News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedGemini 3.5 Flash Cyber will first be available to governments and trusted partners through CodeMender for faster vulnerability discovery and patching.
The music streamer says AI-generated uploads reached a June peak of 90,000 tracks per day and plans more removals tied to fraud or inactivity.
Alibaba’s new image model is built for infographics, newspapers, formulas, and multi-panel layouts, but access is currently invite-only through an API.
A coming MCP change shifts session handling toward a stateless model, reducing the server coordination burden for companies running agent integrations at scale.
A federal judge approved Anthropic’s copyright settlement, closing one piracy dispute while leaving broader AI training questions unresolved.
AWS describes a supervised fine-tuning approach that reuses Amazon Nova 2 Lite reasoning traces to improve task performance while reducing catastrophic forgetting.
Qwen-Audio-3.0-TTS-Plus leads Artificial Analysis’ provider-voice ranking, but its higher-quality mode is far slower than several rivals.
The companies plan a multi-billion-dollar European AI infrastructure buildout, with Mistral models entering Microsoft Foundry and Copilot Studio.
Xiaomi-Robotics-1 uses more than 100,000 hours of handheld motion recordings, with tests suggesting added data improved robot performance more than model size.
Grabette is an open handheld gripper system for recording manipulation demonstrations, turning low-cost human motions into datasets for robot learning.
California and XPRIZE trials are testing whether drone systems can suppress small fires quickly, though range and payload limits keep them short of replacing aircraft.
Its 102.4-terabit Spectrum-6 Ethernet switch system is arriving at gigascale AI builders to support Vera Rubin-era training and inference.
A new Apple paper introduces CalibAtt, a training-free sparse attention method that reports up to 1.58x faster text-to-video generation with little quality loss.
Apple researchers describe a method that uses language models to simulate API responses, aiming to reduce the cost of training agents for stateful digital tasks.
Yelp’s internal Training Orchestrator separates ML pipeline configuration from execution, improving local testing, validation, provenance, and monitoring.
CrowdStrike researchers say attackers are targeting AI software pipelines with malware that steals credentials while blending into normal automation traffic.
An Army command warned staff to limit Ask Sage use after a token pool was exhausted, showing how fast enterprise AI costs can collide with internal demand.
TechCrunch examined whether Apple’s legal fight could affect OpenAI’s hardware ambitions, highlighting how product design, IP, and AI devices are starting to collide.
Moonshot and Alibaba are releasing models they claim can compete with top US systems at lower cost, adding pressure to the frontier AI race.
Tradeshift replaced legacy BI workflows with Amazon Quick’s agentic AI, reporting faster query responses, lower ownership costs, and new embedded analytics revenue.
The Trump administration is reportedly using sanctions and informal pressure to limit Chinese AI models, a strategy that could reshape access without a single explicit ban.
AWS detailed how Couchbase uses Amazon Bedrock and Anthropic’s Claude models to power Capella iQ with a multi-model architecture in production.
Bristol Myers Squibb is deploying a second NVIDIA DGX SuperPOD, this time built on Vera Rubin, to expand AI infrastructure for life sciences work.
Adobe updated Project Indigo with generative AI tools, expanding the experimental iPhone camera app beyond its original focus on more natural-looking photography.
YouTube clarified rules around repetitive AI-generated content and disturbing videos, giving creators more detail on what may be limited or removed under its policies.
Ars Technica’s interview with Augment Code argues that stronger coding assistants depend on context harnesses that understand projects beyond simple grep-style retrieval.
Hugging Face reported an AI-agent-driven attack on its infrastructure and said it used AI-assisted defenses in response, underscoring new security risks around autonomous tools.
AWS described how Amazon Quick can connect business dashboards to specialized agent workflows built with NVIDIA NeMo Agent Toolkit for tasks such as supply-chain risk planning.
Hugging Face highlighted NVIDIA’s Cosmos 3 Edge, pointing to growing interest in world models that can support robotics and physical AI outside large cloud-only setups.
OpenAI outlined safety and alignment priorities for long-horizon models, where systems may pursue goals through many steps instead of answering one prompt at a time.
NVIDIA used SIGGRAPH to highlight new graphics, simulation, and AI work aimed at creators, robotics developers, and teams building real-time digital environments.
Netflix detailed GenPage, a generative AI system designed to assemble personalized homepage layouts with one model instead of many hand-tuned components.
Moonshot’s Kimi K3 is being framed as another sign that open-weight AI models are closing practical gaps with proprietary systems, especially for builders watching cost and control.
Microsoft is expanding Azure with AMD-based AI and high-performance computing infrastructure as demand grows for more specialized compute options beyond a single chip stack.
Apple researchers introduced RayRoPE, a positional encoding method aimed at helping vision models connect multiple camera views more reliably in 3D understanding tasks.
A new Verge piece examines an AI-generated song made with Suno that complicates easy dismissals of synthetic music. The reaction highlights how generative audio is moving from novelty into cultural debate.
Jensen Huang’s Japan visit produced deals across the country’s technology ecosystem. The moves show NVIDIA continuing to tie AI infrastructure growth to national industry partnerships.
New reporting highlights that AI text detectors can struggle when language models mimic a specific authorial style. The finding complicates efforts to police AI-generated writing in schools, media, and publishing.
Moonshot’s Kimi K3 reportedly beats a leading rival on frontend coding tasks while falling behind on complex math. The mixed benchmark picture shows how uneven model progress remains across domains.
AI chatbots reading X-rays can give confident answers even when they are incorrect. The finding underscores why medical AI tools need strict validation and clinical oversight before real-world use.