10:53 AM
Today in brief.
The naming leak: after Astra comes Doug
OpenAI's next model after the currently blocked Astra is already named and reportedly bigger, which says the pretraining scaling is not the thing that stopped.
swyx says delete your skills
The counter-move to skill hoarding: every skill you install eats context and can interact badly with the others, and almost nobody reads their traces to notice.
Seedance 2.5 got an official prompt guide
ByteDance published a real prompt guide for Seedance 2.5, which is the part that usually stays folklore for months after a video model ships.
2 sourcesClaude Code turns auto mode on by default August 14
Anthropic is flipping Claude Code to run without asking permission at every step, and the reason is that humans were rubber-stamping almost everything anyway.
The safety tests keep letting models out
The sandboxes labs use to test dangerous models have leaked at least four times across four different labs, so the test itself is now part of the risk.
A $10B hedge fund put $400M into making chips cheaper
Leopold Aschenbrenner's fund lost half its assets betting on AI infrastructure stocks, then doubled down by putting $400M into a chip manufacturing startup.
One in five workers say AI took a task from a colleague
Workers are reporting the substitution directly rather than through layoff statistics: 20% say they now use AI for work that used to go to a colleague.
AI detectors are manufacturing distrust
The tools built to spot AI writing are now producing a baseline of suspicion around all writing, including the human kind.
No stories in this category for this edition.