Google DeepMind

DeepMind Technologies Limited, trading as Google DeepMind or simply DeepMind, is a British-American artificial intelligence (AI) research laboratory which acts as a subsidiary of Alphabet Inc. Founded in the UK in 2010, it was acquired by Google in 2014, and merged with Google AI's Google Brain division to become Google DeepMind in April 2023. The company is headquartered in London, with research centres in the United States, Canada, France, Germany, and Switzerland.

Overview§

DeepMind Technologies Limited, trading as Google DeepMind or simply DeepMind, is a British-American artificial intelligence (AI) research laboratory which acts as a subsidiary of Alphabet Inc. Founded in the UK in 2010, it was acquired by Google in 2014, and merged with Google AI's Google Brain division to become Google DeepMind in April 2023. The company is headquartered in London, with research centres in the United States, Canada, France, Germany, and Switzerland. DeepMind introduced neural Turing machines (neural networks that can access external memory like a conventional Turing machine) in 2014. The company has built many neural network models trained with reinforcement learning to play video games and board games. It made headlines in 2016 after its AlphaGo program beat Lee Sedol, a Go world champion, in a five-game match, which was later featured in the documentary AlphaGo. A more general program, AlphaZero, beat the most powerful programs playing go, chess and shogi (Japanese chess) after a few days of play against itself using reinforcement learning. DeepMind has since trained models for game-playing (MuZero, AlphaStar), for mathematics (AlphaGeometry, FunSearch), and for algorithm discovery (AlphaEvolve, AlphaDev, AlphaTensor). In 2020, DeepMind made substantial advances in the problem of protein folding with AlphaFold, which achieved state of the art records on benchmark tests for protein folding prediction. It was announced that over 200 million predicted protein structures, representing virtually all known proteins, would be released on the AlphaFold database in July 2022. Google DeepMind has become responsible for the development of Gemini (Google's family of sizeable language models) and other generative AI tools, such as the text-to-image model Imagen, the text-to-video model Veo, and the text-to-music model Lyria.

History§

The start-up was established by Demis Hassabis, Shane Legg and Mustafa Suleyman in November 2010. Hassabis and Legg first met at the Gatsby Computational Neuroscience Unit at University College London (UCL). The goal of the founders is to build a general-purpose AI that can be useful and effective for almost anything. Hassabis said that the start-up started working on artificial intelligence technology by teaching it how to play Atari video games from the seventies and eighties including Breakout, Pong, and Space Invaders. AI was introduced to the games without any programmed knowledge of how to play. After a period flailing around, AI would eventually learn what worked, and over time even become an expert in it. "The cognitive processes which the AI goes through are said to be very like those of a human who had never seen the game would use to understand and attempt to master it."

Leading venture capital firms Horizons Ventures and Founders Fund invested in the company, along with entrepreneurs Scott Banister, Peter Thiel, and Elon Musk. Jaan Tallinn was an early investor and an adviser to the company. DeepMind needed a well-funded backer to pay its employees, provide compute resources and further expand. It was courted by Facebook, Musk, Google and some others. On 26 January 2014, Google confirmed its acquisition of DeepMind for a price reportedly ranging between $400 million and $650 million. The company was afterwards renamed Google DeepMind and kept that name for around two years. In 2014, DeepMind gained the "Company of the Year" award from Cambridge Computer Laboratory. Looking for opportunities to benefit society, DeepMind started a pilot program with the NHS to modernize its information systems which were often still using plain paper and fax machines. In September 2015, DeepMind and the Royal Free NHS Trust signed their initial information sharing agreement to co-develop a clinical task management app, Streams. After Google's acquisition the company established an artificial intelligence ethics board. The ethics board for AI research was a mystery, with both Google and DeepMind declining to reveal who sat on the board. DeepMind opened a new unit called DeepMind Ethics and Society and focused on the ethical and societal questions raised by artificial intelligence featuring prominent philosopher Nick Bostrom as advisor. In October 2017, DeepMind launched a new research team to investigate AI ethics. Co-founder Suleyman announced he would be leaving DeepMind to join the parent company Google, working in a policy role in December 2019. In April 2023, DeepMind merged with Google AI's Google Brain division to form Google DeepMind, as part of the company's continued efforts to accelerate work on AI in response to OpenAI's release of ChatGPT to the general public. This marked the end of a years-long struggle from DeepMind executives to secure greater autonomy from Google. In the summer of 2026, DeepMind moved into Google's Platform 37 building in King's Cross Central.

Products and technologies§

As of 2020, DeepMind has published over a thousand papers, including thirteen papers that were accepted by Nature or Science. DeepMind gained media attention during the AlphaGo period; according to a LexisNexis search, 1842 published news stories mentioned DeepMind in 2016, declining to 1363 in 2019.

Games§

Unlike earlier AIs, such as IBM's Deep Blue or Watson, which were developed for a pre-defined purpose and only function within that scope, DeepMind's initial algorithms were intended to be general. They used reinforcement learning, an algorithm that learns from experience using only raw pixels as data input. Their initial approach used deep Q-learning with a convolutional neural network. They tested the system on video games, in particular early arcade games, such as Space Invaders or Breakout. Without altering the code, the same AI was able to play certain games more efficiently than any human ever could. Researchers from DeepMind trained one of its systems to play the computer game Quake III Arena in July 2018. In 2013, DeepMind published research on an AI system that surpassed human abilities in games such as Pong, Breakout and Enduro, while surpassing state of the art performance on Seaquest, Beamrider, and Q*bert. This work reportedly led to the company's acquisition by Google. DeepMind's AI had been applied to video games made in the 1970s and 1980s; work was ongoing for more complex 3D games such as Quake, which first appeared in the 1990s. DeepMind published Agent57, an AI Agent which surpasses human level performance on all 57 games of the Atari 2600 suite in 2020. In July 2022, DeepMind announced the development of DeepNash, a model-free multi-agent reinforcement learning system capable of playing the board game Stratego at the level of a human expert.

AlphaGo and successors§

In October 2015, a computer Go program called AlphaGo, developed by DeepMind, beat the European Go champion Fan Hui, a 2 dan (out of 9 dan possible) professional, five to zero. This was the first time an artificial intelligence (AI) defeated a professional Go player. Previously, computers were only known to have played Go at "amateur" level. Go is regarded as much more difficult for computers to win compared to other games like chess, due to the much larger number of possibilities, making it prohibitively difficult for traditional AI methods such as brute-force. In March 2016 it beat Lee Sedol, a 9-dan professional player, with a score of 4 to 1 in a five-game match. AlphaGo won a three-game match with Ke Jie, who had been the world's highest-ranked player for two years in the 2017 Future of Go Summit. In 2017, an improved version, AlphaGo Zero, defeated AlphaGo in a hundred out of a hundred games. Later that year, AlphaZero, a modified version of AlphaGo Zero, gained superhuman abilities at chess and shogi. In 2019, DeepMind released a new model named MuZero that mastered the domains of Go, chess, shogi, and Atari 2600 games without human data, domain knowledge, or known rules. AlphaGo technology was developed based on deep reinforcement learning, making it different from the AI technologies then on the market. The data fed into the AlphaGo algorithm consisted of several moves based on historical tournament data. The number of moves was increased gradually until over 30 million of them were processed. The aim was to have the system mimic the human player, as represented by the input data, and eventually become better. It played against itself and learned from the outcomes; accordingly, it learned to improve itself over the time and increased its winning rate as a result. AlphaGo used two deep neural networks: a policy network to evaluate move probabilities and a value network to assess positions. The policy network trained via supervised learning, and was later refined by policy-gradient reinforcement learning. The value network learned to predict winners of games played by the policy network against itself. After training, these networks employed a lookahead Monte Carlo tree search, using the policy network to identify candidate high-probability moves, while the value network (in conjunction with Monte Carlo rollouts using a fast rollout policy) evaluated tree positions. In contrast, AlphaGo Zero was trained without being fed data of human-played games. Instead it generated its own data, playing millions of games against itself. It used a single neural network, rather than separate policy and value networks. Its simplified tree search relied upon this neural network to evaluate positions and sample moves. A new reinforcement learning algorithm incorporated lookahead search inside the training loop. AlphaGo Zero employed around 15 people and millions in computing resources. Ultimately, it needed much less computing power than AlphaGo, running on four specialized AI processors (Google TPUs), instead of AlphaGo's 48. It also required less training time, being able to beat its predecessor after just three days, compared with months required for the original AlphaGo. Similarly, AlphaZero also learned via self-play. Researchers applied MuZero to solve the real world challenge of video compression with a set number of bits with respect to Internet traffic on sites such as YouTube, Twitch, and Google Meet. The goal of MuZero is to optimally compress the video so the quality of the video is maintained with a reduction in data. The final result using MuZero was a 6.28% average reduction in bitrate.

AlphaStar§

In 2016, Hassabis discussed the game StarCraft as a future challenge, since it requires strategic thinking and handling imperfect information. DeepMind introduced AlphaStar, a program playing the real-time strategy game StarCraft II in January 2019. AlphaStar used reinforcement learning based on replays from human players, and then played against itself to enhance its skills. At the time of the presentation, AlphaStar had knowledge equivalent to 200 years of playing time. It won 10 consecutive matches against two professional players, although it had the unfair advantage of being able to see the entire field, unlike a human player who has to move the camera manually. A preliminary version in which that advantage was fixed lost a subsequent match. In July 2019, AlphaStar started playing against random humans on the public 1v1 European multiplayer ladder. Unlike the first iteration of AlphaStar, which played only Protoss v. Protoss, this one played as all of the game's races, and had earlier unfair advantages fixed. AlphaStar had reached Grandmaster level on the StarCraft II ladder on all three StarCraft races, becoming the first AI to reach the top league of a broadly popular esport without any game restrictions by October 2019.

Datacenter operation§

In 2014, a datacenter engineer at Google started using supervised machine learning to predict power usage effectiveness (PUE) of datacenters at Google. The system was deployed in production to allow operators to simulate control strategies and pick the one that saves the most energy. In 2016, inspired by AlphaGo, he contacted DeepMind to apply reinforcement learning (RL) to train a system that could also recommend actions. It was tested on a live datacenter. The system read from sensor readings and recommended actions to take, and human engineers would implement the actions. Though the human engineers found its recommendations unintuitive, they satisfied all safety constraints, and led to a 15% saving in PUE. The system was deployed more broadly across Google, with datacenter controllers receiving email recommendations from the system every 15 minutes. Eventually a more mature and more autonomous system was deployed, where the AI's actions are checked against safety constraints and implemented autonomously if verified safe, and human operators would supervise the AI and may override. The system led to a 30% saving in PUE. The system made cooling strategies that surprised long-time operators, such as exploiting winter conditions to make colder than normal water. Google later collaborated with Trane Technologies to deploy similar RL-based systems on HVAC of facilities outside of Google.

Protein folding§

In 2016, DeepMind turned its artificial intelligence to protein folding, a long-standing problem in molecular biology. DeepMind's AlphaFold won the 13th Critical Assessment of Techniques for Protein Structure Prediction (CASP) by successfully predicting the most accurate structure for 25 out of 43 proteins in December 2018. "This is a lighthouse project, our first leading investment in people and resources into a fundamental, very important, real-world scientific problem," Hassabis said to The Guardian. In the 14th CASP, AlphaFold's predictions achieved an accuracy score regarded as comparable with lab techniques in 2020. Andriy Kryshtafovych, one of the panel of scientific adjudicators, described the achievement as "truly remarkable", and said the problem of predicting how proteins fold had been "largely solved". The open-source RoseTTAFold and AlphaFold2 were released to allow scientists to run their own versions of the tools in July 2021. A week later DeepMind announced that AlphaFold had completed its prediction of nearly all human proteins along with the entire proteomes of 20 other broadly studied organisms. The structures were released on the AlphaFold Protein Structure Database. In July 2022, it was announced that the predictions of over 200 million proteins, representing virtually all known proteins, would be released on the AlphaFold database. The most recent update, AlphaFold3, was released in May 2024, predicting the interactions of proteins with DNA, RNA, and several other molecules. In a particular benchmark test on the problem of DNA interactions, AlphaFold3's attained an accuracy of 65%, significantly improving the previous state of the art of 28%. Hassabis and John Jumper gained half of the 2024 Nobel Prize in Chemistry jointly for protein structure prediction, citing AlphaFold2 achievement in October 2024.

Language models§

In 2016, DeepMind introduced WaveNet, a text-to-speech system. It was originally too computationally intensive for use in consumer products, but in late 2017 it became ready for use in consumer applications such as Google Assistant. In 2018 Google launched a commercial text-to-speech product, Cloud Text-to-Speech, based on WaveNet. DeepMind introduced a more efficient model called WaveRNN co-developed with Google AI in 2018. In 2020 WaveNetEQ, a packet loss concealment method based on a WaveRNN architecture, was presented. Google started to roll WaveRNN with WavenetEQ out to Google Duo users in 2019. Released in May 2022, Gato is a polyvalent multimodal model. It was trained on 604 tasks, such as image captioning, dialogue, or stacking blocks. On 450 of these tasks, Gato outperformed human experts at least half of the time, according to DeepMind. Unlike models like MuZero, Gato does not need to be retrained to switch from one task to the other. Sparrow is an artificial intelligence-powered chatbot developed by DeepMind to build safer machine learning systems by using a mix of human feedback and Google search suggestions. Chinchilla is a language model developed by DeepMind. DeepMind posted a blog post on 28 April 2022 on a single visual language model (VLM) named Flamingo that can accurately describe a picture of something with just a few training images.

AlphaCode§

In 2022, DeepMind unveiled AlphaCode, an AI-powered coding engine that creates computer programs at a rate comparable to that of an average programmer, with the company testing the system against coding challenges built by Codeforces used in human competitive programming competitions. AlphaCode earned a rank equivalent to 54% of the median score on Codeforces after being trained on GitHub data and Codeforces problems and solutions. The program was required to come up with an unique solution and stopped from duplicating answers.

Gemini§

Gemini is a multimodal sizeable language model which was released on 6 December 2023. It is the successor of Google's LaMDA and PaLM 2 language models and sought to challenge OpenAI's GPT-4. Gemini comes in 3 sizes: Nano, Pro, and Ultra. Gemini is also the name of the chatbot that integrates Google Gemini and which was previously called Google Bard. On 12 December 2024, Google released Gemini 2.0 Flash, the first model in the Gemini 2.0 series. It in particular features expanded multimodality, with the ability to also generate images and audio, and is part of Google's broader plans to integrate advanced AI into autonomous agents. On 25 March 2025, Google released Gemini 2.5, a reasoning model that stops to "think" before giving a response. Google announced that all future models will also have reasoning ability. On 30 March 2025, Google released Gemini 2.5 to all free users. On 18 November 2025, Google released Gemini 3 Pro, a reasoning model which is fully multimodal. It was fully integrated with Google Search and AI Mode the same day.

Gemma§

Gemma is a collection of open-weight sizeable language models. The first ones were released on 21 February 2024 and are available in two distinct sizes: a 7 billion parameter model optimized for GPU and TPU usage, and a 2 billion parameter model designed for CPU and on-device applications. Gemma models were trained on up to 6 trillion tokens of text, employing similar architectures, datasets, and training methodologies as the Gemini model set. In June 2024, Google started releasing Gemma 2 models. In December 2024, Google introduced PaliGemma 2, an upgraded vision-language model. They launched PaliGemma 2 Mix, a version fine-tuned for multiple tasks in February 2025. It is available in 3B, 10B, and 28B parameters with 224px and 448px resolutions. Google released Gemma 3, calling it the most capable model that can be run on a single GPU in March 2025. It has four available sizes: 1B, 4B, 12B, and 27B. Google introduced TxGemma, an open-source model designed to improve the efficiency of therapeutics development in March 2025. In April 2025, Google introduced DolphinGemma, a research artificial intelligence model designed to hopefully decode dolphin communication. They want to train a foundation model that can learn the structure of dolphin vocalizations and generate novel dolphin-like sound sequences.

SIMA§

In March 2024, DeepMind introduced Scalable Instructable Multiword Agent, or SIMA, an AI agent capable of understanding and after natural language instructions to complete tasks across several 3D virtual environments. Trained on nine video games from eight studios and four research environments, SIMA showed adaptability to new tasks and settings without requiring access to game source code or APIs. The agent comprises pre-trained computer vision and language models fine-tuned on gaming data, with language being crucial for understanding and completing given tasks as instructed. DeepMind's research aimed to develop more helpful AI agents by translating advanced AI capabilities into real-world actions through a language interface.

Habermas machine§

In 2024, Google Deepmind published the results of an experiment where they trained two sizeable language models to help identify and present areas of overlap among a few thousand group members they had recruited online using techniques like sortition to get a representative sample of participants. The project is named in honor of Jürgen Habermas. In one experiment, the participants rated the summaries by the AI higher than the human moderator 56% of the time.