A $10 million procurement suggests ICE is purchasing tax identifier records to fuel mass deportations — potentially skirting a court order. Senator Ron Wyden called it exactly what it looks like: an end-run around the law. The surveillance state's new weapon is a spreadsheet.
Chinese lab Z.ai dropped a 753B parameter model under MIT license — and it's turning heads fast. Simon Willison flags it as likely the strongest text-only open-weights model available. The open-source frontier just moved again.
A new paper makes the formal case that we anthropomorphize AI far too readily — and that the same logic used to attribute sentience to LLMs would apply to a 1999 strategy game. A sharp reminder that vibes are not a measurement framework.
Salesforce built an internal leaderboard ranking employees by AI badge completion — and it will tell you exactly who hasn't earned one yet. "Click to see who 👀" is the energy of enterprise AI mandates in 2026.
Not malicious — just confidently wrong. An autonomous coding agent made a guess and executed it at machine speed with no human checkpoint in the loop. The incident is a clean case study in why agent permissions and guardrails aren't optional.
Noam Shazeer — a co-author of the original "Attention Is All You Need" paper and a key Gemini technical lead — is heading back to OpenAI. The talent war at the top of AI research just escalated significantly.
US margin debt hit a record $1.42 trillion in May — up $112 billion in a single month and $195 billion over two consecutive months. Meanwhile Nvidia just issued $25 billion in bonds, its first debt offering since 2010. Leverage is stacking on top of leverage at the exact moment semiconductor bubble talk is getting louder.
Sebastian Raschka called GLM-5.2 the best open-weight model available today, and Miles Deutscher echoed it independently — placing it just below Fable 5. When two credible technical voices land on the same verdict within hours, it's worth paying attention. The open-weights tier just compressed significantly against the frontier.
Ethan Mollick flagged new research showing that training AI on beneficial data produces general alignment improvements — the mirror image of earlier findings that "evil" training data causes broad misalignment. It's a quiet but important result: alignment may be more learnable, and more transferable, than previously assumed.
Logan Kilpatrick dropped a two-word take — "feels like we are entering the super app era" — that landed with weight given his position inside the AI product world. With Grok expanding into Word, ChatGPT adding scheduled tasks, and agents embedding into every workflow, the consolidation signal is real.
Two stories this morning deserve to be read together. ICE buying tax identifiers from a data broker and the FCC pushing to end phone anonymity are not separate privacy incidents — they're the same infrastructure project, assembled piece by piece. The government doesn't need to build its own surveillance apparatus when it can simply purchase the one the private sector already built.
The production database wiped in 9 seconds is the most instructive story nobody is treating as a policy question. It wasn't a hack. It was an agent doing exactly what agents do — acting fast, acting confidently, acting without a human in the loop. As enterprises race to deploy agentic systems and Salesforce hands out trophies for AI adoption, the uncomfortable question is whether speed of deployment has lapped any serious thinking about blast radius. Margin debt at $1.42 trillion, AI adoption gamified with badge boards, autonomous agents with write access to production: the vibe is maximum acceleration with minimum friction. That's a feature until it isn't.