Learning from games
DeepMind began in 2010 as a research lab developing general AI methods. One early contribution, DQN, learned to play Atari games from pixels and reward feedback. The lab's history spans several methods and research programs; it should not be reduced to one unchanged algorithm applied to every later problem. [1]
Google acquired DeepMind in January 2014. The lab continued research under Google ownership before the 2023 organizational combination. It was therefore not an independent company throughout the entire period covered by this historical record. [2][6]
AlphaGo beat Lee Sedol four games to one in 2016, combining learning and search in Go. Subsequent self-play work generalized to other board games: the AlphaZero paper described learning chess and shogi without human game examples. Those results established particular advances in game-playing systems, rather than an unrestricted ability to solve every domain. [3][4]
Protein structures and the successor lab
AlphaFold 2's CASP14 result in 2020 applied deep learning to protein-structure prediction. It was a different scientific system from the board-game agents. Its accuracy result concerned predicting structures, not a complete simulation of the physical folding process or a solution to every biological problem. [5]
In April 2023, Google combined DeepMind and the Brain team into Google DeepMind. This record ends at that organizational boundary. Google DeepMind's later work, including Gemini, belongs to the successor lab's chronology. [6]