Google's latest 'Deep Research' initiative counters GPT-5.2

1. Unstructured Information Understanding: Extract first leverages Gemini’s vision-language capabilities to process low-quality scanned documents. It not only recognizes text but also interprets the semantic meaning of handwritten annotations—for instance, distinguishing ‘approval date’ from ‘application date’ in margin notes—with a date recognition accuracy rate of 94%. 2. Visual Reasoning and Polygon Extraction: This represents the most critical technical breakthrough. Gemini comprehends the visual symbol language on maps—for example, differentiating between ‘solid red lines’ denoting property boundaries and ‘dashed blue lines’ indicating drainage channels. Once the target area is identified, the system invokes computer vision tools such as OpenCV and SAM to extract geographic polygons from pixel images with surgical precision; its shape matching accuracy (IoU) reaches 90%. 3. Spatiotemporal Feature Matching: Historical maps often differ from modern satellite imagery in scale and reference frame. Extract employs the LoFTR algorithm to identify common feature points—such as ancient churches or road intersections—between old and new maps, computing an accurate transformation matrix that precisely overlays hand-drawn redlines from decades ago onto today’s digital map coordinate system. 4. End-to-End Automation: Through this pipeline, Extract reduces the average processing time for a complex planning document from two hours to just 40 seconds–3 minutes. This means a local council can digitize hundreds of backlogged archival documents daily—boosting efficiency by over one hundredfold. Currently, Extract is undergoing pilot deployment across four regions, including Westminster and Hillingdon. The UK government plans to roll it out nationwide to all local councils by spring 2026. This will not only free up thousands of hours of administrative labor, but—more importantly—establish a unified national digital planning database, providing the foundational data infrastructure for the UK government’s pledge to build 1.5 million new homes. This stands as the best exemplar of DeepResearch technology applied in a vertical domain—transforming general multimodal reasoning capabilities into tangible administrative productivity. Scientific Infrastructure for the New Era: From AlphaFold to Automated Materials Laboratories In fundamental science, DeepMind’s collaboration with the UK government aims to accelerate the flywheel effect of scientific discovery through AI. DeepMind has announced plans to establish its first automated AI science laboratory in the UK by 2026. ・ Closed-Loop Discovery System: The lab will operate a closed-loop system powered by Gemini and GNoME (Graph Networks for Materials Exploration). AI designs novel crystal structures grounded in quantum chemical principles and predicts their stability. ・ Robotic Synthesis: Design instructions are sent directly to fully automated robotic platforms, which handle material weighing, synthesis, sintering, and testing. ・ Data Feedback Loop: Experimental results are fed back to the AI in real time, refining predictions for subsequent rounds. The goal is to shorten the discovery cycle for new materials—such as room-temperature superconductors or high-efficiency battery electrolytes—from decades to mere months—or even days. This initiative directly supports the UK’s Net Zero strategy and energy security objectives.

Google DeepMind releases world model Genie 3; OpenAI launches open-source model series gpt-oss; Anthropic releases Claude 4.1 Opus [AI Morning Brief, 2025-08-06]

Juyā AI Morning Brief — 2025-08-06 Overview ・ Google DeepMind releases world model Genie 3 #1 ・ OpenAI launches open-source model series gpt-oss #2 ・ Anthropic releases Claude 4.1 Opus and updates Claude Code #3 ・ OpenBMB releases edge-side multimodal model MiniCPM-V 4.0 #4 ・ Intelligent Internet releases open-source research agent model II-Search #5 ・ JetBrains launches Kineto, a no-code AI application development platform #6 ・ ElevenLabs unveils Eleven Music, an AI music generator #7 ・ Qwen launches Flash-series APIs; Qwen-Image now natively supported in ComfyUI #8 ・ Ollama introduces Turbo service to accelerate large model execution #9 ・ Google releases MLE-STAR, a machine learning engineering agent #10 ・ Google Gemini app introduces ‘Storybook’ creation feature #11 ・ Google NotebookLM supports batch URL uploads #12 ・ OpenRouter adds PDF URL upload capability and tool-calling test UI #13 ・ ManusAI opens Wide Research functionality to Plus and Team users #14 ・ Google Gemini is testing a ‘guided learning’ feature #15 Google DeepMind releases world model Genie 3 #1 Google DeepMind has launched the general-purpose world model Genie 3, capable of generating interactive, dynamic virtual environments in real time from text prompts. Google DeepMind has announced Genie 3—a general-purpose world model that generates highly diverse, interactive, and dynamic environments from text prompts. The model can generate navigable worlds in real time at 24 frames per second, at 720p resolution, and sustain coherence for several minutes. Genie 3 marks a major milestone in Google DeepMind’s world model research—it is their first world model enabling real-time interaction, while also improving consistency and realism over its predecessor, Genie 2. Its core capabilities include simulating physical properties (e.g., water and lighting), natural phenomena (e.g., animal behavior and plant growth), animated and fictional scenes, and exploring diverse geographical locations and historical settings. Technically, Genie 3 achieves two key breakthroughs. First, real-time interactivity: while autoregressively generating each frame, the model must account for increasingly long prior trajectories over time—and respond rapidly to user input.

Google DeepMind and Cloud Grant European Researchers Access to Frontier AI Models

15 hours ago

Google DeepMind and Google Cloud will grant 200 European researchers access to frontier models like AlphaFold 3 alongside $3 million in compute credits. This initiative aims to accelerate tackling complex challenges such as mapping molecular structures.

reuters

2026-05-19

AI Companies Introduce New Agent-Based Tools for Scientific Discovery

2026-10-13

Google DeepMind and Cloud Grant European Researchers Access to Frontier AI Models

Google DeepMind and Cloud Grant European Researchers Access to Frontier AI Models

FAQ