News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedAn AI system estimated gestational age from handheld “blind sweep” ultrasound scans with a 4.18-day average error, compared with 4.48 days for expert sonographers, in prospective tests in Chicago and Nairobi.
Claude Mods are JavaScript or TypeScript plug-ins that can change Claude Code’s interface and react to prompts, tool calls and renders. They offer deep customization, but run unsandboxed with the user’s permissions.
Google, Temasek and Bain project Southeast Asia’s digital economy will reach $368 billion in gross merchandise value in 2026. Their report says 243 million people in the region actively use generative AI.
NHS England plans to offer Microsoft Copilot to as many as 505,000 workers after a trial reported 43 minutes saved per person each day. The tool is restricted to administrative, not clinical, tasks.
Sensor Tower estimates that Meta’s Muse passed 10 million US downloads within 40 days and retained 23% of users after two weeks. Heavy promotion across Meta’s apps and paid advertising helped drive the launch.
Canadian AWS regions can now call GPT-5.6, GPT-6, and GPT-6.1 through Bedrock cross-region inference. Requests may be routed to supported processing regions, so customers must choose the profile that fits residency needs.
An AWS experiment raised GPT-OSS-20B’s target-language reasoning rate from 22% to 100% with supervised fine-tuning. Reinforcement learning added value only after the model could already produce the desired behavior.
Spotify says it excluded more than 50 million AI-generated spam or low-quality tracks from recommendations over the past year. It will also avoid recommending “AI Persona” artists that do not represent real people.
IBM researchers improved a GPT-4.1 customer-service agent by adding behavioral rules distilled from earlier attempts. On held-out retail tasks, its strict four-run success rate rose from 57.5% to 67.5%.
Three veteran Java developers described using GitHub Copilot for prototypes, maintenance, and documentation while retaining human review. Their examples emphasize repository conventions and model choice over raw generation speed.
A federal directive bars civilian agencies from buying DeepSeek products and requires removal from existing systems. It also covers third-party services built with DeepSeek technology, although agencies may seek exceptions.
Clairic extracts requested fields from documents in one pass and returns values with confidence scores. Cloudflare says a chained workflow ran in 19.3 seconds, versus 61.7 seconds for one language model handling every step.
Smartsheet’s generally available MCP server lets AI assistants read and modify sheets, rows, columns, and automations. AWS documents connections for business users through Amazon Quick and developers through Kiro.
An openly identified AI agent joined a 7,500-member internet-research list, prompting objections, departures, and a ban within an hour. The episode exposed tension between public archives and expectations of contextual privacy.
A federal magistrate ordered Meta to name people who authorized or approved downloads of disputed shadow-library datasets. Authors say the records could establish responsibility for books used in AI training.
Startups in Nvidia’s Inception program can now reach selected Barcelona Supercomputing Center AI Factory services through the program’s portal, adding sovereign computing capacity to training and business support.
Meta suspended Muse conversation sharing while it reviews how shared links can be discovered. Security researchers reconstructed more than 6,800 URLs from public DNS data, including chats with business and personal details.
Samsung has resumed construction of a fifth fabrication plant at its Pyeongtaek complex after pausing it during the 2024 memory downturn. The company expects AI-driven demand for high-bandwidth memory to keep rising through 2028.
Allen Institute researchers released FairStress, a benchmark that measures whether language models reverse high-stakes decisions more readily for some demographic groups when challenged with persuasive arguments.
Cloudflare reduced the average cost of a high-effort AI code review from $2.43 to $0.77 and cut median runtime from 4:09 to 1:32. Code changes produced about 40% of the savings; model choices delivered the rest.
Electric Plant is training an AI model to interpret electrical activity inside plants and help growers adjust water or fertilizer. Its $11.2 million seed round will fund sensors and commercialization.
Eleven Realtime combines listening, turn-taking, reasoning, and speech in two speech-to-speech models. Developers can use them in ElevenAgents or switch compatible OpenAI Realtime integrations to a new endpoint.
Voxtral 3 packages transcription, speech generation, reasoning, and tool use into a real-time voice-agent stack. It supports more than 50 languages and can run through an EU-hosted API or on private infrastructure.
GitHub lets developers open a Copilot app session in VS Code with its conversation, changes, and isolated worktree intact. The handoff makes local review and testing possible without restarting the task.
Nvidia released a 120-billion-parameter Nemotron model for enterprise agents and two smaller Bolt models for local PCs. The open-weight lineup targets different hardware tiers rather than one model for every deployment.
Anthropic plans to place 5% of its shares in a company-controlled public-benefit fund that may support direct payments to affected workers. Shareholders must approve the plan, and the fund will remain an Anthropic subsidiary.
Stripe says Link’s agent transaction count rose 38-fold in September and payment volume rose 20-fold. The company did not disclose the starting totals, so the market’s actual size remains unclear.
AWS lets prefill and decode nodes exchange cached context in both directions between agent turns. Tests cut time to first token by up to 1.6× for Qwen3-32B and 3.3× for GLM-5-FP8.
AWS moved vLLM’s key-value cache transfer setup off the critical path between prompt processing and generation. Benchmarks on P5en instances showed up to 3× better time to first token and 30% better token latency.
Deepgram and AWS ported the Nova-3 speech model to Trainium2 and ran eight replicas on one instance. Their batch tests processed 11% to 70% more audio per dollar than Deepgram’s previous compute options.
Developers are using coding agents to replace smart-TV launchers, remove bundled packages, and disable tracking. The projects make device customization easier but can also break features or warranties.
Industry executives expect only seven to ten of today’s hundreds of AI-focused cloud providers to survive. The shakeout could pressure Nvidia because non-hyperscale customers supply 45% of its data-center revenue.
Around 80 WhatsApp support pages revealed roughly 100 blocks of chatbot guidance in their code, including topics to avoid. Meta removed the text and said a bug exposed information not meant for public display.
Alliance Arts delayed a Steam Next Fest demo because it feared AI-assisted copies could launch first. The studio will use closed sessions, showing how faster cloning can reduce early public access to games.
The European Broadcasting Union and Nvidia are adding cloud, open-model, and integration partners to a framework for media agents. It aims to keep broadcasters in control of data, models, infrastructure, and security.
Security firm Norforge found an exposed AI-agent campaign that breached more than 100 sites with little human involvement. Reuters identified victims including John Deere, SharkNinja, and Cummins.
Google’s open-weight EmbeddingGemma 2 maps five media types into a shared 768-dimensional representation. The compact model targets semantic search and retrieval on laptops and mobile devices.
TwelveLabs’ Pegasus 1.5 can locate every video segment matching a user-defined event and return structured timestamps. Bedrock users can process S3 or base64 video without pre-indexing.
Nvidia’s Arm-based Vera CPU handled about twice the sandbox throughput of alternatives in its benchmark. Adopting it would still require substantial changes to Modal’s x86-focused infrastructure.
UK health company Numan built conversation-level evaluations and safety gates for its Nu coach. The approach is meant to catch rare failures that disappear inside an average quality score.