GDM

GDM is Google DeepMind, the AI organization formed in April 2023 by combining DeepMind and the Brain team from Google Research. Its work includes the Gemini multimodal model family and AlphaFold 3, developed with Isomorphic Labs to predict joint molecular structures. The combination brought DeepMind’s reinforcement-learning and protein-prediction research together with Brain’s work on Transformers and machine-learning infrastructure.[1][5][7][8][10]

The two predecessor labs began more than a decade before Google DeepMind’s formation. Demis Hassabis founded DeepMind in 2010 with Shane Legg and Mustafa Suleyman. It combined ideas from machine learning, neuroscience, mathematics and engineering to build general AI systems. Google Brain started in 2011 at X, Google’s moonshot factory, exploring how AI could improve Google’s products and services.[6][7]

DeepMind used games to test learning systems. DQN combined deep learning and reinforcement learning, learning 49 Atari games from screen pixels with the objective of maximizing score. AlphaGo and its successors extended the lab’s game research; in his December 2019 retrospective, Hassabis described MuZero as mastering chess, Go, shogi and Atari without first knowing their rules. DeepMind also developed WaveNet, the text-to-speech model used as the voice of Google Assistant. Brain’s contributions included TensorFlow, JAX and the Transformer architecture introduced in 2017, which became a foundation for large language models.[7][6]

By December 2019, Hassabis described DeepMind as publishing hundreds of papers annually and collaborating across Alphabet, including work with Waymo on evolutionary selection for training self-driving systems. He credited Chief Operating Officer Lila Ibrahim and VP of Research Koray Kavukcuoglu with developing the organization’s leadership, and Suleyman with collaborations on data-center energy consumption, Android battery performance and Google Play. That month, Suleyman left DeepMind for a new role at Google.[6]

DeepMind’s scientific work also produced a shared research resource. DeepMind and 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, EMBL-EBI announced its expansion from nearly one million to over 200 million predicted structures, with access under a CC-BY 4.0 license. It reported that, in just over a year, more than 500,000 researchers from over 190 countries had accessed the database, viewed over two million structures and cited it in more than 1,000 scientific papers.[11]

On April 20, 2023, Google announced that DeepMind and the Brain team would become Google DeepMind, led by Hassabis. His message to employees argued for greater speed, collaboration and execution, including a need to “simplify the way we make decisions.” It also announced a Scientific Board led by Kavukcuoglu to oversee research progress and direction, with representatives from across the organizations.[5]

Google introduced Gemini 1.0 on December 6, 2023, crediting teams across Google, including Google Research. Gemini was pretrained across modalities and then refined with additional multimodal fine-tuning, covering text, code, audio, images and video. Its three sizes served different purposes: Ultra for highly complex tasks, Pro for a broad range of tasks and Nano for on-device use.[8]

The initial rollout placed a fine-tuned Gemini Pro in Bard and Nano on Pixel 8 Pro, powering features such as Recorder summaries. Google announced developer and enterprise access to Pro through the Gemini API in AI Studio and Vertex AI starting December 13. Ultra remained in safety testing and refinement ahead of broader availability. Gemini 1.0 was trained on Google’s TPU v4 and v5e infrastructure.[8]

Announced on February 15, 2024, Gemini 1.5 introduced a mixture-of-experts architecture. Rather than activating one large network uniformly, it selectively activated relevant expert pathways according to the input. Google presented this specialization as a way to improve training and serving efficiency, initially releasing the mid-size multimodal 1.5 Pro for limited early testing.[9]

The announcement set a standard context window of 128,000 tokens for wider release, while a limited group could test up to one million tokens in private preview through AI Studio and Vertex AI. Google separately reported research tests reaching ten million tokens. Its demonstrations included reasoning across 402 pages of Apollo 11 mission transcripts and analyzing events and small details in a 44-minute silent Buster Keaton film. Early testers were told to expect longer latency with the experimental million-token window.[9]

Google DeepMind and Isomorphic Labs introduced AlphaFold 3 on May 8, 2024. Building on AlphaFold 2’s protein-structure work, it generated joint 3D structures from lists of molecules, including proteins, DNA, RNA, ligands and chemical modifications. An improved Evoformer module processed the inputs; a diffusion network then began with a cloud of atoms and iteratively formed the predicted molecular structure.[10]

The launch included AlphaFold Server, a free platform for non-commercial research supporting proteins, DNA, RNA and a selection of ligands, ions and chemical modifications. A November 11, 2024 update announced the release of model code and weights for academic use. These access milestones differed from the openly licensed protein-structure database developed with EMBL-EBI.[10][11]

On February 13, 2026, Latent.Space reported that Google was shipping an upgraded Gemini 3 Deep Think reasoning mode to Google AI Ultra subscribers in the Gemini app, alongside an early-access program through Vertex AI and the Gemini API for selected researchers and enterprises. The reported applications included checking mathematical papers, modeling physical systems and optimizing semiconductor crystal growth.[4]

Developer tooling became another part of the organization’s work. In its March 20, 2026 coverage, Latent.Space described Google DeepMind as having bought the team that became Antigravity the previous July. The same article reported an upgraded Google AI Studio experience incorporating an Antigravity coding agent and Firebase integrations for backend services, authentication and persistent app builds.[1]

In August 6, 2026 reporting, Latent.Space described a leadership reshuffle: Hassabis was moving from CEO to Chair of Google DeepMind and Chief Scientist of Alphabet, while Kavukcuoglu was taking operational control as SVP, overseeing Gemini, frontier research and product teams. The article reproduced Hassabis’s August 5 announcement: “I’m stepping into a new role as Chair of Google DeepMind & Chief Scientist of Alphabet.” It also reported that Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le were leaving to found Discovery Loop, a Public Benefit Corporation focused on automating machine learning, science and engineering.[2]