How Google Got Its AI Groove Back --- Powerful new Gemini model allowed it to catch up to OpenAI's ChatGPT

“In the wee hours of an August morning, an artificial-intelligence project manager at Google loaded the newest creation from its DeepMind lab onto a platform that ranks AI models.

 

Google had fallen behind in the AI race, while its rival OpenAI had attracted hundreds of millions of users to its ChatGPT chatbot. Google researchers were hoping that a new feature -- a lightning-fast image generator -- would give the search titan an edge in a weak spot for ChatGPT.

 

Naina Raisinghani, known inside Google for working late into the night, needed a name for the new tool to complete the upload. It was 2:30 a.m., though, and nobody was around. So she just made one up, a mashup of two nicknames friends had given her: Nano Banana.

 

Within days, Nano Banana had the top spot in performance rankings on the platform LM Arena, was trending on X and had far exceeded Google's usage expectations. By September, Google's Gemini AI app had become the most downloaded app in Apple's app store.

 

Two months later, Google launched its most powerful Gemini model yet, which surged past competitors to become the most capable AI chatbot. With that, the Alphabet-owned company had leapfrogged OpenAI to the front of the AI pack.

 

Google's deep roots in science and research, willingness to pour billions of dollars into developing custom hardware, and leadership changes in recent years that cleared the way for faster experimentation are now paying off. It also has managed to protect its all-important search business -- at least for now -- from the surging popularity of chatbots, which are changing how consumers use the internet.

 

Google's AI work has begun generating substantial revenue through search ads, paid versions of Gemini for consumers and business and sales of new computer chips developed in-house. The November release of Google's latest Gemini model outperformed ChatGPT on a variety of measures, sending Alphabet's stock soaring and triggering a Code Red inside OpenAI. That company has since narrowed the race with the launch of a more powerful version of ChatGPT, which still has far more users than Google's Gemini.

 

Google Chief Executive Sundar Pichai talked up the magnitude of the company's AI push on the day the new Gemini model launched. "Love to see that we're launching at the scale of Google," he told employees in an internal memo.

 

When Pichai rose to the top job at Google in 2015, AI was a technology of keen interest to computer science researchers and almost no one else. The following year, he declared that the company known to consumers for its search engine, maps and productivity tools was going all in on AI.

 

In a memo posted to the company's blog, Pichai wrote that the previous decade had been all about a smartphone-oriented world. "But in the next 10 years," he predicted, "we will shift to a world that is AI-first, a world where computing becomes universally available."

 

Google already had laid the foundation with an AI research division called Google Brain, which was co-founded in 2011 by Jeff Dean, a computer scientist who helped develop the neural-network technology that underpins today's large language models. A few years later, Google acquired DeepMind, the London-based AI research lab co-founded by Demis Hassabis, a chess prodigy who would later share a Nobel Prize for work on an AI system that aids biomolecular research.

 

In a move that drew less attention at the time, Google also started designing its own AI chips, believing it would need vastly more computing power to support applications such as voice recognition. Those chips, called tensor-processing units, or TPUs, were designed to draw less power than the central-processing units in computers or the graphics-processing units in videogame cards. They would prove a game changer, for Google and the industry.

 

Early on, though, the company took a cautious approach to developing its own chatbots. Some of its executives and researchers had concerns about the safety of such technology, which has the potential to produce inaccurate, biased or otherwise problematic information.

 

Julia Winn, a former Google Brain employee, said chatbots weren't initially seen as core to the company's broader AI ambitions, and that in tests of early models, it proved easy to prompt racist or sexist responses.

 

"Those kinds of risks Google just took way more seriously than any place I've worked, and for understandable reasons," she said. Such caution frustrated a number of company researchers, some of whom decamped.

 

In August 2022, Google introduced a chatbot model with a range of conversational abilities, making it available to a limited number of people through an app called AI Test Kitchen, a proving ground of sorts. Google named it LaMDA and allowed users to test three functions: "Imagine It," "List It," and "Talk About It (Dogs Edition)," which enabled users to have a conversation only about dogs.

 

Three months later, OpenAI made ChatGPT available to the public following its own multiyear effort to create a chatbot that could not only engage in dialogue but produce stories, jokes, computer code and more.

 

Within five days, a million people signed up to test it. Unlike with Google's LaMDA, users didn't face significant restrictions in how they used it.

 

Some Google employees who had spent years working on the technology seethed at being lapped. Others were stunned at how quickly the public engaged with ChatGPT.

 

Analysts and investors wondered whether Google was missing technology's next big wave. They were asking how quickly the company could launch its own AI products, and whether the rise of chatbots would erode Google's search and advertising businesses, which had brought in $254 billion in revenue in 2022.

 

Dean and Hassabis, Google's two veteran AI scientists, and James Manyika, a roboticist who joined in 2022, worked to unite the DeepMind and Brain divisions in training AI. In January 2023, they presented to Alphabet's board of directors their plan for building the company's smartest model yet.

 

In the meantime, Google needed a chatbot to offer users -- and fast. The following month, it launched Bard, built off its LaMDA model. It botched the introduction.

 

In a video promoting Bard, Google showed it responding to a question about the James Webb Space Telescope. The chatbot inaccurately responded that the telescope took "the very first pictures" of a planet outside the solar system. The stumble sent Alphabet shares down 8%.

 

Around that time, Google co-founder Sergey Brin, who had recently retired, was at a party chatting with a researcher from OpenAI named Daniel Selsam, according to people familiar with the conversation. Why, Selsam asked him, wasn't he working full time on AI. Hadn't the launch of ChatGPT captured his imagination as a computer scientist?

 

ChatGPT was on its way to becoming a household name in AI chatbots, while Google was still fumbling to get its product off the ground. Brin decided Selsam had a point and returned to work.

 

For much of 2023, Google executives labored to coordinate and align its AI development efforts. The cultures of the Brain and DeepMind divisions were different, with the former more focused on research and the latter on building products, according to former employees, creating tension after they were combined.

 

Still, Google possessed one overwhelming advantage over its big rival. OpenAI had to raise money from investors; Google could fund research and development out of its multibillion-dollar profits. But Google also had to find a way to keep generative AI from killing its golden goose -- its 90% share of the web search market, the foundation of its advertising business.

 

To figure out what AI-driven search should look like, the company began a multiteam effort called Project Magi, led by Liz Reid, who became Google's vice president of search in 2024. The group's challenge, she explained in an interview, was to figure out how to revamp the search system to quickly present a clear answer to a question when the answer wasn't contained on a single webpage.

 

"People don't just use search, they rely on search," she said. "If you screw up, you're going to hear from your mom, you're going to hear from your friend, you're going to hear from your child."

 

Google released its first Gemini model before the end of 2023. While OpenAI initially had trained ChatGPT primarily on text, Google had trained Gemini on text, code, audio, images and video, which is one reason it took longer to develop, former employees said.

 

The first version of Gemini still lagged behind ChatGPT in many ways, but Google's technically more ambitious approach would pay dividends over time, just as its early research in neural networks had.

 

"I do think we still have benefited from that long history," Brin observed in December at a Stanford University event.

 

In May 2024, Google introduced AI Overviews -- short, AI generated summaries that often appear at the top of search results. What followed was the biggest overhaul of Google's search engine in years: the development of AI Mode, a search option that answers queries in a chatbot-style conversation. Internally, demo after demo showed what could be possible, but also how difficult it was to reprogram search to become chatbot-like while retaining speed and quality, Reid said.

 

Finally, after many iterations, Reid said, she and others on the team began seeing enough value to roll it out publicly. "We started to see ourselves seek it out, not just for testing, but being like, 'Oh, I want to use this,'" Reid recalled.

 

Google launched AI Mode last May. It also introduced Gemini 2.5, a more powerful version of its AI model, but it didn't generate as much buzz as many employees expected. Alphabet's share price, which had fallen since the start of the year, continued to languish over the summer.

 

The August introduction of Nano Banana boosted Google's shares. Josh Woodward, who oversees the Gemini app as well as Google Labs -- a proving ground of sorts for new AI applications -- called the launch a "success disaster." When people around the world began generating millions, and then billions, of images, Google was hard-pressed to find enough computing power to meet the demand. The company, he said, used emergency loans of server time to get more computing capacity.

 

By October, Gemini had more than 650 million monthly users, up from 450 million in July.

 

 

The November launch of Gemini 3 triggered another bottleneck in computing capacity. It is a problem that Google has been anticipating for more than a decade, and its solution -- the AI computer chips it developed -- is looking like a competitive edge. Its latest chip, called Ironwood, has helped significantly reduce the cost of running its AI models.

 

In an internal memo to employees this December, Pichai sounded a triumphant note. "We're ending 2025 in a great position," he wrote. "Thinking back to where we were as a company even just a year ago, it's incredible to see the progress."” [1]

 

1. How Google Got Its AI Groove Back --- Powerful new Gemini model allowed it to catch up to OpenAI's ChatGPT. Blunt, Katherine.  Wall Street Journal, Eastern edition; New York, N.Y.. 08 Jan 2026: A1.