Skip to content
Back/Forecasts

Jevons 悖论与 AI 工作范围扩张

View in Graph
Updated 2026-09-23
2 min read
464 words

Jevons 悖论与 AI 工作范围扩张

常见误解

大多数人认为:

AI = 同样的工作更便宜更快 = 更少人力或更多空闲时间

Levie 的反直觉观点

Box CEO Aaron Levie 认为 AI 实际上意味着更多的总工作量(Jevons 悖论在知识工作中的应用)。

核心机制

一旦能自动化某项任务,你就会扩大范围:

  • 更大的软件项目
  • 更全面的研究
  • 更多创意产出

"软件项目范围会成倍增长,因为你知道你可以构建更多。"

历史类比

Jevons 悖论最初来自能源经济学:

  • 蒸汽机效率提升 → 煤炭使用量增加(而非减少)
  • 因为效率提升使更多应用场景变得经济可行

对知识工作者的启示

传统观点 Levie 观点
AI 替代人力 AI 扩大工作范围
更少工作时间 更多产出要求
守成思维 扩张思维

Hassabis: 推理永远不会免费 (2026-05-02)

来源:Demis Hassabis YC 访谈 2026-05-02

Google DeepMind CEO Demis Hassabis 在 YC 访谈中将杰文斯悖论直接应用于 AI 推理:

"即使核聚变、超导降低能源成本,芯片物理制造仍是瓶颈。未来几十年内推理端仍有配额限制。"

Hassabis 的论证:

  • 效率提高会推高总需求:百万级智能体集群、单智能体多方向并行思考综合
  • 即使能源成本趋近于零,芯片制造的物理限制(晶圆厂产能、光刻技术、材料科学)仍是硬约束
  • 这与 Aaron Levie 的"AI 扩大工作范围"形成互补:Levie 强调需求侧扩张,Hassabis 强调供给侧的物理天花板

Levie: DeepSeek + Grok 同日更新的实例 (2026-08-13)

来源:AI 简报 2026-08-13 晚间

Levie 在 DeepSeek-V4-Pro 与 Grok 新模型同日发布后再次引用 Jevons 悖论:极低成本上的巨大能力跃升,会让企业端 agent 用例(扫描代码库查安全、审文档、处理工作流)远超当前预算;成本下降只会打开更多用例、推高需求。对应用层(跨模型路由与优化)是结构性利好。

Rauch: OpenAI Sol price drop and gateway demand surge (2026-08-24)

来源:AI Builders Digest 2026-08-24

Vercel CEO Guillermo Rauch reported that OpenAI Sol price reductions, combined with Vercel AI Gateway discounts, made Sol the fastest-growing frontier model on Vercel. His reading supports the Jevons mechanism: demand for intelligence is highly elastic; when inference cost falls, usage expands rapidly. He also notes that without a gateway or router, users miss the price swings that would otherwise flow through to them, which explains the growing strategic importance of model routing and gateway layers.

Levie and Uber: unit cost reduction expands agent workload (2026-08-31)

来源:AI Briefing 2026-08-31 Morning

Aaron Levie restated the enterprise mechanism more directly: enterprises have many tasks that only become ROI-positive when token price, setup cost and change-management cost fall. Cheaper inference can therefore expand usage into contracts, logs, streams of data and background workflows. His example says a 50% token price drop can produce a much larger increase in consumed tokens.

The same briefing summarized an Uber engineering case in which agent use expanded while per-session cost was reduced through planning/execution model split, longer context caching, summary compression, searchable MCP tool gateways and reduced polling. This pairs demand expansion with cost governance. The durable pattern is not "AI gets cheaper, spending falls." It is "AI gets cheaper per unit, more work becomes economically plausible, and governance shifts to waste removal."

2026-09-23 afternoon:前沿定价双降与「每任务成本」collapse

来源:AI Briefing 2026-09-23 Afternoon + AI Builders Digest 2026-09-23

Levie 当日判断:Opus 5.5 降价叠加 GPT-6 Sol/Luna token 价格永久下调 50%,「同类任务成本」的下降速度史上绝无仅有;每次降价都戏剧性扩大 agent 可部署场景——处理全部数据、扫描代码安全、读日志做决策、workflow 里跑 agent swarm。这是应用于 agent 的 Jevons 悖论,将直接推动 AI 在经济中的更大渗透。

同日数据面(Pawel Huryn 的 105-bug 基准,含 API 等价成本):GPT-6 Sol 修 29.3 个 / $9.93,是单位成本最低的一档(Opus 5.5:41.7 / $58.53;GPT-5.6 Sol:43.5 / $95.25;GPT-6 Astra:45 / $33.04);小米 MiMo-V2.6-Pro 以 $0.86 修 22.7 个,构成强帕累托前沿;Qwen 音频线降价最高 95%。反面信号:theo 指出 $200/月编程订阅的「近乎无限」时代已结束——价格下降的另一面是配额运营成为产品面(预存重置、effort 档位默认值、限额算法)。

Counterpoints & Gaps

  • 成本数据来自单一研究者的 planted-bug 基准,n 值小,跨仓库泛化未知。
  • 「降价→场景扩张」的 Jevons 机制在 agent 上尚缺企业级采用数据支撑,目前主要是供给端推理与单点案例。

Pending Jean review

Sources

Synthesized from 8 sources

Evolution

1 event
  1. absorbed

    Derived from source material

    This page is currently synthesized from 8 sources.

    From AI 简报 2026-03-30, Demis Hassabis YC 访谈 2026-05-02, AI Builders Digest 2026-08-24, AI Briefing 2026-08-31 Morning, 2026-09-23-17-58To Jevons 悖论与 AI 工作范围扩张
    Sources: raw/briefing/AI Briefing/2026-03-30.md · raw/social-triage/2026-05-02 Demis Hassabis - AGI 还缺什么 智能体到底行不行.md · raw/briefing/AI Builders Digest/2026-08-24.md · raw/briefing/AI Briefing/2026-08-31-08-11.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-23-17-58.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Builders Digest/2026-09-23.md · raw/briefing/AI Briefing/2026-08-13-22-56.md · raw/briefing/AI Briefing/2026-09-23-17-58.md

Linked from