Google DeepMind

Google DeepMind is an artificial intelligence research and model-development organization at Google. It was formed on April 20, 2023 by combining the London-based DeepMind lab, founded in 2010 and acquired by Google in 2014, with Google Brain, the deep learning group that began inside Google X in 2011 and later sat within Google Research. The organization was led by co-founder Demis Hassabis as CEO from the merger until August 5, 2026, when Koray Kavukcuoglu took over day-to-day leadership as SVP of Google DeepMind and Hassabis became Chair of Google DeepMind and Chief Scientist of Alphabet.^[165] It states that its mission is "to build AI responsibly to benefit humanity." Its work spans reinforcement learning, large language models, biology, weather forecasting, mathematics, robotics, agents, and generative media.^[1]

Google DeepMind develops the Gemini family of multimodal models, which sits behind Google's consumer AI products, its developer platform, and its Cloud offerings, along with the Gemma open-weight models and a long line of scientific systems including AlphaFold, AlphaGo, GraphCast, GNoME, and AlphaProof. Hassabis and AlphaFold director John Jumper shared half of the 2024 Nobel Prize in Chemistry for protein structure prediction. The other half went to David Baker for computational protein design.^[2]

The organization carries two research traditions at once. The DeepMind side made its name on deep reinforcement learning and game-playing systems, and later on structural biology. The Google Brain side produced TensorFlow, much of Google's large-scale distributed training work, and the Transformer architecture that underlies nearly every modern language model. Since the merger both threads have run in parallel, sharing infrastructure, staff, and increasingly the same base models. The balance between them has shifted: through 2025 and 2026 the scientific programme has been progressively rebuilt around Gemini rather than bespoke architectures, and in July 2026 the Financial Times reported that the dedicated AlphaFold team had been dissolved and its researchers reassigned, mostly to Gemini and to other science areas.^[138]

Overview

Scroll sideways for more →

The name refers both to an operating organization inside Google and to a continuing UK legal company. DeepMind Technologies Limited, company number 07386350, was incorporated on September 23, 2010 and remained an active UK private company in 2026. Its registered office is a law firm's filing address, not a workplace, and should not be treated as evidence of Google DeepMind's headquarters or office footprint.^[1][3]

History

Founding

The company that became DeepMind was incorporated on September 23, 2010 under the name FRIARS 2022 LIMITED, a shelf company whose only officers were a nominee director and a corporate secretarial service. It was renamed DeepMind Technologies Limited on November 15, 2010, and Hassabis was appointed a director on December 7, 2010, the same day the nominee officers resigned. Mustafa Suleyman became company secretary on the same date and a director on December 23, 2011; Shane Legg was appointed a director on December 23, 2011. The three are consistently described as co-founders, and the company itself dates its founding to 2010, but the statutory record shows the appointments staggered over a year rather than made together.^[1][3]

The register also documents the early investors more reliably than press accounts do. Luke Nosek, who co-founded Founders Fund with Peter Thiel, was a director from February 23, 2011; Jaan Tallinn, a co-founder of Skype, was a director from December 23, 2011; and Bart Swanson, an adviser at Horizons Ventures, was a director from June 11, 2013. Each of those appointments corroborates the corresponding investor's involvement and dates it. Other names that recur in press coverage of DeepMind's early funding, including Elon Musk and Peter Thiel personally, do not appear in the filings, and no primary record of their individual investments has been published. Contemporary reporting put total pre-acquisition funding at roughly $50 million, a figure Google and DeepMind never confirmed.^[3][130]

The founders brought together work in machine learning, neuroscience, engineering, mathematics, and simulation. Hassabis became chief executive, Legg became chief scientist, and Suleyman led collaborations that applied the company's research in Google products and health care.^[1]

An early technical focus was deep reinforcement learning. In 2013, researchers led by Volodymyr Mnih described a deep Q-network that learned Atari games from pixels and game scores. The peer-reviewed 2015 Nature paper evaluated the approach on 49 games using the same algorithm, network architecture, and hyperparameter settings across the set. The authors reported performance comparable to a professional human games tester across the collection, while also documenting games on which the system remained weak.^[4] DQN is the ancestor of most of the Alpha-prefixed systems that followed, and the pattern it established, of one learning algorithm evaluated across a whole benchmark suite rather than tuned per task, recurs throughout the lab's work.

The 2014 Google acquisition

The acquisition completed on January 24, 2014. On that date all six of the company's directors, the three founders plus Nosek, Tallinn, and Swanson, resigned, and Google appointed its own officers, including corporate development executive Kenneth Yi, who was still a director of DeepMind Technologies Limited in 2026. Google confirmed the deal publicly two days later, on January 26, 2014, so the transaction closed before it was announced.^[3][130]

Google has never disclosed the price. Re/code, which first confirmed the deal with Google, reported about $400 million; The Information reported more than $500 million; UK outlets converted these into a widely repeated figure of £400 million. Higher figures such as $650 million circulate but have no contemporaneous attribution. The honest statement is that the price was never officially disclosed and contemporaneous reports ranged from roughly $400 million to over $500 million.^[5][130] DeepMind had roughly 75 to 100 staff at the time.^[135]

Reporting at the time also said Google had agreed to establish an ethics board covering DeepMind's technology as a condition of the sale. That board's membership, terms of reference, and output have never been made public. Suleyman said in 2016 that it was "ongoing" and internal, and DeepMind declined to comment when journalists asked about it in 2019. It should be described as a reported condition of the acquisition rather than a documented governance body, and it should not be confused with Google's Advanced Technology External Advisory Council, a separate eight-member body announced on March 26, 2019 and dissolved on April 4, 2019 after employee objections to one appointee.^[131][132]

Access to Google's infrastructure and products gave DeepMind opportunities to apply its research outside games. In 2016, DeepMind reported that a machine-learning control system reduced the energy used for cooling at a Google data center by 40 percent and improved overall power-usage effectiveness by 15 percent. Those figures were company measurements from one deployment, not a general efficiency guarantee for data centers.^[6]

AlphaGo and successor game systems

AlphaGo combined policy and value networks with tree search. Its 2016 Nature paper reported a 5-0 win over European Go champion Fan Hui. In March 2016, a later version defeated Lee Sedol 4-1 in a public five-game match in Seoul. The result established that a learned system could defeat an elite professional in full-board Go, a domain long considered difficult for conventional game-search methods.^[7]

AlphaGo Zero then removed the need for human game records. Starting from the rules, it learned through self-play and defeated the published AlphaGo version 100-0 in the authors' evaluation.^[8] AlphaZero generalized this self-play approach to chess, shogi, and Go, using the rules of each game but no human examples.^[9]

MuZero learned a model useful for planning without being given the environment's transition rules or reward function. It matched AlphaZero's performance in Go, chess, and shogi in the reported experiments and was also evaluated on 57 Atari games.^[10] AlphaStar addressed partial information and long action sequences in StarCraft II. The published system reached Grandmaster level with all three playable races and ranked above 99.8 percent of human players active enough to be ranked on the European server, about 90,000 players in the authors' evaluation period.^[11]

These systems are related but should not be collapsed into a single claim that DeepMind "solved" games. Each used a particular interface, training regime, compute budget, and evaluation protocol. Their broader significance was methodological: they advanced value learning, self-play, planning with learned models, multi-agent training, and the use of games as controlled research environments.

AlphaFold and scientific AI

DeepMind entered the Critical Assessment of protein Structure Prediction, or CASP, with the first AlphaFold system in 2018, placing first overall and first in the hardest free-modelling category at CASP13.^[151] AlphaFold 2 produced its major result at CASP14 in 2020, reaching a median GDT score of 92.4 across all targets and 87.0 in the free-modelling category. CASP co-founder John Moult said at the time that "a score of around 90 GDT is informally considered to be competitive with results obtained from experimental methods," which is the origin of the widely repeated claim that AlphaFold matched experiment.^[152] The 2021 Nature paper introduced an architecture that jointly processed multiple-sequence and pairwise residue representations. A predicted structure remains a model output with confidence estimates, not a replacement for every experiment or a direct prediction of all protein behavior.^[12]

DeepMind and EMBL-EBI launched the AlphaFold Protein Structure Database on July 22, 2021 with more than 350,000 structures, including the entire human proteome, and expanded it in July 2022 to over 214 million. In November 2025 Google DeepMind said more than 3 million researchers in over 190 countries had used AlphaFold, including over a million in low- and middle-income countries, and that more than 35,000 papers cited it. The database distinguishes predictions from experimentally determined structures and exposes confidence information that researchers need when interpreting a result.^[13][152]

Health, data protection, and the move into Google

DeepMind Health, established under Suleyman, was the lab's first large applied programme outside research. Its Streams application, built with the Royal Free London NHS Foundation Trust, alerted clinicians to acute kidney injury. It became the subject of the organization's most serious governance controversy, described below, and was also the setting for an unusual accountability experiment: in June 2016 DeepMind appointed a panel of external reviewers, drawn from government, medicine, and academia, to scrutinise DeepMind Health and publish annual reports. The panel published two, in 2017 and 2018. Google announced in November 2018 that DeepMind Health would move into a new Google Health division, and the panel was wound up as part of that reorganisation on the stated basis that a UK-focused review structure did not suit an international effort. The transfer completed in September 2019, and Google confirmed in August 2021 that Streams was being decommissioned.^[51][122]

Merger with Google Brain

On April 20, 2023, Google combined DeepMind and the Brain team from Google Research as Google DeepMind. Sundar Pichai announced it alongside a companion post from Hassabis, who became CEO of the combined group. Pichai wrote that Jeff Dean would take "the elevated role of Google's Chief Scientist," serving in that capacity to both Google Research and Google DeepMind; Hassabis wrote that a new Scientific Board would oversee research direction, led by Koray Kavukcuoglu. Zoubin Ghahramani, who had led Google Brain, joined the research leadership under Kavukcuoglu, and Eli Collins became VP of Product for the combined unit. Kavukcuoglu was not named CTO at that point; that title came later.^[14][133]

The merger joined two different research histories. DeepMind was particularly associated with deep reinforcement learning, game-playing systems, and AlphaFold. Google Brain, which began in 2011 inside Google X, contributed major work on distributed learning, TensorFlow, and the Transformer. The 2017 paper "Attention Is All You Need," written by Google researchers, introduced the Transformer architecture that later became central to modern language and multimodal models.^[15]

The context was competitive. OpenAI had released ChatGPT in November 2022 and GPT-4 in March 2023, and Google's own conversational product, Bard, had launched in March 2023 on a version of LaMDA. Consolidating two research organizations that had been building large models separately, on separate infrastructure and with separate publication cultures, was the structural answer to that pressure. It also ended a longer argument about DeepMind's independence: the Wall Street Journal reported in May 2021 that DeepMind's leadership had spent years negotiating for a separate legal structure with its own governance, and that Google had ended those talks that month.^[146]

The Gemini era

The merged organization shipped Gemini 1.0 on December 6, 2023, and has released a new generation or major point release every few months since. The pattern of 2024 through 2026 is a steady migration of research output into Google products: reasoning models into Search and the Gemini app, robotics models into partner hardware, generative media models into YouTube and Google's creative tools, and scientific systems increasingly built on Gemini rather than on task-specific architectures. By the second quarter of 2026, Google was reporting 950 million monthly active users for the Gemini app and about 22 billion tokens per minute across its model APIs.^[56]

The same period brought a run of senior departures. Tim Brooks, hired from OpenAI in October 2024 to lead world-model research, left for Meta Superintelligence Labs in September 2025.^[149] Noam Shazeer, a co-author of the Transformer paper and a Gemini co-lead whom Google had brought back from Character.AI in 2024, announced on June 18, 2026 that he was joining OpenAI to lead AI architecture research.^[136] John Jumper announced the following day that he was leaving after nearly nine years to join Anthropic; AlphaFold co-authors Jonas Adler and Alexander Pritzel followed him.^[137] On July 29, 2026 the Financial Times reported that Google DeepMind had dissolved AlphaFold's dedicated team over the preceding year, reassigning most of the original paper's authors to Gemini work or to enzyme design, fusion, and genomics, with others moving to Isomorphic Labs. Roughly a quarter of the full-time Google DeepMind authors on the original AlphaFold papers had left the company. The AlphaFold database, server, and AlphaFold 3 academic access remained operational.^[138]

The August 2026 leadership transition

On August 5, 2026, Pichai announced the organization's largest leadership change since the 2023 merger. Hassabis handed over his day-to-day operational responsibilities at Google DeepMind and took a new strategic role as Chair of Google DeepMind and Chief Scientist of Alphabet, which Pichai described as letting him "put his full attention on actively shaping the future of AGI"; he continues to lead Isomorphic Labs, and wrote to staff that he would keep advising Kavukcuoglu and the Google DeepMind leads "from our awesome new London Platform 37 offices". Kavukcuoglu, until then Google DeepMind's CTO and Google's Chief AI Architect, stepped up as SVP of Google DeepMind, reporting to Pichai and overseeing Gemini model development, Frontier AI research, and the Gemini app and developer teams, in addition to the Chief AI Architect role. Pichai's note said the Gemini app had reached more than 950 million monthly users and that Gemma models had surpassed 900 million downloads; Hassabis's note cited "the great progress we're making with our new models including Gemini 4".^[165]

The same announcement said that Jeff Dean, chief scientist of both Google DeepMind and Google Research, was leaving after what Pichai called "an incredible 27-year run", together with Google Senior Fellow Sanjay Ghemawat, to launch an independent public benefit corporation, with Google continuing to work with them "as a founding investor and Cloud partner".^[165][167] Dean announced the company the same day as Discovery Loop, co-founded with Sanjay Ghemawat, Google DeepMind VP of Research Oriol Vinyals, and Quoc Le, with the stated mission "to automate machine learning, science, and engineering to accelerate discoveries and progress".^[166] Wired reported that Dean will serve as Discovery Loop's CEO, that Khosla Ventures and Radical Ventures invested alongside Google on undisclosed terms, and that Google agreed to supply the startup with compute for its first year.^[167]

Organization and structure

Google DeepMind operates within Google and ultimately Alphabet. It conducts research, trains models, and works with Google product and infrastructure teams. The 2023 announcement described one combined Google team, while the active DeepMind Technologies Limited filing shows why brand, operational, and legal-company descriptions should be kept distinct.^[3][14]

Legal structure and finances

A common claim, that DeepMind became an Alphabet subsidiary separate from Google in 2015, is not supported by the filings. DeepMind Technologies Limited's accounts name Google Ireland Holdings Unlimited Company as its immediate parent through the 2010s, and DeepMind Holdings Limited, a UK company incorporated on August 30, 2019, from 2019 onward. Alphabet Inc. has been the ultimate controlling party since the Alphabet restructuring, but the ownership chain has always run through Google entities, and recent accounts describe Google LLC as an intermediate parent.^[129]

The filed accounts are the only public financial record for any part of Google DeepMind, and they need careful reading. The accounts state that "turnover represents research and development remuneration from other group undertakings": it is intra-group cost-plus transfer pricing, not revenue from customers. The accounts also state that the company's workforce is directly employed by other Alphabet group companies, so they disclose no headcount.

Scroll sideways for more →

Financial year 2020 was the first in which the company recorded a profit. The 2019 accounts record that Google Ireland Holdings waived repayment of intercompany loans and accrued interest amounting to £1.1 billion, booked as a capital contribution. The fall in turnover in 2024 alongside a rise in profit reflects a reduction in recharges from group undertakings rather than a contraction in activity.^[129]

Leadership

Scroll sideways for more →

Kavukcuoglu was named Chief AI Architect of Google in June 2025, a Google-wide role reporting to Pichai, while retaining his Google DeepMind CTO title.^[17][147] On August 5, 2026 he became SVP of Google DeepMind, still reporting to Pichai, with responsibility for Gemini model development, Frontier AI research, and the Gemini app and developer teams, while Hassabis moved to Chair of Google DeepMind and Chief Scientist of Alphabet. Jeff Dean and Oriol Vinyals left in the same restructure to co-found Discovery Loop with Sanjay Ghemawat and Quoc Le.^[165][166] Lila Ibrahim, DeepMind's founding COO from 2018, moved in early 2026 to a newly created Chief AI Readiness Officer role covering global affairs, public engagement, and the responsibility organization; Google DeepMind's responsibility page still lists her under the COO title as co-chair of the Responsibility and Safety Council, and no successor as COO has been named.^[16][139] Helen King, as VP of Responsibility, co-chairs that council with her.^[16]

Google does not publish a headcount for Google DeepMind. The best-sourced figures come from press reporting on internal data: roughly 2,600 in May 2024, roughly 5,600 in March 2025, and roughly 6,000 in August 2025. Vendor databases quoting figures near 9,000 are scraped estimates without methodology and should not be used.^[135] Google DeepMind's careers page lists ten locations: London, the Bay Area, Bangalore, Cambridge in Massachusetts, Montreal, New York City, Paris, Tokyo, Toronto, and Zurich. Google's AI research laboratory in Accra, Ghana is a Google Research site, not a DeepMind one.^[134]

Relationship to other Alphabet organizations

Google Research remains a distinct research organization; Jeff Dean served as Chief Scientist of both groups until his August 2026 departure for Discovery Loop.^[165][166] Google Cloud exposes Gemini models through Vertex AI, while custom tensor processing units support model training and serving.^[18][21]

Isomorphic Labs is a separate Alphabet company, incorporated on February 24, 2021 and announced publicly that November, also led by Hassabis, with Google DeepMind's Chief Business Officer Colin Murdoch as its president. It applies AI to drug discovery and collaborated with Google DeepMind on AlphaFold 3. It raised $600 million in its first external round on March 31, 2025, led by Thrive Capital with GV and Alphabet participating, and released a technical report on its proprietary drug design engine, IsoDDE, on February 10, 2026. As of mid-2026 it had not begun human trials; Hassabis said in January 2026 that the company expected its first clinical trials by the end of the year, a slip from an earlier end-2025 target. Isomorphic should not be described as a Google DeepMind department, nor should its financing, drug pipeline, or clinical timelines be treated as Google DeepMind results.^{[22][53][163]}

Google DeepMind, OpenAI, Anthropic, and other frontier-model developers differ in organization, product reach, publication practice, funding, and infrastructure. Relative-size or research-breadth claims depend on a defined metric and a common date; no single headcount, capital, or compute estimate establishes a general ranking.

Reinforcement learning, games, and agents

Reinforcement learning is the thread that runs from DeepMind's founding to its current agent work. The games were never the point; they were tractable environments with clear reward signals in which to develop methods later pointed at protein folding, algorithm discovery, plasma control, and robot manipulation.

Scroll sideways for more →

Agent57 closed a benchmark that DQN had opened. Earlier agents posted strong average scores across the Atari suite while failing completely on a handful of hard-exploration games such as Montezuma's Revenge and Pitfall. Agent57 combined the intrinsic-reward machinery of Never Give Up with separate action-value estimates for extrinsic and intrinsic rewards and a bandit meta-controller that selected among exploration and discount settings. It was the first algorithm to exceed the benchmark's standard human reference score on all 57 games. The result is about generality rather than peak score: on the paper's own table, Agent57's raw mean and median human-normalised scores are lower than MuZero's, and its claim rests on the capped mean of 100.00 and on covering all 57 games where MuZero cleared 51. The authors say as much, noting that MuZero obtains the highest uncapped scores while failing catastrophically on games such as Venture.^[101]

Gato, published in Transactions on Machine Learning Research in 2022, used one set of Transformer weights for 604 tasks including Atari play, image captioning, dialogue, and real-robot control. The unifying feature was the common token-and-action interface, not state-of-the-art performance on every task, and the model was deliberately sized at about 1.2 billion parameters so it could run a real robot in real time rather than because of a compute limit. The paper reports that prompt-based in-context learning on new environments did not improve significantly over prompt-free evaluation, so Gato adapted by fine-tuning rather than in context. It was trained by supervised behaviour cloning, not reinforcement learning.^[19]

RoboCat, described in June 2023, extended the generalist idea to robot manipulation: a single agent that operated several real and simulated arms, learned new tasks from a modest number of demonstrations, and then generated its own training data to improve further.^[108] SIMA, first reported in March 2024, took the opposite tack of accepting only what a human player sees. It observes on-screen pixels and produces keyboard and mouse actions, so it can in principle be attached to any 3D title without access to game internals.^[109] SIMA 2, announced on November 13, 2025, replaced the earlier behaviour-cloning agent with a Gemini model at its core; Google described it as a research preview, so its demonstrations should not be read as a generally available autonomous agent.^[20]

Language and multimodal research before Gemini

The lab's language-model work predates Gemini and shaped it. Gopher, described in December 2021, was a 280-billion-parameter autoregressive Transformer evaluated across 152 tasks; it was never released as weights or an API but was documented in unusual depth.^[103] The follow-up mattered more. Chinchilla, published in March 2022, argued from a large sweep of training runs that frontier models of the era, including Gopher itself and OpenAI's GPT-3, were substantially undertrained, and that for a fixed compute budget model size and training tokens should be scaled roughly in proportion. The 70-billion-parameter Chinchilla outperformed Gopher on most evaluations despite being four times smaller. The result changed how the industry allocated training compute and is still cited as the Chinchilla scaling law.^[102]

Other pre-Gemini work fed directly into the current models. Perceiver and Perceiver IO, presented in 2021, attacked the problem of a single architecture that could ingest images, audio, video, and point clouds by attending inputs into a small latent bottleneck rather than scaling attention with input size.^[107] Flamingo, introduced in April 2022, was a visual language model that accepted arbitrarily interleaved images, video, and text and adapted to new vision-language tasks purely from a few in-context examples.^[105] Sparrow, introduced on September 22, 2022, was the lab's main public demonstration of reinforcement learning from human feedback combined with an explicit rule set for dialogue safety; it was a research artefact and was never released as a product.^[106] AlphaCode, published in Science in December 2022, was the first system to reach roughly the median human level on Codeforces competitive-programming contests, by sampling up to a million candidate programs per problem and filtering them to ten submissions; AlphaCode 2, announced alongside Gemini on December 6, 2023, used a fine-tuned Gemini model and was reported to solve close to twice as many problems.^[104]

Science and mathematics

Protein structure and the 2024 Nobel Prize

AlphaFold is the clearest case in Google DeepMind's portfolio of a claim independently checked before it was publicised, and the reason is CASP. The Critical Assessment of protein Structure Prediction is a blind biennial experiment run by an outside community: organisers release sequences whose experimentally determined structures are not yet public, entrants submit predictions, and independent assessors score them. AlphaFold 1 placed first at CASP13 in 2018 and AlphaFold 2 produced accuracy competitive with experiment for most targets at CASP14 in 2020, both judged by assessors with no connection to DeepMind. That is a stronger form of evidence than an internal benchmark table, and it explains why the AlphaFold result was accepted quickly while several later Alpha-prefixed claims were not.^[12]

The Royal Swedish Academy of Sciences awarded half of the 2024 Nobel Prize in Chemistry jointly to Hassabis and Jumper "for protein structure prediction." The recognition concerned the development of AlphaFold; it did not mean that every protein interaction, folding pathway, or biological function had been experimentally established by the model.^[2]

AlphaFold 3, developed by Google DeepMind and Isomorphic Labs, expanded structure prediction to complexes involving proteins, nucleic acids, small molecules, ions, and modified residues. Its May 2024 Nature paper uses a diffusion-based structure module and reports improved accuracy on several interaction tasks. Predictions still require domain-specific confidence checks and experimental validation where decisions depend on molecular behavior.^[22]

Genomics and protein design

AlphaMissense, published in Science on September 19, 2023, adapts AlphaFold to variant effect prediction, fine-tuned on population frequency data rather than clinical labels. Google DeepMind released a catalogue covering roughly 71 million possible human missense variants, classifying about 89 percent as likely pathogenic or likely benign. Those are model classifications, not clinical determinations, and the distinction has practical consequences. A 2024 review in Disease Models and Mechanisms found that on 18 experimentally confirmed benign IRF6 variants AlphaMissense misclassified 15 as pathogenic, and the broader published consensus is that its clinical utility has not been validated and that predictions should not drive clinical decisions without functional or clinical corroboration.^[110][159]

AlphaProteo generates candidate proteins designed to bind chosen molecular targets. Google DeepMind reported stronger binding affinities than prior design methods on seven tested targets, with gains ranging from 3 to 300 times depending on the target. It did not find successful binders for an eighth target, TNF-alpha. AlphaProteo is therefore an experimental design system with mixed target-level results, not a general solution to protein design.^[23]

AlphaGenome takes up to one million DNA base pairs as input and predicts thousands of molecular tracks at single-base resolution across eleven modalities including gene expression, splicing, chromatin accessibility, and three-dimensional contacts. The model was previewed on June 25, 2025 and described in a Nature paper published in January 2026. The paper reported matched or better performance than the strongest available external model on 25 of 26 variant-effect evaluations.^[24]

Mathematics, algorithms, and program search

AlphaTensor, reported in Nature on October 5, 2022, recast the search for matrix-multiplication algorithms as a single-player game played by an AlphaZero-derived agent. Its headline result was a decomposition multiplying two 4x4 matrices with 47 scalar multiplications in arithmetic modulo 2, the first improvement on Strassen's two-level method for that case since 1969. The modulo-2 restriction is the detail most often dropped in summaries: the 47-multiplication algorithm does not carry over to real or complex arithmetic. Human researchers also improved on several of the paper's other results almost immediately. Three days after publication, Kauers and Moosbauer reduced AlphaTensor's 96-multiplication algorithm for 5 by 5 matrices over the same field to 95, and later work using flip graphs pushed several formats further, over arbitrary fields rather than only modulo 2. That is the normal outcome when an automated result opens a problem rather than closing it.^[29][164]

AlphaDev, published in Nature on June 7, 2023, searched directly in CPU assembly instructions for small sorting and hashing routines. Google reported improvements of up to 70 percent for short fixed-size sorts of three to five elements and about 1.7 percent for sequences over 250,000 elements. Several routines were reverse-engineered into C++ and merged into the LLVM libc++ standard library, the first change to that part of the library in over a decade, and a hashing routine went into Google's Abseil library. The deployment is verifiable in public code review records, though the LLVM patch was filed in January 2022, about 17 months before the paper appeared, so the merge did not follow publication. Critics on technical forums characterised the sorting result as an instruction-level saving in an unrolled insertion sort rather than a new algorithm; that criticism was informal and never published as a peer-reviewed rebuttal.^[30]

FunSearch, published in Nature on December 14, 2023, paired a language model with an automated evaluator to search program space. It found a cap set of size 512 in eight dimensions, improving on a previous best of 496, and produced tailored heuristics for online bin packing that used fewer bins than standard first-fit and best-fit rules.^[31] AlphaEvolve, announced on May 14, 2025, generalised the approach by combining Gemini models with automated evaluators and evolutionary search. Google reported an algorithm multiplying two 4 by 4 complex matrices with 48 scalar multiplications, describing it as the first improvement on Strassen's algorithm in that setting in 56 years, and said the system continuously recovers about 0.7 percent of Google's worldwide compute, sped up a Gemini training kernel by 23 percent, and proposed a Verilog simplification for an upcoming TPU. Those statements are company-reported. The mathematical result was also short-lived as an AI-only achievement: within a month, Dumas, Pernet, and Sedoglavic showed the same 48-multiplication count could be obtained with rational coefficients, which removes the restriction to complex arithmetic and makes it valid over almost any ring.^[32][154] Google published a one-year impact report on May 7, 2026 listing further deployments, all likewise company-reported.^[155]

AlphaGeometry, published in Nature on January 17, 2024, paired a language model that proposes auxiliary constructions with a symbolic deduction engine. It solved 25 of 30 problems on the IMO-AG-30 benchmark within the standard time limit, against an average of 25.9 for human gold medallists.^[116] The paper put the previous automated record at 10, and that comparison did not survive scrutiny: independent researchers showed in April 2024 that Wu's method, a classical algebraic technique, solves 15 of the 30 on its own, that Wu's method combined with conventional synthetic solvers reaches 21 on a laptop in under five minutes per problem, and that combining Wu's method with AlphaGeometry reaches 27. Their work is a preprint rather than a peer-reviewed rebuttal, but it substantially weakens the baseline against which AlphaGeometry was measured.^[153] AlphaGeometry 2, described in a February 2025 preprint, raised the solve rate on all IMO geometry problems from 2000 to 2024 from 54 percent to 84 percent.^[117]

In 2024, AlphaProof and AlphaGeometry 2 solved four of six International Mathematical Olympiad problems for 28 of 42 points, one point below the gold-medal threshold of 29, which 58 of 609 contestants reached that year. The problems were manually translated into a formal language before the systems worked on them; one was solved within minutes and the others took up to three days, well beyond the human time limit. The solutions were graded by Timothy Gowers and Joseph Myers rather than by IMO officials.^[33] AlphaProof was published in Nature on November 12, 2025.^[118] In 2025, an advanced Gemini Deep Think system produced natural-language solutions to five of six problems for 35 of 42 points within the 4.5-hour official time limit and without any formalisation. This time IMO coordinators graded the solutions under the same criteria applied to students, and IMO president Gregor Dolinar confirmed the gold-medal score. An OpenAI model reported the same score in the same competition.^[34]

AlphaProof Nexus, described in a May 2026 preprint, formalizes and attempts open mathematics problems in Lean. The authors reported solving 9 of 353 selected Erdos problems and 44 of 492 selected OEIS sequence problems, with inference costs of a few hundred dollars per problem in their main configuration. These are preprint results, not a peer-reviewed claim that the system can autonomously solve arbitrary research mathematics.^[35] A related system, AI Co-Mathematician, described in a paper posted on May 7, 2026, is an interactive agentic workbench in which a mathematician steers a team of agents that search the literature, run code, and attempt proofs.^[126]

Materials, weather, and the physical sciences

GNoME applied graph neural networks to materials discovery. The Nature paper reported 2.2 million predicted stable crystal structures, including 381,000 candidates on the authors' updated convex hull, and noted that 736 had been independently experimentally realised. These are computational predictions, not 381,000 synthesized materials, and the framing drew a substantial published critique discussed below.^[28]

GraphCast is a graph-neural-network weather model trained on reanalysis data. Its Science paper reports global 10-day forecasts at 0.25-degree resolution in under a minute on a TPU and better results than the authors' deterministic operational baseline on more than 90 percent of 1,380 evaluated targets.^[25] GenCast, published in Nature on December 4, 2024, is the probabilistic counterpart: a diffusion model producing 15-day ensembles at the same resolution. The authors reported that it beat the ECMWF ENS ensemble on 97.2 percent of 1,320 evaluation targets and could generate a 50-member ensemble in about eight minutes on a single Cloud TPU v5.^[26] WeatherNext 2, announced in November 2025, uses a Functional Generative Network to produce many plausible trajectories, and Google reported hundreds of scenarios in under a minute on one TPU with improvements over its earlier operational AI model across most evaluated variables and lead times. Those are first-party evaluation results.^[27]

Weather Lab, launched on June 12, 2025, applies a stochastic model producing 50 scenarios out to 15 days to tropical cyclones. Google reported that its five-day track predictions were on average 140 km closer to the true cyclone position than the ECMWF ensemble on 2023-2024 North Atlantic and East Pacific data, roughly a day and a half of extra lead time. This is one of the few Google DeepMind claims validated by an outside operational agency: the US National Hurricane Center's 2025 verification preview records that the season was the first in which it incorporated AI-based models into real-time operations and that the Google DeepMind model, identified as GDMI, "was very useful," while noting that such systems remain under development and were not always available in time for routine forecasting.^[160]

AlphaEarth Foundations, announced in July 2025, extends the approach to Earth observation. It fuses optical imagery, radar, elevation, and lidar into a single 64-band embedding per location, published as a Satellite Embedding dataset in Google Earth Engine so that mapping tasks such as forest carbon estimation or land-cover change can be done without training a bespoke model. In July 2026 Google Maps Platform opened a private preview of Custom Satellite Embeddings built on it.^[120]

DeepMind also worked with the Swiss Plasma Center on magnetic control of the Tokamak a Configuration Variable. The 2022 Nature paper reported that a deep reinforcement-learning controller maintained several plasma shapes and configurations by controlling the device's magnetic coils.^[54] AlphaQubit, published in Nature on November 20, 2024 with Google Quantum AI, is a neural decoder for surface-code quantum error correction. It made 6 percent fewer errors than tensor-network methods and 30 percent fewer than correlated matching on data from Google's Sycamore processor, but the paper stated plainly that it was too slow to correct errors in real time, since the budget is about one microsecond per cycle. A December 2025 preprint reported a successor that decodes faster than that budget on commercial accelerators for the 105-qubit Willow chip, about 9.6 times faster than the original at code distance 11.^[111][161] AlphaChip, the reinforcement-learning method for chip floorplanning first published in Nature in 2021, has been used in the design of several TPU generations and is also the subject of a long-running scientific dispute.^[112]

Google DeepMind has also begun packaging these systems for outside institutions. On December 18, 2025 it announced support for the US Department of Energy's Genesis Mission, giving all 17 national laboratories accelerated access to the AI co-scientist and Gemini for Government, with AlphaEvolve, AlphaGenome, and WeatherNext to follow. On July 16, 2026 it published a bioresilience programme with Isomorphic Labs covering biosecurity prevention, detection, and response, including applying SynthID-style watermarking to screen AI-generated DNA sequences at synthesis providers.^[162]

TacticAI, built with Liverpool FC and published in Nature Communications on March 19, 2024, models corner kicks as a graph problem and predicts receivers and shot outcomes. In a blind evaluation, football experts preferred its suggested setups to existing tactics in about 90 percent of cases; the pool of expert raters was small and the evaluation was designed by the authors.^[119]

Research assistants

The AI co-scientist, unveiled by Google Research on February 19, 2025 and built on Gemini 2.0, uses a multi-agent process of generation, debate, and refinement to produce and rank research hypotheses. Its best-known validation came from Jose R. Penades at Imperial College London, whose laboratory had spent years establishing an unpublished mechanism by which capsid-forming phage-inducible chromosomal islands acquire diverse phage tails to widen their host range. Given published background, the system proposed the same hypothesis in about two days. Penades confirmed with Google that it had no access to the unpublished work. The achievement was synthesis from the existing literature rather than discovery from nothing, since the component facts were already published, but the specific hypothesis was not. The paired papers appeared in Cell in 2025, and a broader evaluation covering acute myeloid leukaemia drug repurposing, liver fibrosis, and antimicrobial resistance was published in Nature on May 19, 2026. All of those results are preclinical.^[36][156]

A second line of 2026 work aims at autonomous rather than assistive research. Aletheia, described in a February 2026 preprint, wraps an advanced Gemini Deep Think model in a generator, verifier, and reviser harness and applies it to open problems. The authors report one fully autonomous AI-generated mathematics paper, four autonomous solutions among 700 open Erdos problems drawn from a public database, and roughly 92 percent on the IMO-ProofBench benchmark, with the system returning "no solution found" rather than fabricating a proof when it failed. These are preprint results.^[157]

Domain-tuned models occupy the space between the general Gemini line and these research systems. LearnLM, announced in 2024, is a family of Gemini derivatives fine-tuned on learning-science principles for teaching, and its behaviours were later folded into Gemini itself rather than shipped as a permanently separate model.^[150] Med-PaLM and Med-PaLM 2 preceded Gemini as Google's medical question-answering models, and their successors in the open-weight line are MedGemma for clinical text and imaging and TxGemma for therapeutic development. The pattern across all of them is the same: a specialised checkpoint is used to establish that a capability is reachable, and the capability then migrates into the general model.^[75]

The Gemini model family

Gemini is Google DeepMind's general-purpose multimodal model family and the commercial centre of the organization. Google announced Gemini 1.0 on December 6, 2023 in three sizes: Ultra for the most demanding tasks, Pro for general use, and Nano for on-device work. Google described the family as trained from the start across text, code, audio, image, and video rather than as a text model with modality adapters added afterwards.^[37]

The release cadence since then has been fast, and the naming has not been linear. Point releases have sometimes leapfrogged the tier above them, and a version number does not by itself indicate that a model is still available.

Scroll sideways for more →

How the generations differ

The 1.0 generation established the family and the three-size structure. Its most repeated claim, that Gemini Ultra scored 90.0 percent on MMLU and became the first model to exceed a reported human expert baseline, rests on a specific evaluation protocol: Google's own technical report obtained that figure with chain-of-thought prompting and 32 samples, and reported 83.7 percent under the five-shot setting other labs had been using for their published numbers. It is a useful reminder that a Gemini benchmark figure should always be read with its evaluation conditions.^[68]

The 1.5 generation moved to a mixture-of-experts architecture and made very long context the selling point, shipping a one million token context window later extended to two million in preview. Google's technical report described near-perfect retrieval on synthetic needle-in-a-haystack tests at those lengths, a weaker property than reasoning across the whole context, and the report separates the two.^[38]

The 2.0 generation was presented in agentic terms: native tool calling, native image and audio output, and the research prototypes Project Astra and Project Mariner shown alongside it.^[21] The 2.5 generation folded explicit reasoning into the base models rather than shipping it as a separate thinking variant, and added a developer-controlled thinking budget so latency and cost could be traded against deliberation.^[38][58]

Gemini 3 arrived on November 18, 2025 and was the first publicly accessible model to pass 1500 Elo on the LMArena text leaderboard, debuting at 1501. Google shipped it simultaneously across the Gemini app, Google AI Studio, Vertex AI, Gemini Enterprise, and Search, a change from earlier generations that reached consumers weeks after developer surfaces.^[39]

The 3.x line has progressed by point release rather than a fourth generation. Gemini 3.1 Pro, released on February 19, 2026, kept Gemini 3 Pro's price and architecture and reported 77.1 percent on ARC-AGI-2 against 31.1 percent for Gemini 3 Pro, a company-reported figure on a benchmark whose scores depend heavily on the compute allowed per task.^[40] Gemini 3.5 Flash, announced at Google I/O on May 19, 2026, was positioned unusually: Google claimed the low-latency Flash tier now beat the larger Gemini 3.1 Pro on most coding and agentic evaluations while producing output about four times faster.^[18]

As of August 1, 2026 the most recent release is the July 21, 2026 group of Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber. Flash Cyber is fine-tuned for vulnerability detection and patching and, because of its dual-use character, Google is distributing it only to governments and trusted partners through a limited-access pilot rather than the public API. The same announcement said Gemini 3.5 Pro was still testing with partners and that Google had begun what it called its most ambitious pre-training run so far, for Gemini 4. Neither statement was a release, and Pichai repeated both on the July 22 earnings call.^[41][56]

Reading Gemini benchmark claims

Google publishes benchmark tables with each Gemini launch, and those tables are the source of most figures that circulate afterwards. Three caveats recur. Evaluation conditions vary between the launch table and the technical report, as the MMLU example shows. Several benchmarks used since 2025, including ARC-AGI-2 and agentic suites such as SWE-bench and OSWorld, are sensitive to scaffolding and to the compute budget allowed per task, so a number is comparable only against another obtained under the same harness. And Google's comparison set is chosen by Google; independent leaderboards such as LMArena measure blind human preference, a different quantity, and a model can lead one while trailing the other.^[39][40][41]

Gemma open-weight models

Gemma is Google DeepMind's family of open-weight models, built from research shared with Gemini but released as downloadable checkpoints anyone can run and fine-tune.

Scroll sideways for more →

Specialised variants include CodeGemma for code, PaliGemma for vision-language tasks, RecurrentGemma for memory-efficient inference, ShieldGemma for content safety classification, MedGemma and TxGemma for medical and therapeutic-development work, and DolphinGemma, announced on April 14, 2025, which models the vocalizations of wild Atlantic spotted dolphins with the Wild Dolphin Project and Georgia Tech.^[75]

Licensing is worth stating precisely. Gemma 1 through Gemma 3 were released under Google's own Gemma Terms of Use, which impose usage restrictions and are therefore not an open source licence in the sense used by the Open Source Initiative; "open weights" is the accurate description. Gemma 4 was the first release in the family distributed under Apache 2.0. At the Gemma 4 launch Google said the family had been downloaded more than 400 million times and that developers had built more than 100,000 variants from it.^[73] In his August 5, 2026 note to staff, Pichai said Gemma models had surpassed 900 million downloads.^[165]

Generative media and world models

Google DeepMind's generative media work descends from audio and image research that predates Gemini. WaveNet, introduced in 2016, generated raw audio waveforms sample by sample; a faster production version shipped in Google Assistant US English and Japanese voices in 2017.^[42] SoundStream and AudioLM extended that line into neural audio coding and language-model-style audio generation, and MusicLM into text-conditioned music.

Scroll sideways for more →

The image line runs from the original Imagen research model of 2022 through Imagen 2, released to developers on December 13, 2023, and Imagen 3, announced at Google I/O on May 14, 2024, to Imagen 4 in May 2025. Since late 2025 the practical successor has been the Nano Banana line, which is not a separate model family at all but the image-generation capability of the Gemini models themselves, so a Gemini generation and an image generation now advance together.^[65][77]

Veo 2, announced December 16, 2024, extended video generation to higher resolutions and longer clips.^[43] Veo 3, announced at Google I/O on May 20, 2025, was the first commercially available video model to generate synchronized audio, including dialogue and sound effects, in the same pass as the picture.^[78] Lyria 2, announced April 24, 2025, powers the Music AI Sandbox and an interactive real-time music mode.^[79]

Genie is the world-model line: systems that generate an environment a person or an agent can act inside, rather than a fixed clip. Genie 2, unveiled December 4, 2024, produced navigable 3D environments from a single image.^[80] Genie 3, announced August 5, 2025, generates them from a text prompt and runs in real time at 720p and 24 frames per second, with environmental consistency holding for several minutes in the demonstrations Google published. Those numbers describe the announced configuration, not a guarantee for arbitrary prompts.^[44] On January 29, 2026 Google opened a consumer research prototype, Project Genie, to Google AI Ultra subscribers aged 18 and over in the United States.^[81]

SynthID runs across this work. First announced in August 2023 for Imagen output, it has been extended to audio, video, and text, and a 2024 Nature paper describes the text watermarking scheme and its evaluation. Watermark survival depends on content type and on what transformations are applied, so SynthID is a provenance signal rather than a universal detector for synthetic media.^[48]

Robotics

Google DeepMind's robotics programme absorbed Google Brain's Robotics at Google team, whose RT-1 and RT-2 models applied Transformer architectures to robot control, and the Open X-Embodiment dataset collaboration that pooled robot demonstrations across dozens of institutions.

The current line began on March 12, 2025 with Gemini Robotics, a vision-language-action model built on Gemini 2.0 that emits robot actions directly, and Gemini Robotics-ER, an embodied-reasoning model intended to be used with a roboticist's own controllers and planners.^[45] Gemini Robotics On-Device, announced June 24, 2025, runs locally on supported hardware and was released first through a trusted-tester programme with an accompanying SDK.^[46] On September 25, 2025 Google announced Gemini Robotics 1.5 and Gemini Robotics-ER 1.5, adding an explicit thinking step before action and native tool calling, with ER 1.5 opened to all developers and the vision-language-action model limited to selected partners.^[82]

On July 30, 2026 Google DeepMind announced Gemini Robotics 2, a set of three models: a vision-language-action model of the same name, Gemini Robotics ER 2 built on Gemini 3.5 Flash, and Gemini Robotics On-Device 2. Google describes it as the first generation in which one learned policy drives a whole humanoid robot, which it characterises as control from feet to fingertips rather than the upper-body tabletop manipulation of earlier releases. The launch shipped two model cards, a safety technical report, and an open safety benchmark called ASIMOV-Agentic.^[83]

Products and platforms

Google DeepMind builds models; most of them reach the public through Google product and Cloud teams.

Developer access. Google AI Studio is the free browser tool for prototyping against the Gemini API, combining a prompt playground, multimodal file upload, and code export; since late 2025 it also has a Build mode that generates working applications.^[84] Gemini CLI, released June 25, 2025 under Apache 2.0, puts a Gemini agent in the terminal.^[85] Antigravity, launched in public preview alongside Gemini 3 Pro on November 18, 2025, is an agent-first development environment built on a fork of Visual Studio Code.^[86]

Enterprise access. Vertex AI is Google Cloud's machine learning platform and the route by which enterprises consume Gemini with the governance, region, and compliance controls Cloud customers expect.^[87] Gemini Enterprise, announced October 9, 2025, sits above it as a workplace front end combining the models with a no-code agent workbench and connectors to business data.^[88]

Consumer surfaces. The Gemini app began as Bard on March 21, 2023 and took the Gemini name on February 8, 2024, when Gemini Advanced launched inside a $19.99 per month Google One AI Premium plan.^[55] Gemini Live, launched August 13, 2024, is its spoken conversational mode.^[89] NotebookLM is a separate research tool that grounds answers in documents a user uploads rather than the open web; it was shown as Project Tailwind at Google I/O in May 2023, released in limited testing that July, and later added Audio Overviews, which turn a source set into a synthetic two-host discussion.^[90]

Research prototypes and their successors. Project Astra, shown at Google I/O in May 2024, explored a universal assistant that can see and hear through a device camera and microphone, hold a low-latency spoken conversation, remember what it has just encountered, and act on a user's behalf. Its research fed Gemini Live rather than shipping as a product of its own.^[91] Project Mariner, first shown with Gemini 2.0 in December 2024, was a browser-using agent released as a Chrome extension in a research preview; Google reported an 83.5 percent result on the WebVoyager benchmark in a single-agent configuration and required user confirmation for sensitive actions in the demonstrated system.^[21] Google discontinued Project Mariner as a standalone product on May 4, 2026 and said its technology had moved into other Google products, including the Gemini API and the Gemini Agent feature in the Gemini app.^[92]

Deep Research. Deep Research, launched inside Gemini Advanced on December 11, 2024, is an agentic mode that plans a multi-step search, reads and cross-references sources, and returns a cited report rather than a chat answer. It was among the first products of its kind from a major lab and has since been rebuilt on each successive Gemini generation. The pattern is characteristic of how DeepMind research reaches users: a capability first demonstrated in a research prototype becomes a mode inside an existing product rather than a separate application.^[21]

Search. AI Overviews, the generative summaries that appear above conventional results, reached general availability in the United States in May 2024 after a year of testing as the Search Generative Experience. AI Mode, a conversational search surface, was announced in March 2025 and expanded globally in October 2025. At Google I/O on May 19, 2026 Pichai said AI Overviews had passed 2.5 billion monthly active users and AI Mode 1 billion, and that Google was processing more than 3.2 quadrillion tokens a month.^[57] On Alphabet's second-quarter 2026 earnings call he put the Gemini app at 950 million monthly active users and the model APIs at roughly 22 billion tokens per minute.^[56] These are company-stated figures with company-defined denominators. An AI Overviews "user" is a Search user who was shown a generated summary, which is not the same kind of engagement as a Gemini app session.

Compute

Google DeepMind trains and serves on Google infrastructure, above all on tensor processing units, the custom accelerators Google has built since the mid-2010s. The relationship runs in both directions: DeepMind's model architectures shape what the chips are optimised for, and Google says the eighth-generation parts were designed in partnership with Google DeepMind.^[93]

Scroll sideways for more →

Ironwood, the seventh generation, was unveiled at Google Cloud Next on April 9, 2025 and described as the first TPU designed specifically for inference; a full 9,216-chip superpod is rated at 42.5 FP8 exaflops.^[94] At Google Cloud Next on April 22, 2026 Google split the eighth generation into two chips for the first time: TPU 8t for training and TPU 8i for serving. Google states that a TPU 8t superpod links 9,600 liquid-cooled chips with about two petabytes of shared high bandwidth memory for roughly 121 FP4 exaflops, close to three times Ironwood's per-pod compute, and that TPU 8i triples on-chip SRAM to 384 MB and delivers up to 80 percent better performance per dollar than Ironwood for low-latency serving. Both figures are Google's own.^[93] Alongside them Google introduced the Virgo Network fabric, which Pichai said lets customers connect a million AI accelerators across multiple data centre sites.^[56][95]

How Google DeepMind's compute differs from an external customer's is mostly a matter of access rather than of different silicon. Google DeepMind trains on the same TPU generations Google Cloud sells, but it typically gets them first and at a scale not offered externally: Ironwood ran Gemini and AlphaFold workloads internally before general availability, and the eighth-generation parts were co-designed around DeepMind's workloads and are scheduled to reach customers outside Google later.^[93][94] External customers buy capacity through Google Cloud under published quotas and regional availability, and the largest such arrangement is with a competitor: in October 2025 Anthropic agreed to expand its use of Google Cloud TPUs to up to about one million chips.^[96] Google has also reportedly explored silicon outside the main TPU line; in July 2026 The Information reported an internal inference chip codenamed Frozen v2 that would hardwire parts of the Gemini architecture directly into hardware, a report Google has not confirmed.^[97]

Safety, alignment, and governance

The responsibility organization

Google DeepMind's published governance structure has two standing bodies. The Responsibility and Safety Council, co-chaired by Lila Ibrahim and VP of Responsibility Helen King, evaluates research, projects, and collaborations against Google's AI Principles. The AGI Safety Council, led by Shane Legg, works on extreme risks from more capable future systems. Google DeepMind also says it maintains dedicated teams for technical safety, ethics, governance, security, and public engagement, and it is a member of the Frontier Model Forum and a co-founder of the Partnership on AI.^[16]

The Frontier Safety Framework

The Frontier Safety Framework is Google DeepMind's public risk-management policy for severe risks from frontier models. It was first published on May 17, 2024, updated to version 2.0 in February 2025, and to version 3.0 on September 22, 2025.^[49][98] The framework is built around Critical Capability Levels: capability thresholds which, if a model is assessed to have crossed them, trigger specified evaluation, security, and deployment mitigations. Version 3.0 sharpened the CCL definitions, added a harmful manipulation CCL covering models capable of systematically changing beliefs and behaviour in high-stakes settings, and expanded the treatment of misalignment, including scenarios in which a model interferes with an operator's ability to direct, modify, or shut it down. Version 3.1, published on April 17, 2026, added Tracked Capability Levels, an earlier-warning layer intended to flag less extreme risks before a CCL is reached, and gave more detail on how holistic risk assessments are conducted.^[49]

The framework is a company policy document. It describes processes Google DeepMind commits to; it is not an independent certification, and comparisons ranking one developer's framework above another's require a stated methodology rather than a reading of the documents.

Technical safety and interpretability

Google DeepMind's alignment work is published rather than only implemented. An April 2025 paper set out the organization's approach to technical AGI safety and security, organised around misuse and misalignment as the two risk areas its technical work targets.^[99] The interpretability programme released Gemma Scope in July 2024, a set of sparse autoencoders trained on the activations of Gemma 2 models and published openly so outside researchers could study model internals without training their own probes, an unusually concrete contribution to mechanistic interpretability from a frontier lab.^[100] SynthID covers the provenance side, and model cards and safety evaluations accompany each Gemini release. Anca Dragan leads the AI safety and alignment organization.^[16]

Publications and open releases

Google DeepMind's public record splits into two halves that are worth distinguishing. Its scientific work is published in the ordinary academic way, in Nature, Science, and their sister journals, with peer review and, since the AlphaFold 3 dispute, generally with code. The list of papers on its own publications page ran continuously through July 2026.^[145] Its frontier model work is documented instead through technical reports, model cards, and launch posts, which describe evaluations and capabilities but not training data or architecture in a form another group could reproduce. Both halves are normal for their genre; conflating them produces the common mistake of treating a Gemini technical report as though it carried the same evidentiary weight as a Nature paper.

The organization also maintains a substantial open-release programme that is easy to overlook next to the proprietary models. The AlphaFold Protein Structure Database is free to use and covers more than 200 million predicted structures.^[13] The Gemma weights are downloadable, and Gemma 4 moved the family to Apache 2.0.^[73] Gemma Scope published sparse autoencoders for interpretability research.^[100] Open X-Embodiment pooled robot demonstration data across dozens of laboratories, and the Gemini Robotics releases have shipped model cards, safety reports, and the open ASIMOV benchmarks alongside restricted model access.^[47][83]

Notable people

Scroll sideways for more →

Criticism and controversies

Health data and the Royal Free ruling

DeepMind's Streams project with the Royal Free London NHS Foundation Trust used patient records to support an acute kidney injury application. On July 3, 2017 the UK Information Commissioner's Office concluded that the Trust, as data controller, had failed to comply with the Data Protection Act 1998 in four respects, under the first, third, sixth, and seventh data protection principles. The ICO proceeded on the basis that DeepMind was a data processor, and its remedies were directed at the Trust. The regulator found the processing of approximately 1.6 million partial patient records for the purpose of clinical safety testing excessive and disproportionate, and Commissioner Elizabeth Denham said patients "would not have reasonably expected their information to have been used in this way."^[50][140]

Two points are commonly misreported. No fine was issued: the Trust signed an undertaking, given in consideration of the Commissioner not serving an enforcement notice, committing it to a privacy impact assessment, evidence of lawful processing conditions, and a third-party audit. And the finding was against the Trust, not DeepMind. Academic criticism, notably by Julia Powles and Hal Hodson in Health and Technology in 2017, went further, questioning the lawfulness and transparency of the arrangement as a whole.^[140][141] DeepMind's own response acknowledged that "we underestimated the complexity of the NHS and of the rules around patient data" and that "we got that wrong."^[50]

A representative action followed. Andrew Prismall sued Google UK Limited and DeepMind Technologies Limited on behalf of a class of about 1.6 million people, claiming misuse of private information in respect of an October 2015 transfer and a live data feed running to September 2017. Mrs Justice Heather Williams struck the claim out and gave reverse summary judgment on May 19, 2023, and the Court of Appeal dismissed the appeal on December 11, 2024. The reasoning is often summarised too loosely: because the representative-action rule requires all class members to share the same interest, the claim had to be pleaded on an irreducible minimum basis, and on that basis the court held no class member had a realistic prospect of establishing a reasonable expectation of privacy or of recovering more than trivial damages. The Court of Appeal added that "a representative class claim for misuse of private information is always going to be very difficult to bring." No further appeal has been recorded.^[121]

Governance structures that did not last

Three governance mechanisms associated with DeepMind have been criticised for being announced more prominently than they were operated. The AI ethics board reported at the time of the 2014 acquisition was never documented publicly.^[131] The DeepMind Health Independent Review Panel, appointed in June 2016 with a mandate to publish annual reports, published two and was wound up in 2018 and 2019 when DeepMind Health moved to Google Health; its 2018 report had warned of the risk of DeepMind acquiring excessive market power through bundled data access and infrastructure.^[122] And DeepMind's own attempt to secure an independent legal structure, negotiated with Google over several years, was ended by Google in May 2021.^[146]

Workplace conduct

Suleyman was placed on leave in August 2019 after employee complaints about his management. Google and DeepMind later confirmed that an external law firm had investigated, and reporting in January 2021 described complaints of bullying and settlements with former staff. Suleyman apologized for driving people too hard and for a management style that was at times not constructive. He moved to a Google policy role as VP of AI Policy, co-founded Inflection AI in 2022, and in March 2024 became CEO of Microsoft AI.^[52]

Military use, the AI Principles, and unionisation

In August 2024, Time reported that nearly 200 Google DeepMind employees had signed an internal letter dated May 16, 2024 asking the company to investigate whether militaries and weapons manufacturers were Google Cloud customers, to terminate military access to DeepMind technology, and to create a governance body to prevent such use in future. Google responded that it complies with its AI Principles and that its Israeli government cloud contract was not directed at military or intelligence workloads.^[143]

On February 4, 2025 Google published a revised set of AI Principles, co-signed by James Manyika and Hassabis, built around three tenets. The revision removed a section listing applications Google would not pursue, which had included weapons, surveillance technologies violating international norms, and technologies likely to cause overall harm. The blog post did not mention the removal; journalists noticed it by comparing versions. Hassabis wrote in the same post that "democracies should lead in AI development."^[142]

Union organising followed. About 300 London staff approached the Communication Workers Union in April 2025. In April and May 2026 the CWU said 98 percent of its members at DeepMind backed recognition, and the CWU and Unite wrote to Google UK seeking joint recognition for roughly 1,000 London employees, with demands including an end to US Department of Defense and Israeli military use of Google AI, restoration of the weapons and surveillance pledge, an independent ethics body, and a right to refuse work on moral grounds. Google's response noted that at that stage there had been no vote to unionise. The 98 percent figure was an internal union poll, not a statutory recognition ballot, and the outcome was unresolved as of August 2026.^[144]

Reproducibility and disclosure

Google DeepMind's science publications have drawn several substantive challenges.

The most consequential concerns AlphaChip. The 2021 Nature paper on reinforcement-learning chip floorplanning was followed by a dispute over whether the method beat conventional tools. A UC San Diego group led by Andrew Kahng reported at ISPD 2023 that their reimplementation did not outperform existing techniques, and Igor Markov, then at Synopsys, published a critical meta-analysis the same year. The formal record is easily garbled, so it is worth setting out. Nature published an author correction in March 2022, added an editor's note on September 20, 2023 saying the performance claims had been called into question, and on September 21, 2023 retracted the accompanying News and Views commentary at the request of its author, Kahng, who said new information about the methods had changed his assessment. The research paper itself was never retracted, and no Matters Arising was published against it. On September 26, 2024 Nature concluded its post-publication review in the authors' favour, published an addendum supplying previously missing methodological detail, and removed the editor's note. The same day Google gave the method the name AlphaChip and released a pre-trained checkpoint; a rebuttal preprint followed in November 2024. Markov's critique appeared in Communications of the ACM in October 2024. The technical disagreement is unresolved: the paper stands with an addendum, and critics maintain that the full training and evaluation inputs needed for independent replication were never released.^[112][113]

GNoME drew a published critique in Chemistry of Materials on April 8, 2024. Anthony Cheetham and Ram Seshadri of UC Santa Barbara reported "scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility," characterising many proposals as trivial dopant variants, symmetry-broken duplicates, or chemically implausible high-element-count compositions, and objecting that the outputs were crystalline compounds rather than materials in the sense of demonstrated function. Their sample was small and they said so: they examined the first 250 entries of the GNoME Explorer database and ten randomly selected entries from the much larger stable-structure listing, noting that a comprehensive review would have been too time-consuming. Google DeepMind's response was that it stood by all the claims in the paper and that hundreds of its predicted materials had already been independently synthesized. Google's original claim was of computationally predicted stable structures rather than synthesized materials, so part of the disagreement is about how the result was framed in publicity; the critique nonetheless remains the standard citation against reading GNoME as 380,000 new materials.^[28][114]

A related dispute concerned the autonomous synthesis laboratory paper published alongside GNoME by a Lawrence Berkeley National Laboratory group, which is often wrongly attributed to Google DeepMind. Its claim of 41 successful syntheses was challenged by Leeman and colleagues in PRX Energy in 2024, and the authors issued a correction in Nature on January 19, 2026 that changed the title from "novel materials" to "inorganic materials," clarified that the compounds were new to the prediction platform rather than to science, and reduced the confirmed successes to 36 with four inconclusive.^[158]

AlphaFold 3 prompted a dispute about disclosure rather than accuracy. The May 2024 Nature paper was published without the model code, and access was initially limited to a web server with usage caps and restrictions on ligand modelling. More than a thousand scientists signed a letter to Nature arguing that the omission compromised peer review and made the paper's claims impossible to test independently, since the code had been withheld even from the reviewers. Nature's editor-in-chief cited biosecurity concerns and offered pseudocode as a reproducibility substitute, and Google DeepMind committed to a release within six months. The code appeared on November 11, 2024 under a non-commercial licence, with model weights available by application to academic organisations. The source has since been relicensed under Apache 2.0, while the weights remain separately gated.^[115]

Publication practice more broadly has been questioned. Google DeepMind still publishes prolifically, and its research publications page was active through July 2026.^[145] But the Financial Times reported in April 2025 that the organization had imposed a roughly six-month embargo on selected generative-AI papers plus additional internal approval layers, with some papers blocked outright, particularly those describing Gemini in detail or making comparisons unfavourable to it. Google DeepMind did not publicly dispute the report. The pattern is visible in the published record: Gemini technical reports describe capabilities and evaluations without reproducible architecture or training-data detail, a change from the Gopher and Chinchilla era when DeepMind documented both at length.^{[145][68][102][103]}

Product-level criticism

AI Overviews attracted widespread criticism shortly after its May 2024 US rollout for confidently presenting satirical or low-quality web content as fact, including a widely circulated suggestion involving glue on pizza. Google described the examples as edge cases, said some were fabricated screenshots, and made changes to the system. Because AI Overviews sits above conventional Search results at very large scale, errors of this kind are more visible than equivalent errors in a chat product, and publishers have separately raised concerns that generated summaries reduce click-through to source sites.^[127]

See also

• Artificial intelligence
• Deep learning
• Reinforcement learning
• Gemini
• Gemini 3
• Gemini 3 Pro
• Gemini 3.1 Pro
• Gemini 3.5 Flash
• Gemma
• Gemma 4
• AlphaFold
• AlphaFold 3
• AlphaGo
• AlphaZero
• MuZero
• AlphaStar
• Agent57
• AlphaProof
• AlphaProof Nexus
• AlphaEvolve
• GraphCast
• GenCast
• GNoME
• AlphaGenome
• Genie 3
• Gemini Robotics
• Gemini Robotics 2
• Veo
• Imagen
• SynthID
• Frontier Safety Framework
• Tensor Processing Unit
• Google Brain
• Google Research
• Isomorphic Labs
• Waymo
• ChatGPT
• OpenAI
• Anthropic

References

1. ^Google DeepMind. "About Google DeepMind." Accessed August 1, 2026. deepmind.google/about ; Demis Hassabis. "From unlikely start-up to major scientific organisation: Entering our tenth year at DeepMind." Google DeepMind, December 5, 2019. deepmind.google/...ring-our-tenth-year-at-deepmind
2. ^The Nobel Prize. "The Nobel Prize in Chemistry 2024: Summary." October 9, 2024. nobelprize.org/...summary
3. ^UK Companies House. "DeepMind Technologies Limited, company number 07386350: company overview, officers, and filing history." Accessed August 1, 2026. find-and-update.company-information.service.gov.uk/...07386350 ; officers register: find-and-update.company-information.service.gov.uk/...officers
4. ^Volodymyr Mnih et al. "Human-level control through deep reinforcement learning." *Nature* 518, 2015. nature.com/...nature14236
5. ^Reuters. "Google to buy artificial intelligence company DeepMind." January 27, 2014. finance.yahoo.com/...ntelligence-company-022140062
6. ^Rich Evans and Jim Gao. "DeepMind AI reduces Google data centre cooling bill by 40%." Google DeepMind, July 20, 2016. deepmind.google/...-data-centre-cooling-bill-by-40
7. ^David Silver et al. "Mastering the game of Go with deep neural networks and tree search." *Nature* 529, 2016. nature.com/...nature16961 ; Google DeepMind. "AlphaGo." Accessed August 1, 2026. deepmind.google/...alphago
8. ^David Silver et al. "Mastering the game of Go without human knowledge." *Nature* 550, 2017. nature.com/...nature24270
9. ^David Silver et al. "A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play." *Science* 362, 2018. science.org/...science.aar6404
10. ^Julian Schrittwieser et al. "Mastering Atari, Go, chess and shogi by planning with a learned model." *Nature* 588, 2020. nature.com/...s41586-020-03051-4
11. ^Oriol Vinyals et al. "Grandmaster level in StarCraft II using multi-agent reinforcement learning." *Nature* 575, 2019. nature.com/...s41586-019-1724-z
12. ^John Jumper et al. "Highly accurate protein structure prediction with AlphaFold." *Nature* 596, 2021. nature.com/...s41586-021-03819-2
13. ^EMBL-EBI and Google DeepMind. "AlphaFold Protein Structure Database." Accessed August 1, 2026. alphafold.ebi.ac.uk ; Google DeepMind. "Demis Hassabis and John Jumper awarded Nobel Prize in Chemistry." October 9, 2024. deepmind.google/...warded-nobel-prize-in-chemistry
14. ^Demis Hassabis. "Announcing Google DeepMind." Google DeepMind, April 20, 2023. deepmind.google/...announcing-google-deepmind
15. ^Ashish Vaswani et al. "Attention Is All You Need." *Advances in Neural Information Processing Systems 30*, 2017. proceedings.neurips.cc/...ttention-is-all-you-need
16. ^Google DeepMind. "Responsibility & Safety." Accessed August 1, 2026. deepmind.google/...responsibility-safety
17. ^Google. "Koray Kavukcuoglu." Accessed August 1, 2026. blog.google/...koray-kavukcuoglu
18. ^Koray Kavukcuoglu, Jeff Dean, Oriol Vinyals, and Noam Shazeer. "Gemini 3.5: frontier intelligence with action." Google, May 19, 2026. blog.google/...gemini-3-5
19. ^Scott Reed et al. "A Generalist Agent." *Transactions on Machine Learning Research*, 2022. arxiv.org/...2205.06175
20. ^Google DeepMind. "SIMA 2: An agent that plays, reasons, and learns with you in virtual 3D worlds." November 13, 2025. deepmind.google/...s-with-you-in-virtual-3d-worlds
21. ^Google. "Introducing Gemini 2.0: our new AI model for the agentic era." December 11, 2024. blog.google/...google-gemini-ai-update-december-2024
22. ^Josh Abramson et al. "Accurate structure prediction of biomolecular interactions with AlphaFold 3." *Nature* 630, 2024. nature.com/...s41586-024-07487-w
23. ^Google DeepMind. "AlphaProteo generates novel proteins for biology and health research." September 5, 2024. deepmind.google/...for-biology-and-health-research
24. ^Ziga Avsec et al. "Advancing regulatory variant effect prediction with AlphaGenome." *Nature* 649(8099):1206-1218, published online January 28, 2026. doi.org/...s41586-025-10014-0 ; Google DeepMind. "AlphaGenome: AI for better understanding the genome." June 25, 2025. deepmind.google/...better-understanding-the-genome
25. ^Remi Lam et al. "Learning skillful medium-range global weather forecasting." *Science* 382, 2023. science.org/...science.adi2336
26. ^Ilan Price et al. "Probabilistic weather forecasting with machine learning." *Nature* 637, 2025. nature.com/...s41586-024-08252-9
27. ^Google. "WeatherNext 2: Google DeepMind's most advanced forecasting model." November 17, 2025. blog.google/...weathernext-2
28. ^Amil Merchant et al. "Scaling deep learning for materials discovery." *Nature* 624, 2023. nature.com/...s41586-023-06735-9
29. ^Alhussein Fawzi et al. "Discovering faster matrix multiplication algorithms with reinforcement learning." *Nature* 610, 2022. nature.com/...s41586-022-05172-4
30. ^Daniel J. Mankowitz et al. "Faster sorting algorithms discovered using deep reinforcement learning." *Nature* 618, 2023. nature.com/...s41586-023-06004-9
31. ^Bernardino Romera-Paredes et al. "Mathematical discoveries from program search with large language models." *Nature* 625, 2024. nature.com/...s41586-023-06924-6
32. ^Google DeepMind. "AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms." May 14, 2025. deepmind.google/...r-designing-advanced-algorithms
33. ^Google DeepMind. "AI achieves silver-medal standard solving International Mathematical Olympiad problems." July 25, 2024. deepmind.google/...-problems-at-silver-medal-level
34. ^Google DeepMind. "Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical Olympiad." July 21, 2025. deepmind.google/...rnational-mathematical-olympiad
35. ^George Tsoukalas et al. "Advancing Mathematics Research with AI-Driven Formal Proof Search." arXiv preprint, submitted May 21, 2026, revised June 8, 2026. arxiv.org/...2605.22763
36. ^Google Research. "Accelerating scientific breakthroughs with an AI co-scientist." February 19, 2025. research.google/...hroughs-with-an-ai-co-scientist ; Google DeepMind. "Co-scientist: A multi-agent AI partner to accelerate research." May 19, 2026. deepmind.google/...-partner-to-accelerate-research
37. ^Sundar Pichai and Demis Hassabis. "Introducing Gemini: Google's most capable AI model yet." Google, December 6, 2023. blog.google/...google-gemini-ai
38. ^Sundar Pichai and Demis Hassabis. "Our next-generation model: Gemini 1.5." Google, February 15, 2024. blog.google/...next-generation-model-february-2024 ; Google. "Gemini 1.5 Pro updates, 1.5 Flash debut and 2 new Gemma models." May 14, 2024. blog.google/...ni-gemma-developer-updates-may-2024 ; Koray Kavukcuoglu. "Gemini 2.5: Our most intelligent AI model." Google, March 25, 2025. blog.google/...i-model-thinking-updates-march-2025
39. ^Google. "Introducing Gemini 3." November 18, 2025. blog.google/...gemini-3
40. ^Google. "Gemini 3.1 Pro: A smarter model for your most complex tasks." February 19, 2026. blog.google/...gemini-3-1-pro
41. ^Google. "Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber." July 21, 2026. blog.google/...lash-3-5-flash-lite-3-5-flash-cyber
42. ^Google DeepMind. "WaveNet launches in the Google Assistant." October 4, 2017. deepmind.google/...aunches-in-the-google-assistant
43. ^Google DeepMind. "Veo." Accessed August 1, 2026. deepmind.google/...veo
44. ^Google DeepMind. "Genie 3: A new frontier for world models." August 5, 2025. deepmind.google/...a-new-frontier-for-world-models
45. ^Carolina Parada. "Gemini Robotics brings AI into the physical world." Google DeepMind, March 12, 2025. deepmind.google/...ings-ai-into-the-physical-world
46. ^Carolina Parada. "Gemini Robotics On-Device brings AI to local robotic devices." Google DeepMind, June 24, 2025. deepmind.google/...ngs-ai-to-local-robotic-devices
47. ^Google DeepMind. "Gemini Robotics." Accessed August 1, 2026. deepmind.google/...gemini-robotics
48. ^Google DeepMind. "SynthID." Accessed August 1, 2026. deepmind.google/...synthid ; Sumanth Dathathri et al. "Scalable watermarking for identifying large language model outputs." *Nature* 634, 2024. nature.com/...s41586-024-08025-4
49. ^Four Flynn, Helen King, and Anca Dragan. "Strengthening our Frontier Safety Framework." Google DeepMind, September 22, 2025, updated April 17, 2026. deepmind.google/...g-our-frontier-safety-framework
50. ^UK Information Commissioner's Office. "Google DeepMind and class action lawsuit." Accessed August 1, 2026. ico.org.uk/...le-deepmind-and-class-action-lawsuit ; Google DeepMind. "The Information Commissioner, the Royal Free, and what we've learned." July 3, 2017. deepmind.google/...oyal-free-and-what-weve-learned
51. ^Dominic King. "DeepMind's health team joins Google Health." Google DeepMind, September 18, 2019. deepmind.google/...health-team-joins-google-health
52. ^Sam Shead. "DeepMind co-founder was investigated by law firm after staff complaints." CNBC, January 27, 2021. cnbc.com/...ed-by-law-firm-after-staff-complaints- ; Microsoft. "Mustafa Suleyman, DeepMind and Inflection Co-founder, joins Microsoft to lead Copilot." March 19, 2024. blogs.microsoft.com/...s-microsoft-to-lead-copilot
53. ^Demis Hassabis. "Introducing Isomorphic Labs." Isomorphic Labs, November 1, 2021. isomorphiclabs.com/...introducing-isomorphic-labs
54. ^Jonas Degrave et al. "Magnetic control of tokamak plasmas through deep reinforcement learning." *Nature* 602, 2022. nature.com/...s41586-021-04301-9
55. ^Sissie Hsiao. "Bard becomes Gemini: Try Ultra 1.0 and a new mobile app today." Google, February 8, 2024. blog.google/...bard-gemini-advanced-app
56. ^Sundar Pichai. "Alphabet Q2 2026 earnings call remarks." Google, July 22, 2026. blog.google/...alphabet-earnings-q2-2026
57. ^Sundar Pichai. "Google I/O 2026: Sundar Pichai's opening keynote." Google, May 19, 2026. blog.google/...sundar-pichai-io-2026
58. ^Google Developers Blog. "Start building with Gemini 2.5 Flash." April 17, 2025. developers.googleblog.com/...-with-gemini-25-flash
59. ^Google Developers Blog. "Introducing Gemini 2.5 Flash Image, our state-of-the-art image model." August 26, 2025. developers.googleblog.com/...emini-2-5-flash-image
60. ^Google. "Try Deep Think in the Gemini app." August 1, 2025. blog.google/...gemini-2-5-deep-think
61. ^Google. "Introducing the Gemini 2.5 Computer Use model." October 7, 2025. blog.google/...gemini-computer-use-model
62. ^Google. "Nano Banana Pro: Gemini 3 Pro Image model from Google DeepMind." November 2025. blog.google/...nano-banana-pro
63. ^Google. "Introducing Gemini 3 Flash: Benchmarks, global availability." December 17, 2025. blog.google/...gemini-3-flash
64. ^Google. "Gemini 3 Deep Think: Advancing science, research and engineering." February 12, 2026. blog.google/...gemini-3-deep-think
65. ^Google. "Nano Banana 2: Combining Pro capabilities with lightning-fast speed." February 26, 2026. blog.google/...nano-banana-2
66. ^Google. "Start building with Nano Banana 2 Lite and Gemini Omni Flash." 2026. blog.google/...gemini-omni-flash-nano-banana-2-lite
67. ^Gemini Team, Google. "Gemini: A Family of Highly Capable Multimodal Models." arXiv:2312.11805, December 2023. arxiv.org/...2312.11805
68. ^Google. "Gemma: Introducing new state-of-the-art open models." February 21, 2024. blog.google/...gemma-open-models
69. ^Google. "Google launches Gemma 2, its next generation of open models." June 27, 2024. blog.google/...google-gemma-2
70. ^Google. "Gemma 3: Google's new open model based on Gemini 2.0." March 12, 2025. blog.google/...gemma-3
71. ^Google Developers Blog. "Announcing Gemma 3n preview: powerful, efficient, mobile-first AI." 2025. developers.googleblog.com/...introducing-gemma-3n
72. ^Google. "Gemma 4: Byte for byte, the most capable open models." April 2, 2026. blog.google/...gemma-4
73. ^Google. "DiffusionGemma: 4x faster text generation." June 10, 2026. blog.google/...fusion-gemma-faster-text-generation
74. ^Google. "DolphinGemma: How Google AI is helping decode dolphin communication." April 14, 2025. blog.google/...dolphingemma
75. ^Google DeepMind. "Imagen." Accessed August 1, 2026. deepmind.google/...imagen
76. ^Google. "Generative media models at I/O 2025: Veo 3, Imagen 4 and Flow." May 20, 2025. blog.google/...generative-media-models-io-2025
77. ^Google DeepMind. "Music AI Sandbox, now with new features and broader access." April 24, 2025. deepmind.google/...new-features-and-broader-access
78. ^Google DeepMind. "Genie 2: A large-scale foundation world model." December 4, 2024. deepmind.google/...ge-scale-foundation-world-model
79. ^Google. "Project Genie: AI world model now available for Ultra users in the U.S." January 29, 2026. blog.google/...project-genie
80. ^Google DeepMind. "Gemini Robotics 1.5 brings AI agents into the physical world." September 25, 2025. deepmind.google/...-agents-into-the-physical-world
81. ^Carolina Parada. "Gemini Robotics 2." Google DeepMind, July 30, 2026. deepmind.google/...gemini-robotics-2
82. ^Google. "Google AI Studio." Accessed August 1, 2026. aistudio.google.com
83. ^Google. "Gemini CLI: your open-source AI agent." June 25, 2025. blog.google/...ing-gemini-cli-open-source-ai-agent
84. ^Google. "Google Antigravity." Accessed August 1, 2026. antigravity.google
85. ^Google Cloud. "Vertex AI." Accessed August 1, 2026. cloud.google.com/vertex-ai
86. ^Google Cloud. "Introducing Gemini Enterprise." October 9, 2025. blog.google/...gemini-enterprise
87. ^Google. "Gemini Live." Accessed August 1, 2026. gemini.google/...gemini-live
88. ^Google. "NotebookLM." Accessed August 1, 2026. notebooklm.google
89. ^Google DeepMind. "Project Astra." Accessed August 1, 2026. deepmind.google/...project-astra
90. ^Google Labs. "Project Mariner." Accessed August 1, 2026. labs.google/mariner
91. ^Google. "Our eighth generation TPUs: two chips for the agentic era." April 22, 2026. blog.google/...eighth-generation-tpu-agentic-era
92. ^Google. "3 things to know about Ironwood, Google's latest TPU." 2025. blog.google/...ironwood-google-tpu-things-to-know ; Google Cloud. "TPU7x (Ironwood)." Accessed August 1, 2026. docs.cloud.google.com/...tpu7x
93. ^Google Cloud. "Cloud TPU system architecture." Accessed August 1, 2026. docs.cloud.google.com/...system-architecture-tpu-vm
94. ^Anthropic. "Anthropic and Google Cloud expand their partnership." October 23, 2025. anthropic.com/...ding-our-use-of-google-cloud-tpus
95. ^The Information. "Google is developing a Gemini-specific inference chip." July 20, 2026. theinformation.com
96. ^Google DeepMind. "Introducing the Frontier Safety Framework." May 17, 2024. deepmind.google/...g-the-frontier-safety-framework
97. ^Rohin Shah et al. "An Approach to Technical AGI Safety and Security." Google DeepMind, April 2025. arxiv.org/...2504.01849
98. ^Google DeepMind. "Gemma Scope: helping the safety community shed light on the inner workings of language models." July 31, 2024. deepmind.google/...ner-workings-of-language-models
99. ^Adria Puigdomenech Badia et al. "Agent57: Outperforming the Atari Human Benchmark." *Proceedings of the 37th International Conference on Machine Learning*, PMLR 119:507-517, 2020. arxiv.org/...2003.13350
100. ^Jordan Hoffmann et al. "Training Compute-Optimal Large Language Models." arXiv:2203.15556, March 2022. arxiv.org/...2203.15556
101. ^Jack W. Rae et al. "Scaling Language Models: Methods, Analysis & Insights from Training Gopher." arXiv:2112.11446, December 2021. arxiv.org/...2112.11446
102. ^Yujia Li et al. "Competition-level code generation with AlphaCode." *Science* 378, December 2022. science.org/...science.abq1158 ; Google DeepMind. "AlphaCode 2 Technical Report." December 6, 2023. storage.googleapis.com/...AlphaCode2_Tech_Report.pdf
103. ^Jean-Baptiste Alayrac et al. "Flamingo: a Visual Language Model for Few-Shot Learning." arXiv:2204.14198, April 2022. arxiv.org/...2204.14198
104. ^Amelia Glaese et al. "Improving alignment of dialogue agents via targeted human judgements." DeepMind, September 22, 2022. arxiv.org/...2209.14375
105. ^Andrew Jaegle et al. "Perceiver: General Perception with Iterative Attention." *Proceedings of the 38th International Conference on Machine Learning*, 2021. arxiv.org/...2103.03206
106. ^Google DeepMind. "RoboCat: A self-improving robotic agent." June 20, 2023. deepmind.google/...-a-self-improving-robotic-agent
107. ^SIMA Team, Google DeepMind. "Scaling Instructable Agents Across Many Simulated Worlds." arXiv:2404.10179, March 2024. arxiv.org/...2404.10179
108. ^Jun Cheng et al. "Accurate proteome-wide missense variant effect prediction with AlphaMissense." *Science* 381, September 19, 2023. science.org/...science.adg7492
109. ^Johannes Bausch et al. "Learning high-accuracy error decoding for quantum processors." *Nature* 635, November 20, 2024. nature.com/...s41586-024-08148-8
110. ^Azalia Mirhoseini et al. "A graph placement methodology for fast chip design." *Nature* 594, 2021, with addendum published September 26, 2024. nature.com/...s41586-021-03544-w ; Google DeepMind. "How AlphaChip transformed computer chip design." September 26, 2024. deepmind.google/...ransformed-computer-chip-design
111. ^Igor L. Markov. "The False Dawn: Reevaluating Google's Reinforcement Learning for Chip Macro Placement." arXiv:2306.09633, 2023; published in *Communications of the ACM*, October 2024. arxiv.org/...2306.09633 ; Anna Goldie, Azalia Mirhoseini, and Jeff Dean. "That Chip Has Sailed: A Critique of Unfounded Skepticism Around AI for Chip Design." arXiv:2411.10053, November 2024. arxiv.org/...2411.10053
112. ^Anthony K. Cheetham and Ram Seshadri. "Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery." *Chemistry of Materials* 36(8):3490-3495, April 8, 2024. pubs.acs.org/...acs.chemmater.4c00643
113. ^Ewen Callaway. "AI protein-prediction tool AlphaFold3 is now more open." *Nature*, November 2024. nature.com/...d41586-024-03708-4 ; *Nature* editorial. "AlphaFold3: why did Nature publish it without its code?" May 2024. nature.com/...d41586-024-01463-0
114. ^Trieu H. Trinh et al. "Solving olympiad geometry without human demonstrations." *Nature* 625, January 17, 2024. nature.com/...s41586-023-06747-5
115. ^Yuri Chervonyi et al. "Gold-medalist Performance in Solving Olympiad Geometry with AlphaGeometry2." arXiv:2502.03544, February 2025. arxiv.org/...2502.03544
116. ^AlphaProof Team, Google DeepMind. "Olympiad-level formal mathematical reasoning with reinforcement learning." *Nature* 651(8106):607-613, published online November 12, 2025. doi.org/...s41586-025-09833-y
117. ^Zhe Wang et al. "TacticAI: an AI assistant for football tactics." *Nature Communications* 15, March 19, 2024. nature.com/...s41467-024-45965-x
118. ^Google DeepMind. "AlphaEarth Foundations helps map our planet in unprecedented detail." July 2025. deepmind.google/...-planet-in-unprecedented-detail
119. ^*Prismall v Google UK Ltd* [2023] EWHC 1169 (KB), High Court (King's Bench Division), May 19, 2023. caselaw.nationalarchives.gov.uk/...1169 ; *Prismall v Google UK Ltd and DeepMind Technologies Ltd* [2024] EWCA Civ 1516, Court of Appeal (Civil Division), December 11, 2024. caselaw.nationalarchives.gov.uk/...1516
120. ^Natasha Lomas. "Google completes controversial takeover of DeepMind Health." TechCrunch, September 19, 2019. techcrunch.com/...sial-takeover-of-deepmind-health ; Andrew Orlowski. "DeepMind Health told to explain business model, relationship to Google." The Register, June 15, 2018. theregister.com/...ationship_google_business_model ; Natasha Lomas. "Google confirms it's pulling the plug on Streams, its UK clinician support app." TechCrunch, August 26, 2021. techcrunch.com/...ams-its-uk-clinician-support-app
121. ^Daniel Zheng et al. "AI co-mathematician: Accelerating mathematicians with agentic AI." arXiv:2605.06651, May 7, 2026. arxiv.org/...2605.06651
122. ^Liz Reid. "AI Overviews: About last week." Google, May 30, 2024. blog.google/...ai-overviews-update-may-2024
123. ^DeepMind Technologies Limited. "Annual report and financial statements," financial years 2015 to 2024, filed at UK Companies House. find-and-update.company-information.service.gov.uk/...-history
124. ^Ingrid Lunden and Catherine Shu. "Google acquires artificial intelligence startup DeepMind." TechCrunch, January 26, 2014. techcrunch.com/...google-deepmind
125. ^Sam Shead. "DeepMind's mysterious ethics board will reportedly control AGI if it's ever created." Forbes, March 14, 2019. forbes.com/...edly-control-agi-if-its-ever-created
126. ^Karen Hao. "Google has cancelled its AI ethics board after a backlash from staff." *MIT Technology Review*, April 5, 2019. technologyreview.com/...fter-a-backlash-from-staff
127. ^Sundar Pichai. "Google DeepMind: Bringing together two world-class AI teams." Google, April 20, 2023. blog.google/...april-ai-update
128. ^Google DeepMind. "Careers: our locations." Accessed August 1, 2026. deepmind.google/...careers
129. ^Epoch AI. "AI companies staff reports dataset." Accessed August 1, 2026. epoch.ai/...ai_companies_staff_reports.csv
130. ^Jennifer Elias. "Google Gemini co-lead Noam Shazeer leaves for OpenAI." CNBC, June 18, 2026. cnbc.com/...co-lead-noam-shazeer-leaves-for-openai
131. ^"John Jumper to leave Google DeepMind for Anthropic." CNBC, June 19, 2026. cnbc.com/...to-leave-google-deepmind-for-anthropic ; "Nobel laureate John Jumper is leaving DeepMind for rival Anthropic." TechCrunch, June 20, 2026. techcrunch.com/...ing-deepmind-for-rival-anthropic
132. ^"Google DeepMind disbands its Nobel-prize winning AlphaFold team." Engadget, July 30, 2026, reporting *Financial Times*, July 29, 2026. engadget.com/...google-shuts-down-alphafold
133. ^"Google DeepMind's Lila Ibrahim is helping the world get ready for AI." *Fast Company*, 2026. fastcompany.com/...ping-the-world-get-ready-for-ai
134. ^UK Information Commissioner's Office. "Royal Free - Google DeepMind trial failed to comply with data protection law." Press release and undertaking, July 3, 2017. web.archive.org/...comply-with-data-protection-law
135. ^Julia Powles and Hal Hodson. "Google DeepMind and healthcare in an age of algorithms." *Health and Technology* 7(4):351-367, 2017. doi.org/...s12553-017-0179-1
136. ^James Manyika and Demis Hassabis. "Responsible AI: Our 2024 report and ongoing work." Google, February 4, 2025. blog.google/...onsible-ai-2024-report-ongoing-work
137. ^Billy Perrigo. "Google DeepMind staff call on company to drop military contracts." *Time*, August 22, 2024. time.com/...-ai-deepmind-military-contracts-israel
138. ^Beatrice Nolan. "Google DeepMind staff seek union recognition over military AI contracts." *Fortune*, May 5, 2026. fortune.com/...entagon-deal-israeli-defense-forces
139. ^"Google's DeepMind slows research publications to maintain AI edge." Investing.com, April 1, 2025, reporting *Financial Times*. investing.com/...s-to-maintain-ai-edge--ft-3959172 ; Google DeepMind. "Publications." Accessed August 1, 2026. deepmind.google/...publications
140. ^Parmy Olson. "Google unit DeepMind tried and failed to win AI autonomy from parent." *Wall Street Journal*, May 21, 2021. wsj.com/...win-ai-autonomy-from-parent-11621592951
141. ^Reed Albergotti. "Google names new chief AI architect to advance developments." Semafor, June 11, 2025. semafor.com/...i-architect-to-advance-developments
142. ^Google DeepMind. "Models." Accessed August 1, 2026. deepmind.google/models
143. ^Harry Booth. "Google AI researcher joins Meta's world-models push." *Time*, October 21, 2025. time.com/...meta-google-ai-researcher-world-models
144. ^Google. "How Google's LearnLM generative AI models support teachers and learners." 2024. blog.google/...google-learnlm-gemini-generative-ai ; Google. "I/O 2025: LearnLM in Gemini 2.5 and more AI updates to help people learn." May 2025. blog.google/...google-gemini-learnlm-update
145. ^Andrew W. Senior et al. "Protein structure prediction using multiple deep neural networks in the 13th Critical Assessment of Protein Structure Prediction (CASP13)." *Proteins* 87(12), 2019. doi.org/...prot.25834
146. ^Google DeepMind. "AlphaFold: a solution to a 50-year-old grand challenge in biology." November 30, 2020. deepmind.google/...-old-grand-challenge-in-biology ; Google DeepMind. "AlphaFold: Five Years of Impact." November 25, 2025. deepmind.google/...alphafold-five-years-of-impact ; EMBL-EBI. "AlphaFold Protein Structure Database launch." July 22, 2021. ebi.ac.uk/...alphafold-database-launch
147. ^Shiven Sinha et al. "Wu's Method can Boost Symbolic AI to Rival Silver Medalists and AlphaGeometry to Outperform Gold Medalists at IMO Geometry." arXiv:2404.06405, April 9, 2024. arxiv.org/...2404.06405
148. ^Jean-Guillaume Dumas, Clement Pernet, and Alexandre Sedoglavic. "A non-commutative algorithm for multiplying 4x4 matrices using 48 non-complex multiplications." arXiv:2506.13242, June 16, 2025. arxiv.org/...2506.13242
149. ^Google DeepMind. "AlphaEvolve: Gemini-powered coding agent scaling impact across fields." May 7, 2026. deepmind.google/...alphaevolve-impact
150. ^Jose R. Penades et al. "AI mirrors experimental science to uncover a mechanism of gene transfer crucial to bacterial evolution." *Cell* 188(23):6654-6665, 2025. doi.org/...j.cell.2025.08.018 ; Google DeepMind. "Accelerating scientific discovery with Co-Scientist." *Nature* 655(8122):487-496, published online May 19, 2026. doi.org/...s41586-026-10644-y
151. ^Google DeepMind. "Towards Autonomous Mathematics Research." arXiv:2602.10177, February 10, 2026. arxiv.org/...2602.10177 ; Google DeepMind. "Gemini Deep Think: Redefining the Future of Scientific Research." February 11, 2026. deepmind.google/...iscovery-with-gemini-deep-think
152. ^Joshua Leeman et al. "Challenges in High-Throughput Inorganic Materials Prediction and Autonomous Synthesis." *PRX Energy* 3(1):011002, 2024. doi.org/...PRXEnergy.3.011002 ; "Author Correction: An autonomous laboratory for the accelerated synthesis of inorganic materials." *Nature* 650(8100), published online January 19, 2026. doi.org/...s41586-025-09992-y
153. ^Sarah J. Horton et al. "Making sense of missense: challenges and opportunities in variant pathogenicity prediction." *Disease Models & Mechanisms* 17(12):dmm052218, December 2024. pmc.ncbi.nlm.nih.gov/...PMC11683568
154. ^Google DeepMind. "How we're supporting better tropical cyclone prediction with AI." June 12, 2025. deepmind.google/...ical-cyclone-prediction-with-ai ; US National Hurricane Center. "2025 NHC Verification Report Preview: Atlantic Basin." nhc.noaa.gov/...rification_Report_2025_Preview.pdf
155. ^Andrew Senior et al. "A scalable and real-time neural decoder for topological quantum codes." arXiv:2512.07737, December 8, 2025, revised March 11, 2026. arxiv.org/...2512.07737
156. ^Google DeepMind. "Google DeepMind supports the US Department of Energy on the Genesis Mission." December 18, 2025. deepmind.google/...department-of-energy-on-genesis ; Google DeepMind. "Our approach to bioresilience." July 16, 2026. deepmind.google/...our-approach-to-bioresilience
157. ^Isomorphic Labs. "Isomorphic Labs announces $600m external investment round." March 31, 2025. isomorphiclabs.com/...0m-external-investment-round ; Ewen Callaway. "DeepMind spin-off unveils secretive drug-design engine." *Nature*, February 19, 2026. doi.org/...d41586-026-00365-7
158. ^Manuel Kauers and Jakob Moosbauer. "The FBHHRBNRSSSHK-Algorithm for Multiplication in Z_2^{5x5} is still not the end of the story." arXiv:2210.04045, October 8, 2022. arxiv.org/...2210.04045 ; Manuel Kauers and Jakob Moosbauer. "Flip Graphs for Matrix Multiplication." arXiv:2212.01175, December 2, 2022. arxiv.org/...2212.01175
159. ^Sundar Pichai and Demis Hassabis. "The next chapter of our AI momentum." Google, August 5, 2026. blog.google/...next-chapter-ai-momentum
160. ^Jeff Dean (@JeffDean). "Announcing Discovery Loop!" X, August 5, 2026. x.com/...2085034604172603724
161. ^Steven Levy. "4 of Google's Top AI Brains Are Leaving, and Launching Their Own AI Startup." Wired, August 5, 2026. wired.com/...jeff-dean-google-discovery-loop-startup