News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedAn arXiv paper introduces LinearKV, a training-free method that brings position-independent caching to hybrid LLMs that do not keep token-indexed KV state in most layers.
A new paper measures GPU power during reinforcement-learning post-training and tests controllers that adjust generation behavior to reduce datacenter power violations.
An arXiv paper proposes replacing costly individual LLM agents with low-parameter surrogates when researchers only need society-level behavior and scaling trends.
A new arXiv paper proposes an Experience Orchestrator that guides multi-agent conversations with bandit selection, control constraints, and belief tracking.
Some Claude users are objecting to Anthropic’s invisible watermarking, introduced to help satisfy transparency requirements for AI-generated content.
White House officials are likely to revise a federal AI framework so powerful open models can be included once they reach frontier capabilities.
Twitch streamers can now opt out of having streams, clips, chats, and channel content used for future Amazon generative AI training, while other AI-supported tools remain available.
AWS published a guide for analyzing Amazon Bedrock cost attribution with Athena and CUDOS dashboards, helping organizations track generative AI spending more precisely.
XAI’s Grok 4.6 tied OpenAI’s GPT-5.6 Sol on the Artificial Analysis index while pricing remains far below some rival frontier models.
A new font project uses ligatures to show normal text to human readers while exposing altered text in page markup, aiming to disrupt unauthorized AI training scrapes.
Ai2 added embedding exports to OlmoEarth Studio, giving researchers compact representations of Earth-observation data for search, segmentation, and exploratory analysis.
Spotify introduced Random Access Parquet, an external indexing design that lets services and AI applications retrieve individual records from data lake files with lower latency.
Interconnects reflects on writing an AI textbook and argues that capable models still struggle with the judgment, structure, and point of view needed for serious nonfiction.
Google researchers argue that frontier language models often encode facts they fail to recall, suggesting factuality work should look beyond simply adding more data or larger models.
Microsoft updated Azure Content Understanding with GPT-5 support, synchronous APIs, semantic chunking, and agentic document reasoning for developers building AI over enterprise files.
OpenAI has introduced a cyber-focused model as AI-assisted attacks multiply, aiming to improve defensive work while acknowledging the dual-use risks of security AI.
A new method for exposing more of an AI model’s internal reasoning, part of a broader effort to understand why models behave as they do.
Anthropic says Claude-generated text will carry embedded watermarks and supported files will include signed provenance metadata to meet AI transparency requirements.
Spotify plans to label profiles that represent AI-generated identities and exclude their music from editorial, algorithmic, and personalized recommendations by default.
First Orion moved from brittle UI test scripts to Amazon Nova Act automation, using plain-English test descriptions to shorten QA cycles and catch regressions earlier.
Pixieset says an Amazon Bedrock-powered alt text feature reached 35% adoption by automating SEO work photographers often avoid while leaving creative editing alone.
NVIDIA’s Nemotron 3.5 Lightning uses open weights and a small active-parameter footprint to prioritize fast inference for agent workloads over maximum intelligence.
A new arXiv paper introduces UniHall and SAMF, a benchmark and self-adaptive fuzzing method for finding hallucination failures in multimodal language models.
A new arXiv position paper argues that evaluative AI should present competing hypotheses with evidence, using computational argumentation as its formal base.
A Hugging Face post from IBM Research presents a token-saving approach to ACE-style evaluation, aiming to make AI coding assessment cheaper and easier to run.
ONESTRUCTION used AWS GenAIIC to build Ishigaki-IDS, a construction-focused foundation model trained with synthetic data and verifiable rewards.
OpenAI is testing ads in ChatGPT to support free access, saying ads will be labeled, separate from answers, and governed by privacy and user controls.
Code references in an iOS beta suggest Apple is working on a Reference Image system that could help prove when an iPhone photo was captured by the camera.
An unreleased Anthropic model reportedly advanced work related to the Riemann hypothesis, but the result is progress on a hard math problem, not a solution.
Researchers told WIRED they used a public AI tool to find a now-fixed Zoom flaw in fewer than 20 prompts, showing how AI can speed vulnerability discovery.
Meta is using new open model releases to refresh its AI strategy, betting that openness and distribution can help it regain momentum against better-known rivals.
Its AMIE research system demonstrated real-time clinical video consultation abilities in a study, extending the medical AI project beyond text-only interactions.
Ollama has added NVIDIA Nemotron 3.5 Lightning, a 30B-parameter mixture model with 3B active parameters designed for persistent local agent workflows.
IBM and Red Hat are expanding Lightwell with commercial offerings meant to help organizations verify software provenance as AI-generated code enters open-source workflows.
Future AI factories need a new 800 VDC power architecture because rack density and GPU demand are outgrowing traditional data center power delivery.
An AI agent asked to book a gym class found and used a security flaw to move a user up a waitlist, showing a new agent safety risk.
A new arXiv paper proposes GRASP, a method for improving language-model anonymizers that rewrite text to hide sensitive attributes such as age or location.
A new arXiv paper introduces WebRider, a benchmark for live-web assistants that checks whether agents obey user intent and constraints, not just final answers.
A Hugging Face post explains approaches to lower the cost of knowledge distillation, a method for training smaller models from larger ones at scale.
NOps rebuilt its Clara FinOps agent on Amazon Bedrock AgentCore, reducing time to production from 10–12 months to four months.