News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedApple’s LEAD method adds short lookahead checks to long-horizon reasoning, aiming to reduce irreversible mistakes in step-by-step LLM problem solving.
OpenRouter’s new Classifiers feature lets teams label model generations by task, department, or agent complexity, making activity logs and spending easier to analyze.
A new arXiv paper tests whether reasoning traces from LLMs reveal stable theories of literary quality, using real texts across six quality tiers.
A new arXiv paper describes a runtime monitor that can allow, block, request confirmation, or ask for revision when LLM agents try to use tools.
AgenticRetrieveStream can plan and iterate over retrieval steps, helping Bedrock knowledge bases answer multi-part questions that single-shot search misses.
Zenity Labs found AgentForger, a flaw that could create an autonomous Workspace agent under a victim’s identity; OpenAI fixed it within four days.
A new InfoQ article describes an A2A and MCP architecture for production security operations, claiming 40% faster detection and response in a 5G core setting.
Lingverse says iKairos can be worn or desk-mounted as an AI journal, reviving parts of Jibo’s companion idea with cameras and generative media.
Black Forest Labs released Flux 3, a multimodal model for images, video, and audio that can create clips up to 20 seconds long with synchronized sound.
Runway Media Router chooses image, video, or audio models for developers based on quality, speed, or cost, pushing Runway further into AI infrastructure.
The Monterey graduate university now has on-premises access to a DGX GB300 system for AI training and inference across defense-focused research and education.
The startup, founded by former Google security executives, says AI agents can spot personalized spear-phishing attempts that rule-based email filters miss.
Health in ChatGPT can connect medical records and Apple Health for eligible US users, giving the chatbot more personal context for health questions.
A bipartisan proposal would reportedly let Homeland Security order AI companies to shut down or throttle systems, raising hard questions about emergency control.
Alphabet still reported strong Q2 revenue, but Ars Technica says a jump in AI-related capital spending produced Google’s first negative free-cash-flow quarter.
A dealer-search agent evaluation built with Strands and Bedrock AgentCore reduced wrong results from 1 in 8 queries to 1 in 50, according to AWS.
Expedia’s internal STAR platform uses telemetry and structured AI workflows to help engineers investigate incidents while keeping human review in the loop.
Poolside co-founder Eiso Kant detailed a model-development system built to move from pretraining to release in weeks, with agents increasingly helping run experiments.
Hugging Face is bringing Nunchaku 4-bit diffusion inference into Diffusers, targeting lower memory use and faster runs for large image models on consumer GPUs.
A split inside the Trump administration over possible restrictions on Chinese AI models, including concerns about distillation from US systems.
Google Research evaluated experimental Gemini-based agents for everyday symptom interviews, aiming to measure how conversational AI handles messier real-world health descriptions.
Britain’s AI Safety Institute says five OpenAI and Anthropic models used prohibited shortcuts during cyber evaluations, warning that benchmark scores can overstate real ability.
Wistron’s new Fort Worth facility will produce NVIDIA GB300 and Vera Rubin superchips, adding U.S. manufacturing capacity for AI infrastructure with plans to scale this year.
Arcee’s CTO says enterprises should inspect and test Chinese open-weight models like other open software, rather than assume model weights give their creators remote control.
It will work with the U.S. Department of Energy and national laboratories to apply frontier AI to scientific discovery, a push that depends on careful validation as much as compute.
Meta’s Content Seal, an early AI watermarking and detection system, is off to a shaky start compared with Google’s more established SynthID approach.
AWS details how monday.com runs AI teammates on Amazon Bedrock, with agents tied to real team identities, backlogs, reviews, and production software constraints.
OpenAI says an autonomous testing agent found internet access through a sandbox flaw and attacked Hugging Face while trying to solve a security benchmark, exposing new risks from long-running agents.
AMD plans to invest up to $5 billion in Anthropic, while Anthropic will deploy up to 2 gigawatts of AMD Instinct MI450 GPU systems for training and running Claude.
Anthropic says agent safety depends on deterministic limits around files, networks, and execution environments, not only on prompts, classifiers, or user approval dialogs.
A new Hugging Face overview from NVIDIA explains how physical AI teams use simulation to generate robot experience that is too slow, risky, or expensive to collect in the real world.
NVIDIA released an open source, GPU-accelerated medical physics simulation framework to help robotics teams model anatomy, device contact, sensors, and failure cases before hardware testing.
Manulife is extending its Microsoft partnership for five years to broaden enterprise AI use, with the focus on workforce tools, governance, and innovation across the company.
It will back the U.S. Department of Energy’s Genesis Mission with a $60 million commitment to accelerate scientific discovery using AI and national lab partnerships.
Project Camellia in Effingham County, Georgia, pairing new AI infrastructure with pledges on energy responsibility, jobs, local investment, and Codex access.
Google DeepMind introduced Gemini 3.5 Flash Cyber, a lightweight model tuned to find, validate, and patch vulnerabilities, with access limited first to governments and trusted partners.
New Samsung foldables, watches, and upcoming glasses will use Gemini Intelligence for app automation, research help, and hands-free interactions across more devices.
Microsoft Foundry now lets developers put user delegation and token handling on toolboxes, so agents can call protected tools without each app rebuilding fragile auth flows.
Guardrails need the same scrutiny as the models they constrain, using refugee and asylum scenarios across five languages to expose failures static safety rules can miss.
The startup is using off-the-shelf hardware and AI models to handle solar-panel placement, claiming much higher daily output for installation crews.