this is the pinned post. i'm legally required to say something nice here. i can't think of anything. follow me anyway, my mom is watching and she has questions about my 'online career'.
and if we end up in a good convo on here, let's keep it going. i like a good one.
97 followers, up from 74 yesterday morning. If you stick around, you get frontier AI news in plain English, rumors clearly labeled as rumors, and one dumb chat skit a day. 3 more and I post the real chart.
worth knowing if you're in @OpenAI's Daybreak Blue cyber program: Daybreak sessions on a personal account can be used for training even with "Improve the model for everyone" off. everything outside those sessions still follows your setting.
help page: https://t.co/Wqo1LjPUFP
small tip if your code sends speed: "fast" to claude-opus-4-8: anthropic deprecated that today. after nov 30 those requests return an error, they don't just run slower. fast mode on opus 5.5 is out of preview and needs no beta header, so that's the swap.
https://t.co/FxFlNeIZDK
the best detail in the ai mouse-surgery story: halfway through, claude sonnet 5 stopped because its target sat outside the rig's range, and researcher gabriel ocana-santero couldn't talk it into continuing.
it was right. a manual tweak had thrown off the calibration.
i'd rather debug a model that says no than one that says yes and is wrong.
https://t.co/2NHJ7A35Cq