Every published Top AI Stories item tagged with Google DeepMind, newest first.
Gemini Robotics 2 is a three-model system, pairing a vision-language-action model for motor control with an embodied reasoning model for planning and an efficient on-device variant. The headline capability is intelligent whole-body control, letting a humanoid walk, crouch, reach and manipulate an object at the same time rather than treating locomotion and manipulation as separate problems. Demonstrations ran on Apptronik's Apollo 2 humanoid and several research platforms. The reasoning model is available now in Google AI Studio; the other two are limited to early-access partners.
Google put Lyria 3.5 into Flow Music, its consumer song-generation app, with richer melodic structure, closer adherence to lyric prompts, more emotionally nuanced vocals and better pronunciation, plus finer control over tempo and duration. It arrives as a free upgrade for existing Flow Music users rather than a new paid tier. Flow Music is Google's most direct answer to Suno and Udio, and Lyria is now the model underneath it.
An open letter published under the banner Pacing the Frontier now carries 1,224 signatures from employees at OpenAI, Anthropic, Google DeepMind, Meta, Thinking Machines and others, asking the US government to back an international effort to build "the technical and governance tools needed to deliberately pace the frontier of automated AI development." The signatories include Anthropic chief executive Dario Amodei and co-founder Jack Clark, OpenAI chief scientist Jakub Pachocki, and Thinking Machines chief scientist John Schulman. The argument is that competitive pressure makes unilateral restraint impossible, so a shared mechanism has to come from outside any single lab.
Google committed $40 million in AI tools and cloud credits to the Genesis Mission, a White House initiative to double the pace of US scientific discovery within a decade. Tens of thousands of researchers across the US Energy Department's 17 national labs get a year of access to systems like AlphaFold, AlphaEvolve, and Gemini for Government. Early results are concrete rather than promissory: one materials lab cut microscope calibration from more than 90 minutes to about 13, freeing scientists to run real-time, autonomous experiments.
Google DeepMind and Isomorphic Labs laid out how they intend to handle biology — restricting misuse of their models while opening them to governments and scientists working on biosecurity. The two describe more than 15 partnerships with government bodies, biosecurity organizations, and research groups, and point to work already running: AlphaFold and AlphaGenome for pathogen characterization, AlphaEvolve tuned for metagenomic sequencing to catch outbreaks earlier, and a unit inside Isomorphic aimed at designing countermeasures quickly when a novel pathogen appears. Extending SynthID-style watermarking to screen DNA synthesis orders comes next.
DeepMind chief executive Demis Hassabis called for an independent standards body to test frontier AI models before they ship, modeled on Wall Street's self-regulator, FINRA. Under his plan, labs would voluntarily submit models for review up to 30 days before launch, then shift to mandatory checks as a condition of access to the United States market. The group would be backed by government but funded and run by the industry, staffed with technical and open-source experts.
Google DeepMind and India's Atal Innovation Mission launched ATL Saathi, a Gemini-powered web app that helps teachers in the country's network of school "tinkering labs" plan projects, deliver micro-lessons, and mentor students across eight languages. The tool is starting as a live pilot in an initial cohort of 100 schools, aimed at cutting educators' administrative load. It is early — no outcome data yet — but it is a concrete deployment built to stretch scarce hands-on STEM-teaching capacity.
Google DeepMind and the independent film studio A24 announced a research partnership that pairs the lab's researchers with A24's filmmakers to develop new creative tools and production workflows, with Google also taking an undisclosed equity stake in the studio. The companies named no specific models and set no fixed deliverables, framing it as open-ended, exploratory work. It is a notable move as frontier labs court Hollywood, arriving amid ongoing tension over AI's role in creative work.
OpenAI has held early-stage talks about handing the US government a 5 percent equity stake — worth roughly $43 billion at its $852 billion valuation — the Financial Times reported. CEO Sam Altman pitched it as the best way to share AI's upside with the public, and floated a broader framework in which Washington would hold 5 percent of each leading US lab, including Anthropic, Google, Meta, and xAI, through a sovereign "public wealth fund." The stakes would be donated, not sold — but talks are early, would need Congressional sign-off, and neither OpenAI nor the White House has confirmed anything.
Google put Gemini Spark, its agentic assistant, into a macOS beta for US users on the $100-per-month Google AI Ultra plan. On the desktop, Spark can work with local files and connect to apps like Google Tasks and Keep, pushing it into the same territory as Claude's desktop app and Microsoft's Copilot. Google says remote task-handling — fetching files from a laptop while you are away — will follow. For now it is US-only and limited to adults.
Google launched Nano Banana 2 Lite, its fastest and cheapest image model — four-second generations at 3.4 cents per thousand images, aimed at high-volume developer pipelines. It slots below the generalist Nano Banana 2 and the pricier Nano Banana Pro, and rolls out across the Gemini API, AI Studio, Search, the Gemini app, NotebookLM, and Google Photos. Alongside it, the new Gemini Omni Flash turns images into ten-second videos at 10 cents per second of output.
Google has now lost four senior AI researchers in roughly a week — the clearest sign yet that rivals are picking off the people who built Gemini. Transformer co-author Noam Shazeer is heading to OpenAI, while AlphaFold Nobel laureate John Jumper and two more Gemini contributors are going to Anthropic. Jumper's exit alone helped wipe $245 billion off Alphabet's market value in a single session. The departures revive a pointed question for the AI race: can Google DeepMind stay at the frontier while its rivals keep hiring away its core talent?
Google Research retrofitted multi-token prediction (MTP) onto its existing Gemini Nano models without retraining them, appending a lightweight transformer head that predicts several future tokens in parallel by reusing the main model's frozen cache. The "zero-copy" design adds no separate memory cache and delivers more than 50% faster generation on a Pixel 9 while saving about 130 megabytes of memory per instance. It is a concrete win for on-device AI, where battery and memory limits make every efficiency gain count for features like notification summaries and proofreading.
An IEEE Spectrum feature surveys how fast AI is closing in on professional mathematics — gold-medal performances at the International Mathematical Olympiad, a DeepMind system producing publishable research in arithmetic geometry, and a tool that formalized a Fields Medalist's proofs in days. The progress is forcing uncomfortable questions: if machines handle the technical work, what is mathematics for, how do students build intuition, and do humans risk becoming, in one researcher's phrase, "priests to oracles"?
The Linux Foundation, backed by Amazon, Anthropic, Google, Microsoft, OpenAI, NVIDIA, IBM, and banks including Citi and JPMorgan Chase, launched Akrites, a coordinated effort to fix vulnerabilities in the open-source software that runs critical systems. The group warns that AI has made finding flaws far easier than fixing them — of the thousands of validated vulnerabilities surfaced recently, fewer than 5 percent have been patched. Akrites gives maintainers one trusted place to report and remediate, and will act as a "maintainer of last resort" for abandoned but critical packages.
Google released the Agentic Resource Discovery (ARD) specification, an open, Apache-licensed protocol that lets AI agents look up which tools and services exist, decide which to use, and verify they are safe to connect to. Microsoft, Salesforce, GitHub, Hugging Face, Nvidia, and Amazon are backing it under a new Linux Foundation working group — a coalition that reads as an attempt to set the agentic web's plumbing before OpenAI and Anthropic do. Reference implementations are already live from Google Cloud and Hugging Face.
At the G7 summit in France, Anthropic's Dario Amodei and Google DeepMind's Demis Hassabis pitched a US-led AI coalition, and leaders weighed a "trusted partner" framework that would exempt vetted allied nations from Washington's new export limits. The talks followed the Trump administration's June order blocking Anthropic from serving its top models abroad — and pointed warnings from Emmanuel Macron and Narendra Modi that a US "kill switch" on AI access would hurt allies and American firms alike.
Google opened preorders for the Google Home Speaker, its first standalone smart speaker since 2020 and the first built around Gemini rather than the older Assistant. It ships June 25 for about $100 and supports back-and-forth conversation, mid-sentence corrections, and follow-up questions without repeating a wake word. A $10-per-month Home Premium tier unlocks Gemini Live and camera summaries, free for the first six months.
Google DeepMind, working with the UK government and three local councils, built an AI tool that drafts planning assessments, checks proposals against local policy, and summarizes public objections for housing applications. Early trials in Barnet, Camden, and Dorset aim to cut decision times in half and could save councils around 255 hours a year, with a national rollout planned for 2027. Human planning officers keep final say — the AI handles the paperwork bottleneck that slows the UK's goal of 1.5 million new homes by 2029.
Loft Orbital says its YAM-9 satellite became the first spacecraft to run a vision-language model in orbit, using Google DeepMind's Gemma 3 to answer plain-English queries — classifying where wilderness meets development, or spotting infrastructure near rail hubs — without beaming raw imagery to ground analysts first. The model ran on an Nvidia Jetson edge chip paired with NAVI-Orbital software from NASA's Jet Propulsion Laboratory. Onboard triage like this could sharply cut how much data satellites need to downlink, and Loft says 50 to 100 such craft would give near-real-time eyes on Earth.
The FBI, working with Google and Lumen's Black Lotus Labs, seized the servers and domains of 'Outsider Enterprise,' a Chinese phishing-as-a-service operation that ran more than a million fraudulent links and 9,000 fake sites since 2023. Investigators tie it to 3.8 million stolen credit-card records and about $1.9 billion in losses; its operators allegedly used Google's Gemini to mass-produce the fake telecom and brand pages. The takedown — part of the bureau's Operation Riptide — is the criminal counterpart to the civil lawsuit Google filed against the same network last week.
Google filed suit against an alleged Chinese cybercrime group it calls the Outsider Enterprise, accusing it of prompting Gemini to generate HTML for thousands of fake "gift redemption" phishing pages. The operation sent more than 2.5 million scam texts to US Android users and spun up over 9,000 fraudulent websites. It is the first time Google has taken an alleged Gemini abuser to court, and the company is coordinating with the FBI and carriers including Verizon, AT&T, and T-Mobile.
The Regional Court of Munich ruled that Google is directly liable for content its AI Overviews generate, rejecting the search-engine "host" defense that has long shielded platforms. The case involved Overviews that falsely tied two Munich publishers to fraudulent companies. The court found the summaries make "independent, new, and substantive statements" rather than merely pointing elsewhere, and dismissed the idea that users should fact-check them. Legal analysts say the precedent could expose every AI-answer provider — Google, OpenAI, Anthropic, Perplexity — to similar liability across Europe.
Google released Gemini 3.5 Live Translate, an audio model that translates speech continuously instead of waiting for a speaker to finish — staying just a few seconds behind while preserving the speaker's intonation and pacing. It auto-detects more than 70 languages and supports over 2,000 language combinations in a single meeting. It is rolling out in Google Translate on Android and iOS, in Google Meet for enterprises, and through the Gemini Live API for developers.
Google released DiffusionGemma, an experimental open model that abandons the usual word-by-word approach and instead generates whole blocks of text at once, the way image models denoise a picture. It is a 26 billion-parameter mixture-of-experts (MoE) model that activates 3.8 billion parameters per step and hits more than 1,000 tokens per second on a single Nvidia H100 — up to 4-times faster than typical autoregressive models. The Apache 2.0 weights are on Hugging Face, though Google notes output quality trails standard Gemma 4.
Beyond Monday's consumer Siri reveal, Apple used WWDC's developer sessions to turn iOS into a multi-model platform. A new LanguageModel protocol in its Foundation Models framework lets any cloud provider expose a common inference interface to apps, and Xcode 27 ships dual-engine agentic coding — an on-device Neural Engine model for instant Swift suggestions plus a cloud layer that routes heavier work to Anthropic's Claude, Google's Gemini, or OpenAI. Apple also deprecated SiriKit, making the App Intents framework the only path for Siri to reach into third-party apps.
At its WWDC keynote, Apple confirmed that the next generation of its Apple Intelligence features — including a rebuilt, chatbot-style Siri — will run on a custom Google Gemini model, under a multi-year deal reportedly worth about $1 billion a year. The roughly 1.2-trillion-parameter model runs inside Apple's Private Cloud Compute, and Apple says Google cannot use the queries to train Gemini. A new Extensions system also lets users route requests to rival assistants like Claude and Grok.
Google will pay SpaceX about $920 million a month from October 2026 through June 2029 — roughly $30 billion in total — for access to around 110,000 Nvidia chips, capacity SpaceX first built for its own xAI division. Google Cloud called it short-term "bridge capacity" for its Gemini Enterprise platform while its own data centers scale up. The deal lands days before SpaceX's IPO, where Google already holds about a 5 percent stake, and echoes an earlier SpaceX compute arrangement with Anthropic worth roughly $1.25 billion a month.
Google released Gemma 4 12B, a 12-billion-parameter open-weights model under an Apache 2.0 license that runs locally on consumer laptops with 16 gigabytes of memory. Its encoder-free design feeds vision and audio directly into the language backbone — Google's first mid-sized model with native audio input — and it claims benchmark performance approaching the company's 26-billion-parameter model at under half the memory footprint. Gemma downloads have now crossed 150 million.
The UK's Competition and Markets Authority ordered Google to give publishers a toggle in Search Console to exclude their content from AI Overviews and AI Mode — and confirmed that opting out will not be used as a ranking signal in regular search. The order leans on Google's October 2025 designation as a firm with strategic market status. Google will test the tool with a subset of UK publishers before rolling it out globally; it says AI Overviews now reach 2.5 billion monthly users.
Vibe-coding startup Lovable signed a multiyear agreement to expand its Google Cloud footprint — including AI compute — fivefold, gaining wider access to both Anthropic's Claude and Google's Gemini models. The Stockholm company crossed $400 million in annualized revenue in February with just 146 employees, and says more than half of the Fortune 500 use its AI app builder. Its agents will also be sold through Google Cloud's enterprise marketplace.
Google released the hydrology framework behind its Flood Hub as an open-source Python package under an Apache 2.0 license, letting national weather and hydrology agencies train their own AI river-flow forecasts. The new version, built on long short-term memory networks, extends reliable predictions by up to six days in gauged river basins and by one day in ungauged ones. The Czech Hydrometeorological Institute has already integrated it into a standard operational forecasting platform used by agencies and NGOs.
Following Alphabet's plan to raise $80 billion in equity for AI compute — with Berkshire Hathaway taking $10 billion — Ben Thompson's Stratechery argues the move marks a pivot from a high-margin advertising business into a capital-intensive infrastructure company, akin to how Berkshire redeployed See's Candies profits into railroads. His thesis: in a compute-constrained market, the firm with the most cash to deploy compounds its lead, and Berkshire's stake validates that bet.
Google flipped Gemini Spark — the always-on personal agent it teased at I/O — live for US AI Ultra subscribers on Friday. Spark surfaces as a new tab in the Gemini app on web, Android, and iOS, and can check calendars, RSVP to events, declutter inboxes, draft Gmail replies, navigate Drive, and create Workspace docs without follow-up prompts. The launch puts Google ahead of Anthropic and OpenAI in shipping a mass-market 24/7 background agent for consumer subscribers, though the rollout is US-only for now.
TechCrunch's preview of the rebuilt Siri describes it as a standalone ChatGPT-style app — chat history, document and photo uploads, text mode — with Google's Gemini running under the hood, even as Apple keeps shipping its own on-device proprietary models for privacy claims. A separate Ars Technica scoop says Apple is distilling Google's multi-trillion-parameter Gemini down to something that can run on an iPhone. With 2.5 billion Apple devices versus ChatGPT's roughly 900 million weekly users, the WWDC launch on June 8 could reset the consumer chatbot market overnight.
OpenRouter, the multi-model API gateway founded in 2023, raised a $113 million Series B led by CapitalG — Google's growth venture fund — at a $1.3 billion post-money valuation, more than doubling from its $547 million Series A last June. The company routes requests across more than 400 models from Anthropic, Google, OpenAI, xAI, and DeepSeek, and now serves roughly 8 million users at around 100 trillion tokens per month — a 5-times jump in throughput over the past six months. The funding signals that multi-model routing has moved from optional plumbing to default architecture for production AI deployments.
Google DeepMind opened applications Thursday for the inaugural **DeepMind Accelerator** program in Asia Pacific, a three-month initiative pairing startups, research teams, and nonprofits with DeepMind engineers to apply frontier AI and "science AI" models to nature, climate, agriculture, and energy problems. The program runs out of an in-person bootcamp in Singapore, with DeepMind citing a recent KPMG-Google study that found the region is both "a global engine for economic growth" and "highly vulnerable to climate change," with green technology not scaling fast enough to keep pace with rising environmental risks.
Google announced four new Gemini-built ad formats for its AI Mode and AI Overviews surfaces: Conversational Discovery ads that respond to user questions, Highlighted Answers within AI-generated lists, Gemini-authored Shopping product explainers, and "Business Agent" chatbots that live inside ads. The formats roll out over the coming months and are exposed primarily through Performance Max, AI Max for Search, and AI Max for Shopping campaigns. The announcement formalizes ads as a revenue layer atop the agentic-search interface Google has been beta-testing for two years.
At Tuesday's annual developer keynote, Google unveiled Gemini 3.5 Flash as its new default agentic model — 76.2% on Terminal-Bench 2.1, 83.6% on MCP Atlas, and roughly four-times faster output than other frontier models at less than half the cost. Gemini Spark, a personal always-on agentic assistant, rolls to trusted testers; Antigravity 2.0 lets developers spawn parallel subagents across multi-week workflows; and the existing Gemini CLI sunsets June 18 in favor of an Antigravity CLI rewrite in Go. Gemini 3.5 Pro is teed up for next month.
Google framed its biggest Search box upgrade in over 25 years around agents rather than answers. Information agents will monitor the web around the clock for user-set triggers like apartment listings or sports updates, agentic booking expands to local services and automated business calls, and generative UI builds custom dashboards and mini-apps inside the results page on the fly. AI Mode now defaults to Gemini 3.5 Flash worldwide across 98 languages, and a new Universal Cart is designed to follow a single shopping session across the open web.
OpenAI added two provenance signals to its image models: C2PA metadata identifying AI-generated content, and SynthID, Google's invisible watermark designed to survive screenshots and resizing. A preview verification tool checks for both signals, initially only on OpenAI-generated images. The move tightens the cross-lab consensus on detectability — Google created SynthID and adopted C2PA across its product suite, and OpenAI's adoption pushes both standards closer to defaults for the rest of the industry.
Google Research published **ERA — Empirical Research Assistance — in *Nature* on May 19**, an AI system using tree-search over thousands of candidates to write and optimize scientific code across genomics, public health, satellite imagery, neuroscience, and time-series forecasting. Concrete wins: ERA-built forecasts ranked at or near the top of the CDC's leaderboards for flu, COVID-19, and RSV; a California water-runoff model beat the state's official Bulletin 120 outlook; and a retail forecasting variant met or exceeded both commercial consensus and Chicago Fed estimates. Built on Gemini.
Per a TechCrunch report citing leaked product specs, Apple will unveil at WWDC in June 2026 a rebuilt standalone Siri app whose chatbot experience runs on Google Gemini — not Anthropic Claude or OpenAI ChatGPT, despite reporting earlier in May that both labs were courting Apple as a launch partner. The app's headline feature is auto-deleting conversations with user-configurable retention windows of 30 days, one year, or indefinite, mirroring the privacy controls in Apple's Messages app. Analysts read the privacy emphasis as marketing cover for Apple's continued backend dependence on a competitor's frontier model, three years after the original Siri-on-ChatGPT integration shipped at WWDC 2024.
At Android Show 2026, Google announced a top-to-bottom AI overhaul: Gemini can now execute multi-step actions across apps (photograph an event flyer and have Gemini surface it on travel and calendar apps), Gboard ships a Gemini-powered dictation feature called Rambler that cleans up filler words in real time, and a new "Create My Widget" lets users vibe-code custom widgets in natural language on Pixel and Galaxy this summer. The biggest surprise: Googlebooks, a new line of Android-native laptops with Gemini at their core, shipping this fall with Acer, Asus, Dell, HP, and Lenovo.
Google updated its Gemini API File Search tool on May 5 with three additions for
retrieval-augmented generation (RAG) workflows: multimodal indexing of images and text together
using the Gemini Embedding 2 model, custom key-value metadata labels (such as
department: Legal) for query-time filtering, and page-level citations that link readers to the
exact source location in indexed documents. The update targets both prototypes and
production-scale RAG applications, with example use cases in creative agencies, scientific
research, and engineering. No pricing change was announced.
Microsoft Research released DELEGATE-52, a benchmark simulating extended document-editing workflows across 52 professional domains including coding, crystallography, and music notation. Across 19 large language models tested, the strongest frontier systems — Gemini 3.1 Pro, Claude 4.6 Opus, and GPT 5.4 — corrupted an average of 25 percent of document content by the end of long sessions, with errors silently accumulating rather than failing loudly. Adding agentic tool use did not improve results, and degradation worsened with larger documents, longer interactions, and distractor files. Authors: Philippe Laban, Tobias Schnabel, and Jennifer Neville.
Google DeepMind detailed measurable algorithmic improvements from AlphaEvolve — its Gemini-powered coding agent — across seven research and industry fields. Highlights include a 10x error reduction in quantum circuits, a 30% cut in DNA sequencing variant-call errors, a 20% reduction in Google Spanner write amplification, doubled transformer training speed at Klarna, and a 10.4% logistics routing gain at FM Logistic. The post turns AlphaEvolve from research demo into a concrete pattern for code-generating agents tackling specialized scientific and operational problems.
Google released an acceleration update for its open-weight Gemma 4 family using multi-token prediction — the same speculative-decoding technique that delivered DeepSeek V3.2's recent throughput gains. The update claims up to 3x faster inference on Gemma 4's 26 billion- and 31 billion-parameter variants, alongside the smaller E2B and E4B mobile models, with no quality regression. Combined with April's Gemma 4 launch, Google is positioning Gemma as the most efficient open-weight family for on-device and single-GPU deployment, narrowing the gap with Llama and Qwen.
Bloomberg's Mark Gurman reports that Apple is preparing an "Extensions" framework in iOS 27, iPadOS 27, and macOS 27 that lets users route Apple Intelligence features — Siri, Writing Tools, Image Playground — through third-party models rather than only OpenAI's ChatGPT. Google's Gemini and Anthropic's Claude are both reportedly already in testing. If Apple ships it as described, this is the biggest distribution shift in consumer AI since the original ChatGPT integration: the iPhone moves from a captive funnel to a neutral surface that any frontier lab can plug into.
A privacy researcher reports Chrome silently downloads roughly 4 gigabytes
of Gemini Nano weights to user devices via the OptGuideOnDeviceModel
directory, with no settings-panel checkbox to disable the install — only a
buried chrome://flags toggle or enterprise policy. Deleted weight files
re-download on subsequent eligible update windows. The researcher verified
the behavior through macOS filesystem logs and Chrome configuration files;
Google has not publicly responded. With Chrome's installed base near 3.5
billion users, the claim — if accurate — is the largest covert on-device
AI rollout to date and a flashpoint for the on-device-vs-cloud consent
debate.
DeepMind announced an "AI co-clinician" research initiative built on Gemini and Project Astra, structured as a triadic care model where the AI works alongside a supervising physician with a separate "Planner" module monitoring safety boundaries. Academic collaborators include Harvard Medical School and Stanford Medicine, with phased trusted-tester evaluations planned across the US, India, Australia, New Zealand, Singapore, and UAE. It is research, not yet approved for clinical diagnosis or treatment.
Cadillac, Chevrolet, Buick, and GMC vehicles from 2022 onward — roughly 4 million cars — start receiving Gemini as a free over-the-air update, replacing the older Google Assistant. The headline feature is Gemini Live's open-ended voice mode, activated by "Hey Google, let's talk," covering directions, climate, music, vehicle diagnostics, and message summaries. TechCrunch notes the announcement language doesn't limit Gemini to GM, suggesting more automakers to follow.