Topic
Explore the latest artificial intelligence news and research.
A test configuration of RTX Spark with 18 cores and 5,120 GPU cores has appeared in Geekbench 7. Reading the performance gap against the 20-core version and the limits of a single benchmark run, we examine the product tiers NVIDIA may be considering.
Across six LLMs assigned occupational roles, compliance rates with inappropriate requests from a higher-status role rose by 2.0 to 3.7 percentage points. The findings are limited to synthetic English dialogues, and questions remain about statistical procedures and reproducibility.
AWS reportedly asked internal engineers to conserve EC2 CPU capacity. As AI agents drive up demand for general-purpose CPUs, we examine the supply constraints Amazon itself has acknowledged and the conditions for safely reclaiming spare capacity.
Intel's continuation patent application outlines a two-tier architecture in which a small number of high-orbit NOCs control a LEO satellite network numbering in the thousands. Product specifications and flight demonstrations remain unpublished, and inter-orbit link performance will determine whether the concept becomes practical.
Five companies, including Amazon and OpenAI, have released Agent Plugins 1.0.0, which packages Agent Skills and MCP servers in a common format. Repackaging work is reduced, but permissions and authentication remain up to each product's implementation.
Cloudflare has released a beta of Kitesurf, a browser for AI agents that runs on Workers. It's a purpose-built design that trades speed and compatibility for reduced CPU and memory usage compared to Chromium.
Global semiconductor sales reached $403.3 billion in Q2 2026, marking a 35.1% quarter-over-quarter increase—the highest ever recorded in over 40 years of WSTS statistics. This piece examines how AI demand pulled forward the $1 trillion annual market milestone, once projected for around 2032, to 2026, and untangles the uncertainties surrounding whether this growth can be sustained.
Reports say ByteDance is pretraining an AI model with up to 10 trillion parameters. Looking at Kimi K3's official specs and Anthropic's non-disclosure policy reveals why total parameter counts alone can't measure performance differences.
Claude Code v2.1.224 adds text communication between independent sessions. It automates coordination for parallel development while keeping permissions and conversation history separated per session.
BQP and Modovolo have introduced a search engine that doesn't use quantum computers into the design of 3D-printed drone propellers. The improvement rate in actual performance and comparison conditions remain undisclosed, and verified figures are needed to evaluate commercialization.
