起源故事 · 一篇死后才发表的遗稿
Origin Story · A Posthumously Published Manuscript
英国长老会牧师 Thomas Bayes 一生研究概率,却从未发表这条以他命名的公式。1761 年他去世后,好友 Richard Price
整理他的遗稿,于 1763 年代为发表。当时几乎无人重视。两个多世纪后,这条"用新证据更新旧信念"的公式,
竟成了现代世界的隐形引擎 —— 垃圾邮件过滤、医学诊断、自动驾驶、大语言模型的推断,背后都是 Bayes 当年的思想。
Thomas Bayes, an English Presbyterian minister, spent his life studying probability but never published the rule that bears his name.
After Bayes died in 1761, his friend Richard Price edited the manuscript and published it on his behalf in 1763. Almost no one noticed at the time.
Two centuries later, this little rule for updating beliefs with new evidence became the hidden engine of the modern world —
spam filters, medical diagnosis, self-driving cars, and large-language-model inference all trace back to what Bayes wrote down.
1 把 10000 人摊在面前数一数 Lay 10,000 People Out and Just Count
自然频率natural frequencies2 只看"阳性"这群人:真假各占几成 Zoom Into the Positives: True vs False, Side by Side
把所有阳性的人拎出来排成一条 —— 真患者在左,虚惊者在右。病越罕见,橙色越长 —— 这就是为什么单看"准确率 99%"会骗人。 Pull every positive person into a single bar — true patients on the left, false alarms on the right. The rarer the disease, the longer the orange — which is exactly why "99% accurate" alone is misleading.
3 现实里的贝叶斯 Bayes in the Real World
医学筛查:罕见病初筛阳性≠确诊,必须复检 —— 否则大量虚惊会带来过度治疗与恐慌。
Medical screening: a positive on a rare-disease screen is not a diagnosis. Always confirm with a second test — otherwise false alarms drive overtreatment and panic.
垃圾邮件过滤:贝叶斯过滤器根据词频不断更新"是垃圾"的后验概率,越用越准。
Spam filtering: Bayesian filters keep updating the posterior "this is spam" from word frequencies — the more you use them, the sharper they get.
刑侦证据:DNA 匹配"百万分之一"听着铁证,但在巨大人群里也可能误指无辜者(检察官谬误)。
Forensic evidence: a "one in a million" DNA match sounds airtight, yet across a huge population it can still finger an innocent person — the prosecutor's fallacy.
AI 推断:从贝叶斯网络到大模型,本质都是用证据更新先验、输出后验判断。
AI inference: from Bayesian networks to large language models, the core move is the same — update a prior with evidence and emit a posterior.
一句话In One Line
人脑天生忽略基础概率(先验):一听"99% 准"就把它当成患病概率,却忘了"病有多罕见"。
贝叶斯定理把它扳正 —— 后验 = 先验 × 似然,再归一化。罕见的事,即使证据强,也未必成真;常见的事,弱证据也足以确信。
把概率换成自然频率(10000 人里多少人),这个反直觉的悖论立刻一目了然 —— 这正是教医生读检测报告的标准方法。
Human intuition routinely ignores the base rate (the prior): hear "99% accurate" and we treat it as the probability of being sick,
forgetting how rare the disease actually is. Bayes' theorem snaps this back into place — posterior = prior × likelihood, then normalize.
For rare events, even strong evidence may not be enough; for common ones, even weak evidence can be convincing.
Swap probabilities for natural frequencies (how many out of 10,000) and the paradox dissolves —
that's exactly how clinicians are now taught to read test reports.
常见误用Common Mistakes
把准确率当成患病概率。P(阳│病) ≠ P(病│阳),两者会差很远(条件颠倒谬误)。
Treating accuracy as the probability of disease. P(+│D) ≠ P(D│+) — they can be wildly different. This is the inverse-conditional fallacy.
忽略先验(基础概率)。病越罕见,阳性里假阳越多,先验必须纳入计算。
Ignoring the prior (base rate). The rarer the disease, the more positives are false alarms — the prior has to be in the math (the base-rate fallacy).
用模糊概率讲,听者算不清。改用自然频率"10000 人里…",悖论瞬间清晰。
Explaining with vague probabilities. Switch to natural frequencies — "out of 10,000 people…" — and the paradox becomes obvious.