AI news5
xAI says Grok V9-Medium has finished training, with a public release weeks away
Elon Musk said xAI has completed training of Grok V9-Medium, a 1.5 trillion-parameter model, roughly three times the scale of the previous V8. Supervised fine-tuning is underway and reinforcement learning starts within days, with a public release expected in two to three weeks. Musk said a large amount of Cursor coding data was added in supplementary training, signalling a heavy bet on developer and coding use.
techloy.com ↗Why it matters
it keeps the frontier-model race moving and points xAI squarely at the coding market where Anthropic and OpenAI currently lead.
Anthropic is in talks to run Claude on Microsoft's custom Maia 200 chips
Anthropic is in early-stage talks with Microsoft to rent Azure servers powered by Microsoft's custom Maia 200 AI accelerator, according to CNBC. If the deal closes, Anthropic would be the first major external customer for Maia, which Microsoft says delivers more than 30 percent better tokens per dollar than its current fleet. It would add a fourth silicon supplier alongside Anthropic's AWS Trainium, Google, and NVIDIA arrangements.
cnbc.com ↗Why it matters
it shows frontier labs deliberately spreading across chip vendors so no single supplier controls their cost base.
Fujitsu builds AI agents that tune themselves as the work changes
Fujitsu has developed a multi-agent system whose agents identify why tasks succeed or fail, extract the lessons, and apply only the changes that test out, taking over work like prompt tuning and evaluation updates that experts used to do by hand. In testing it lifted the accuracy of Takane, Fujitsu's business-specialised model, by an average of 28 points. Fujitsu is working with Carnegie Mellon researchers to shrink the system's memory and power needs for cloud, on-premises, and edge use.
global.fujitsu ↗Why it matters
self-tuning agents could cut the ongoing human maintenance that quietly makes most enterprise AI expensive to run.
Microsoft Research open-sources Webwright, a web agent that writes its own code
Microsoft Research has released Webwright, an open-source web agent that, instead of guessing one browser click at a time, writes Playwright code and bash commands and saves reusable scripts. It scores 60.1 percent on the Odysseys benchmark, up from 33.5 percent for the base GPT-5.4 model it runs on. The codebase is deliberately small, around 1,000 lines, with no multi-agent orchestration.
microsoft.com ↗Why it matters
turning web automation into inspectable, rerunnable code makes agent behaviour easier to debug and trust than opaque click-prediction.
Kore.ai launches Artemis, an enterprise platform for governing AI agents
Kore.ai has launched Artemis, a rebuilt version of its agent platform aimed at letting enterprises ship production multi-agent systems in days with governance and oversight enforced before any agent goes live. It introduces a compiled Agent Blueprint Language for defining and validating agents, plus an AI agent architect and a design that mixes reasoning with fixed workflows. It launches first on Microsoft Azure, with AWS and Google Cloud versions due by the end of 2026.
businesswire.com ↗Why it matters
governance and observability, not raw capability, are becoming the gating factor for putting agents into real operations.
AI in the nonprofit sector5
A community health network's AI now handles 80 percent of patient calls
Capital Area Health Network, a nonprofit federally qualified health centre with 34 providers, has deployed an AI contact centre called healow Genie connected to its eClinicalWorks records. The system handles about 80 percent of total call volume, answering routine questions, medication refills, and appointment queries around the clock, and has removed wait times that used to reach 45 minutes. Staff are freed to focus on urgent and complex needs.
marketminute.com ↗Why it matters
phone access is a real barrier to care for low-income patients, and this is a working fix rather than a pilot.
A nonprofit aging agency uses a companion robot to support isolated seniors
The Olympia Area Agency on Aging, a Washington nonprofit, has been deploying ElliQ, an AI companion robot from Intuition Robotics, to support seniors living alone, working with Pacific County Fire District No. 1. Over about three years the pilot has grown to more than 20 people across Grays Harbor, Pacific, Clallam, and Jefferson counties. The robot provides daily check-ins and company for remote, hard-to-reach clients.
chinookobserver.com ↗Why it matters
rural aging services struggle to reach isolated clients, and low-cost companion AI extends a thin workforce.
Indian nonprofit builds a multilingual AI assistant for tribal communities
The Magicterrain Foundation, a nonprofit working across ecology, indigenous knowledge, and rural livelihoods, has entered the final community-validation phase for Mahul Vani AI, a multilingual platform for tribal communities, built with Centurion University. Developed over two years of field research, it gives advice on crops and local knowledge in community languages. The design centres indigenous knowledge rather than imposing outside content.
timesofindia.indiatimes.com ↗Why it matters
most AI assistants ignore tribal languages and contexts entirely, so a community-validated tool is a rare fit-for-purpose example.
Rotherham NHS trust cuts IT helpdesk calls 28 percent with an AI agent
Rotherham NHS Foundation Trust has cut IT helpdesk call volumes by 28 percent in about three months after deploying an autonomous AI agent through Netcall's Liberty Converse, with a pilot that started on 28 January. Around 41 percent of IT queries are now handled by self-service or AI, which creates pre-categorised tickets and routes complex issues to specialists. A second phase will target out-of-hours support.
resultsense.com ↗Why it matters
freeing internal support capacity inside a stretched health system returns staff time to clinical and frontline work.
Brazil's Landless Workers' Movement builds a collective AI for agroecology
Brazil's Landless Workers' Movement (MST) and the World March of Women have developed IARAA, an AI for agrarian reform and agroecology built around collective, community use rather than individual queries. The project frames AI as shared infrastructure to spread agroecological practice and counter the environmental costs of industrial agribusiness. It is positioned as a tool for technological sovereignty in the Global South.
capiremov.org ↗Why it matters
it is a rare example of a grassroots movement building AI on its own terms instead of adopting a vendor's.