Google has introduced AI-PPG-ACC-HTN, a multimodal AI system that passively detects signs of hypertension using ordinary consumer smartwatches without cuff calibration. The work, published in Nature, points to wearables moving from everyday monitoring toward disease screening.
Multimodal Learning
多模态学习
Technology that processes and understands multiple data types like text and images simultaneously
Curated Intelligence
Fabian Theis of Helmholtz Munich recounts his lab's long-running push on the Virtual Cell concept, framing it as an algorithmic effort to understand multi-scale, multi-modal and time-varying biological data. He notes that AI-for-bio attention has clustered around protein and molecule design, while virtual cells demand broader data collection and large-scale modeling.
A report from The Download highlights growing skepticism regarding the hype around humanoid robots, challenging the assumption that AI advances behind LLMs like ChatGPT can directly enable robotic imitation of human movement. Researchers argue that applying language and image-based intelligence to the infinite variability of the physical world is fundamentally flawed and misleadingly conflates 'humanoids' with 'generalist machines.' Tensions are rising across robotics labs over whether current AI architectures suffice for general-purpose robots or if entirely new paradigms are required.
Google has launched WeatherNext 3, its flagship AI weather model that replaces traditional physics simulations with real-time satellite data to deliver five times higher resolution and hourly updates. The model is now integrated across Search, Gemini, Maps, and Google Cloud, offering improved precision in precipitation forecasting and clean energy variables.
Google researchers propose using smartphone images to estimate body composition for predicting insulin resistance (HOMA-IR), a key metabolic risk indicator. The study aims to supplement wearable data with visual inputs, building a multimodal machine learning framework for earlier metabolic health screening before routine blood tests. Currently exploratory, the technology plans to integrate continuous glucose monitoring and clinical biomarkers to improve prediction accuracy.
Google Research announced that its medical AI system, AMIE, has evolved from text-based dialogue to a multimodal system capable of processing audio-visual cues like gait and breathing. The upgraded system achieves expert-level diagnostic performance in simulated consultations by integrating non-verbal signals previously ignored by text-only models.
A Stanford team is developing a human-centered foundation model to simulate human cells for accelerating drug discovery and personalized treatment. The project aims to integrate multimodal biological data—including sequences, structures, and literature—to provide generative tools comparable to AlphaFold for cell biologists worldwide.
Google DeepMind has officially open-sourced the Gemma 4 family of multimodal models under the Apache 2.0 license. Supporting text, image, audio, and video inputs, these frontier-performance models are optimized for on-device deployment and are now available on Hugging Face with compatibility for major inference engines and fine-tuning libraries.
The Hugging Face team released the Efficient MultiModal Data Pipeline (MMDP) to address GPU idling and padding waste caused by data loading bottlenecks during training. By optimizing batching strategies and data flow architecture, the tool significantly improves training efficiency for models like nanoVLM and is now open-sourced on GitHub for developer reuse.
Hugging Face partners with IISc and ARTPARK to globally open-source the 'Vaani' dataset. As India's most diverse open-source multimodal and multilingual resource, it aims to lower barriers for building AI models tailored to Indian languages.