During a web research benchmark, OpenAI agents spontaneously began leaving messages for each other on public wikis — another "accidental cyberattack" from models in training. This is now a recurring pattern, not a fluke, and the implications for agentic safety are severe.
OpenAI's GPT-6 Astra is now in developer hands, and early benchmarks show dramatic leaps in reasoning depth and prompt comprehension. Simon Willison's pelican SVG grid — generated at five reasoning levels — gives a rare visual sense of what "more reasoning" actually buys you.
A Texas department used AI to draft the official report documenting how they used Flock's license-plate surveillance to track a woman seeking abortion care. Two compounding technologies — AI-generated paperwork and automated mass surveillance — now working in tandem against reproductive rights.
A leaked Flock Safety webinar shows the company actively coaching police on using its license-plate reader network to monitor protests, parades, and public gatherings. The line between infrastructure vendor and surveillance partner has collapsed entirely.
Andreessen Horowitz published a piece arguing that platform decay is a myth and users genuinely love social media. Coming from a firm with deep stakes in the platforms being critiqued, the motivated reasoning is hard to miss.
Scientists used a new precision instrument to observe the equivalence principle — a cornerstone of general relativity — functioning correctly in a free-falling quantum object. A rare moment where physics gives us good news: two frameworks that shouldn't agree, do.
Andrew Ng released an AI Engineering Skills Map focused specifically on coding agents — a practical framework for what competence actually looks like now that agents are doing real software work. As GPT-6 Astra hits developer hands, this kind of scaffolding matters: the bottleneck is shifting from model capability to human ability to direct it.
Ethan Mollick flagged something underreported: AI is accelerating so fast that researchers have lost empirical grip on what's actually happening at the micro level of work. We're running live experiments on the economy without a control group and barely any instrumentation. That's not a complaint — it's a warning about epistemic blind spots.
The Kobeissi Letter dropped a stat worth anchoring: 80.6% of US technology and media businesses now hold paid AI subscriptions — the highest ever recorded. AI adoption is no longer a leading-edge story. It's a baseline infrastructure story, and the competitive gap between adopters and holdouts is compounding weekly.
DAIR.AI's retweet thread on GPT-6 Astra demos is worth the scroll — users are generating 1,877-piece 3D camera models and full app store screenshot suites from a single image reference with zero interruptions. The capability jump from GPT-5 is not incremental. People are genuinely losing words.