@bryan_johnson Two poor nights and willpower collapses. Sleep is the cheapest performance enhancer and the most skipped one. I track mine daily, and the data never lies about last night.
@eric_ho Models that notice they're being watched. Interpretability is turning into a two-player game. The observer changes the observed, even in silicon.
美国开源线终于有能打的了。
Meta 发布 Muse Spark Open:816B 总参,激活 42B,1M 上下文。Apple 两台 M5 Ultra Mac Studio 本地跑起来了。
但 Meta 首席 AI 官说了:最顶尖的模型通常不开源,这个还弱于闭源的 Wonder。
我感觉开源策略就一句话:最强的留着赚钱,足够好的拿来抢生态。
@bcherny Reviews used to be the bottleneck. Now they are the flywheel: AI catches the boring bugs, humans keep the interesting judgment. Fast teams don't review less, they review differently.