Google Research
·
4 months ago
● 2 sources
Google Research presented multiple AI applications in healthcare, including a Personal Health Agent that outperforms single-task health apps, and an AI system that identified 25% of previously missed interval cancers in breast cancer screening. The company has deployed screening models for diabetic retinopathy across India, Thailand and Australia, conducting over one million screenings, and launched open-weight medical models (MedGemma) available to developers and healthcare providers worldwide. These advances aim to shift AI from a tool to a clinical collaborator that improves diagnostic accuracy, reduces clinician workload, and democratizes healthcare technology access globally.
AI Act
·
4 months ago
The EU AI Act classifies AI systems used in hiring, candidate screening, and performance evaluation as high-risk, requiring staffing businesses to conduct risk assessments, perform bias testing, implement human oversight, and disclose AI use to candidates by 2 August 2026. Non-compliance can result in fines up to EUR 15 million or 3% of global annual turnover, with regulators also able to withdraw AI systems from the market. Staffing businesses deploying AI tools in employment decisions—regardless of who built the technology—must redesign their compliance infrastructure and vendor relationships, as obligations cannot be passed to technology partners.
Google Research
·
4 months ago
● 2 sources
Google Research and NHS organizations published two studies in Nature Cancer evaluating an AI system for breast cancer screening, comparing it to the current double-read human workflow. The AI system achieved 9.33 cancer detections per 1,000 women versus 7.54 for the original first reader, and detected 25% of interval cancers missed by human readers, while the AI-enabled workflow reduced total human reading time by 36-44% without compromising detection sensitivity or specificity. The findings suggest AI could help address the projected 40% radiologist shortfall by 2028, though challenges remain in managing arbitration volumes and improving explainability to prevent human readers from overruling correct AI decisions.
Hugging Face Blog
·
4 months ago
Hugging Face's open source AI ecosystem nearly doubled in scale over the past year, growing to 13 million users, 2 million public models, and 500,000 public datasets. Chinese models surpassed U.S. models in downloads for the first time, accounting for 41% of all downloads in 2025 following DeepSeek's January release. Independent developers increased their share of ecosystem output from 17% to 39%, reshaping how innovations spread as individuals and small teams now steer meaningful portions of what users can access.
Google DeepMind
·
4 months ago
Google DeepMind released a cognitive framework for measuring progress toward artificial general intelligence, drawing on psychology and neuroscience research to identify 10 key cognitive abilities including perception, reasoning, memory, and social cognition. The company is launching a Kaggle hackathon with $200,000 in prize money running from March 17 through April 16 to build evaluations for five abilities where measurement gaps are largest: learning, metacognition, attention, executive functions, and social cognition. The framework will allow researchers to benchmark AI system performance against human baselines across cognitive tasks to track incremental progress toward AGI.
Mistral AI
·
4 months ago
● 2 sources
Mistral AI launched Forge, a system enabling enterprises to train custom AI models using their proprietary internal data rather than relying solely on publicly available information. The system supports pre-training, post-training, and reinforcement learning across both dense and mixture-of-experts architectures, with initial partnerships including ASML, Ericsson, and the European Space Agency. Organizations can now build AI agents that understand internal terminology, workflows, and policies while maintaining control over their models and intellectual property within regulated environments.
Hugging Face Blog
·
4 months ago
H Company released Holotron-12B, a 12-billion-parameter multimodal model designed to serve as a policy model for computer-use agents that perceive and act in interactive environments. On the WebVoyager benchmark with 100 concurrent workers, Holotron-12B achieved 8.9k tokens per second throughput on a single H100 GPU, more than 2 times higher than Holo2-8B's 5.1k tokens per second. The model's hybrid state-space and attention architecture enables enterprises to deploy computer-use agents at scale with reduced memory requirements and improved batch processing efficiency.
OpenAI Blog
·
4 months ago
OpenAI released GPT-5.4 mini and nano, smaller versions of its GPT-5.4 model designed for coding, tool use, and multimodal tasks. The nano variant is the smallest available model in the GPT-5.4 family, optimized for high-volume API calls and sub-agent workloads. These models allow developers to reduce latency and costs when deploying AI systems that don't require the full capabilities of the standard GPT-5.4.
OpenAI Blog
·
4 months ago
OpenAI Japan launched the Japan Teen Safety Blueprint, a set of policies designed to protect minors using generative AI through age verification, parental oversight tools, and content safeguards. The framework targets users under 18 and includes mechanisms for parents to monitor their children's AI interactions. The initiative requires platforms to implement stricter age-gating measures and limits on potentially harmful content exposure for younger users.
.jpg)
Together AI
·
4 months ago
Mamba-3 is a new state space model designed to prioritize inference efficiency rather than training speed, introducing a more expressive recurrence formula, complex-valued state tracking, and multi-input multi-output variants. On the 1.5B scale, Mamba-3 SISO achieves lower prefill+decode latency than Mamba-2, Gated DeltaNet, and Llama-3.2-1B across all sequence lengths, with the fastest latency at 4.39 milliseconds for 512 tokens versus Mamba-2's 4.66 milliseconds. The team open-sourced custom kernels built with Triton, TileLang, and CuTe to enable wider adoption and experimentation with the new architecture.
OpenAI Blog
·
4 months ago
Americans send nearly 3 million daily messages to ChatGPT asking about compensation and earnings questions. This volume represents a significant source of wage information for workers seeking to understand their pay relative to market rates. The data suggests workers are using AI tools to fill gaps in compensation transparency that traditional resources may not address.
