Google’s DeepMind-Brain merger: tech giant regroups for AI battle
Start-up founder Demis Hassabis trades independence for greater influence over the future of artificial intelligence
Madhumita Murgia, Artificial Intelligence Editor
11 HOURS AGO
Financial Times
Demis Hassabis will lead the newly formed ‘Google DeepMind’ unit following the merger of the tech group’s two AI teams © FT montage/AFP via Getty Images
--------------------------
Google founder Larry Page convinced Demis Hassabis to sell his artificial intelligence company DeepMind with a promise. The London-based start-up would be shielded from pressure to make money in order to focus on a single goal: creating computer software that equals or surpasses human intelligence.
Since the £400mn deal in 2014, Hassabis has fought to maintain Page’s pledge and according to three people with knowledge of the efforts, has gone further still. DeepMind pushed for an independent legal status akin to a non-profit, with an independent governance board overseeing the powerful technology it was trying to build.
Such moves came to a close last week, however, when Sundar Pichai, Page’s successor as chief of Google parent Alphabet, announced that DeepMind would merge with Google Brain, the tech giant’s own AI lab headquartered in California.
The move means Hassabis will cede DeepMind’s cherished independence in return for greater power and influence over the future of AI. The newly formed “Google DeepMind” unit will be led by him, with a clear mission to develop “general AI systems” that are even more “capable and responsible”, and can be integrated into new products and services, according to Pichai.
Google’s reorganisation was triggered by the rise of OpenAI, the Microsoft-backed group which last November launched ChatGPT, a chatbot that provides plausible and nuanced text responses to questions.
The sudden popularity of ChatGPT, used by more than 100mn people as of January, shattered Google’s belief that it had a strong lead in the race to build and commercialise AI.
Microsoft has since announced it will embed OpenAI-built tools across its productivity software. Deep-pocketed rivals from Meta to Alibaba have announced large investments in AI products. Several well-resourced start-ups, including a nascent effort from Tesla and Twitter chief Elon Musk, are set to provide yet more competition.
Google also faces an existential dilemma. Its search business, which accounted for the majority of Alphabet’s annual revenues of $283bn in 2022, appears to be under direct threat. Large language models like ChatGPT can respond to queries with comprehensive written answers, rather than a set of links. These tools could provide millions of users with a route to bypass online advertising.
Seeking to respond to the threat, Google has reorganised so that Hassabis is more clearly in charge of the company’s fightback, according to people with knowledge of the move.
To accelerate the search giant’s efforts, Hassabis said his new team would “need to work with greater speed, stronger collaboration and execution, and to simplify the way we make decisions to focus on achieving the biggest impact”.
To take on OpenAI, the newly formed Google DeepMind unit will need to put aside years of rivalry, according to several former Google and DeepMind employees and their collaborators.
Although Google Brain was also staffed by elite AI researchers, the two teams had distinct cultures. London-based DeepMind was considered the company’s crown jewel, largely closed off from the rest of the group and taking a top-down approach, according to four former employees across both organisations.
It often worked on projects in stealth, although that became harder to do as they began to consume ever more computing power in Google’s data centres. Its mission to “solve” intelligence was sacred. Success for its employees was measured through publishing work in top-tier scientific journals like Nature. Three people close to DeepMind said its leaders are focused on “Nobel-level problems” or problems that, if solved, would be worthy of a Nobel Prize.
Meanwhile, California-headquartered Brain was more open and unstructured. It has contributed significantly to Google’s bottom line over the years, according to multiple former employees. Brain’s output has been integrated into search, ads, translation and other parts of Google’s business.
Despite these successes, employees say they had started to feel aimless. “I felt the organisational set-up [at Brain] was actually, from the get go, one that couldn’t scale,” said one former Google employee.
Brain’s leader Jeff Dean, a long-serving and highly respected AI scientist at the company, was loath to make the hard calls required to focus and scale ideas, such as addressing personnel issues or killing failing projects, the person said.
“If I had to sum up why OpenAI is where it is with such fewer resources and people compared to Google, it’s because they provided a clear mission. Brain had lost its way,” they added.
According to Pichai, last week’s shake-up would allow Dean to focus on research, leaving Hassabis in charge of an expanded group of scientists.
While the changes would probably “ruffle lots of feathers” and cause some churn in staff, this might be Google’s best chance at running “a tighter ship” and catching up to its competitors, according to people with knowledge of the merger.
Two AI teams are already responsible for major breakthroughs in artificial intelligence. DeepMind’s contributions include AlphaFold, software that can predict the shape of almost every known protein. Brain’s include the transformer, an AI technique that is the foundation of the new wave of generative-AI products, including ChatGPT and Google’s own chatbot Bard.
Over the past two years, Hassabis came to accept the need for greater collaboration with Google, according to people close to the company.
The amount of computing power and investment required to build so-called large language models, which can generate sophisticated and plausible text, has rapidly expanded. Recruiting AI talent is increasingly expensive. Microsoft announced this year that it would invest up to $10bn in OpenAI over multiple years to meet these challenges.
Hassabis’s elevation also means that the heart of Google’s AI operation shifts to London, a move with profound implications.
“How much influence DeepMind keeps?.?.?. greatly affects how responsible [Google] will be,” said Dan Hendrycks, research director of the Center for AI Safety. He noted that Shane Legg, a DeepMind co-founder, has been focused on AI safety and risk for more than a decade, while Google Brain did not have a dedicated safety or alignment team. Google said it had hundreds of people working on responsible AI across the company.
People close to the new division said they believed the merger would be beneficial for AI research overall. Pichai said the most “critical and strategic” AI project would involve building a series of “powerful, multimodal AI models”. These would be similar to GPT-4, which powers part of Microsoft’s Bing search engine, and LamDA, Google’s model that powers its Bard chatbot.
But it was not yet clear how the reorganisation would result in more ambitious commercial products, the people said. After all, DeepMind had never actively advocated the need to make money from its work. Those close to Hassabis argue that he understands now is the moment for change.
“Demis is very pragmatic, and highly, highly competitive,” one former DeepMind employee said. “He is ideological in terms of the vision, but he wants to be on the frontier of this. He’s not anti-commercial. This is the best way for him to do the job, or he wouldn’t have agreed to it.”
Google’s DeepMind-Brain merger: tech giant regroups for AI battle | Financial Times (archive.ph) |
