Regression to the Mean: on LLMs and the quiet death of the new
摘要
标题为 Regression to the Mean: on LLMs and the quiet death of the new。作者指出 LLM 承诺「每人一台机器，每人一名协作伙伴」，带来「思想爆炸」，实际却带来「向中心回归」。它给出最可能的延续，纠正未见过的内容；把答案喂回问题循环，分布收窄；过去发现（移动地球、未见细菌、大陆漂移）曾被共识视为错误，现在模型把新事物判定为错误。稀缺的是偏差位置，而非平均数。
荐读理由
把训练数据喂回模型做下一轮问题，每次迭代让分布收窄至均值点，收敛到共识而非新意，这正是 AI 工程里你当前项目要避开的收敛陷阱
原文
the promise of more ideas · and the pull to the center
Regression to the Mean
We were handed a machine that could think alongside us, and told it would set off an explosion of new ideas. It may do the opposite — so gently that we mistake the flattening for progress.
scroll
the promise · no. 01
A model on every desk; a collaborator for every mind. The pitch was a Cambrian bloom — a thousand directions explored at once, by everyone. More minds thinking, surely, means more thoughts worth thinking.
an explosion of the new
the mean · no. 02
But ask it anything and it returns the most probable continuation — the center of mass of everything already written. Trained on the past, it answers in the past tense of thought. Not what is true. What is typical.
the average · returned at the speed of certainty
the correction · no. 03
Offer it something it has never seen, and it doesn't light up. It corrects you. To a system built to predict the expected, the genuinely new is indistinguishable from a mistake.
The pushback is soft, and constant:
Did you mean —
the familiar thing, offered in place of yours.
Not a standard term
the new name returned as a typo.
Most sources agree
consensus handed back as fact.
Are you sure?
conviction sanded down to the mean.
the collapse · no. 04
And we feed its answers back in as the next questions. Each pass, the spread narrows; the strange tails thin out. Variance leaks out of the culture. We converge — not on what is right, but on what is average.
output becomes input · the curve sharpens to a spike
the cost · no. 05
Yet every discovery was, at the moment it was made, out of distribution. It disagreed with the consensus of its day — moving earth, unseen germs, drifting continents, each first filed as error. A model of consensus is, by construction, a machine for telling you the new thing is wrong.
discovery has only ever lived in the tail
the deviation · no. 06
So the scarce thing inverts. The average is now free, infinite, identical — worth little precisely because everyone holds it. What is priceless is the deviation: the position the model marks as wrong, kept anyway. Not the answer it was sure of — the one it would not stop correcting.
guard the tail
what the machine can't hand back
It will give you the average of all that has been thought. The new was never in there.
A tool that returns the most likely sentence is a comfort and a quiet narrowing at once. The work it cannot do for you is the only work that ever mattered: to stand, on purpose, where the curve runs thin — and stay there long enough to be right.
not the most probable answer. the one it tried to correct.
written off-distribution · on purpose
这条对你有帮助吗？