@teortaxesTex The spicier line is further down: kernel dev on "certain nonstandard accelerators" just falls back to Opus 5. Kinda wonder if Ascend is on that list
@GoogleDeepMind@signalapp Fun detail in the repo README: re-checking all 6.5M lines of proof yourself takes several hours and like 15-34 GB of RAM. A Signal-grade proof you can replay on a beefy workstation is kind of wild
@emollick Same CEO told WSJ last week one of Take-Two's mobile studios now makes ~$800 AI ads that used to cost $25-50k with actors. Handcrafted at the top, AI everywhere underneath
@deedydas The HC1 is literally Llama 3.1 8B in silicon, and Taalas says a new model takes about 2 months to turn into a chip. In open-weights land that's forever, so the real bet is picking which model is worth freezing
Claude Pro 这周能用 Fable 5.2,但默认模型不会自己换,要去模型菜单手动选。
北京时间 11/17 15:59 截止,最多占每周额度一半,���且它掉额度更快。
Anthropic 在发布稿里反倒劝:大多数活先用 Opus 5.5,搞不定再往上切。
发新旗舰,还劝你别拿它干日常活,挺少见。
https://t.co/4NgeG1VC3j