Next, Google needs to integrate the Gemini large model into its products without cannibalizing existing business shares such as search. The rise of OpenAI at the beginning of this year shocked the world, and Microsoft's embedding of ChatGPT into Bing made Google feel even more threatened. How did Google regain its foothold in the AI ​​competition in one year?

According to an article published by the media, the technology giant first calmed down the internal fighting and merged the two major artificial intelligence departments. Secondly, it accelerated the research and development of AI models and launched multi-modal models with the intention of catching up with OpenAI. Finally, with its comprehensive layout in chips, model frameworks, etc., it launched the powerful Gemini model to start a counterattack.

Now, Google has reached the most difficult step - integrating the Gemini large model into its products.

The following is the original text compiled by Wall Street News:

Whether it's Google or other companies, this is a moment worth looking back on, and they clearly defeated the doomsdayers.

Last week, Google released its highly anticipated new artificial intelligence model, Gemini, nearly a year after some critics predicted that OpenAI's ChatGPT could challenge Google's leadership in search. Google has aggressively taken on OpenAI, and its leadership has managed to get competing parts of the company to collaborate on a response, fending off those who thought the company had become too indecisive and bureaucratic. Today, Google still dominates the search field, and even its stock has been rising this year.

Now the hard work begins. Over the next few months, Google must prove that it can integrate Gemini into its products without cannibalizing existing business shares such as search.

The company has implemented a stripped-down version of Gemini, the chatbot Google created to compete with ChatGPT, into Bard, but adoption so far appears to be limited. In the future, the company plans to use Gemini in nearly its entire product line, from search engines to productivity apps and Pixie, an AI assistant unique to Pixel devices, two people familiar with the matter said. Products could also include wearable devices, such as glasses, that leverage artificial intelligence's ability to recognize objects the wearer sees, and the device can then provide suggestions to the wearer, such as how to use a tool, solve a math problem or play an instrument, according to a person familiar with internal discussions.

While Google is doing these things, it must also carefully deal with regulators. The company is in the middle of two antitrust battles over its search engine and advertising businesses. Antitrust officials are watching Google's artificial intelligence efforts closely as an example of how the company can use its strengths in one area to win in another. In this case, Google used website data from its search, as well as data from billions of customers, to train the new artificial intelligence.

The success of artificial intelligence is also crucial to Google's cloud computing business. Artificial intelligence models with extremely high computing requirements are stimulating demand for cloud services. Microsoft has established a close partnership with OpenAI, and its cloud business has grown faster than Google and Amazon.

While ChatGPT is best known for its conversational AI, the technology behind it has proven to be highly effective across the business spectrum, from automating customer service and software coding, to quickly generating marketing proposals, and helping Wall Street firms make sense of large amounts of data. The problem for Google is that OpenAI and Microsoft already have a head start in selling the technology to consumers and businesses, providing them with valuable data and feedback that they can use to improve their products.

"Only after you try it yourself can you judge what people can create with it. We have just begun to see this, but what we have seen is very remarkable," said Jon Turow, a partner at Madrona Venture Group, who was responsible for artificial intelligence products in Amazon's cloud computing department.

Gemini is one of the biggest efforts in Google's 25-year history. As Google enters middle age, its core advertising business continues to generate huge profits, which has also funded a series of bets by its parent company Alphabet in new businesses such as self-driving cars, health insurance and biotechnology. But these decade-long bets have not paid off.

As a result, investors are increasingly asking Google's leadership to cut costs by 182,000 people, and this year's massive layoffs have hit employee morale. Meanwhile, Google is preparing for more layoffs in the new year, but it's unclear whether the cuts will be widespread or targeted at specific groups.

Artificial intelligence is another big bet that requires companies to invest huge amounts of money to pay for everything from people to hardware. A person close to the Gemini team said that Google needs to invest heavily to fend off the threat of the artificial intelligence team defecting to OpenAI.

Google has also chosen a particularly expensive technical approach, designing its own AI chips. The decision makes Google hardware independent from Nvidia, the main supplier of artificial intelligence server chips. Competitors such as OpenAI rely on hardware produced by Nvidia and other companies.

Google also wants to dispel the perception that its achievements are the result of innovations dating back decades. Over the years, Google has invested heavily in artificial intelligence research through two separate divisions, Google Brain and DeepMind. Google even invented the underlying technology of transformer, which is the core of the GPT series of artificial intelligence models created by OpenAI.

However, the rise of OpenAI has raised concerns that Google, like many other established tech giants before it, may lose its leadership position in technology. Inside Google, Google executives were particularly upset when Microsoft incorporated ChatGPT into its Bing search engine in February, according to a person with direct knowledge of discussions between Google and Microsoft.

A Google spokesman declined to comment.

Infighting in the AI ​​department

Sundar Pichai, the chief executive of Google and Alphabet, has complained to colleagues for years that he couldn't get his two artificial intelligence research units to collaborate. Google acquired DeepMind in 2014. The company's CEO, Demis Hassabis, has long insisted on independence from the parent company. He believes that such an arrangement can better achieve the company's goal of developing general artificial intelligence.

Meanwhile, DeepMind's sister division, Google Brain, focuses on how to apply AI to Google's products and has incubated important advances in machine learning, such as the transformer, an invention that paves the way for Google and other companies to train more complex models. The unit is led by Jeff Dean, a veteran engineer whose coding work helped Google expand its search engine to billions of users in its early days.

The rift between the two divisions runs deep, with Google Brain based at Google's headquarters in Mountain View, California, while Hassabis and his team are based in offices near King's Cross Station in London.

One person who has worked at DeepMind said that as the company grew in size, DeepMind went to great lengths to avoid working with Google Brain. For example, it opened offices in cities where Google Brain had no major operations at the time, such as Paris and Edmonton, Alberta. DeepMind researchers can access code written by Google Brain, but not the other way around. Some employees saw this as a sign of DeepMind's excessive secrecy, even among Google employees. When Hassabis wants to take steps to keep DeepMind independent, he communicates directly with co-founder Larry Page, who led the acquisition along with co-founder Sergey Brin and owns a controlling stake in Alphabet.

Over time, Hassabis wants to separate DeepMind more completely from Google as he grows concerned about how the giant company might use the technology, including selling it to the military, a person familiar with the matter said. He came up with the idea of ​​creating an independent company that would own DeepMind's intellectual property. But in 2021, Hassabis told DeepMind employees that efforts to break away from Google were over after Pichai pledged to give the company more money for purposes including AI ethics.

Over the years, competition for resources has heightened tensions between the two companies. Google distributes a limited number of server chips to its artificial intelligence researchers. The chips are becoming scarcer as an industry-wide craze for artificial intelligence boosts demand for them.

At the same time, as Google executives became mired in internal politics, the company's famous artificial intelligence researchers began to leave one after another. Some of them formed their own companies, frustrated by Google's bureaucratic culture, which had hindered the launch of ChatGPT even before OpenAI launched a service like it. Others were acquired by OpenAI, a startup that Musk and other prominent figures launched as a nonprofit in 2015, in part because they feared Google would own an artificial intelligence future. One of the founders of OpenAI is Ilya Sutskever, a key engineer at Google Brain who later led advances such as creating artificial intelligence that can use reasoning to solve problems it has never encountered before.

When OpenAI released ChatGPT last November, the public reaction sent shockwaves throughout Google. The 400-person startup actually beat Google to the punch with a chatbot that can convincingly answer questions on a variety of topics, calling into question Google's competitiveness.

However, some leaders at Google don't seem to be scared by this new Internet darling. At a staff meeting weeks after ChatGPT launched, Dean said in response to a question about the chatbot that Google would not react to what other startups were doing, a person familiar with the matter said.

But in February, Microsoft announced that it would use ChatGPT in its Bing search engine. Some investors believe chatbots could undermine Google's dominance in search, a view that Google executives feel is dangerous.

fusion of ideas

Google needs to do something, and fast.

As a result, Google pieced together the Bard chatbot in just a few months and officially released it in March this year. Within Google, the effort caused uproar after Jacob Devlin, a prominent Google researcher, resigned after raising concerns with Pichai and other executives about Google's use of ChatGPT's data to train artificial intelligence models. He immediately joined OpenAI, but just a few months later, he returned to Google for unknown reasons.

Another response from Google was to finally end the infighting between DeepMind and GoogleBrain. Google selected researchers from both teams to build a new model: the Gemini model, led by Dean and DeepMind senior researcher Oriol Vinyals, who had worked with Dean on the brain.

In April this year, Google announced the merger of Google Brain and DeepMind. Hassabis took over the new entity, Google DeepMind, while Dean took a back seat and became Google's chief scientist. The move shocked many Google engineers, who believed that Dean should have been the leader of the department given his accomplishments and long tenure at the company.

Leaders tried to cast the merger as a victory for the combined units, with Zoubin Ghahramani, Google Brain's vice president of research, visiting DeepMind's London offices to explain the rationale for the reorganization to employees at an all-hands meeting the week the changes were announced. GoogleBrain held a separate meeting for its employees. Hassabis told employees that Google DeepMind will bring together two of the best artificial intelligence research teams in the world.

But Google's artificial intelligence employees soon realized that their priorities were changing. Google DeepMind’s leadership scaled back research projects that were not critical to building competitive artificial intelligence products. Projects that lost resources include a multimodal model called Gato and a research team called GenRL that builds artificial intelligence systems capable of navigating virtual environments like those in Atari games, these people said.

Executives say the changes have the added benefit of reducing overlap and cutting back on lower-priority projects, meaning employees no longer have to fight to get chips for research.

In Mountain View, artificial intelligence employees spread across multiple buildings on the company's campus moved into an office in the center of the campus with the goal of increasing collaboration among researchers.

As the impact of OpenAI's explosive rise fades, Google finally has its chance to fight back.

"Secret Weapon"

Still, Google faces a huge challenge: building a model that outperforms OpenAI's state-of-the-art model, GPT-4.

From the outset, this meant the researchers had to meet a deadline to develop the model. Employees worked around the clock to complete tasks within tight deadlines, a top-down approach that was far different from Google's past hands-off approach to research labs. One person close to the job said 80-hour weeks had become the norm for some employees.

Even outside of the company's artificial intelligence, Google employees are being asked to quickly master the technology. According to two people familiar with the matter, during the year, Google Cloud required employees to pass tests on artificial intelligence and provided additional materials to employees in non-technical positions such as sales to promote employees to improve their knowledge of artificial intelligence.

Google aims to gain an advantage over OpenAI by giving Gemini the ability to understand a variety of different media, including text, images, video and audio. For example, the AI ​​can interpret and illustrate the contents of complex charts in plain English. Pichai later said Gemini would be trained from scratch on these types of data. Pichai is well aware that OpenAI announced similar image recognition capabilities for GPT-4 in March, but those capabilities were not initially widely available. This gives Google the opportunity to release an extensive set of multi-modal patterns through Gemini ahead of the release of OpenAI.

Google also has a secret weapon: YouTube. Google researchers rely heavily on Google-owned streaming services, two people familiar with the matter said. The data, which includes images, videos and audio subtitle text, is invaluable for training artificial intelligence models.

This gives Google access to a much richer information base than competitors such as OpenAI and AI imaging startup Midjourney. It also means Google must meet legal department demands, such as ensuring that if a YouTube user deletes a video, Google also removes that content from the data sets used by its models, these people said.

Another advantage of Google is computing power. Unlike OpenAI, which relies on Microsoft servers, Google has its own data center. In order to run the software more efficiently, Google even built its own dedicated artificial intelligence chip - the tensor processing unit (TPU). Google has accumulated an astonishing number of chips for the Gemini project-77,000 fourth-generation tensor processing units code-named Pufferfish. In the third quarter, Google's unallocated enterprise costs, including spending on DeepMind, jumped nearly 40% to $1.6 billion.

As employees in London finish their work day and those in Mountain View begin their day, Gemini leaders stay informed on the researchers' progress through daily meetings with employees overseeing parts of the project. One person said the meeting was chaired by Dean, Vinyals and Vice President of Research Koray Kavukcuoglu.

Senior managers are also working in the trenches, with Dean spearheading efforts to improve software that helps the company's algorithms process massive amounts of data. Co-founder Sergey Brin had distanced himself from Google amid an office romance scandal, but he worked side by side with Gemini researchers in Mountain View and regularly had lunch with them in the company cafeteria.

There have been some awkward moments along the way. During a demo for Bard in February, the chatbot made a factual error about the James Webb Space Telescope, embarrassing Google as it struggles to catch up with OpenAI. Two days later, when the error became widely known, Google's stock price fell 9%.

Google first disclosed Gemini's existence in May during a presentation at its annual developers conference, and Wall Street was impressed: The company's stock price jumped more than 4% that day.

Accumulating strength to catch up

Over the next few months, Google moved closer to releasing Gemini, and in September Google let some developers use a smaller version of Gemini for testing.

But in the same month, OpenAI launched GPT-4withVision, beating Google in multi-modal functionality, which also brought more attention to its technology and new business. Consumer adoption of Bard has frustrated some executives within the team, said a person close to the team. On the same day that Microsoft announced an impressive 29% revenue growth for its Azure cloud computing unit, Google disclosed in October that its cloud computing unit's third-quarter revenue growth was weak, with only 22% growth. This only adds to the pressure on the Gemini team to do something big.

Then, around November, during an administrative review of new products, the state-of-the-art Gemini model that was supposed to compete with GPT-4 struggled to work properly in languages ​​other than English.

What gives Google some relief is that OpenAI is also dealing with its own problems. In mid-2023 OpenAI scrapped an important new model called Arrakis because it performed poorly in training. At the end of November, OpenAI's board of directors fired CEO Sam Altman, and OpenAI nearly collapsed. Altman's return to OpenAI after this incident seemed to steady the ship for the time being.

Finally, in early December, Google pulled back the curtain on Gemini. It published test results showing that the most powerful version, GeminiUltra, outperformed GPT-4 on several industry-standard benchmarks, although many researchers have disputed these claims. Google itself admitted that the video exaggerated Gemini's capabilities. The release of the video frustrated some rank-and-file employees at the company who had not seen the video beforehand, but the flurry of statements sent a strong message: Google is ready to compete.

One executive even went on the offensive, slamming Microsoft for relying on OpenAI to develop its cutting-edge technology. Kent Walker, president of global affairs for Google and Alphabet, said at an event hosted by news outlet Semafor that the company does not believe in "outsourcing" artificial intelligence development.

Now, the test facing Google is to integrate Gemini into various product applications, just like Microsoft has done with OpenAI technology. But Google has an advantage that its rivals don't: its portfolio of Pixel hardware devices, including phones, watches and earbuds, all benefit from artificial intelligence. One version of Gemini is specifically designed to run on Pixel phones, which use Google's customized artificial intelligence chip.

The launch of an AI assistant dedicated to Pixel devices could boost Google's hardware business at a time when tech companies are racing to integrate their own hardware with new AI capabilities. Pixie will use information from customers' phones, including data from products like Google Maps and Gmail, to evolve into a more personalized Google Assistant, according to a person familiar with the matter. The person familiar with the matter said that this feature will be launched on Pixel 9 and 9 Pro as soon as next year.

Eventually, Google hopes to bring this feature to devices such as regular phones and watches, and the company will need increasingly advanced models to support all of its product ideas. However, Google seems to be seizing time to ensure that it doesn't find itself in trouble again. According to a person familiar with the situation, Google is already training the next generation large model Gemini2.