 'ready for an extremely persuasive interaction,'" Krawczyk says. Last fall, "we started to see some signals," he says, skimming over the popularity of ChatGPT with some embarrassment.
Today, Krawczyk is the senior product director for the Bard team. Though the product draws on years of Google research, Bard did complete development of ChatGPT soon after its launch, on Feb. 6, when Google unveiled its new chatbot, and a few days later when Microsoft launched Bing Chat. Google would not say how many employees were involved in the project. But there are some signs that the company is under pressure.
One of the secrets to ChatGPT's smooth response is fine-tuning through a process known as Reinforcement Learning with Human Feedback (RLHF). The specific idea is to have humans rate the chatbot's responses, and the AI gradually learns to tweak the answers in an effort to get closer to the high-scoring version. The more conversations a company can train, the better the chatbot is likely to perform.
In just two months ChatGPT gained 100 million users, and OpenAI achieved a distant second place by having a lot of conversations. To catch up, Google hired contract evaluators. A number of contractors working for outsourcing firm Appen later filed complaints with the National Labor Relations Board about being fired for publicly talking about low earnings and unreasonable deadlines. One of them told the Washington Post that raters had only five minutes to score even Bard's lengthy responses to complex topics such as the origins of the Civil War. The contractors were concerned that the time pressure could lead to flawed scoring and compromise Bard's security. Google says the problem is mostly between Appen and the staff, and that the scoring is just one piece of the many data points used to train and test the Bard; training continues apace. There were also reports that Google tried to train Bard using answers from competitor ChatGPT, as a number of users had previously posted conversations on the ShareGPT site. Google denied utilizing the data in question.
Unlike the new Bing, Bard is not a search tool, although it can provide links to relevant sites. Krawczyk said that Bard is intended to be a "creative collaborator". According to him, Bard's main role is to retrieve ideas from users' own brains. "Get the key information in your head, abstract the concept, and then expand it," he said. "Ultimately, it's about enhancing the imagination." Google search is like a telescope, and Bard is like a mirror, Krawczyk said.
Exactly what people see in Bard's mirror is hard to say at this point. The chatbot's debut was shaky: In the blog posting Bard's output, a screenshot contained the erroneous claim that the James Webb Space Telescope, which launches in 2021, took the first picture of an exoplanet in our solar system. (It was in fact taken by an Earth-based telescope in 2004.) The mistake proved to be worth $100 billion: the market capitalization of Alphabet that shrunk in the 48 hours after reporters reported the error. Meanwhile, Google warned employees not to put too much faith in Bard: Google issued a memo in June warning employees not to rely on Bard or other chatbots for advice on code without double-checking.
Since Bard's debut, Google has upgraded the chatbot's AI to PaLM 2 LLM. according to tests released by Google, PaLM 2 outperforms OpenAI's top model, GPT-4, on benchmarks in a number of areas of reasoning, math, and translation. (Some independent evaluators did not come up with the same results.) Google also made some tweaks that greatly improved Bard's response to math and coding queries. Some of those tweaks reduced the likelihood of Bard hallucinating, Krawczyk said, though the problem is far from solved. "There are no best practices that produce 'x,' " he says. "That's why Bard is still a trial run."
Google declined to disclose the number of Bard users. But there are signs to be seen in third-party data: the number of visits to the Bard site increased from about 50 million in April to 142.6 million in June, according to Similarweb, well behind the 1.8 billion visits to ChatGPT in the same month. (In July, Google extended Bard to the European Union and Brazil, increasing the range of languages answered by 35, including Chinese, Hindi and Spanish.) The numbers in question are dwarfed by Google's main search engine, which receives 88 billion visits per month and 8.5 billion daily searches. Since the launch of Bing Chat, Google's search market share has increased slightly to 93.1 percent, while Bing's share of the search market has remained essentially unchanged at 2.8 percent, according to StatCounter.
Bing is clearly not the biggest threat to search from AI. in May, a Bloomberg Intelligence survey of 650 people in the U.S. showed that 60 percent of 16- to 34-year-olds said they preferred asking ChatGPT questions to searching on Google. "The younger demographic could drive a radical shift in the way online searches are conducted," said Mantip Singh, senior technology analyst at Bloomberg Industry Research.
That's where SGE comes in. Google's new generative AI tool can provide users with answers to multi-step queries that are more complex than traditional Google searches, said Elizabeth Reid, vice president of Google's search business.