Google DeepMind

Google DeepMind develops AI models and scientific tools, including Gemini and AlphaFold. Google announced the organization on April 20, 2023, combining DeepMind with Google Research’s Brain team under Demis Hassabis. The combination brought together research teams to build increasingly capable and general AI systems and products across Google and Alphabet.[3][5][6]

In the 2023 announcement, Hassabis argued that developing more capable AI required greater speed, stronger collaboration and simpler decision-making. The organization also announced a Scientific Board led by Koray Kavukcuoglu to oversee research progress and direction; its final composition was still pending. Eli Collins was joining Hassabis’s leadership team as VP of Product, while Zoubin Ghahramani was joining the research leadership team reporting to Kavukcuoglu.[3]

On April 18, 2024, Sundar Pichai announced a further consolidation: model-building teams across Google Research and Google DeepMind would sit within Google DeepMind. He described concentrating compute-intensive development in one place and giving Google’s product teams single access points for models. Research’s Responsible AI teams would also move into Google DeepMind, closer to model development. Google Research retained a distinct mandate in computing systems, foundational machine learning and algorithms, and applied science and society, with continued collaboration between the organizations.[4]

Google introduced Gemini 1.0 on December 6, 2023, crediting collaboration across Google, including Google Research. Gemini was pretrained across different modalities and then fine-tuned with additional multimodal data, supporting text, code, audio, images and video. Google trained the models using its TPU v4 and v5e infrastructure.[5]

The first generation had three variants: Ultra for highly complex tasks, Pro for a broad range of tasks, and Nano for on-device use. The rollout differed by variant. Bard began using a fine-tuned Gemini Pro, with English availability in more than 170 countries and territories; Nano powered features on Pixel 8 Pro. Google announced Pro API access through Google AI Studio and Vertex AI starting December 13. Ultra remained under safety testing and further refinement before broad release.[5]

In July 2025, Google DeepMind reported that an advanced research version of Gemini Deep Think solved five of six International Mathematical Olympiad problems, scoring 35 of 42 points, a gold-medal-standard result. Thang Luong led the technical direction and coordinated the effort with Edward Lockhart, supported by a wider research collaboration. IMO review confirmed the submitted solutions, but did not validate the system, processes or underlying model.[9]

The team described a change from its 2024 AlphaProof and AlphaGeometry 2 approach, which required expert translation into formal languages and two to three days of computation. The advanced Gemini system produced natural-language proofs within the 4.5-hour competition limit. It used parallel thinking to explore and combine possible solutions, additional reinforcement learning, access to a curated corpus of mathematics solutions, and general problem-solving hints in its instructions.[9]

The Gemini Robotics team introduced two Gemini 2.0-based models in March 2025. Gemini Robotics added physical actions as an output modality for direct robot control. Gemini Robotics-ER supplied spatial and embodied reasoning that roboticists could connect to existing controllers. Robotics was trained primarily on ALOHA 2 data and demonstrated on a platform using Franka arms. Google DeepMind also announced a partnership with Apptronik, whose Apollo humanoid was a target for model specialization; Robotics-ER was available to selected trusted testers.[10]

These developments followed more than a decade of work by the predecessor DeepMind, including reinforcement-learning agents, AlphaGo and protein-structure prediction.[2][8][6]

Shane Legg left academia to co-found DeepMind Technologies in 2010, alongside Demis Hassabis and Mustafa Suleyman. In his 2014 conference account, Legg described their vision as combining machine learning and systems neuroscience to create artificial agents with general intelligence. Investments from entrepreneurs including Peter Thiel and Elon Musk preceded the assembly of a research team spanning systems neuroscience, deep learning, reinforcement learning and Bayesian statistics. Google acquired DeepMind in early 2014.[2]

DeepMind’s Deep Q-Networks combined neural networks with reinforcement learning. A Q-network estimated the total reward expected after an action; experience replay randomly sampled stored experiences to provide more diverse, less correlated training data and stabilize learning. In the Atari experiments described by David Silver in June 2016, agents observed screen pixels and game-score rewards, then selected joystick directions. DeepMind trained separate agents for 50 games without prior knowledge of their rules.[8]

AlphaGo combined a policy network that selected actions, a value network that predicted the winner, and tree search. Self-play trained the policy and value networks. In October 2015, AlphaGo became the first program to defeat a professional Go player; in March 2016, it defeated Lee Sedol four games to one.[8]

AlphaFold placed first at the CASP13 protein-structure prediction assessment in December 2018. AlphaFold2 won CASP14 in November 2020. DeepMind published its detailed methodology in Nature in July 2021 and open-sourced the code, followed by the launch of the AlphaFold Protein Structure Database with EMBL’s European Bioinformatics Institute, EMBL-EBI.[6]

The database began with more than 350,000 predictions, including the entire human proteome. On July 28, 2022, DeepMind and EMBL-EBI expanded it to over 200 million predicted structures, covering nearly all catalogued proteins known to science. Structure access was openly licensed under CC-BY 4.0. EMBL-EBI reported that over 500,000 researchers from more than 190 countries had accessed the database, viewing over two million structures, and that more than 1,000 scientific papers had cited it.[7]

On May 8, 2024, Google DeepMind and Isomorphic Labs introduced AlphaFold 3, extending prediction to interactions between proteins and other biomolecules. Google DeepMind also launched AlphaFold Server with free access for non-commercial research. Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for their AlphaFold work, alongside David Baker for computational protein design. Google DeepMind’s November 2025 timeline reported AlphaFold use by over three million researchers in more than 190 countries.[6]

In a 2026 biology panel, Google DeepMind’s Pushmeet Kohli explains that AlphaFold’s team worked with structures deposited by experimental scientists in the Protein Data Bank. Scientific knowledge also shaped the model: amino acid residues affect one another, so the team incorporated that relationship instead of requiring the model to learn everything from scratch. Kohli contrasts this with work on a virtual cell, where he says available cell-by-gene data was insufficient for that broader ambition. His point is to choose modeling and data collection around the scientific problem and its constraints.[11]

Kohli cautions that reproducing a deposited structure does not settle how a protein changes shape in different contexts or reveal its full dynamics. He describes extracting distributions and motion from cryo-electron-microscopy images as a promising direction that still needs work. He also distinguishes understanding a model’s internal reasoning from knowing its behavior: calibrated confidence and a clear account of strengths and limitations help scientists decide when a prediction merits trust.[11]

Nicholas Carlini, a DeepMind research scientist interviewed in August 2024, advocated personal benchmarks built from tasks users had actually asked models to solve. His domain-specific language ran generated code and checked results; a multi-turn Git task executed model-proposed commands in Docker, returned the outputs to the model, and checked the final commit history. These were his evaluation practices rather than an organization-wide policy.[1]

Carlini also described practical attacks on deployed AI systems. Expired domains referenced by an image dataset could be bought and used to replace training images. A model-extraction experiment recovered one layer, with OpenAI’s permission, according to his account. He said the team notified affected providers, including Google and OpenAI, and allowed 90 days for fixes before publication; a mitigation prevented returning log probabilities when a logit bias was supplied.[1]

In separate training-data extraction work, co-author Milad Nasr discovered that asking ChatGPT to repeat a word indefinitely could cause it to diverge into other text. Carlini reported outputs matching previously published internet documents verbatim for at least 50 consecutive words, sometimes hundreds. This concerned leakage of memorized text, distinct from recovering model weights.[1]