DeepMind

DeepMind is an AI research organization whose work spans game-playing systems, protein-structure prediction and general-purpose multimodal models. In April 2023, it combined with Google Brain to form Google DeepMind. The combined organization introduced Gemini later that year through collaboration across Google, including Google Research, while continuing scientific work such as AlphaFold with external collaborators.[5][6][7][11]

Before the combination with Brain and the launch of Gemini, DeepMind developed reinforcement-learning systems and tools for biology. That earlier work supplies the research context for the combined organization.[5][6][7]

Demis Hassabis, Shane Legg and Mustafa Suleyman founded DeepMind Technologies in 2010. Its interdisciplinary approach brought together machine learning, neuroscience, engineering, mathematics, simulation and computing infrastructure to build general AI systems. An early system, DQN, combined deep learning with reinforcement learning. DeepMind’s retrospective describes it learning 49 Atari games from raw screen pixels and a signal to maximize the score.[4][5]

AlphaGo combined deep neural networks with search algorithms. A policy network selected moves, while a value network predicted the winner. Training began with thousands of expert Go games, followed by reinforcement learning through games against different versions of itself.[8]

AlphaGo defeated European champion Fan Hui 5–0 in October 2015, then defeated Lee Sedol 4–1 in Seoul in March 2016. Two moves became distinctive moments of the latter match: AlphaGo’s Move 37 helped it win game two, and Sedol’s Move 78 helped him win game four. Sedol described his reaction to AlphaGo’s play: “I thought AlphaGo was based on probability calculation and that it was merely a machine. But when I saw this move, I changed my mind. Surely, AlphaGo is creative.”[8]

DeepMind subsequently developed AlphaZero and MuZero, and demonstrated AlphaStar against a top professional StarCraft II player in 2019. Its work also reached consumer products: WaveNet, a realistic text-to-speech model, was used as the voice of Google Assistant.[5]

AlphaFold predicts a protein’s three-dimensional structure from its amino-acid sequence. The connection matters because a protein’s shape is closely linked to its function. AlphaFold 2 made a breakthrough in protein-structure prediction in 2020; DeepMind subsequently released its code and published detailed papers in Nature.[9][11]

DeepMind and EMBL’s European Bioinformatics Institute (EMBL-EBI) launched the AlphaFold Protein Structure Database in July 2021 with more than 350,000 predictions, including the entire human proteome. On July 28, 2022, they announced its expansion from nearly one million to over 200 million predicted protein structures. The structures were freely available under a CC-BY 4.0 licence.[10][9]

EMBL-EBI reported in July 2022 that over 500,000 researchers from over 190 countries had accessed the database to view over two million structures, and that more than 1,000 scientific papers had cited it in just over a year. These were predictions supporting research: for example, DeepMind described researchers combining experimental outlines of the nuclear pore complex with AlphaFold predictions to complete and interpret unclear regions.[10][9]

On April 20, 2023, Sundar Pichai announced the combination of DeepMind and the Brain team from Google Research. Hassabis became CEO of Google DeepMind, and Jeff Dean became Google’s Chief Scientist, serving both Google Research and Google DeepMind. Google Research continued separately. The combination brought together distinct research histories: Brain began in 2011 at X and developed the Transformer architecture in 2017, alongside work on TensorFlow, JAX and language models.[6][5]

Google introduced Gemini 1.0 on December 6, 2023. Hassabis presented it on behalf of the Gemini team and credited collaboration across Google, including Google Research. Gemini was pre-trained on different modalities and then fine-tuned with additional multimodal data, covering text, code, audio, images and video. Training used Google’s TPU v4 and v5e infrastructure.[7]

The first generation had three sizes: Ultra for highly complex tasks, Pro for a broad range of tasks, and Nano for on-device use. The December announcement began a Bard rollout using fine-tuned Gemini Pro in English in more than 170 countries and territories; it announced Pro API access through Google AI Studio and Vertex AI starting December 13. Nano powered features on Pixel 8 Pro, including Recorder summaries. Ultra’s broad availability remained prospective while safety checks and further refinement continued.[7]

Google DeepMind and Isomorphic Labs, a sister Alphabet company, introduced AlphaFold 3 on May 8, 2024. It expanded prediction to joint structures involving proteins, DNA, RNA, ligands and chemical modifications. An improved Evoformer processed the inputs, followed by a diffusion network that refined a cloud of atoms over multiple steps. The launch offered AlphaFold Server free for non-commercial research; a November 11 update reported the release of model code and weights for academic use.[9][11]

DeepMind’s research also included machine-learning security. In an August 2024 interview, research scientist Nicholas Carlini explained his personal approach to evaluation: “Benchmarks tell you how well the model solves the task the benchmark is designed to solve.” He built a domain-specific language for tests drawn from tasks he had actually wanted models to perform. One multi-turn test ran model-suggested commands in Docker, fed results back to the model, and checked whether the final Git history showed a correct cherry-pick.[2]

Carlini also described a production-model extraction demonstration conducted with OpenAI’s permission and coordinated legal agreements. His team notified affected providers, including Google and OpenAI, and allowed 90 days for remediation before publication. The demonstrated attack extracted one layer, rather than a complete production model. Carlini described the narrow mitigation as withholding log probabilities when a query supplied logit bias.[2]

In separate training-data extraction work, Carlini credited co-author Milad Nasr with discovering that asking ChatGPT to repeat a word indefinitely could eventually produce other text, sometimes matching earlier internet documents. Carlini said the team checked outputs for verbatim matches of at least 50 consecutive words, with some matches extending to several hundred words.[2]

In July 2025, Google DeepMind reported that an advanced research version of Gemini Deep Think solved five of six International Mathematical Olympiad problems for 35 points, reaching gold-medal-standard performance. Unlike its 2024 AlphaProof and AlphaGeometry demonstration, which required expert translation into formal languages and two to three days of computation, the team reported natural-language proofs within the 4.5-hour competition limit. The setup used parallel thinking to explore and combine possible solutions, additional reinforcement learning, a curated mathematics-solution corpus and general instruction hints. IMO review confirmed the submitted solutions; it did not validate the system, processes or underlying model.[12]

Thang Luong led technical direction for the IMO effort and co-led coordination with Edward Lockhart. Quoc Le and Pushmeet Kohli advised the project, with program management from Kristen Chiafullo and Alex Goldin, alongside a large contributing team.[12]

Axios reported on August 5, 2026 that Hassabis was leaving the CEO role to become chairman. Latent.Space’s August 6 account reported his transition to Chair of Google DeepMind and Chief Scientist of Alphabet, with Koray Kavukcuoglu taking operational control as SVP, overseeing Gemini, frontier research and product teams. The same account reported that Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le departed to co-found Discovery Loop.[14][1]

The Gemini 3.8 Flash model card, published September 2, 2026, identifies Gemini 3.7 Flash as its basis. It specifies text, image, audio and video inputs with context up to one million tokens, and text output up to 64,000 tokens. Customizable effort levels control the balance of quality, cost and latency; listed limitations include hallucinations, occasional slowness or timeouts, and greater token use at higher effort levels.[15]