Why did Google buy DeepMind in 2014, and how much did it pay?

On Sunday, 26 January 2014, the technology news site Re/code reported that Google had agreed to buy DeepMind, a London company most people had never heard of. The next day, Google confirmed the deal. It did not say how much it was paying, and it did not say what it planned to do with DeepMind's work.

The purchase looked small next to Google's other deals that month, yet it has proved one of the most consequential in AI. It placed a young research lab with a very large goal, building general-purpose learning systems, inside one of the richest technology companies, and it signaled that big tech firms were competing hard for AI talent.

What DeepMind was

DeepMind Technologies was founded in London in September 2010 by three people. Demis Hassabis had been a chess prodigy and a video game designer, co-designing the hit game Theme Park at 17, before earning a PhD in cognitive neuroscience at University College London in 2009. Shane Legg, a machine learning researcher, met Hassabis at UCL's Gatsby Computational Neuroscience Unit. Mustafa Suleyman was the third co-founder.

The company described its approach as combining machine learning with systems neuroscience, the study of how the brain works as a whole, to build general-purpose learning algorithms. Its backers included the venture firms Horizons Ventures and Founders Fund, as well as investors Peter Thiel, Elon Musk and Jaan Tallinn.

Its best-known work was a paper posted in December 2013 showing a single neural network learning to play several Atari video games from the screen pixels alone, beating human experts on some. According to later accounts, that research helped draw Google's interest.

The deal and the price

Here is how the main reports compared in the days after the news:

The figures do not agree, and they cannot be checked against an official number, because neither company published one. The Guardian's own story mixes the figures, calling it £400 million in the text and around $400 million in a photo caption. The safest summary is that Google paid somewhere in the range of $400 million to $650 million.

Other details came out in the reporting. The Guardian said Google's chief executive Larry Page was said to have led the deal himself, and that Facebook had reportedly tried and failed to buy DeepMind earlier. Citing the news site The Information, it also reported that Google agreed to set up an AI ethics board as part of the deal. In a statement, Hassabis said joining Google would let DeepMind "turbo-charge" its mission. The company became Google DeepMind.

Why Google wanted it

The deal fit a pattern. In 2013 Google had bought seven robotics companies, including Boston Dynamics, and in January 2014 it announced a $3.2 billion purchase of Nest, which makes learning thermostats. The New York Times framed DeepMind as part of Google's plans for driverless cars and a new generation of robots.

There was also a talent race. Murray Shanahan, a professor at Imperial College London, told The Guardian that DeepMind focused on reinforcement learning and deep learning, which suit Google's huge stores of data. In March 2013 Google had hired Geoffrey Hinton, whose Toronto team had won the 2012 ImageNet contest. Buying DeepMind brought in a whole team of people with the same rare skills at once.

What has DeepMind done since joining Google?

DeepMind is still based in London, and in 2015 it became a subsidiary of Alphabet, Google's new parent company. Its first big public moment came in October 2015, when its program AlphaGo beat the European Go champion Fan Hui five games to none. In March 2016, AlphaGo beat Lee Sedol, one of the world's top players, 4 to 1. Go had long been considered much harder for computers than chess, because the number of possible positions is vastly larger.

Since then, AlphaZero taught itself Go, chess and shogi through play against itself, and in 2020 Agent57 beat the human benchmark on all 57 Atari games, closing the loop on the work that made DeepMind famous. In late 2020 DeepMind also announced major progress on protein folding, the problem of predicting a protein's 3D shape from its sequence.

The partnership has had friction too. The membership of the ethics board agreed in 2014 has never been made public, and in 2018 privacy advocates objected when DeepMind's health app Streams was folded into Google. Those debates, about who controls powerful AI research and the data it uses, are likely to grow as AI labs inside big companies become more influential.