News
Short posts prepared from approved sources with clear links back to the original articles.
RSS feedIBM’s Granite time-series models are in early access on Confluent Cloud, where they can forecast and flag anomalies without moving live data to a separate ML system. On-premises support is planned.
Google’s MAPL-EMIT model finds methane plumes and estimates their sources in hyperspectral satellite data. The team reports 84% recall on expert-labeled plumes and has released a global database.
REFACTOR-VLA turns action sequences into reusable, typed motor programs for longer robot tasks. Tests also found that a larger world model alone made performance worse across every benchmark suite.
CrowdStrike’s SafeMind pairs offensive and defensive agents so each side can improve the other. The system uses Nvidia Nemotron models and will ship inside the Falcon security platform.
Power, memory and networking limits make raw AI capacity a poor measure of progress. It wants the industry to track how much useful work each layer actually produces.
Amazon's Update Me When feature sends personalized notices about new products, books, tours and shows, extending its shopping assistant from answers toward proactive alerts.
Andrew Bailey says high valuations, borrowed-money investing and cross-holdings among AI firms and cloud providers could spread losses beyond a single company.
Sonos 27 adds conversational music requests, custom automations and connection diagnostics, but some older S1-controlled devices will miss the new features.
The generally available service gives companies one searchable record for agents, tools, skills and MCP servers, with ownership, versioning and approval controls.
Projects including Flue and tldraw now close external pull requests, while Vercel and Astro use in-house agent pipelines to triage issues, reproduce bugs and propose fixes.
The free beta connects household communication channels to produce checklists and upcoming-event summaries, with broader app integrations and paid plans expected later.
The startup discovers workplace agents, checks their skills, plugins and MCP servers for changes, and blocks components that do not meet company security rules.
Nvidia's generative graphics system debuts in NBA 2K27 on September 3, but official support is limited and the company estimates a 50% to 60% performance hit.
AI Overviews had advised emergency action for queries about being alone with some nationalities while offering small talk for a British person; Google says it is fixing the issue.
The research system renders a 720p interface frame by frame and responds to clicks, drags or speech, but unreliable text and missing screen-reader support block practical use.
The x402-secure service scores paid endpoints before settlement and combines that gate with isolated credentials and session budgets for autonomous micropayments.
A serverless system measures per-user AI spending every 15 minutes, removes access to expensive models at set thresholds and restores it after the daily reset.
Across 2,005 guideline and cancer-case decisions, none of nine tested models answered 42.1% correctly, challenging the idea that one model can safely guide care alone.
A 300-task benchmark across ten languages found the best of six frontier models solved 63.1%, suggesting English coding scores do not capture multilingual performance.
The Sequoia-incubated startup tracks changes and their dependencies, allowing low-risk updates while escalating potentially dangerous ones for human review.
The JD assistant uses field, machine and operations data to answer questions about equipment, fuel and harvest timing, with an early test limited to selected US customers.
AlgorithmWatch's 4,480-query study found inconsistent election coverage, concentrated citations and partisan-sounding descriptions, though clear factual errors were rare.
HashiCorp's proposed workflow lets agents draft and test infrastructure changes while HCP Terraform retains approval, identity, policy and audit controls.
Google's new image tool can isolate objects, edit or translate text inside graphics and generate multiple variations, with access rolling out to paid AI and Workspace customers.
A new open-source kernel library lets developers load optimized machine-learning operations from the Hugging Face Hub, while a browser benchmark gathers evidence across real GPUs.
Glassdoor found that 98 percent of AI-related reviews from claims adjusters were negative. Workers describe misclassified cases and faulty summaries that they must correct while entry-level hiring contracts.
Washington is restricting foreign drones and advanced robots, but Chinese manufacturers hold large cost and volume advantages. Analysts expect competition to shift toward regional markets and allied supply chains.
Circleback’s new free plan includes unlimited transcription, transcript search and mobile access but keeps only 30 days of history. Paid plans begin at $15 monthly when billed annually.
MediaTek will use Nvidia’s NVLink Fusion technology to connect custom accelerators with Nvidia-based data centers. The deal lets Nvidia stay in the infrastructure stack as cloud companies build alternatives to its GPUs.
Clipto raised an all-equity round at a $250 million valuation after reaching $15 million in annual recurring revenue. Its software searches local video, audio, images and documents and exposes selected results to AI tools.
Instagram is renaming its profile label and reducing recommendations for accounts that present an AI-generated person as real. Ordinary accounts that merely post AI-made content are outside the new rule.
Debian developers rejected a ban on generative AI and will apply existing quality and legal standards to assisted work. Disclosure is encouraged rather than required, and contributors must review and test output.
OpenAI and other labs are reportedly using large Mac fleets to train agents on desktop tasks. Strong demand and a memory shortage have left top configurations scarce, while new tools cluster Macs for local AI.
OpenClaw’s largest release simplifies setup, lets teams collaborate in live agent sessions and can run work on local, paired or rented machines. The open-source update incorporates more than 16,000 pull requests.
Some OpenAI customers can reportedly pay only when an AI completes an agreed task. The model aligns fees with results but creates disputes over whether software, staff or other changes produced the outcome.
CXMT is reportedly producing small quantities of HBM3E, the stacked memory used beside AI processors. Low yields and a one-generation gap mean China’s advanced-memory constraint is easing, not disappearing.
Cloudflare AI Search now gives agents one managed pipeline for crawling, embeddings, retrieval and reranking across custom data. The beta is free, with some model-dependent charges planned at general availability.
Microsoft has broadened Foundry Model Router and refreshed its default pool with Claude Opus 4.8 and GPT-5.6. Automatic pool changes can alter behavior even when the API endpoint stays the same.
DoorDash’s Flux platform shifts coding agents from laptops into isolated cloud sandboxes. It handled 130,000 tasks in one month and now supports more than 25,000 automated code reviews each week.
Amazon OpenSearch Service can now identify 17 causes of red or yellow cluster health and recommend specific fixes. The diagnostics are free on supported versions in 11 AWS regions.