metir Blog

Google is rolling out Gemini skills in Workspace and the Gemini app and phasing out Gems. Skills vs Gems vs custom GPTs, the SKILL.md standard, and admin steps.

A Defense Department official told the BBC the Pentagon has stopped using Anthropic's Claude. What a phaseout means for embedded models, and for vendor concentration.

Goldman Sachs data shows AI-linked leveraged finance at $88B year to date vs $20B a year ago. How high-yield debt, leveraged loans and GPU collateral work.

OpenAI is adding an invisible text watermark called textGrain to ChatGPT and Codex in the EU. How it works, where it breaks, and how it compares with SynthID.

Anthropic's Claude Frontier Academy commits $100M to train 10,000 Frontier Deployed Engineers by end of 2027. How the model works and what it signals.

Amazon stopped using NDAs with government agencies on data center projects and pledged $1B over five years. How it compares with capex, tax breaks and rivals.

#TeamHuman, led by YouTube creators with 300M+ subscribers, wants a global AI slowdown, chip tracking and treaties. How it differs from earlier open letters.

Trump's Super Intelligence executive order swaps AI for SI in federal language. What changes in law, contracts and procurement, and what stays the same.

Altman says there is "a lot of daylight" between OpenAI and Anthropic on AI regulation. A sourced side-by-side plus how aviation and drugs weigh risk.

Apple says AI agents raise the risk of macOS Full Disk Access. How Mac permissions work, what the grant exposes, why agents raise risk, and an audit checklist.

A researcher extracted Meta Muse's system prompt, reportedly with a household authority clause and hourly contact profiles. What it means for privacy law.

SoftBank closed its DigitalBridge purchase and Ganzi calls it SoftBank's third-party infrastructure arm. What it manages, why outside capital, and what to watch.

Anthropic's employee charity match reportedly cost over $660M in non-cash stock expense in six months. How equity matching works, with derived dilution math.

Toshiba will invest about 60 billion yen to double AI data center HDD capacity in the Philippines, aiming for 30% share. Market shares, HAMR roadmap, supply.

The Eighth Circuit paused Minnesota's AI nudify ban while xAI's First Amendment case proceeds. What the law bans, the injunction standard and the TAKE IT DOWN Act.

Cambridge startup Zenithon raised $10M for physics AI models in fusion, aerospace and chips. How surrogate models differ from CFD and FEM, and who else is funded.

AIB Data Centers signed a 12-year, 50 MW lease with Nebius. A neutral look at what was disclosed, how it fits Nebius's deals, and the miner-to-AI-host pivot.

Schneider Electric agreed to buy PTC for $205 a share, a $22.6B equity value. How it fits industrial AI, AVEVA, data center demand and earlier software deals.

Strata, an open-source inference engine, reportedly runs a 125B MoE Qwen model on a 12GB GPU at 60-95 tokens a second. How expert offloading makes that work.

Tavus says Griffin-Lite fooled 26 of 54 people in one-minute video calls. We check the video Turing test method, the confidence interval and fraud implications.

François Chollet says OpenAI set AGI back 5 to 10 years. We explain ARC-AGI-1, 2 and 3, the latest ARC Prize scores, the human gap and both sides of the debate.

CoreWeave put Nvidia Vera Rubin NVL72 into production with Cognition, which reports up to 4.8x GB200 token throughput. What the claim does and does not cover.

Nvidia-backed Reflection AI is reportedly preparing its first open-weight model for Western enterprises, with $7B+ in compute deals. What is confirmed and what is not.

Sean Parker is reportedly rebuilding Stability AI around music tools with Sony, Warner and Universal as backers. Here is the timeline and what it signals for AI music.

The Information reports Beijing token resellers use accumulated identities and USDT-funded cards to run a Claude gray market. What it says about access policy.