Google DeepMind
 ia subsidiary of Alphabet |
|---|
[ Show/Hide ]
|
| Google DeepMind | |
|---|---|
 i
| |
| DeepMind Technologies Limited | |
| Type | Subsidiary of Alphabet Inc. |
| Industry | Artificial intelligence · Research |
| Founded | September 23, 2010 (as DeepMind Technologies, London) |
| Founders | Demis Hassabis · Shane Legg · Mustafa Suleyman |
| Headquarters | King's Cross, London, United Kingdom |
| CEO | Demis Hassabis |
| Parent | Google / Alphabet Inc. |
| Acquired by Google | January 26, 2014 (about $400 million USD) |
| Products | Gemini · AlphaGo · AlphaFold · Imagen · Veo · Lyria · WaveNet |
| Website | deepmind.google |
Google DeepMind, often called just DeepMind, is a British-American artificial intelligence research laboratory based in London. It is a subsidiary of Alphabet Inc., the parent company of Google.
DeepMind is best known for three breakthroughs. In 2016, its AlphaGo program became the first computer system to beat a top human world champion at the ancient board game Go. In 2020, its AlphaFold system solved a 50-year-old scientific puzzle by accurately predicting the 3D shapes of proteins, work that helped earn its lead scientists a share of the 2024 Nobel Prize in Chemistry. And starting in 2023, it built Gemini, Google's family of large language models that powers Google's chatbot, search features, and many other products.
DeepMind was founded in 2010 by Demis Hassabis, Shane Legg, and Mustafa Suleyman. Google bought the company in 2014, and in 2023 it was merged with another Google research team called Google Brain to form Google DeepMind. The combined lab is led by Hassabis and acts as the main AI research arm of Alphabet.
1 History✎
DeepMind's story has three main chapters: its early years as an independent London startup, its long stretch as a Google research lab, and the post-2023 era as Google DeepMind.
1.1 Founding and early research (2010–2014)✎
DeepMind Technologies was founded on September 23, 2010, in London. The three founders came from very different backgrounds. Demis Hassabis was a former chess prodigy and video game designer with a PhD in cognitive neuroscience. Shane Legg was an AI researcher who had studied with Hassabis at the Gatsby Computational Neuroscience Unit at University College London. Mustafa Suleyman was an entrepreneur and policy expert.
Their company motto was "solve intelligence, and then use that to solve everything else." Early funding came from investors including Peter Thiel and Elon Musk.
DeepMind's first big result came in 2013, when researchers trained a single AI system to play dozens of Atari 2600 video games — including Breakout, Pong, and Space Invaders — at human level or better. The system learned each game from scratch, using only the pixels on the screen and the score as feedback. This was an early demonstration of deep reinforcement learning, the technique of combining neural networks with trial-and-error reward learning that would become DeepMind's signature method.
1.2 Google acquisition and the AlphaGo era (2014–2022)✎
On January 26, 2014, Google acquired DeepMind for a reported $400 million. As a condition of the deal, Google agreed to set up an internal AI ethics board.
After the acquisition, DeepMind kept its London headquarters and continued publishing research openly. Its biggest moment came in March 2016, when its AlphaGo program played a five-game match against Lee Sedol, one of the strongest Go players in the world. AlphaGo won 4-1. Go had been considered far too complex for computers to master because of its astronomical number of possible board positions. The match was watched live by hundreds of millions of viewers and was later the subject of a documentary film, AlphaGo.
DeepMind followed up with several other landmark systems:
- AlphaZero (2017) — a single program that taught itself to play chess, shogi, and Go at superhuman level, starting only from the rules.
- AlphaStar (2019) — the first AI to defeat top-ranked human players at the real-time strategy game StarCraft II.
- WaveNet (2016) — one of the first AI systems to generate realistic human-sounding speech. It later became the voice of Google Assistant.
- AlphaFold (2018, with a major upgrade in 2020) — an AI system that predicts the 3D structure of proteins from their amino acid sequences. AlphaFold's predictions are now used by biologists worldwide and helped scientists understand the virus that causes COVID-19. In 2024, Demis Hassabis and DeepMind researcher John Jumper shared the Nobel Prize in Chemistry for the work.
During these years, DeepMind also tried twice — reportedly in 2017 and again in 2020 — to gain more independence from Google, but Alphabet's leadership did not agree to the proposed structures.
1.3 Merger with Google Brain and the Gemini era (2023–present)✎
In April 2023, Alphabet announced that DeepMind would merge with Google Brain, Google's older internal AI research division, to form a single unit called Google DeepMind. The merger was a response to the rapid rise of OpenAI and ChatGPT, which had taken the world by surprise the previous November. Demis Hassabis became CEO of the combined lab.
The first major product of the new Google DeepMind was Gemini, a family of multimodal large language models announced on December 6, 2023. Gemini replaced Google's earlier Bard chatbot and quickly became the AI engine behind many Google products. Later versions — Gemini 1.5 in 2024, Gemini 2.5 in 2025, and Gemini 3 in late 2025 — each added new capabilities such as longer memory, faster responses, better reasoning, and stronger coding abilities.
Alongside Gemini, Google DeepMind released a steady stream of other generative AI models, including the Imagen image generator, the Veo video generator, and the Lyria music generator. By 2026, the lab's work was being used across nearly every Google product, from Search to Gmail to Android.
2 Products and research✎
DeepMind's work falls into two broad categories: research systems that solve scientific and game-playing problems, and generative AI models used in Google's products.
2.1 Gemini✎
Gemini is Google's flagship family of large language models. It is built to be multimodal from the start, meaning a single model can handle text, images, audio, video, and computer code together. The Gemini family includes several sizes:
- Gemini Pro — the standard model used for most everyday tasks.
- Gemini Flash — a faster, cheaper version for quick responses.
- Gemini Flash Lite — an even smaller version for simple jobs.
- Gemini Deep Think — a higher-end mode that takes more time to reason through hard problems.
- Gemini Nano — a tiny model designed to run directly on phones and other devices.
Gemini powers the Gemini app (formerly called Bard), which is Google's main AI chatbot, available on the web and on phones. It also runs inside Google products such as Google Search (AI Overviews), Gmail, Google Docs, and Google Workspace.
2.2 AlphaFold and science✎
AlphaFold is DeepMind's AI system for protein structure prediction. Proteins are the molecular machines that do almost every job in living cells, and their function depends on the precise 3D shape they fold into. Before AlphaFold, working out a single protein's structure could take years of lab work. AlphaFold can do it in minutes.
The AlphaFold Protein Structure Database, run with the European Bioinformatics Institute, contains predicted structures for more than 200 million proteins — almost every protein known to science. It is used for free by millions of researchers worldwide.
DeepMind has also released other science-focused systems, such as GraphCast for weather forecasting and AlphaMissense for predicting harmful genetic mutations. A spinoff company, Isomorphic Labs, uses DeepMind's technology to design new medicines.
2.3 Game-playing AI✎
DeepMind's game-playing programs are some of its most famous research projects, even though they are not products.
- AlphaGo (2016) — beat world Go champion Lee Sedol.
- AlphaZero (2017) — mastered Go, chess, and shogi from scratch.
- MuZero (2019) — learned to play games without being told the rules.
- AlphaStar (2019) — reached grandmaster level at StarCraft II.
These projects are used as testbeds for general AI techniques that are later applied to more practical problems.
2.4 Generative media✎
Beyond Gemini, DeepMind builds specialized generative models for different types of media:
- Imagen — generates images from text descriptions.
- Veo — generates short videos from text or images. Veo 3, released in 2025, can produce high-resolution clips with realistic motion.
- Lyria — generates music. Lyria 3 was released in February 2026, and Lyria 3 Pro followed in March.
- Genie — generates interactive, game-like virtual worlds from a single image or sketch.
2.5 Hardware✎
DeepMind's AlphaChip system uses AI to help design computer chips. According to Google, AlphaChip has been used to lay out every generation of Google's Tensor Processing Unit (TPU) chips since 2020.
3 Structure and leadership✎
Google DeepMind is a wholly-owned subsidiary of Alphabet Inc. Its CEO is Demis Hassabis, who is also one of Alphabet's senior executives. The lab is headquartered in King's Cross in London, with additional research offices in the United States, Canada, France, Germany, and Switzerland.
DeepMind employs thousands of researchers and engineers and publishes much of its work in scientific journals such as Nature and Science.
4 Reception and controversies✎
DeepMind is widely regarded as one of the world's top AI research organizations. AlphaGo, AlphaFold, and Gemini are often listed among the most important AI achievements of the 2010s and 2020s. Hassabis and Jumper's Nobel Prize in 2024 made DeepMind one of the few private research labs to be honored at that level.
The lab has also faced criticism. Major concerns include:
- NHS data deal. In 2017, an investigation in the United Kingdom found that DeepMind's earlier work with the National Health Service had improperly shared the medical records of 1.6 million patients.
- Gemini image generation. In early 2024, Google had to pause Gemini's ability to generate images of people after users complained that the system produced historically inaccurate pictures, such as racially diverse Vikings and Founding Fathers.
- AlphaChip claims. Some independent researchers have questioned whether DeepMind's AI chip-design tool really beats older methods, asking for more detailed benchmarks.
- Independence from Google. DeepMind's founders reportedly tried more than once to win greater independence from Alphabet, and several senior staff members have left over concerns about the lab's direction.
- Energy use. Like other large AI labs, DeepMind's models require huge amounts of electricity and water to train and run.
5 See also✎
- Artificial intelligence
- Large language model
- Generative AI
- Gemini
- OpenAI
- Anthropic
- Google Brain
- Alphabet Inc.