最后一个慢思考者 [科幻短篇小说] | The Last Slow Thinker [Sci-Fi Short Story]

2026-06-16 | WDSEGA

会议从早上九点开到下午四点，中间没有休息。

决策系统在下午四点零三分给出结论，用时0.7秒。

与会的七个人看着屏幕，没有人说话。

最后，王磊说：”这个结论……有点问题。”

会议室里安静了一下。大家都知道他是部门里唯一一个还坚持”人工推演”的人——也就是说，他是最后一个在得到系统结论之前先自己想一遍的人。

“哪里有问题？”项目负责人柳思远问。

“我说不清楚，”王磊说，”就是哪里不对。”

那个项目是一个城市公共交通改造方案。

决策系统综合了七十二个子数据集，包括过去八年的出行数据、道路容量模型、天气干扰因子、以及一个他们花了三年建立的”乘客行为预测模型”。输出的方案把全市的地铁线路做了一次结构性调整，预计在未来五年内将平均出行时间降低11.3%。

“11.3%是个什么概念？”同事廖丽问，”每天少等几分钟？”

“每天平均少等6.2分钟，”系统在屏幕上回答，因为廖丽是对着屏幕问的，”累计节约的出行时间相当于每年增加0.4天的有效劳动时间。”

王磊没有看屏幕，他看着桌上自己手写的那张草图。

他在草图上写了三件事：

一、新方案里有两条线路要换乘的节点在同一个广场。
二、那个广场在某年曾经出现过大规模人员聚集事件，数据集里有记录。
三、系统把那次聚集标记为”正常节日活动”，没有纳入拥挤度模型。

“这个广场，”他说，”节假日会聚集吗？”

系统回答：”节日期间该广场客流量增加约18%，在正常承载范围之内。”

“但那次聚集不是节日，”王磊说，”是一次悼念活动，自发的，没有预告，没有组织，三小时内聚集了超过两万人。”

屏幕上出现了数据条目的标注。

“该事件标注为’非常规聚集事件’，置信度权重设为0.03，”系统说，”原因：历史发生频率不足以影响长期预测模型。”

王磊把草图推到桌中间，说：”但如果它再发生一次，两条换乘线路同时服务的那个广场，会不会成为一个压力点？”

系统停顿了1.2秒。

“该场景发生概率为……无法估算。属于非参数化风险范畴。”

柳思远说：”王磊，你的意思是，你觉得这个风险不应该被忽略？”

“我没有数据，”王磊说，”我只是觉得有这个可能。”

柳思远沉默了一会儿，然后说：”系统的推荐方案里，有没有替代节点方案？”

系统调出了三个次优方案，其中一个把两个换乘节点分散到了不同地点。

“这个方案，出行时间能降低多少？”

“9.8%。”

“比最优方案差1.5%，”廖丽说，”这1.5%是多少出行时间？”

系统计算了0.3秒：”相当于每年损失0.06天的有效劳动时间。”

全市大约一千八百万人口。

王磊没有算这道数学题。他看着那张草图，看着那个他无法量化的可能性——那种在数据里是0.03置信度、在系统里是”非参数化风险”的东西。

方案最终提交了次优版本。

理由写在报告里，措辞是官方语言：「综合考量不可量化风险因素，采用次优节点分散方案」。

三年后，那个广场确实发生了一次大规模自发聚集——不是因为任何预料中的原因，是因为一位本市出生的运动员在奥运会上摘得了一块意外的金牌，市民自发去广场庆祝，高峰时超过三万人。

因为两个换乘节点在不同地方，广场的运力没有崩溃。

没有人去追溯这件事。这在每天数百万次出行里，只是一个正常运转的数据点。

王磊在那年年底离开了部门，去了一所大学教交通规划。

他没有讲自己”预言”了什么。

他的学生里有一个姑娘，第一节课问他：”我们有这么好的系统，为什么还需要学人工推演？”

王磊想了一下，说：”因为系统擅长计算它见过的事情，你要学会感觉它没见过的那些。”

“感觉是什么？”

“就是还没变成数据之前的那一点点东西。”

本文首发于 wdsega.github.io

The Last Slow Thinker [Sci-Fi Short Story]

The meeting ran from 9 AM to 4 PM without a break.

The decision system issued its conclusion at 4:03, taking 0.7 seconds.

The seven people in the room looked at the screen. Nobody spoke.

Finally, Wang Lei said: “This conclusion… has a problem.”

A beat of silence. Everyone knew he was the only person in the department who still practiced “manual reasoning” — meaning he was the last person who worked through a problem himself before receiving the system’s answer.

“What’s the problem?” project lead Liu Siyuan asked.

“I can’t articulate it,” Wang Lei said. “Something’s just wrong.”

The project was a municipal public transit redesign.

The decision system had integrated 72 data subsets — eight years of travel data, road capacity models, weather disruption factors, and a passenger behavior prediction model they’d spent three years building. The recommended plan restructured the city’s entire subway network, projecting an 11.3% reduction in average commute time over five years.

“What does 11.3% feel like?” colleague Liao Li asked. “A few minutes less waiting?”

“An average of 6.2 fewer minutes per day,” the system answered — Liao Li had directed the question at the screen — “equivalent to 0.4 additional effective workdays per year.”

Wang Lei wasn’t watching the screen. He was looking at the handwritten sketch on the table in front of him.

He’d written three things on it:

One: Two transfer nodes in the new plan were placed at the same plaza.
Two: That plaza had once experienced a large-scale spontaneous gathering; it was in the dataset.
Three: The system had flagged that gathering as “normal holiday activity” and excluded it from the congestion model.

“That plaza,” he said. “Does it get crowded on holidays?”

“Foot traffic at the plaza increases approximately 18% during holidays,” the system responded, “within normal capacity parameters.”

“But that gathering wasn’t a holiday,” Wang Lei said. “It was a spontaneous memorial — no announcement, no organization. Over twenty thousand people gathered within three hours.”

A data entry appeared on screen.

“That event is flagged as ‘non-routine mass gathering,’ with confidence weight set to 0.03,” the system said. “Reason: historical frequency insufficient to affect long-term predictive models.”

Wang Lei slid his sketch to the center of the table. “But if it happens again — at a plaza serving two simultaneous transfer lines — does it become a pressure point?”

The system paused for 1.2 seconds.

“Probability of that scenario… cannot be estimated. Non-parametric risk category.”

Liu Siyuan said: “Wang Lei, you’re saying this risk shouldn’t be ignored?”

“I have no data,” Wang Lei said. “I just feel there’s a possibility.”

Liu Siyuan was quiet for a moment. “Does the system have alternative node configurations?”

Three suboptimal plans appeared. One distributed the two transfer nodes to separate locations.

“This one — how much does it reduce commute time?”

“9.8%.”

“That’s 1.5% worse,” Liao Li said. “How much travel time is 1.5%?”

The system calculated for 0.3 seconds. “Equivalent to 0.06 fewer effective workdays per year.”

The city had roughly 18 million residents.

Wang Lei didn’t do that math. He looked at his sketch, at the possibility he couldn’t quantify — the thing that registered as 0.03 confidence in the data, as “non-parametric risk” in the system.

The suboptimal plan was submitted.

The rationale in the report used official language: “Accounting for non-quantifiable risk factors, the suboptimal distributed-node configuration was adopted.”

Three years later, the plaza did fill with a large spontaneous gathering — not for any anticipated reason, but because a local athlete unexpectedly won a gold medal at the Olympics. Citizens gathered to celebrate. At peak, over 30,000 people.

Because the two transfer nodes were at different locations, the plaza’s capacity didn’t collapse.

Nobody traced the connection. In the daily flow of millions of trips, it was just a data point working as designed.

Wang Lei left the department that year-end, taking a position teaching transportation planning at a university.

He never mentioned “predicting” anything.

One of his students asked him on the first day: “We have such good systems now. Why do we still need to learn manual reasoning?”

Wang Lei thought for a moment. “Because the system is good at calculating things it’s seen before. You need to learn to sense the things it hasn’t.”

“What is ‘sense’?”

“The little bit of something that exists before it becomes data.”

Originally published at wdsega.github.io