AI Lab
Official Google DeepMind announcements — model releases, product launches and research, each one summarized with every source covering it, by AIBriefs. RSS
Google announced a new research hub in Singapore plus initiatives aimed at supporting clinicians, advancing research, and helping Singaporeans live healthier lives. The announcement came at the Google for Health Summit in Singapore.
GPU inference via ML Drift runs about 1.4x faster than the legacy GPU delegate, exposed through the new CompiledModel API. Apps switching from bundled LiteRT shrink by roughly 8 MB, with over 100K Android apps still bundling their own copy.
Ads in AI Mode roll out to Australia, Canada, India, Kenya, Malaysia, New Zealand, Nigeria, Pakistan, the Philippines and Singapore, with Brazil, Mexico and the UK planned for Q4 2026. Google says AI Mode queries run three times longer than traditional searches and sponsored click-through rates already match organic product listings.
Agents CLI v1.8.0 ships eval generate, grade, compare, analyze and optimize commands to score agent traces and auto-tune prompts. The walkthrough's LLM-as-judge metric grades responses 1-5 on accuracy, relevance and clarity using gemini-3.8-flash.
Google launched a library of interactive, layer-by-layer visual explorations in the Gemini student hub, covering topics from microscopic systems to deep space. Available in English, Indonesian, Hindi, Brazilian Portuguese, French and simplified Chinese; other languages get English only.
Google DeepMind says Carbon is state-of-the-art in cybersecurity and long-horizon capabilities, a step up across most benchmarks with strong autonomous coding and cyber defense. It is being released first to a limited set of users.
XREAL AURA is the first wired XR glasses powered by Android XR, positioned for watching, working, and playing anywhere. Google says Gemini is available alongside the glasses.
Chrome's enhanced autofill suggests order numbers, tracking info, and flight bookings from Gmail, powered by Gemini with Personal Intelligence. Rolling out in the U.S. on desktop, Android, and iOS for Gemini in Chrome users with Personal Intelligence enabled; Google Photos support is planned.
Google's developer blog walks through a working prototype, HappyDelivery, built on LangGraph, Gemini, FastAPI and Google Cloud across three decoupled layers: Ingress, Memory and Cognition. It includes deterministic financial tools and an automated verification suite, plus webhook handling for Meta's X-Hub-Signature-256 and Google's X-Goog-Signature.
Google Research proposes a "sleep" phase for continual learning, alternating active interaction with offline memory consolidation so LLMs keep new facts without catastrophic forgetting. It retains long-context knowledge up to 10M tokens and improves complex reasoning.
A2UI Express compiles compact Python-style syntax into standard A2UI JSON via Agent SDKs, using up to 3x fewer tokens than Direct JSON. Google says decision models like TypeSafe's Jev can cut layout-decision latency by up to 80x.
Genie 4 generates real-time interactive worlds with on-the-fly ambient sound, music, and dialogue, steerable mid-simulation via natural language. It builds on Gemini models — Gemini Omni anchors appearance and behavior — and is a research preview for trusted testers.
Google's newest AI Quests lesson, "Polar Peaks," teaches AI literacy through a neuroscience framing about mapping the human brain. It is the latest addition to the education program, announced on Google's blog by Ronit Levavi Morad.
Google Research detailed at IDWeek 2026 how open foundation models like MedGemma can be tuned for chest X-ray TB screening at the point of care. TB kills an estimated 1.23 million people annually, and scarce radiologists in high-burden settings leave millions of active cases undetected.
Logan Kilpatrick said Google processed ~25 trillion tokens in the first 72 hours after Gemini 4 Argon launched. Builders used it inside Antigravity with Blender MCP to generate 3D scenes, including a Gothic cathedral and a multi-tiered pagoda sanctuary.
Google Cloud's reference architecture, codenamed Project Chrysalis, cut time-to-first-mitigation from 15–45 minutes to 1.8 seconds in simulated Black Friday cache-stampede outages. It also dropped HTTP 5xx error rates from 72.7–80.0% to 0.0% and P99 latency from 519.8 ms.
Agents CLI 1.5.0 introduces an Extension System that overrides or adds commands without forking the tool; 1.8.0 ships a first-party LangChain extension replacing framework-coupled commands. Teams keep the unified deploy, eval, and observability workflow without adopting ADK.
A prospective multi-center study found non-sonographer operators using low-cost probes plus AI blind-sweep interpretation matched trained sonographers on gestational age and fetal presentation. WHO estimates 250,000+ preventable pregnancy-related deaths yearly, and two-thirds of people lack access to diagnostic imaging.
Google's e-Conomy SEA report projects the region's digital economy will reach $368 billion in 2026 and $1 trillion by 2035, with widespread AI adoption cited as the driver.
Personal Intelligence, which pulls context from Gmail and Calendar, is rolling out to Android Auto today and to Google built-in vehicles in the next few months. It stays optional; a driver can say "take me to my doctor's appointment" and Gemini pulls the event from a connected calendar.
Announced at PyTorch Conference North America 2026, TorchTPU makes TPUs a native torch.device("tpu") in PyTorch, co-developed with Meta's PyTorch team. AWS's TorchNeuron, in public beta, does the same for Trainium. Both are on GitHub.
Arm built open-source adapters for TOSA, VGF, and PTE formats, letting engineers inspect Arm-specific IRs and hardware profiling traces in Model Explorer's hierarchical graph view. Arm used it to catch unintended TRANSPOSE operators introduced by an ExecuTorch Arm backend change in an exported NSS model.
Google Research's AMIE is an AI system patients chat with before appointments, securely collecting information so doctors can spend more time on collaborative care. The post is framed as a Q&A on findings from a clinical study of the system.
Google announced new Gemini tools aimed at small business owners worldwide, including a dedicated business landing page for the Gemini app.
Google Research added a Matrix Execution Unit to its 32-bit open-source Coral NPU RISC-V architecture, delivering up to 1 TOPS while targeting 10 mW average power for edge inference. The release also ships the v1.0 open-source IREE/MLIR toolchain, and Synaptics is unveiling chips for wearables and industrial devices built on Coral.
The C++ authoring layer lets developers build computation graphs op by op and run the same graph on CPU, GPU (via open-source ML Drift) and NPU, or directly on CPU through XNNPACK without a model file or runtime. Google recommends the standard convert-and-run path for most projects, positioning this for custom architectures and new lowering logic.
Google's Gemini Notebook now gives users more creative control when generating Slide Decks and Video Overviews, per a post on the Google blog.
Switching from a blocking tool to a streaming tool on gemini-3.8-live cut time to the agent's first words from 14.4 seconds to 1.2 seconds in a benchmark sales-report workflow. The guide covers non-blocking FunctionTool definitions and response_scheduling in the open-source Agent Development Kit.
Google Research's VERID framework links AI agent reasoning to liquid-handling robots, closing the loop between in-silico design and physical experiments. The team reports discovering single-digit nanomolar protein binders for cancer targets.
Google rolled out three study features in Gemini Notebook: 60-second Short Video Overviews, Interactive Learning Overviews that weave summaries with slides, quizzes and flashcards, and new quiz formats including short-answer, multi-select and fill-in-the-blank.
Google DeepMind's four Gemini 4 Argon agents defeated the Ender Dragon in a sandboxed survival world in 27 hours, writing 2,783 programs with no human help. Each controlled a player via the open-source Mineflayer library and coordinated through in-game chat and shared memory, with no roles assigned.
Long-press the home button or navigation bar to activate Circle to Search, then circle a suspicious message to get an AI Overview with guidance and suggested next steps. Google advises deleting the message and blocking the number if red flags appear.
Google's playbook covers building a continuous eval pipeline for multi-turn agents using Gemini Enterprise Agent Platform, Google ADK, and OpenInference standards. It accompanies an AI Agent Clinic episode that builds an eval pipeline for a LangGraph agent in 60 minutes.
The repo splits ADK resources into core/ reference patterns, plugins/ vertical skills conforming to the Agent Plugins v1.0.0 spec, and contrib/ deployable services across five languages. Every recipe ships a manifest.yaml, AGENTS.md, and for contrib/ and plugins/ a Dockerfile plus runnability test, with CI enforcing standards on each PR.
DeepMind used its agentic science workbench, Google Antigravity, to identify an antimicrobial peptide with predicted activity against respiratory tract infection pathogens; the lead is now in wet-lab validation. The platform scans millions of uncharacterized UniProt proteins via ProtNLM, which predicts biological function from amino acid sequences.
Google says more than 400 publications now participate in its News AI pilot, up from about 100 reported in September, with new partners including Axios, CNN, The Financial Times and Forbes. The program pays publishers for content used in AI features like article overviews and audio briefings.
Google is updating AI Mode and AI Overviews to add clearer links and context about sources. The change targets how Search surfaces the web within AI-generated answers.
SynthID has watermarked more than 180 billion images and videos and 240,000 years of audio, embedding invisible watermarks at creation without altering quality. Google recently announced SynthID Bio for synthetic biology, and the watermark is also used by OpenAI, Nvidia, Kakao, and soon Apple.
Universal Cart is rolling out in the U.S. across Google Search and the Gemini app, aimed at making holiday shopping easier.
Apigee hybrid 1.17.0, released September 14, 2026, brings native Model Context Protocol support so the MCP data plane runs inside a customer's own Kubernetes cluster. APIs described with OpenAPI 3.0.x specs become MCP tools via a Discovery Proxy, with no MCP server code to write or operate.
Deep Review, an agentic AI system, launched publicly after deployments across 16,000+ submissions at STOC, ICML, NeurIPS, and ICLR. It expands from academic peer review to verifying Google Docs, presentations, legal and finance documents.
Google DeepMind and Google Cloud are giving up to 200 European scientific researchers direct access to frontier models plus $3 million in compute credits. Tools include AlphaFold 3, WeatherNext, AlphaGenome, AlphaEvolve, and AlphaEarth Foundations, delivered via the EU's RAISE initiative.
Google DeepMind and GovTech Singapore fine-tuned a lightweight Gemma model on local scam conversations to flag social-engineering fraud on-device, with zero cloud data egress. Scams account for over half of Singapore's reported crimes and more than S$4 billion in losses since 2019, with social engineering driving up to 80% of losses.
An 11-hour unattended run on google/adk-python (~205,000 lines, ~800 files, ~17,000 tests) cut in-memory session prep ~40% in long sessions and sped cold imports 21%. Four of six proposed optimizations were rejected in human review.
Argon scores 87% on the Embodied Reasoning (ERQA) benchmark, which DeepMind says beats every other frontier model. In an internal test it watched a cornhole video, built its own MuJoCo simulation, and taught Apptronik's Apollo humanoid to sink 7 throws in a row.
Gemini 4 Argon is rolling out globally across the Gemini app, API, AI Studio, Android Studio, and Google Antigravity, targeting coding, enterprise workflows, and cybersecurity defense. Scale Labs' Humanity's Sixth Sense benchmark puts it top at 55.4%, edging GPT-6-Astra at 53.6%.
Google says its models and tools are helping scientists pinpoint rare genetic variants and speed up drug discovery, with the stated goal of making disease more detectable, treatable, and preventable.
In a BIDMC primary care study published in The Lancet, 98 patients consulted Google's AMIE diagnostic chatbot before urgent care visits; no conversation was interrupted for safety reasons. Clinicians said AMIE summaries helped them prepare in 75% of cases and changed their care approach in over half.
ML Drift is a cross-platform on-device GPU compute engine released under Apache 2.0, abstracting OpenGL ES, OpenCL, Metal, and WebGPU. It serves as the core GPU acceleration engine inside LiteRT and is also available as a standalone library.
AQuA runs beside a production agent in Google Cloud, sweeping failing trajectories to surface recurring loss patterns that pass health checks and return HTTP 200. Google frames the problem as the last 20% of agent quality, where model, harness, or tool changes silently shift task success.
Announced at Gemini at Work, the Gemini agent takes goals instead of prompts, runs in the cloud for hours or days, and works across Gmail, Docs, Sheets, Drive, Chat and Calendar plus Microsoft 365 and Slack. Managers can create "coworker agents" with their own Workspace email, calendar and directory listing; no pricing or GA date given.
A three-month randomized trial with practicing patent attorneys found AI raised work quality across the board, but senior lawyers using AI for 90 days showed stronger judgment while juniors showed no average skill gain, their scores splitting into more strong and more low scores.
The API serves fresh, Markdown-formatted docs for Google Cloud, Firebase, and Android via semantic and keyword search, chunking, and grounded Q&A. A gcloud CLI surface ships pre-installed in Cloud Shell across Linux, macOS, and Windows.
Google's SynthID Detector is now available globally in English, letting anyone check images, video, and audio for AI watermarks from Google and partners OpenAI, NVIDIA, Kakao, and soon Apple. Google says it has watermarked over 180 billion images and videos and 240,000 years of audio since 2023.
Playground is a Google Labs experiment that turns text prompts into playable, shareable browser games with no coding required. It is live in the US, with game creation tiered by Google AI subscription while anyone can play gallery games.
EmbeddingGemma 2 unifies text, code, images, video, and audio into a shared 768-dimensional space under an Apache 2.0 license, built on the Gemma 4 architecture. It scales from 270M parameters for text-only to 740M for all modalities, and Matryoshka truncation cuts vectors to 128 dimensions for up to 6x storage reduction.
Google Research detailed five partner-driven case studies using its Population Dynamics Foundation Model (PDFM), which compresses privacy-preserving search trends, mobility, and environmental data into location embeddings without task-specific fine-tuning. Partners tested a Geospatial Reasoning agent and a planetary prediction engine during the Ebola outbreak in the Democratic Republic of Congo.
Founders from more than 100 startups chosen from over 2,000 applications will meet at Google's Mountain View HQ in November for a two-day summit. Participants represent 17 countries and get up to $350,000 in Google Cloud credits.
Google Research published a workshop report from the Contextual Agent Privacy and Security (CAPS) workshop, held in late 2025 in New York City, with more than 50 academic and industry leaders. It outlines open research directions at system, model, user, and ecosystem levels, drawing on Contextual Integrity theory.