New Technology– category –
-
Medical AI Benchmarks: Evaluating Clinical Task Performance in 2026
Stanford HAI USA Overview Stanford HAI research reveals a "generative AI chasm" where models excel at medical exams but struggle with real-world clinical execution, with the best agents completing less than half of complex work. While pe... -
FairStress: Measuring LLM Demographic Fairness and Robustness
Ai2 (Allen Institute for AI) USA Overview Ai2 has introduced FairStress, a benchmark that evaluates demographic fairness in LLMs by testing how resistant model decisions are to persuasive "adversarial prefixes." The benchmark reveals tha... -
MASS: Recursive Self-Improvement via Multi-Agent Self-Supervision
arXiv International Overview Multi-Agent Self-Supervision (MASS) enables models to recursively improve on non-verifiable tasks by alternating between evolutionary workflow optimization and supervised fine-tuning. Using Qwen3.6-27B, the m... -
NVIDIA CUDA-X: AI-Defined Fab Tools and Industry Adopters
NVIDIA USA Overview NVIDIA has expanded its CUDA-X library portfolio, including the new cuPotential for atomic-scale simulation, to enable "AI-defined fabs." Industry leaders such as Synopsys, Cadence, and Lam Research are utilizing thes... -
Gemini 4 Argon: Performance Specs and Engineering Use Cases
Google Blog USA Overview Google has released Gemini 4 Argon, a frontier model ranking #1 on the Vals Index for critical economic sectors. The model demonstrates significant gains in complex engineering, reducing code rollbacks by nearly ... -
Microsoft MAI-Base-2: 2.8T Parameter Model Specs vs DeepSeek
BigGo Finance Taiwan Overview Microsoft has pre-trained MAI-Base-2, a 2.8 trillion-parameter foundation model that claims 4.2x to 12.3x higher compute efficiency than DeepSeek-V4-Pro-Base. The model utilizes a Mixture-of-Experts (MoE) ar... -
Google Willow: Magic-State Cultivation and Error Reduction
Science / Google Quantum AI Blog USA Overview Google has demonstrated a magic-state cultivation method on its Willow quantum processor, reducing state-preparation errors approximately 40-fold. The team achieved 99.99% logical fidelity fo... -
GPT-6-Astra: Research-Level Physics Test Performance and Costs
Human Progress (citing arXiv) USA Overview GPT-6-Astra, harnessed in Codex, achieved 98% accuracy on a suite of 50 unpublished research-level theoretical physics problems. The agent performed these tasks autonomously at an average cost o... -
MLX Distributed Inference: Mac Studio M5 Ultra LLM Performance
Apple Machine Learning Research (via Reddit) USA Overview Apple's MLX framework enables the execution of frontier-class models, including the 816B parameter Muse Spark Open, across clusters of Mac Studios using RDMA over Thunderbolt. On ... -
AI Timeline 2026: Cognito-X, OrionNet 2.0 and CausalGraph-Net Specs
AI TECH Overview Recent breakthroughs in October 2026 highlight a shift toward embodied AI and secure edge computing, featuring Synapse AI's Cognito-X with a 1M token context window. Key advancements include 98% success rates in humanoid...
