Foundation Models

Over 100 AI experts are demanding better resources and protections for AI safety testers. They emphasize the need for independent oversight of powerful AI models. A consortium of academics and evaluators, including Geoffrey Hinton, published a letter urging AI providers to ensure third-party evaluators have objectivity, transparency, and independence. This initiative follows recent commitments by AI companies to embrace more rigorous third-party testing, with proposals suggesting “employee-like access” for evaluators to audit cutting-edge models and development processes. The call for standardization aims to build confidence in AI risk assessment.

Meta has launched its new AI personal agent app, “Muse,” built on its advanced Muse Spark foundation models. Designed to be approachable and explainable, Muse assists users with tasks like scheduling and monitoring home security. The app offers flexible pricing, with free and paid tiers, aligning with Meta CEO Mark Zuckerberg’s vision of AI-powered digital assistants. This move comes amidst increasing regulatory scrutiny and public skepticism towards AI, but Meta aims to demonstrate tangible returns on its AI investments and drive future revenue.

Samsung Research America is advancing digital health with two AI foundation models, xMAE and HiMAE, for smartwatches. These models extract deep insights from biosignals, focusing on heart activity, sleep, and exertion. xMAE links PPG and ECG data for cardiovascular analysis, while HiMAE analyzes wearable data across multiple time scales. These innovations aim to provide efficient, precise, and continuous health insights, supporting Samsung’s “Connected Care” vision for personalized, preventive healthcare.

Meta is significantly boosting AI investments, earmarking $115-135 billion for AI infrastructure in 2026. Despite past investor concerns, strong Q4 earnings and 24% revenue growth, largely from advertising, have bolstered confidence. CEO Mark Zuckerberg aims for “personal super intelligence,” focusing on data center expansion and developing advanced AI models, including a successor to Llama, codenamed “Avocado.” The company’s financial strength from advertising provides the latitude for this ambitious AI-driven future.

At the Huawei Cloud AI Summit, agentic AI applications demonstrated autonomous decision-making potential. A Conch Group cement plant utilizes the system to predict clinker strength with over 90% accuracy and autonomously optimize calcination, reducing coal consumption by 1%. Huawei Cloud’s agentic AI platform supports independent planning, decision-making, and execution, requiring robust infrastructure like CloudMatrix384. These systems are being tailored for specific industries, improving efficiency in sectors like cultural tourism and travel management by addressing operational challenges and streamlining processes.