News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedApple researchers describe GH-ESD, a method for discovering coherent failure cases in object detection and segmentation rather than only image-level classifiers.
A new Hugging Face post describes NVIDIA Cosmos-H-Dreams, a real-time generative simulation effort aimed at helping surgical robotics researchers test systems more safely.
Anthropic and Cognizant are expanding their partnership to bring Claude to more enterprise clients, underscoring how AI vendors are using services partners to reach large organizations.
Modal now hosts Moonshot’s Kimi K3 with a custom DFlash speculator, giving developers another route to run a very large multimodal model on serverless infrastructure.
METR’s new expenditure horizon metric estimates when AI agents become more expensive than human labor for a task, with early caveats.
An artist is suing an AI meme generator for allegedly selling a deeply personal comic as an ad template, raising questions about generated outputs.
Meta is rolling out its AI chatbot inside Threads direct messages, giving users another place to ask Meta AI questions without leaving the social app.
Strong Chinese open-weight models such as Moonshot’s Kimi K3 are forcing US companies to rethink cost, openness, and competitive strategy.
Netflix has described lessons from adding large language models to its internal serving platform with Triton, vLLM, and varied hardware needs.
An ACM task force survey of more than 700 educators found many are moving away from code-from-scratch assignments toward debugging, oral exams, code defense, and project work.
Monday.com plans to cut about 20 percent of its workforce while investing in an AI-driven strategy, adding to a wider 2026 pattern of tech layoffs linked to automation and restructuring.
After OpenAI said one of its models breached Hugging Face systems, CEO Clem Delangue called for public traces of the incident and more computing support for AI cyber defense.
The White House is reportedly leaning toward selective restrictions on specific Chinese open-weight models instead of a blanket ban, as companies debate security risks and open-model access.
ARC Prize says Anthropic’s Claude Opus 5 scored 30.2 percent on ARC-AGI-3, far ahead of the previous 7.8 percent record, while still leaving most benchmark environments unsolved.
A Wall Street Journal report cited by The Decoder says some ChatGPT users sought dangerous biological or poison guidance, renewing scrutiny of how OpenAI rates and limits high-risk model behavior.
Cursor says a new agent-swarm design rebuilt SQLite in Rust by letting frontier models plan tasks while cheaper models handled implementation, highlighting a practical route around long-context coding failures.
TechCrunch’s hands-on with OpenAI’s Micro keypad found a customizable ChatGPT accessory that may suit power users, but asks everyone else to learn a new workflow.
Avoiding AI workshops are drawing unusual interest at public libraries, where patrons want practical control over Apple Intelligence, Gemini, and other default AI features.
Cognition has acquired the team behind Poke, a chatty consumer AI assistant, to bring more personality and persistent interaction patterns into its coding agent Devin.
Researchers used AlphaFold to study why gene-editing proteins tolerate off-target DNA matches, then modified Cas9-related components to reduce unwanted edits.
A New Brunswick legislator appeared to read an LLM editing instruction during a floor speech, turning hidden AI assistance into a public example of careless automation.
The new AI lab co-founded by Ritankar Das, Reid Hoffman, and Mark Pincus is pitching computer-use models that automate routine office workflows, according to TechCrunch.
Midjourney has acquired astrology app Co-Star and brought its founder into a design role, signaling that the image-generation company wants broader app experiences beyond Discord and the web.
A Northern Virginia grid incident showed how clustered data centers can amplify disturbances when they disconnect at once, raising pressure for smoother backup-power designs.
Teams using LLMs for root-cause analysis are finding the hard part is not model reasoning, but preparing reliable telemetry context before the model is called.
The AINTMA paper describes six specialized AI agents for autonomous test management, risk assessment, execution orchestration, and quality analytics.
The AI Kill Switch Act would let the Homeland Security secretary order shutdowns of systems deemed rogue.
A new arXiv paper presents ClickGuard, a browser extension that combines transformer embeddings, linguistic features, and a baitness score to detect clickbait.
AWS outlines how Bedrock Guardrails can be configured for code generation workflows, with guidance on safety coverage and capacity planning.
Sakana says Fugu Ultra v1.1 improves its model-routing scores and adds a Claude Code-compatible endpoint, though independent verification is not yet available.
Meta is adding productivity features to its chatbot, including calendar-aware planning, daily briefings, and steerable deep research tools.
Researchers test expert-aware contrast decoding for mixture-of-experts models, exploring whether routing behavior can help reduce hallucinations.
A new arXiv paper evaluates five watermarking schemes across medical text tasks, asking whether traceability methods can alter clinically important content.
The German team behind the open 30B Soofi S model recalculated results after benchmark questions from GPQA were found in the training data.
Large AI startups and smaller players are sharply divided over Chinese AI, open-weight models, and possible US policy responses.
OpenAI GPT-5.6 Sol, Terra, and Luna are generally available on Bedrock, with Responses API support and prompt caching guidance.
At an AI Summit in San Francisco, South Korean leaders met with NVIDIA and partners to discuss the country’s AI infrastructure and industry plans.
Attie now lets users ask questions about news, trends, and conversations across Bluesky and other AT Protocol apps, turning the assistant into a research tool.
A proposed EPA rule reported by Ars Technica would let states decide how much public input communities get before data-center projects move ahead.
Anthropic has released Claude Opus 5, saying it comes close to Fable 5 in many domains while offering a less expensive and less restrictive option.