Topic
Explore English articles about Generative AI.
Ukraine's military intelligence agency has discovered an NVIDIA Jetson Orin NX inside Russia's new S-71M autonomous cruise missile. This $999 consumer-grade edge AI chip has been repurposed for terminal guidance, concretely demonstrating how export controls against Russia are structurally failing.
Anthropic will roll out statistical watermarking to future Claude models worldwide. This article explains, based on primary sources, Google's empirical measurements, the gap in Claude-specific accuracy data, detection limits that weaken with short text and editing, and compliance with EU law.
At its Investor Day, SanDisk presented HBF as the savior of AI inference, but the tech community fired back with three concrete rebuttals: understated HBM bandwidth, outdated quantization assumptions, and omitted write endurance figures. This article sorts out HBF's true value and its limitations.
Z.ai improved GLM-5.3's coding performance and offensive capability through additional training on the same base model alone. The API and weights are not yet available, with weight release after safety evaluation marking the next checkpoint to watch.
PTP from IIT Bombay and Adobe Research achieved a 64.77% exact match rate in reconstructing prompts within the Qwen environment. The rate drops to 11.36% when transferred to GPT-4o, indicating limited applicability.
X has significantly expanded its For You recommendation repository, publishing everything from ranking to visibility filtering. Users can now check the previous month's restriction labels via JSON, though some rules and production experiments remain excluded.
NVIDIA's NeMo Switchyard curbs calls to expensive LLMs during managed evaluation, cutting costs by 74%. Given the accuracy drop and pre-alpha status, we examine the conditions under which enterprise adoption actually pays off.
Google boosted Tensor G6's TPU compute resources by 50%. The claimed up-to-3.5x on-device AI figure includes Gemini Nano, and real-world performance and SDK support await post-launch verification.
DualLane, developed by Alibaba Cloud, is an AI agent that processes routine cases through verified templates while reserving LLM-based planning only for unknown cases. This design suppresses errors and delays that come with inference, achieving both high accuracy and execution speed in production.
NVIDIA has released Nemotron 3.5 Lightning, which activates 3B out of a 30B total parameter count. The design hands planning to large models while routing repetitive execution to a fast, small model.
