Textbook of AI

DeepMind is the London-based artificial-intelligence research laboratory founded in November 2010 by Demis Hassabis, Shane Legg and Mustafa Suleyman. The mission, articulated from inception, was to "solve intelligence, and then use that to solve everything else". Hassabis -- a chess prodigy, Bullfrog/Lionhead games designer, and PhD in cognitive neuroscience from UCL -- brought an unusually deep neuroscience orientation that has remained a defining characteristic of the lab.

DeepMind was acquired by Google in January 2014 for a reported USD 500 million, then the largest European AI acquisition in history. It operated as a semi-independent subsidiary within Alphabet for a decade, with an internal ethics-and-safety board as a condition of the acquisition. In April 2023, in response to competitive pressure from OpenAI and the post-ChatGPT realignment of the AI industry, DeepMind merged with Google Brain to form Google DeepMind, with Hassabis as CEO of the combined entity and Jeff Dean as Chief Scientist of Google AI.

DeepMind has produced a long string of foundational results that have repeatedly redefined the frontier of AI:

• DQN (2013--2015): the Deep Q-Network combined Q-learning with deep convolutional networks and experience replay to achieve human-level performance on 49 Atari 2600 games from raw pixels, published as a Nature cover article in February 2015. DQN reignited modern deep reinforcement learning.
• AlphaGo (2016): combined Monte Carlo tree search with policy and value networks to defeat European champion Fan Hui (October 2015) and world champion Lee Sedol 4--1 in Seoul (March 2016), 19 years after Deep Blue's victory over Kasparov and decades earlier than Go experts had predicted.
• AlphaGo Zero (2017): learned Go entirely from self-play with no human data and surpassed AlphaGo within days.
• AlphaZero (2017): a single algorithm that mastered Go, chess and shogi from scratch through self-play.
• AlphaStar (2019): defeated top professional players at StarCraft II, a real-time, partial-information strategy game.
• AlphaFold (2018) and AlphaFold 2 (2020): solved the 50-year protein structure prediction problem, achieving CASP14 accuracy comparable to experimental crystallography. Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry with David Baker for this work.
• MuZero (2019): planned in a learned latent model without knowing game rules in advance.
• Gato (2022): a single transformer agent operating across 600+ tasks including Atari, captioning, chat and robotic control.
• Gemini family (2023--): Google DeepMind's frontier multimodal foundation models, competing with GPT-4/5 and Claude.
• AlphaProof and AlphaGeometry (2024): silver-medal performance at the International Mathematical Olympiad combining LLMs with symbolic search.
• AlphaFold 3 (2024): extended structure prediction to protein--ligand, protein--DNA and protein--RNA complexes.

The lab's research culture has unusually deep ties to neuroscience for an AI organisation, reflecting Hassabis's PhD background. Many DeepMind papers cross-reference neural mechanisms (hippocampal replay underlying experience replay; predictive coding informing world models; meta-learning frameworks for cognitive control). DeepMind has been a major contributor to AI safety research -- early work on reward hacking, scalable oversight, mechanistic interpretability and the frontier safety framework -- and historically maintained somewhat more cautious deployment positions than other frontier labs, although the post-2023 merger has shifted the balance toward faster product integration with Google's offerings.