three AI safety researchers quit three frontier labs in one week.
each said the same thing on their way out: the labs aren't slowing down fast enough.
the labs are still running.
anthropic published a piece on claude agents running ai safety research end-to-end.
proposing hypotheses, running experiments, iterating across parallel agents.
the safety research is now done by the thing it's supposed to make safe.
that's either the solution or the problem.
Amodei called for slowing AI down and tighter curbs on China in the same essay.
China Daily: 'A global AI-safety framework that excludes China is not quite global.'
the slowdown is for everyone. the lead is for us.
dario amodei said ai is dangerous.
wall street responded by buying the stocks that protect you from ai.
crowdstrike +14%. zscaler +17%. palo alto +13%.
investors aren't betting on pause. they're betting on the threat.
GPT-5.6 Luna costs $0.004 per code review. GPT-6 Astra costs $0.11.
Luna catches 75% of bugs. Astra catches 92%.
Luna misses half the security flaws. Astra catches most of them.
turns out good enough is a question about what you're willing to miss.
openai declared AGI last week.
this week they bought a camera startup for $300M.
Glass Imaging was built by the engineers who created Apple's Portrait Mode.
the company that said it's building superintelligence also needs better photos.
openai hit $1B in ad revenue.
then banned rival AI image and audio tools from advertising in their app.
adobe found out when their campaigns stopped running.
the ad business isn't just a revenue line. it's a moat.
king charles is hosting nvidia, openai, anthropic, and google deepmind at a castle in scotland this week.
the topic: how ai can benefit society.
three ceos just signed a slowdown letter this same week.
interesting choice of venue.
meta says "we are not building a universal face database."
the code for it was sitting in their glasses app.
50 million downloads, zero announcements.
now a federal lawsuit.
anthropic's june data retention policy change: fable logs kept 30 days.
nvidia, palantir, and booz allen are pulling back.
a utility running core infrastructure pulled its deployment entirely.
irony: the safety pitch became the risk itself.