DeepMind History: AlphaGo, AlphaFold & Google Merger

A lab bought for a method, not a product

Demis Hassabis, Shane Legg and Mustafa Suleyman founded DeepMind in London in 2010, with a stated aim of building general-purpose AI. The lab's early work had no product attached to it: it taught a single learning algorithm to play classic Atari games from raw pixels, with no rules programmed in and no game-specific tuning. [3]

Google bought DeepMind in January 2014. Neither company has ever disclosed an exact figure: contemporary reports ranged from a low of $400 million to more than $500 million, for a company of roughly 75 people with no shipped product. Investors at the time included Founders Fund and Horizons Ventures, and Facebook had reportedly also pursued the lab before Google closed the deal. [3][4]

Atlas interpretation: What Google paid for was the research method, not a roadmap. The eight years between the acquisition and the 2023 merger bore that out: DeepMind's headline results were a series of demonstrations that one reinforcement-learning approach, refined, could out-compete specialists in unrelated domains. [3]

Self-play beats the humans, then the need for them

In March 2016, AlphaGo beat Lee Sedol four games to one in Seoul. AlphaGo had trained on human professional games before refining itself through self-play, and its win against a top-ranked human player at a game long treated as too intuitive for machines was DeepMind's first result to reach a general audience. [5]

Eighteen months later, AlphaGo Zero removed the human data entirely. Given only the rules of Go, it trained purely against itself, starting from random play, and beat the version that had beaten Lee Sedol by 100 games to nil after three days of training. AlphaZero then took the same self-play approach and applied it to chess and shogi, learning both well enough to beat the strongest existing engines in each. [6][7]

Atlas interpretation: The progression from AlphaGo to AlphaGo Zero is the more important of the two results. Beating a human champion is a headline; training a system that needs no human examples at all, and still generalizes across three different games, is the demonstration that self-play could be a general method rather than a Go-specific trick. That is the part of DeepMind's work later systems across the field borrowed from. [6]

The same toolkit solves a biology problem

DeepMind entered the CASP protein-structure prediction competition for the first time in 2018. AlphaFold placed first at CASP13 against groups that had spent entire careers on the problem, ahead of specialist structural biology labs on its debut attempt. [8]

Two years later, AlphaFold 2 crossed a different threshold at CASP14, with a median accuracy score of 92.4 GDT and an average error comparable to the width of an atom. CASP's co-founder, John Moult, said the field had been stuck on the protein-folding problem for nearly fifty years and called AlphaFold 2's result a solution; Nobel laureate Venki Ramakrishnan called it a stunning advance decades ahead of where the field expected to be. [9]

Atlas interpretation: AlphaFold is the clearest evidence that DeepMind's research was not a games hobby with a marketing budget. The same deep-learning toolkit that mastered Go and chess produced, on a four-year timeline, a tool structural biologists actually use, which is a harder and slower kind of proof than winning a match. [9]

The independent lab ends in a merger

In April 2023, Google CEO Sundar Pichai announced that DeepMind would combine with the Brain team from Google Research to form Google DeepMind, one unit under Demis Hassabis as CEO, with Jeff Dean moving to the newly created role of Google Chief Scientist. Pichai described the merger as concentrating talent and compute to accelerate progress inside Alphabet, at a moment when the pace of AI releases elsewhere had picked up. [10]

Atlas interpretation: This page covers DeepMind as the independent London lab, from its 2010 founding through that April 2023 merger. Its subsequent work under Google DeepMind, including the Gemini model line, is a different organization's record in this Atlas, not an extension of this one. [10]