AI Lab
Official Google DeepMind announcements — model releases, product launches and research, each one summarized with every source covering it, by AIBriefs. RSS
Google built autofinetune, an autonomous research loop that applies the autoresearch paradigm to LLM post-training (SFT and GRPO reinforcement learning) using Tunix, Gemma, and Cloud TPUs. An agent edits run.py, runs training, keeps winning commits or reverts regressions, and logs results to results.tsv.
ToolGrad reverses the usual pipeline by generating a ground-truth tool-use chain first, then annotating its user prompt in a single LLM step. Google says this yields more complex long-horizon tool-use data at lower cost than DFS-based approaches like ToolBench and ToolACE.
Google says running searches like "run club" and "how to train for a marathon" hit all-time highs this year. AI Mode's Canvas tool builds personalized training schedules, and linking YouTube Music lets it generate custom run playlists.
Google's Gemini desktop app is available globally today for Windows 10 and 11, summoned with Alt + Space over active work. It hands multi-step tasks to Gemini Spark, pulls context from Gmail and Drive, and generates images with Nano Banana and video with Gemini Omni.
The inaugural Creative Intelligence weekend at London's Southbank Centre features four Google Arts & Culture interactive experiments, including Yinka Ilori's "Dreaming with Flamingos" built on Google DeepMind's Lyria music model. "Splash Canvas" uses Gemini and fluid dynamics, while "Learn Everything" turns photos of objects into visual metaphors.
Google Developers Blog argues end-to-end benchmarks like Terminal-Bench and DeepSWE are expensive, slow, and lack root-cause diagnostics when scores shift. It recommends behavioral evaluations that measure discrete, observable agent actions as integration tests for harness engineering.
Google is rolling out AI-powered voice assistant modes for Gmail, Docs, and Keep, called Gmail Live, Docs Live, and Keep Live, allowing hands-free management via conversation. The features, previewed at Google I/O in May, are now available on mobile platforms globally in English starting today.
Google's AI & Economy ATLAS v1.0 study found nearly half of Gemini conversations about civic tasks happen outside business hours. About 40% of users consult Gemini for form-filling queries, and 26% use it for civic participation like jury duty.
Google announced the 1.0 general availability of the Agent Development Kit (ADK) for Kotlin, reaching full feature parity with ADK 1.0 Core. It includes Android-first extensions for on-device agents via LiteRT-LM and ML Kit, and supports Kotlin Multiplatform.
Google announced a €13 billion ($15.1 billion) investment in AI infrastructure in Finland, its largest single investment in Europe. The two-year plan will support over 37,000 jobs during construction and contribute €3.6 billion annually to Finland's GDP.
Google partners with Missouri to give 1.1 million students and nearly 100,000 educators free access to Gemini for Education, AI training, and career certificates. The initiative includes Gemini Notebook, Google AI Educator Series, and professional certificates.
AlphaGenome Atlas precomputes predictions for all 9 billion possible single-letter DNA changes across the human genome, packaged as a 1PB dataset with a new AlphaGenome Variant Impact (AVI) score combining AlphaGenome and AlphaMissense. It is free for academic research via a web portal, Google's Antigravity agentic platform, and the AlphaGenome interface.
Google's 'Make AI Work for You' workshops visited five states along Route 66, training hundreds of small business owners. Data from the U.S. Chamber of Commerce shows 85% of small businesses using AI saw increased sales, and 82% grew their teams.
Google DeepMind selected 16 organizations for its Accelerator: AI for the Planet in APAC, supporting startups, nonprofits, and researchers using frontier AI for climate, agriculture, and biodiversity. The program kicked off with a bootcamp in Singapore.
Google Cloud's Gemini Enterprise DevEx program runs sprint testing of end-to-end developer workflows to find and fix friction. This sprint focused on enterprise AI governance, covering agent identity provisioning, registry, gateway binding, and policy enforcement.
WeatherNext 3 is the first global weather model generating forecasts every hour, drawing directly from raw satellite imagery. It predicts surface variables at 5km resolution and is integrated into Google Search, Maps, and Gemini.
Lyria 3.5, Google's best-sounding music generation model, is now available in the Gemini app, Gemini API, and AI Studio. It offers more expressive vocals, richer arrangements, and higher fidelity, with new templates and short or longer track options.
Google Research evaluated transfer learning from European cohorts to improve polygenic risk score prediction in underrepresented populations, finding it helps small cohorts but degrades accuracy as target sample sizes grow, especially for traits with population-specific genetic architectures.
Google adds monthly GES modules on the first Wednesday, including Gemini's Guided Learning and Deep Research features, and announces a national Badge-a-thon on September 19 for K-12 educators.
WeatherNext 3 generates hourly forecasts at up to 5-kilometer resolution, roughly five times sharper than WeatherNext 2, and is now in Search, Gemini, Maps, and Cloud. It ranks as the best global weather model on Brightband's live leaderboard.
Google for Startups AI Agents Challenge winners relied on foundational engineering patterns, not raw model power. Top submissions used bidirectional MCP for inter-agent communication and mediated database access through tools to keep context small.
Fairwind gives governments, Google Cloud customers, and security partners early access to Gemini 3.8 Flash Cyber paired with the CodeMender harness, generating verified, deployment-ready patches in minutes instead of weeks. Google says it is working with over 650 partners including CrowdStrike, Palo Alto Networks, and Snowflake.
Gemini 3.8 Flash ships at the same introductory price as 3.7 Flash: $0.75 per million input tokens and $3.75 per million output tokens. The Cyber variant targets vulnerability detection and automated patching, gated to trusted defenders via the new Fairwind Program.
Google announced a multi-year partnership with MrBeast's Beast Industries spanning Gemini and Google Health. A September 5 video will show him using Gemini to survive extreme climates, with Fitbit Air integration planned.
MAPL-EMIT, a deep learning model from Google and NASA JPL, detects 50% more methane plumes than human experts and identified over 23,000 additional plumes globally, including 24 of the 25 largest-emitting landfills. Trained on 3.6 million simulated plumes, it uses NASA's EMIT instrument data.
Agentic video understanding is now available for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. It cuts token consumption by up to 88% and analysis costs by up to 66%, while improving accuracy by up to 7%.
Google Pics, built on the Nano Banana model, rolls out over coming weeks to Google AI Pro/Ultra subscribers and most Workspace business customers. It offers object segmentation, in-image text editing, and integrates with Docs, Slides, and Drive.
Antigravity's Teamwork multi-agent framework, paired with Gemini 3.7 Flash, solved seven open problems in mathematics and theoretical computer science, including Knuth's Cycles Conjecture (verified in Lean), and scored 71% on TCSBench. It also produced a cycle-accurate RISC-V CPU simulator and optimizations merged into Eigen and ParlayHash.
TimesFM-3, a 330M-parameter time series foundation model, natively pre-trained for multivariate forecasting on over 1 trillion time points, outperforms other models across major benchmarks in a single forward pass.
Google's new Expert Intelligence feature lets users add purchased Google Play Books ebooks directly to Gemini Notebook, enabling grounded Q&A and generation of infographics, audio overviews, and quizzes. Over 100,000 books from publishers like Penguin Random House and O'Reilly Media are supported, with 15 authors creating Featured Notebooks.
Google's blog post explains how deep learning with Keras is applied to astroparticle physics, analyzing data from observatories to detect rare cosmic signals. It highlights the role of deep learning in improving instrument sensitivity and uncovering hidden patterns.
Gemini 3.5 Transcribe reports 2.6% average WER across 85+ languages. It offers two endpoints: gemini-3.5-transcribe for pre-recorded audio via the Interactions API, and gemini-3.5-transcribe-live for real-time streaming via the Live API.
Google Research's planetary prediction engine (PPE) autonomously executes the full geospatial modeling workflow from data discovery to model training, improving prediction tasks in public health, food security, environmental risk, and socioeconomics. It works directly from natural-language queries.
Gemini Omni 1.1 Flash adds scene extension that analyzes up to 10 seconds of prior context, versus one second in previous models, extending clips in 10-second increments to a 40-second cumulative max. It also generates continuous video between specified first and last keyframes.
Gemini Omni 1.1 Flash adds scene extension that analyzes up to 10 seconds of prior context, up from one second, extending clips in 10-second increments to a 40-second cumulative cap. It also generates continuous video between specified first and last keyframes, available via the Gemini API in Google AI Studio.
Omni 1.1 analyzes up to 10 seconds of prior video context, up from one second in earlier models, and extends clips in 10-second increments to a 40-second cumulative cap. Available via the Gemini API in Google AI Studio.
Omni 1.1 analyzes up to 10 seconds of prior context for scene extension, up from one second in previous models, and extends video in 10-second increments to a 40-second cumulative cap. Available via the Gemini API in Google AI Studio.
Google's AI Mode in Search now lets users track flight prices, see costs in points or miles, and book hotels via conversation. Flight price tracking is available in 180+ countries; hotel booking is rolling out in the U.S. in English with partners like Booking.com, Expedia, and Hilton.
BenchMIRT analyzes benchmarks question by question, finding BBQ's questions distinguish models more by reasoning ability than safety. It extends Item Response Theory to separate signals within a benchmark.
GlucoFM is a 0.72M-parameter self-supervised foundation model that separates slow glucose trends from short-term deviations. It sets new performance standards across seven clinical tasks, including diabetes risk and insulin resistance, evaluated on four cohorts.
Google Cloud integrated native TPU support into vLLM, enabling elastic scaling of embedding pipelines via GKE. The engineering team implemented optimizations for 15K+ token contexts, supporting models like Qwen3-Embedding-8B.
Google for Startups Accelerator chose 28 AI-driven energy startups across North America and Europe, including 12 in the North American cohort. Startups focus on grid modernization, energy efficiency, and demand flexibility.
In a Google DeepMind video, VP Research Zoubin Ghahramani discusses teaching AI systems self-doubt and probability for safer decision-making. He critiques current models for lacking rigid mathematics over probabilities and cause-effect, calling the result "a sort of mishmash."
AgentHands, an LLM-powered XR prototype published at CHI 2026, augments conversational agents with synchronized, expressive hand gestures for spatially grounded guidance. It builds on Project Astra and Gemini 3.1 Flash Live, moving beyond 2D bounding-box overlays to embodied dialogue in Android XR.
Gemini app for macOS now lets users long-press the Fn key to speak naturally into any window, with intelligent dictation that removes "ums" and "ahs" and drops formatted text at the cursor. Opt-in Gemini reasoning enables context-aware tasks like summarizing files, rewriting text, and generating images. Rolling out globally in English.
Google is partnering with the State of Delaware to provide free AI and career training, including Google AI courses and Career Certificates, with no college degree or experience required. Over 70% of Career Certificate graduates report a positive career outcome within six months.
ADK now supports native live evaluation, letting developers test graph-based voice agents against LLM-driven simulated users that speak audio turns. The example uses gemini-live-2.5-flash-native-audio in a three-stage workflow.
Google announced the winners of the Gemma 4 Good Challenge, a Kaggle competition for impactful AI solutions. First place went to GEM-4, a robotic assistant for elderly and disabled individuals, built with a Gemma 4 31B model and a fine-tuned Gemma 4 E2B controller.
Google Research's ME-POIs framework blends text descriptions with anonymized mobility patterns (arrival times, stay durations) to improve language models' predictions of place attributes like opening hours, price levels, and busyness. It uses a self-supervised approach on public benchmark datasets.
Google Research's multi-agent Biomarker Discovery Framework prioritizes biomarker candidates from wearable sensor data via iterative hypothesis generation, statistical analysis, and literature-grounded reasoning. Across three cohorts (N=9,279), it recovered known clinical signals and improved downstream prediction.
Paige Bailey, engineering lead at Google DeepMind, explains full-stack AI as five layers: infrastructure, security, research, models & tooling, and products. A new video walks through how the layers work together to make Google's AI products faster, more secure, and more helpful for users and developers.
Google DeepMind is partnering with game studios like Fenris Creations and Hello Games to prototype new AI-driven gameplay, building on 15 years of research from Atari to EVE Online. The work includes a major research partnership with the EVE Universe unveiled earlier this year.
Google announced that its Gemma family of open models has surpassed one billion downloads, with over 100,000 community-published variants. The milestone is celebrated with an event in San Francisco on August 20.
Google Ads introduces A/B testing for AI Max across multiple Search campaigns, rolling out in September, plus new Performance Planner capabilities to preview and apply budget/bidding changes.
Google offers U.S. college students one year of Google AI Pro free ($19.99/mo value) and international students Google AI Plus, plus a new student hub in Gemini with study notebooks, flashcards, and practice quizzes. Search adds interactive visuals and practice quizzes for tests like SAT and ACT.
Waymo has integrated Gemini as an in-car AI assistant in its purpose-built Ojai vehicles, enabling voice control of cabin features and local information. Gemini operates independently of the Waymo Driver and stays inactive until engaged.
Google's Small Brief project gave creatives Susan Credle, Jayanta Jenkins, and Tiffany Rolfe unlimited access to Flow to craft campaigns for nonprofits Stonewood Farm, South Ferry, and ARCHANGELS. The films were made with Google's AI creative studio to champion local businesses.