absorb.md

Google DeepMind

Chronological feed of everything captured from Google DeepMind.

blog / GoogleDeepMind / Dec 18

Google DeepMind Accelerates US DOE's Genesis Mission with Frontier AI Access for National Labs

Google DeepMind partners with the US Department of Energy to support the Genesis Mission by granting accelerated access to its AI tools, starting with the Gemini-powered AI co-scientist on Google Cloud for all 17 National Laboratories. This multi-agent system synthesizes information to generate hypotheses and has validated drug repurposing for liver fibrosis and predicted antimicrobial resistance mechanisms matching unpublished experiments. Access will expand in early 2026 to AlphaEvolve for algorithm design, AlphaGenome for non-coding DNA analysis, and WeatherNext for forecasting, aiming to boost discoveries in energy, materials, and biomedicine.

ai-in-sciencegoogle-deepmindgenesis-missionnational-labs-partnershipai-co-scientistalphaevolvealphagenome
“Google DeepMind provides immediate accelerated access to AI co-scientist for scientists at all 17 DOE National Laboratories via Google Cloud.”
blog / GoogleDeepMind / Dec 11

Google DeepMind Expands Partnership with UK AI Safety Institute for Foundational AI Safety Research

Google DeepMind has expanded its partnership with the UK AI Safety Institute (AISI) through a new Memorandum of Understanding, moving beyond model testing to foundational safety research. This collaboration aims to accelerate safe AI development by focusing on critical areas such as monitoring AI reasoning processes, understanding socioaffective impacts, and evaluating economic implications of AI. The partnership involves sharing proprietary models and data, joint research, and collaborative problem-solving to mitigate AI risks and ensure beneficial AI progress.

ai-safetyuk-ai-safety-institutegoogle-deepmindai-policyai-ethicsai-monitoringeconomic-impact-of-ai
“Google DeepMind has expanded its partnership with the UK AI Safety Institute (AISI) to foundational safety research.”
blog / GoogleDeepMind / Dec 10

Google DeepMind and UK Government Partner for AI Advancement in Key Sectors

Google DeepMind is strengthening its collaboration with the UK government to leverage AI for national prosperity and security. This partnership focuses on accelerating AI access in science, education, modernizing public services, and enhancing national security. The initiative aims to position the UK as a leader in AI innovation application, setting a precedent for international collaborations.

ai-advancementpublic-private-partnershipuk-governmentscientific-innovationeducation-technologypublic-services-modernizationnational-security-ai
“Google DeepMind will provide priority access to its "AI for Science" models to UK scientists.”
youtube / GoogleDeepMind / Oct 28

Nano Banano: A New Era of AI-Powered Creative Tools

Nano Banano is a multimodal image generation and editing model developed by Google DeepMind. It combines the visual quality of Imagine models with the conversational and multimodal capabilities of Gemini 2.0 Flash. The model empowers creators by automating tedious tasks, allowing them to focus on creative aspects. It signifies a shift towards more personalized and interactive AI tools in artistic and professional fields, emphasizing user intent and control.

ai-image-generationmultimodal-aicreative-toolshuman-computer-interactionai-ethicslatent-diffusion-modelsgenerative-ai-applications
“AI models like Nano Banano empower creators by automating tedious tasks, enabling them to focus on creativity.”
youtube / GoogleDeepMind / Aug 16

Genie 3: How DeepMind Built a Real-Time World Model with Persistent Memory from Scratch

Google DeepMind's Genie 3 represents a convergence of three internal research threads — Genie 2 (3D environment generation), VO2 (high-quality video generation), and GameNGen (the "Doom paper") — into a single real-time, text-prompted world model capable of generating interactive, photorealistic environments. The model's most technically significant capability is persistent spatial memory exceeding one minute, achieved without explicit 3D representations (no NeRFs, Gaussian splatting, etc.), instead emerging from scale and a frame-by-frame generation approach. The team views Genie 3 not as an agent but as a general-purpose environment simulator, positioning it as infrastructure for training embodied AI agents — potentially bridging the sim-to-real gap in robotics by combining data-driven realism with the scalability of simulation. Text-to-world prompting (replacing image prompting from prior Genie versions) was enabled by cross-pollination with the VO project team, and emergent behaviors like terrain-appropriate agent locomotion arose from scale rather than explicit engineering.

world-modelsgenerative-aivideo-generationgoogle-deepmindembodied-airoboticsreinforcement-learning
“Genie 3's persistent spatial memory (1+ minute) was achieved without any explicit 3D world representation such as NeRFs or Gaussian splatting, instead emerging from a frame-by-frame generative approach at scale.”
youtube / GoogleDeepMind / Jun 26

AlphaEvolve: Scaling Evolutionary Search for Superhuman Algorithmic Discovery

AlphaEvolve is an autonomous coding agent that utilizes LLMs and evolutionary search to discover novel, high-efficiency algorithms for computer science and mathematics. By iterating through generations of code and filtering them via strict evaluation functions or simulators, it can navigate vast, non-intuitive search spaces to find superhuman optimizations. Its ability to produce interpretable code rather than black-box models allows for human verification and deployment in critical infrastructure, such as Google's data centers.

ai-coding-agentevolutionary-algorithmsllm-applicationsscientific-discoveryalgorithm-optimizationself-improving-aideepmind-research
“AlphaEvolve has discovered algorithms for data center scheduling and training infrastructure that are currently deployed in Google's production environment.”
youtube / GoogleDeepMind / Nov 14

AlphaProof: DeepMind's Breakthrough in Formal Mathematics

DeepMind's AlphaProof represents a significant advancement in AI for mathematical reasoning, building on the success of AlphaZero. By leveraging reinforcement learning with a neural network and a planning/search component, AlphaProof can generate and verify mathematical proofs in a formal language, entering a self-improving loop. This system has demonstrated remarkable capabilities in solving complex problems, notably at the International Mathematical Olympiad (IMO) level.

deepmind-alphaproofmathematical-reasoningai-systemsreinforcement-learningformal-verificationai-ethicsimo-problems
“AlphaProof utilizes a reinforcement learning architecture inspired by AlphaZero, combining a neural network, large-scale reinforcement learning, and a planning/search component to generate and verify mathematical proofs.”
youtube / GoogleDeepMind / Sep 11

From Scaling Search to Multimodal Intelligence: The Evolution of Google's AI Strategy

Google's AI trajectory has evolved from scaling search infrastructure to developing massively parallel neural networks and eventually natively multimodal models like Gemini. This shift emphasizes the transition from recurrent sequential processing to parallelized attention mechanisms, enabling high-dimensional representations across text, image, audio, and video to create a unified conceptual model of the world.

ai-developmentlarge-language-modelsmultimodal-aigoogle-deepmindai-historymachine-learning-engineeringtransformers
“The Transformer architecture's primary advantage over recurrent models is the ability to process sequences in parallel, increasing efficiency by 10 to 100 times.”
youtube / GoogleDeepMind / May 21

AI as a Cognitive Necessity for Large-Scale Scientific Discovery

The scale of contemporary scientific data has exceeded the cognitive capacity of individual human experts across physics, biology, and mathematics. AI is evolving into a primary engine for discovery by identifying latent patterns and conjectures in high-dimensional data—such as the link between algebraic and geometric knot theory—that were previously invisible to human analysts. This represents a fundamental shift toward a hybrid intelligence model where AI handles large-scale data synthesis and humans focus on higher-order collaboration.

ai-scientific-discoverymathematical-proofbiology-aidata-analysisai-necessitymachine-learning-applications
“AI is transitioning from a tool of convenience to a cognitive necessity for processing modern scientific datasets.”
youtube / GoogleDeepMind / Mar 18

Video as the New Language for Real-World Decision Making

Sherry Yang of Google DeepMind and UC Berkeley argues that video is an underappreciated data format with the potential for a "ChatGPT moment." She proposes that video can serve as a unified representation of information and a unified task interface for real-world decision-making, similar to how text functions for large language models. The research focuses on leveraging video generation models as real-world simulators to overcome limitations of text-based models in applications like robotics and scientific discovery.

video-generationllmsai-researchdeepmindroboticsworld-modelsmultimodal-ai
“Video is a unified data format for information, similar to text for language models.”