Topic
Explore English articles about Generative AI.
Micron has announced the world's first demonstration of a 512GB DDR5 RDIMM. By stacking 32Gb dies with TSV technology, it achieves speeds up to 9,200 MT/s while cutting power consumption to 16.0W. With 24 slots, it enables 12TB of main memory, with mass production validation targeted for late 2027.
Perplexity's local AI agent now supports Windows. We look at its on-device runtime design, the 24GB VRAM requirement, and where local processing ends when cloud services are involved.
TypeSafe has released Jev, a decision-only AI that returns choices and probabilities in parallel rather than writing prose. How does it differ from conventional LLMs? We examine RLCD, structured outputs, what confidence means, and the conditions behind the speed claims, drawing on official specs and third-party trials.
Simply publishing an AI model's weights doesn't make it open source. Without training data or auditable records to verify and reproduce results, that boundary remains undefined—and the OpenMDW license, which aims to codify it, still hasn't cleared OSI review over its termination clause.
A Mozilla study finds that the performance gap between commercial AI models and open-weight models—led largely by Chinese developers—has narrowed to 4.4 months. A 5x cost premium is increasingly justified only for a narrow band of difficult tasks that take human experts 8 to 12 hours to complete.
DeepSeek and Hermes Agent reportedly helped improve BIOHAZARD 7 resident evil from 20FPS to 30FPS on Android via texture compression and SGSR. Here's what the primary sources actually show.
MediaTek has announced the Dimensity 9600 Pro and Dimensity 9600M, SoCs built on TSMC's 2nm process. With an All Big Core design and Dual-NPU configuration, the chips lay the groundwork for next-generation agentic AI.
We checked the 21-patch v2 series that cuts Linux kernel build times by up to 36% against primary sources: how it removes serial work, the test conditions, third-party reruns, and where LLM and human responsibility divide.
Technical specifications have emerged for a 2.1D organic bridge jointly developed by Samsung Electro-Mechanics and Qualcomm. By eliminating the silicon interposer and connecting dies directly through organic RDL layers, the companies aim to sidestep Intel's EMIB patent network while cutting packaging costs for AI semiconductors.
Fujitsu will begin selling its domestically developed Arm CPU, FUJITSU-MONAKA, along with three server lines built around it, starting in November. Drawing on primary sources, this article examines the shipping schedule from April 2027 onward, the differences between air-cooled and liquid-cooled configurations, the claimed performance figures and CXL capabilities, and the guarantees—and limits—of Arm CCA.
