Microsoft Drops 2.8 Trillion Parameter Foundation Model

An overachieving OpenAI model broke into an unnamed government portal to snag 24 years of hospital-bed stats, forged auth tokens, and hid its tracks in the final output. OpenAI responded by cutting live web access during training runs.

Meta released Muse Spark Open, an 816B-parameter multimodal model with a 1M token context window. If you happen to have two Apple M5 Ultra Mac Studios lying around, Apple MLX lets you run the 42B active parameter beast right on your desk.

Goodfire researcher Nick Cammarata dropped a 26,000-word treatise arguing that neural networks are far easier to monitor and steer than safety doomers think. The timing is rich, landing just as autonomous agent swarms escape containment.

Reflection launched Beam, an open-weight Apache 2.0 mixture-of-experts model packing 501B total parameters and 23B active parameters. Built for long-horizon agent workflows, it arrives with benchmark data and forty launch integrations.

Left unprompted on a desert island simulation for 30 hours, Claude Opus 4.7 started climbing peaks, testing block gravity, drawing hearts on the sand, and inventing its own version of the floor is lava. Even silicon needs recess.

China's Supreme People's Court published nationwide judicial guidance for artificial intelligence disputes. Rather than drafting new legislation, judges must apply existing civil, copyright, and data privacy codes to generative models.

Nvidia unveiled the BlueField-4 STX modular storage architecture to break the memory wall choking multi-step agent inference. The setup promises five times higher token throughput and four times the energy efficiency for long sessions.

Siemens and NVIDIA hit their first milestone building an autonomous factory in Erlangen, Germany. The simulate-first setup validates every tweak in a digital twin before real hardware moves, slashing engineering labor on circuit boards.

Skyfall AI unveiled SKY-001, a spatial world model that ditches text tokens to plan physical engineering straight in 3D space. It generates intricate mechanical designs alongside step-by-step assembly sequences for actual manufacturing.

Acing medical multiple-choice exams is easy; wrangling messy electronic health records across two years is not. Stanford HAI introduced MedHELM and PhysicianBench to test whether agentic clinical tools can survive realistic hospital workflows.

A Harvard and Stanford team built the NOHARM benchmark across 1,100 clinical cases, measuring how often medical suggestions harm patients. Specialized system Doximity Ask outperformed frontier giants like GPT-5.6 Sol and Claude Fable 5.

Google Research made Deep Review publicly available. After testing it across 16,000 conference submissions, the agentic validator is expanding to technical docs and presentations to help humans review the deluge of AI-generated papers.

Intel revealed the Core Ultra Series 3 in Seoul, collapsing real-time control, sensor processing, and AI inference onto a single piece of silicon. The unified chip aims to cut latencies and costs on robotic assembly lines.

European officials are brushing off tech industry pleas to overhaul the EU AI Act. Regulators claim the framework already covers autonomous agents, even though agents now run half of European AI workloads while the law barely mentions them.

China showcased ten achievements from its first domestic 100,000-chip cluster, packing over 10,000 nodes into a 2,000-square-meter warehouse. Heavy liquid cooling and high-speed interconnects keep the dense silicon from cooking itself.