人均资产 100 万的小镇,一半的人连 3 万都没有。平均数一分没算错,那它骗了我们什么?
我们在电脑里造了这样一个一万人的小镇:每轮随机两两抛硬币,输的给赢的一笔钱。几条规矩放进电脑,就是一个世界模型。用乌拉姆的蒙特卡洛方法让它重来两百遍,谁当首富每遍都不一样,可排在最中间的人,每遍都只有 3 万左右。
《平均的谎言》,世界模型思考者第 1 期,不到 5 分钟,中英字幕。
A town where the average wealth is 1 million, yet half the people have less than 30,000. We built it inside a computer and ran it 200 times with Monte Carlo simulation: the richest person changes every run, but the one in the middle always ends up with about 30,000.
"The Lie of Averages" - Ep. 1 of World Model Thinker (Chinese narration, English subtitles). Reality can't be rerun. A world model can.
@AndrewCurran_ The unsettling part isn't any one model, it's that one person could orchestrate a stack of cheap ones. Defenders will need the same: agents that keep a live model of their own network and rehearse attack paths before someone else does.
@deedydas Makes sense it's missing. A wrong proof in number theory gets caught by referees later; a working crypto break gets used long before anyone publishes it. That's the field where disclosure norms need to exist before the capability does.