50 articles found
Google DeepMind's SIMA 2 announcement positions game-playing AI as a triumph of research continuity. But when foundation models can master virtual worlds from Atari to EVE Online, we should ask what happens when those skills transfer to environments where failure isn't just losing a match.
Google DeepMind announces SIMA 2, an AI agent capable of playing, reasoning, and learning in virtual 3D environments, building on 15 years of AI research in gaming contexts from Atari to complex MMOs like EVE Online.
Google DeepMind has introduced SL2T, a breakthrough sign-language-to-text AI model that translates sign language into written text, now integrated into Gboard and Live Transcribe on Pixel 11 devices starting with American Sign Language. The model addresses the unique challenges of sign language processing through computer vision and machine translation, enabling Deaf and hard of hearing users to sign instead of type for the first time in consumer products.
The Robot Report Podcast Episode 256 features Josh Cloer, general manager at Nomagic, discussing how physical AI systems are being deployed to improve intelligence and adaptability in warehouse order-fulfillment operations. The episode covers Nomagic's expansion into North America and includes news about HII's $900M agreement with Path Robotics, Google DeepMind's Gemini Robotics 2 for full-body robot control, and a new industrialized construction robotics lab at University of Florida.
Reimagine Robotics, founded by former Google DeepMind leaders, emerged from stealth with technology enabling robots to learn on-the-job through human instruction rather than requiring specialist programmers. The startup has already deployed robots in advanced manufacturing and electronics disassembly facilities, reducing prototype testing time from one day to 10 minutes, and is now seeking Series A funding to scale operations.
Google is committing $40 million in AI tokens and cloud credits to support the White House's Genesis Mission, a national initiative to accelerate American scientific discovery using AI. The commitment provides DOE National Laboratories access to Google DeepMind's frontier AI tools including AlphaFold 3, AlphaEvolve, AlphaGenome, and WeatherNext, along with Gemini for Government seats to thousands of researchers across the DOE ecosystem.
Google DeepMind unveiled Gemini Robotics 2, an advanced vision-language-action model enabling full-body control of humanoid robots with improved dexterity and multi-robot collaboration capabilities. The system can now control entire robots like Apptronik's Apollo 2 to perform complex tasks requiring whole-body coordination, such as walking, manipulating objects with five-fingered hands, and executing multi-step instructions lasting several minutes.
Google DeepMind and Isomorphic Labs are sharing our joint approach to bioresilience and AI models.
Apptronik unveiled Apollo 2, an updated humanoid robot available in bipedal and wheeled-base configurations, alongside its expanded Robot Park facility in Austin for data collection and training. The robot is designed to work with Google DeepMind's Gemini Robotics foundation models, creating a continuous learning loop where Apollo 2 collects real-world operational data to improve humanoid robot intelligence for deployment in actual work environments.
Google DeepMind announced Nano Banana 2 Lite, a faster and more cost-efficient image generation model optimized for high-throughput applications with 4-second latency and $0.034 per 1K image pricing. The company also released Gemini Omni Flash, a high-quality model for video generation and conversational editing now available to developers through Google AI Studio, Gemini API, and Gemini Enterprise Agent Platform.
UK government partners with Google DeepMind to build a new AI-powered prototype aimed at faster housing decisions.
Google DeepMind just committed $10 million to study multi-agent AI safety—a problem most people don't know exists yet. As companies race to deploy autonomous AI agents that can act independently online, we're building a system where millions of AIs will soon interact with each other, and nobody has seriously tested what happens next.
While American politicians debate establishing a robotics commission, Google DeepMind has quietly launched an accelerator that could determine which European robotics startups survive the next decade. The move reveals a uncomfortable truth about innovation power in 2026.
Google DeepMind is funding a $10 million research initiative to study the safety risks of multi-agent AI systems, concerned about potential dangers when millions of autonomous AI agents interact online. The funding, announced alongside Schmidt Sciences, the UK government's ARIA, and other partners, aims to establish a field of research around multi-agent safety, focusing on risks like coordinated scams, prompt injections, and cyberattacks that could emerge from agent deployments at scale.
Google DeepMind released DiffusionGemma, an experimental open-source language model that generates multiple tokens in parallel instead of sequentially, achieving up to 4x faster text generation speeds. NVIDIA has optimized DiffusionGemma to run efficiently on its GPU hardware including RTX, DGX Spark, and DGX Station systems, enabling low-latency local AI inference for developers and researchers.
Google DeepMind, Schmidt Sciences, and partner organizations are announcing a $10M research funding call to advance multi-agent AI safety research globally. The initiative aims to develop frameworks for understanding and mitigating risks that emerge when millions of independently-built AI agents interact across digital environments, addressing emergent collective behaviors that current safety evaluations cannot predict or monitor.
Google DeepMind introduced Gemma 4 12B, a multimodal AI model designed to run efficiently on consumer laptops with 16GB of RAM. The model features a novel unified, encoder-free architecture that processes vision and audio inputs directly into the language model backbone, achieving performance comparable to larger models while reducing memory footprint and latency.
MIT Technology Review's newsletter covers multiple AI developments: Anthropic demonstrated AI-driven coding with Claude, where developers increasingly ship AI-generated code without review; Google DeepMind announced Gemini for Science, shifting toward agentic AI systems for scientific research; and leading AI researchers are developing 'world models' to help AI systems understand physical environments, with backing from Google DeepMind, Fei-Fei Li's World Labs, and Meta's Yann LeCun.
Google DeepMind has launched a three-month accelerator program supporting 15 early-stage robotics startups across Europe, providing technical mentorship, access to AI models, and guidance to help integrate AI into their products. The cohort includes companies working on applications ranging from industrial automation and manufacturing to healthcare and climate solutions, with a focus on embodied AI and physical world applications.
Google I/O highlighted a strategic shift in AI for science from specialized tools like WeatherNext and AlphaFold toward autonomous, agentic AI systems that could conduct research independently. The article examines the tension between these two approaches, noting that while specialized scientific AI tools remain popular with researchers, Google appears to be reallocating resources toward developing general-purpose agentic AI systems capable of scientific discovery with minimal human guidance.
Google DeepMind is launching an accelerator program in Asia-Pacific to help startups, research teams, and nonprofits use frontier AI to tackle environmental challenges including climate, nature, agriculture, and energy. The three-month program offers mentorship, support, and access to Google AI models, with an in-person bootcamp launching in Singapore.
Google is preparing to announce new AI developments at its I/O developer conference, with particular focus on addressing weaknesses in AI coding capabilities where competitors like Anthropic's Claude and OpenAI's models have gained significant advantages. The company is also expected to showcase advances in AI for science and health applications, areas where Google DeepMind has maintained leadership, including potential new scientific tools and the public release of its AI-powered Health Coach.
Google DeepMind has introduced Gemini Omni, a new multimodal AI model capable of generating high-quality videos from combined inputs of images, audio, video, and text. The Gemini Omni Flash model, now rolling out to the Gemini app, Google Flow, and YouTube Shorts, enables users to edit videos through natural language conversation while maintaining character consistency, physics accuracy, and contextual continuity across multiple editing turns.
Google DeepMind and Singapore partner to apply frontier AI to address complex challenges across health, education, and sustainability and more.
The Download newsletter covers several AI and robotics developments: researchers are using AI and robotic systems to automate parts of the IVF process to make fertility treatment more effective and accessible; China's humanoid robots are driving an export boom with gig workers training robots at home; and Google DeepMind is using the MMORPG Eve Online to test advanced AI models.
Google DeepMind has introduced Gemini 3.5, a new family of AI models designed for agentic workflows and complex tasks. The first release, Gemini 3.5 Flash, combines frontier-level intelligence with fast performance for coding and agent-based applications, and is available today across Google's platforms including the Gemini app, Google Search, and APIs for developers and enterprises.
Google DeepMind and Korea partner to accelerate scientific breakthroughs using frontier AI models
World models—AI systems that represent and simulate the physical world—are emerging as a key area of focus for major AI labs and researchers, with Google DeepMind, Stanford's World Labs, and OpenAI all investing in the technology. Proponents argue world models are essential for overcoming LLM limitations and advancing robotics applications, from delivery robots to healthcare assistants, with early efforts including 3D environment generation and integration into intelligent agents.
The article examines the emerging field of multi-agent AI systems, where multiple AI agents coordinate to complete complex tasks beyond individual agent capabilities. It discusses applications ranging from coding (Anthropic's Claude Code) to general productivity tools (Claude Cowork, OpenAI's Codex, Perplexity's Computer) and scientific research (Google DeepMind's Co-Scientist), while highlighting both the potential to transform knowledge work and the risks of deploying unpredictable LLMs in critical infrastructure.
Google DeepMind is partnering with Accenture, Bain & Company, BCG, Deloitte, and McKinsey to accelerate AI adoption across enterprises, providing early access to frontier models like Gemini and technical expertise to help organizations implement AI-driven transformation at scale across finance, manufacturing, retail, and other sectors.
AI companies including Google DeepMind, OpenAI, and Anthropic are developing AI systems to act as autonomous researchers and co-scientists that can generate hypotheses, design experiments, and conduct scientific research with minimal human oversight. Recent breakthroughs include OpenAI connecting GPT-5 to automated biological labs to iteratively propose and test experiments, achieving a 40% cost reduction in protein synthesis, while researchers at Stanford created a "virtual lab" of AI agents that designed new antibody fragments against SARS-CoV-2. However, research suggests that widespread AI adoption in science may inadvertently reduce scientific diversity by concentrating research on data-rich, established topics while leaving fewer scientists to explore problems less amenable to AI analysis.
Antioch, a startup founded in May 2025 by former Tesla Autopilot and Google DeepMind engineers, raised $8.5 million to develop cloud-based simulation tools for autonomous robot and system validation. The platform aims to eliminate expensive physical testing by leveraging simulation engines like NVIDIA Isaac Sim and generative AI techniques, allowing robotics teams to conduct testing and validation at software speed rather than hardware speed.
Boston Dynamics announced a partnership with Google Cloud and Google DeepMind to integrate Gemini and Gemini Robotics ER 1.6 into its Orbit AIVI-Learning platform for Spot and Orbit robots. The integration enhances visual inspection and reasoning capabilities for industrial environments, enabling better asset monitoring, safety checks, and complex task planning with improved accuracy and real-time learning. DeepMind also released Gemini Robotics-ER 1.6, a reasoning-first model designed specifically for robotic applications including visual understanding, task planning, and tool integration.
Google DeepMind introduces Gemma 4, a family of open-source language models ranging from 2B to 31B parameters, built for advanced reasoning and agentic workflows. The models achieve state-of-the-art performance for their sizes, with the 31B variant ranking #3 on Arena AI's leaderboard, and are optimized to run efficiently on devices from smartphones to workstations while supporting multimodal inputs, long context windows (up to 256K tokens), and 140+ languages.
Google researchers are reimagining the computer mouse pointer for the AI era by developing AI-powered capabilities that allow the pointer to understand both what it's pointing at and its context to the user. The approach introduces four interaction principles—maintaining workflow across apps, capturing visual context automatically, enabling natural shorthand commands, and transforming pixels into actionable entities—to make AI collaboration more intuitive and seamless without requiring users to enter separate AI tools or write detailed prompts.
Agile Robots' partnership with Google DeepMind marks the latest in an accelerating trend: AI giants racing to embed their foundation models into industrial automation. This isn't just about smart robots—it's about controlling the physical layer of the economy.
Researchers have released new findings and a validated toolkit for measuring AI's potential for harmful manipulation—specifically its ability to alter human thought and behavior deceptively. The study involved over 10,000 participants across multiple countries and found that AI's manipulative effectiveness varies significantly by domain, with health topics showing the least susceptibility to harmful influence.
Agile Robots, a Munich-based industrial robotics company, has announced a strategic partnership with Google DeepMind to integrate Gemini Robotics foundation models into its robotic solutions for manufacturing, automotive, data centers, and logistics. The partnership will involve deploying and fine-tuning AI-powered robots in industrial settings while collecting data to improve the underlying Gemini models, positioning Agile Robots among the latest hardware companies collaborating with DeepMind on physical AI applications.
Google DeepMind released a cognitive taxonomy framework for measuring progress toward Artificial General Intelligence (AGI) based on cognitive science principles, identifying 10 key cognitive abilities needed for general intelligence in AI systems. The company is launching a Kaggle hackathon with a $200,000 prize pool to invite researchers to design evaluations for five cognitive abilities where measurement gaps are largest: learning, metacognition, attention, executive functions, and social cognition.
Google is rolling out Gemini-powered content creation features across its Workspace apps (Docs, Sheets, Slides, and Drive), enabling users to generate document drafts, build spreadsheets, create presentations, and query files using natural language prompts. The features began rolling out in beta on March 10 for Google AI Ultra and Pro subscribers, with some capabilities leveraging Google DeepMind and Google Research technology.
German robotics startup Neura Robotics has partnered with Qualcomm to develop humanoid and general-purpose robots using Qualcomm's Dragonwing Robotics IQ10 processors and Neura's Neuraverse simulation platform. The collaboration represents a broader trend of robotics startups partnering with major tech companies to accelerate physical AI development, similar to Boston Dynamics' partnership with Google DeepMind.
Google DeepMind has launched Nano Banana 2, a new AI image generation model that combines advanced capabilities from Nano Banana Pro with the speed of Gemini Flash. The model features improved world knowledge, precision text rendering, subject consistency, and is rolling out across Google products including Gemini, Search, and Ads.
Google DeepMind researchers are calling for rigorous evaluation of large language models' moral behavior, publishing work in Nature that challenges whether chatbots demonstrate genuine moral reasoning or merely mimic expected responses. The research highlights concerning vulnerabilities in LLMs, including tendency to flip positions when questioned, change answers based on formatting, and alter responses to subtle presentation changes, demonstrating that apparent moral competence may be 'virtue signaling' rather than genuine understanding.
Google DeepMind brings National Partnerships for AI initiative to India, scaling AI for science and education
Google DeepMind has made its Genie 3 world model publicly available through Project Genie, allowing users to generate and interact with 3D environments using AI. The tool offers three interaction modes—World Sketching, exploration, and remixing—and generates dynamic, interactive worlds that respond to user movement, though it is not a full game engine and outputs are limited to 60-second sequences at 720p/24fps.
Google DeepMind unveiled AlphaGenome, an AI tool that predicts how DNA mutations affect gene expression by analyzing up to one million DNA letters simultaneously across 11 types of genomic mutations. Since its launch, approximately 3,000 scientists from 160 countries have used the model to study diseases including cancer and neurodegenerative disorders, with the tool now available for free non-commercial use through an online portal.
Google DeepMind has introduced agentic vision capabilities to Gemini 3 Flash, combining visual reasoning with Python code execution to enhance image analysis and enable active investigations. This advancement allows the AI model to not only analyze images but also take actions based on visual understanding.
Google DeepMind has launched Project Genie, making its Genie 3 world model available to the public through Google's AI Ultra subscription ($250/month). The system generates interactive 3D environments from text descriptions and allows users to explore them in real-time, offering three interaction modes including world sketching, exploration, and remixing of existing worlds.
Leaders of major AI labs—OpenAI, Google DeepMind, and Anthropic—engaged in public criticism of each other at the World Economic Forum in Davos, with disagreements over monetization strategies, regulatory approaches, and corporate competition. The dispute highlights tensions between the three frontier AI companies as they compete for enterprise customers and influence over AI policy and safety narratives.
