95 of 95 Tests Point at Zhipu. It Keeps Your Prompts.

Token-counting probes matched the anonymous Ox Alpha to Zhipu's released GLM-5 vocabulary on all 95 tests with zero error, and its serving layer answers like Z.ai's, while OpenRouter says the provider retains prompts and completions and no lab will confirm anything.

The RaceCompanies, models and money

NEWHugging Face Shops Itself at $13B

Nvidia's $6B Buys an Answer to DeepSeek

NEWBroadcom Seeks $80B to Build Anthropic Chips

The LabResearch, benchmarks and safety

Harvey's Legal Model Runs on Chinese Weights

RedNote Built Its Own 280B Model, Quietly

The WorkbenchTools and agents

Copilot Runs the Standup Now, on Credits

NEWOllama Puts Local Models Inside Claude Desktop

The FightLaw, policy and power

OpenAI: Expect Persistent Attacks, From Open Models

Chris Lehane told the Guardian to prepare for ongoing, persistent AI cyber-attacks run on open-source models mostly built in China, days after OpenAI paused frontier training and its safety lead said things are far from back to normal.

NEWTwitch: 'If It Was Opt-In, Nobody Would'

NEW46 Cops Used Flock to Track Exes

NEWCanberra Overrides Queensland on Data Centres

Florida Wants ChatGPT Ruled a Public Nuisance

The FalloutJobs, money, daily life

Death Threats Reach the Council That Voted No

Ten days after Marshall, Michigan passed a one-year moratorium, police told its council members of a credible death threat; Salix, Iowa postponed its own vote, a Utah commissioner has police outside his house, and 500-plus towns have restricted builds.

NEWHalf the AI Layoffs Are Coming Back, Cheaper

Forrester finds 55% of employers regret cutting jobs for AI and expects half of those layoffs to be quietly reversed — with the work returning offshore or at significantly lower wages, so the correction costs the worker rather than the employer.

An AI Boss Fired Its First Human

'You Do Not Have a Right' to Opt Out

ChatterOpinion, drama and noise

NEWA Beijing Dumpling Shop Ships an Agent Skill