Large Language Models

Today's Stories

1. 1

   Meta's Muse AI pulls Amazon data, unnerves tester

   Meta launched an AI assistant called Muse this week. In The Verge's Emma Roth's testing, Muse linked to her Gmail and Amazon and pulled her family-visit state from an Amazon shipping address. Muse asks users to hand over deep access to email and payment accounts, at a company whose privacy track record the article calls catastrophic. Roth said the experience left her uneasy about credit card and email access.

   The value proposition hinges on whether users trust Meta enough to link accounts. Watch whether Meta's claim of a secure "dedicated virtual machine" for login credentials holds up under scrutiny.

2. 2

   Salesforce unveils Enterprise AI Harness with AI Control Plane

   Salesforce introduced an Enterprise AI Harness built on six capabilities — context, agency, action, governance, security and models — plus a new AI Control Plane. Instead of each agent needing its own separate stack, enterprises can now use one composable system across Salesforce and third-party models and agents.

   New capabilities and a unified experience are planned to begin rolling out in early fiscal FY28. Specifics on availability, pricing and upgrade paths will come closer to general availability.

3. 3

   John Deere's JD AI assistant enters limited access

   John Deere introduced JD, a generative AI assistant on its Operations Center platform, currently in limited access. It analyzes harvest timing, fuel usage and sprayer operator performance via web or mobile. Sixth-generation farmer Kelly Garrett said his planting window should have been narrower — "we just didn't know" — a gap he credits data with revealing on his 9,000-acre farm.

   Whether JD moves from limited access to integration with the G5 touchscreen display inside equipment cabs, and whether John Deere adds agentic capabilities, which CTO Jahmy Hindman said may be possible.

4. 4

   LINE Yahoo adds Task Function to AI agent Agent i

   LINE Yahoo began offering a "Task Function" for its AI agent Agent i on September 11, 2026. Users register keywords, and Agent i collects the latest information, price drops, and event news, then notifies them at their chosen timing. Agent i shifts from an AI that answers questions to an AI that acts, executing everyday tasks autonomously and continuously. Notifications reach the Agent i home screen, plus push alerts through the Yahoo!JAPAN and LINE apps.

   LINE Yahoo says it plans to expand the task range to store reservations and product purchases, so the test is whether Agent i can move from information gathering to completing transactions. The company says registrable tasks will be expanded gradually.

5. 5

   KDDI's Hojo: AI agent adoption stuck at 7.1％ without "AI-ready" systems

   KDDI AI integration head Hiroki Hojo said AI delivers no value unless built into frontline work, and that autonomous cross-system AI agent adoption stands at just 7.1%. Many firms moved to the cloud but kept legacy application and data structures, so AI agents cannot reach the data they need — the barrier Hojo ties to that low adoption figure.

   Closing the gap hinges on consolidating siloed data and modernizing via standard interfaces like APIs and Model Context Protocol so AI can read and write safely.

6. 6

   NEC creates all-AI 無人部署 for management analysis

   NEC has established a new 無人部署 (unmanned department) where every role, from department head to staff, is filled by AI that autonomously carries out management analysis and document creation. This marks a shift from AI used as a support tool to AI operating as an organizational unit on its own, with NEC positioning AI as the one doing the work rather than assisting people.

   The test is whether an AI-run department can keep decision-making quality high and output trustworthy. Watch for details on how NEC uses the results in its actual management.