作者主张，关键不只是效率，而是技术是否跨越 S 曲线的隐藏门槛。依 Richard Foster 于 1986 年的分析，技术扩散通常呈 S 曲线：底部缓慢、中段陡升、顶部饱和。ChatGPT 并非相对 GPT-3 的能力剧变，而是以足够可用、可近用且低成本的形式，让原本被排除的大众进入市场。结果是需求由平转直：自 2022 年 11 月发布后约一年达到每周活跃用户 1 亿，到今日约 9 亿；连 OpenAI 工程师也对此规模感到意外，显示潜在需求在门槛之下长期被低估。
但文章也强调，AI 不会无限垂直成长。所有 S 曲线终将趋平；喷射引擎在 1950 年代后期跨越门槛后，全球航空旅客量曾连续 20 年以每年 14% 成长，但今日增速已降至其一小部分。更根本的上限来自问题本身：Michael Berry 指出，预测撞球第 9 次碰撞已需计入旁观者重力，第 56 次则需知道可观测宇宙每个粒子的位置；George Soros 也认为，金融市场中的反身性会使预测改变行为本身。因此，AI 当前位于 S 曲线甜蜜区，但下一场革命终将来自另一项仍在曲线底部的技术。
The essay uses the Jevons Paradox to explain AI economics: in 1865, William Stanley Jevons argued that better steam-engine efficiency did not reduce coal use, but increased total demand because lower costs created more uses and users. AI shows the same pattern. Epoch AI says inference prices for large language models at a fixed performance level have fallen at a median rate of about 50x per year; yet OpenAI’s annualized revenue rose from $2 billion in 2023 to more than $20 billion in 2025, while its computing capacity tripled in one year. Costs are collapsing while spending is surging, showing that efficiency gains are expanding rather than shrinking total usage.
The author argues that the key is not efficiency alone, but whether a technology crosses the hidden threshold on an S-curve. Following Richard Foster’s 1986 analysis, technologies usually diffuse in an S-shape: slow at the bottom, steep in the middle, saturated at the top. ChatGPT was not a dramatic capability leap over GPT-3; it made the model usable, accessible, and cheap enough for mass adoption. Demand then turned vertical: from 100 million weekly active users about a year after its November 2022 launch to about 900 million today. Even OpenAI’s engineers were surprised, indicating that latent demand had long been underestimated below the threshold.
But the essay also stresses that AI will not grow vertically forever. Every S-curve eventually flattens; after jet engines crossed their threshold in the late 1950s, global air passenger traffic grew 14% a year for two decades, but today the rate is only a fraction of that. The deeper ceiling comes from the problems themselves: Michael Berry argued that predicting a billiard ball’s ninth collision already requires accounting for spectators’ gravity, and the 56th would require the position of every particle in the observable universe; George Soros likewise held that reflexivity in financial markets makes predictions alter behavior itself. AI may be in the sweet spot of its S-curve now, but the next revolution will come from some other technology still near the flat bottom of its own curve.