Chronological feed of everything captured from Google DeepMind.
Google DeepMind has announced "AI Co-Clinician," a research initiative focused on developing multimodal AI agents to support healthcare workers and patients. The project signals DeepMind's continued push into clinical AI, moving beyond single-modality tools toward integrated, agentic systems. The announcement is framed as a progress snapshot, suggesting ongoing, iterative research rather than a product launch.
Google DeepMind has developed an AI co-clinician system designed to augment medical decision-making with high-quality evidence. Evaluated against the NOHARM safety framework, the system produced zero critical errors in 97 out of 98 primary care queries, and outperformed comparable AI systems in blind evaluations. The result signals meaningful progress in applying safety-aware evaluation frameworks to clinical AI, though real-world deployment readiness remains to be established.
Google DeepMind has developed an AI co-clinician system capable of processing live video and audio to assess physical symptoms in real-time, including gait analysis, respiratory sounds, and dermatological presentation. The system was evaluated in a simulation study of 20 scenarios using patient-actors, designed in collaboration with physicians from Harvard Medical School and Stanford Medicine. This represents a meaningful step toward multimodal, physician-assistive AI that operates on raw sensory input rather than structured clinical data alone.
Google DeepMind's AI co-clinician benchmark study reveals a nuanced performance profile: the model matched or outperformed physicians in 68 of 140 assessed clinical areas (~49%), with competitive results in triage. However, human clinicians retained a clear advantage in identifying critical red flags and directing physical examinations — capabilities that depend on embodied, contextual judgment the AI cannot replicate. The findings suggest AI's near-term clinical value lies in augmentation of physician decision-making rather than substitution, particularly for structured reasoning tasks like triage.
Google DeepMind has implemented a dual-agent architecture in a clinical AI system where a dedicated "Planner" agent continuously monitors a conversational "Talker" agent to ensure it stays within safe clinical boundaries. This separation of concerns — conversational capability vs. safety oversight — reflects a real-time, in-process approach to AI safety rather than relying solely on post-hoc review or model-level guardrails. The design prioritizes patient safety by embedding oversight directly into the system's runtime loop.
Google DeepMind is scaling its health-focused AI research by collaborating with academic institutions and researchers worldwide, while incrementally broadening a clinician-facing trusted tester program. The phased expansion is designed to capture diverse perspectives from health workers and patients across different global contexts. The approach suggests a cautious, feedback-driven deployment strategy prior to any broader clinical rollout.
Google DeepMind is running a community-facing campaign ahead of Google I/O, inviting developers and creators to submit projects built with Gemini App or Google AI Studio for a chance to be featured during the main stage countdown. The initiative targets "vibe coders" and creative technologists, with highlighted examples including protein simulators, physics engines, and math-based art — signaling DeepMind's intent to publicly showcase the breadth of generative and scientific AI use cases possible on its platforms.
Google DeepMind is running a creative submission challenge requiring entries to be built around the numbers 1–10, with a deadline of May 6. Participants are directed to use the Canvas feature within either the Gemini App or Google AI Studio as their creation tool. The content is sparse and promotional in nature, offering little technical depth beyond the submission guidelines.
Ten years after AlphaGo's landmark matches in South Korea, Google DeepMind is formalizing a collaboration with the Korean government to deploy AI capabilities toward scientific research acceleration and regional economic development. The announcement signals a shift from symbolic AI milestones to institutionalized government-level AI partnerships. Specific programs or funding commitments are not detailed in the post.
Google DeepMind, in partnership with Raspberry Pi Foundation, has been running the "Experience AI" program since 2023, providing free educational resources aimed at helping students and teachers build foundational AI literacy. The initiative reflects a broader institutional push to align K-12 and general education curricula with the pace of AI development. Impact metrics were referenced but not detailed in the available content.
Google DeepMind has scaled an AI literacy training program that has reached over 30,000 teachers and 2.9 million students across 180 countries in 19 languages. The program reports strong self-reported efficacy gains among educators, with 93% claiming improved AI conceptual knowledge and 87% reporting greater confidence in teaching AI topics. This represents one of the broader global efforts to embed AI education into K-12 or equivalent pipelines at scale.
Google DeepMind is expanding its AI education initiative into Latin America, backed by $4.6 million in funding from Google.org. The program targets educators as the primary vector for scale, aiming to train 24,000 teachers who will in turn reach 1.25 million students by 2028. This reflects a broader strategy of embedding AI literacy at the institutional level through educator upskilling rather than direct student outreach.
Decoupled DiLoCo integrates Pathways and DiLoCo to enable continuous AI model training across multiple data centers without halting due to chip failures or synchronization issues. It features self-healing capabilities, isolating disruptions from artificial hardware failures and reintegrating recovered units seamlessly. Demonstrated by training a 12B Gemma model over four US regions on low-bandwidth networks and mixing TPUv5p with TPU6e without performance loss, it overcomes geographic, capacity, and hardware constraints for scalable AI infrastructure.
Decoupled DiLoCo combines Pathways and DiLoCo to enable continuous AI model training across multiple data centers without halting due to chip failures or synchronization issues. It features self-healing capabilities, isolating disruptions and reintegrating recovered units automatically. Demonstrated by training a 12B Gemma model over four US regions on low-bandwidth networks and mixing TPUv5p with TPU6e hardware without performance loss.
Decoupled DiLoCo combines Pathways and DiLoCo to enable continuous AI model training across data centers without halting due to chip failures, featuring self-healing capabilities that isolate disruptions and reintegrate recovered units. It supports low-bandwidth networks and mixed hardware generations like TPUv6e and TPUv5p without performance degradation. Demonstrated by training a 12B Gemma model across four US regions, it decouples training from geographic, capacity, and chip constraints.
Decoupled DiLoCo combines Pathways and DiLoCo to enable continuous AI model training across multiple data centers without halting due to chip failures, featuring self-healing capabilities that isolate disruptions and reintegrate recovered units. It supports low-bandwidth networks and mixed hardware generations like TPUv6e and TPUv5p without performance degradation. Demonstrated by training a 12B Gemma model across four US regions, it redefines scalable AI infrastructure unbound by geography or hardware constraints.
Decoupled DiLoCo integrates Pathways and DiLoCo to enable continuous AI model training across multiple data centers without halting due to chip failures. It features self-healing by isolating disruptions and reintegrating recovered units. Demonstrated training a 12B Gemma model over four US regions on low-bandwidth networks using mixed TPU generations (v5p and 6e) without performance loss.
Decoupled DiLoCo combines Pathways and DiLoCo to enable continuous AI model training across multiple data centers without halting due to chip failures or synchronization issues. It features self-healing capabilities, isolating disruptions and reintegrating recovered units automatically. Demonstrated by training a 12B Gemma model over four US regions on low-bandwidth networks and mixing TPUv5p with TPU6e hardware without performance loss.
Demis Hassabis' journey to AGI began in 1988 programming Othello on an Amiga 500, leading to his core insight that software can act autonomously on humans' behalf. This epiphany remains foundational to Google DeepMind's mission. The team now applies this principle to tackle major scientific challenges.
Gemini Robotics-ER 1.6 improves robots' ability to precisely identify, locate, and count specific objects in cluttered images like workshops. It employs multi-view reasoning and fuses live camera streams to build a comprehensive scene understanding. The model autonomously assesses task completion, deciding to retry or proceed accordingly.
Gemini Robotics-ER 1.6 upgrades robots with advanced visual and spatial understanding, enabling precise object detection in clutter, multi-view task completion verification, and sub-tick analog gauge reading. It processes complex scenes like industrial inspections by self-correcting for camera distortions and fuses live camera streams for full-scene comprehension. The model also prioritizes safety through physical constraint awareness and improved human injury risk detection.
Gemini Robotics-ER 1.6 upgrades robots with enhanced visual and spatial understanding, enabling accurate object pinpointing in cluttered environments, multi-view scene fusion for task completion verification, and precise reading of analog instruments like gauges with sub-tick accuracy. It processes complex industrial inspections by self-generating code to correct camera distortions and calculates exact measurements. The model prioritizes safety through physical constraint awareness and 10% improved human injury risk detection in videos.
Gemini Robotics-ER 1.6 upgrades visual and spatial understanding, enabling robots to pinpoint objects in cluttered environments, fuse multi-view camera streams for task completion verification, and read analog instruments with sub-tick accuracy. It addresses industrial challenges like processing distorted images from patrols, self-correcting via code generation for precise measurements. The model also improves safety by respecting physical constraints and boosting human injury risk detection by 10%.
Gemini Robotics-ER 1.6 upgrades robot perception with superior visual and spatial understanding, enabling accurate object detection in clutter, multi-view scene fusion for task completion verification, and precise analog instrument reading. It integrates spatial reasoning, world knowledge, and agentic vision to process complex industrial inspections, including distortion-corrected gauge analysis on robots like Spot. The model also prioritizes safety by respecting physical constraints and improving human injury risk detection by 10%.