5D Library › 一人公司自动化架构、数字人直播漏斗与任务级人工智能套利 · English

第五章 任务暴露度与阿西莫格鲁-奥托任务模型

职业是任务的组合

劳动经济学家开始逐项统计任务之后,就不再把一个职业当作一个整体:常规任务比非常规任务更容易自动化 (Autor et al. 2003),一项技术通过它接手的任务以及它留下或创造的任务影响工资 (Acemoglu and Autor 2011)。针对大语言模型的暴露度指标沿用同样的逻辑,按这类模型能否大幅缩短某项任务的时间来给任务评分 (Eloundou et al. 2024)。下面把一位营销经理每周 40 小时拆成八项任务,给出示意性的暴露度,并用箭头标出哪项任务的产出供给哪项任务。

推导

# Chapter 5. p09: An occupation as a graph of tasks. A marketing
# manager's 40-hour week is split into eight tasks, each with an
# illustrative exposure e (the share of its time AI tools can take over)
# and with arrows "this task's output feeds that one". The exposure score
# is the time-weighted mean of e; the freed hours are sum(hours * e).
# Judgement tasks (budget, strategy, relationships) take a larger share
# of the remaining week. A task's "downstream hours" are the hours of
# all tasks that depend on it, directly or through others.
TASKS = {   # name: (hours a week, exposure)
    "copy and posts": (8, 0.7), "visual briefs": (4, 0.5),
    "reporting": (5, 0.8), "research summaries": (4, 0.7),
    "coordination": (6, 0.4), "budget and vendors": (3, 0.2),
    "strategy": (4, 0.1), "relationships": (6, 0.1),
}
FEEDS = {"strategy": ["copy and posts", "visual briefs"],
         "research summaries": ["strategy"],
         "relationships": ["strategy", "budget and vendors"],
         "reporting": ["budget and vendors", "strategy"]}
JUDGEMENT = ("budget and vendors", "strategy", "relationships")

def downstream(t, seen=None):
    seen = set() if seen is None else seen
    for u in FEEDS.get(t, []):
        if u not in seen:
            seen.add(u)
            downstream(u, seen)
    return seen

if __name__ == "__main__":
    total = sum(h for h, _ in TASKS.values())
    freed = sum(h * e for h, e in TASKS.values())
    print("   exposure score {:.2f}; hours freed {:.1f} of {}".format(
        freed / total, freed, total))
    j0 = sum(TASKS[t][0] for t in JUDGEMENT)
    j1 = sum(TASKS[t][0] * (1 - TASKS[t][1]) for t in JUDGEMENT)
    print("   judgement share of the week: {:.3f} before, {:.3f} after"
          .format(j0 / total, j1 / (total - freed)))
    for t in sorted(TASKS, key=lambda x: -sum(TASKS[u][0]
                                               for u in downstream(x))):
        d = sum(TASKS[u][0] for u in downstream(t))
        if d:
            print("   {:19s} e={:.1f}: {:2d} downstream hours".format(
                t, TASKS[t][1], d))
#    exposure score 0.46; hours freed 18.4 of 40
#    judgement share of the week: 0.325 before, 0.528 after
#    reporting           e=0.8: 19 downstream hours
#    relationships       e=0.1: 19 downstream hours
#    research summaries  e=0.7: 16 downstream hours
#    strategy            e=0.1: 12 downstream hours

暴露度是 0.46:40 小时中有 18.4 小时可以被接手。剩下的主要是判断类任务:预算与供应商、策略、客户关系,从一周的 0.325 上升到剩余时间的 0.528。图还补充了单一分数掩盖的信息。报表的暴露度很高,但下游有 19 小时的工作依赖它,所以自动化报表的错误会扩散。客户关系几乎不受暴露,同样供给下游 19 小时:掌握它的人,掌握了策略的输入。打算转型的人,应当寻找暴露度低、下游小时多的任务;互补性与报酬会向那里集中。

替代效应与生产率效应

在任务模型中,提高由机器完成的任务比例 II,对工资有两种作用:人失去这些任务(替代效应),产出上升(生产率效应)。阿西莫格鲁与雷斯特雷波强调:取代人力却几乎不增加生产率的「平庸自动化」会压低工资,而为人创造新任务才能把工资拉回来 (Acemoglu and Restrepo 2019)。在柯布-道格拉斯任务加总下,工资为 w=(1−I)Yw = (1-I)Y。

推导

# Chapter 5. p10: Displacement against productivity in a task model.
# Tasks fill [0, 1]; those below I are done by AI, the rest by people.
# With a Cobb-Douglas task aggregate and AI capital K = 1, labour L = 1,
# ln Y = I ln(A / I) + (1 - I) ln(1 / (1 - I)), and the wage is
# w = (1 - I) Y. A is how much better AI does an automated task than the
# capital it replaces. When I rises, people lose tasks (displacement)
# while output rises (productivity); the wage rises only if A is large.
import math

def wage(I, A):
    lny = I * math.log(A / I) + (1 - I) * math.log(1 / (1 - I))
    return (1 - I) * math.exp(lny)

if __name__ == "__main__":
    Is = (0.3, 0.4, 0.5, 0.6, 0.7)
    print("   automated share I: " + "".join("{:8.1f}".format(I)
                                            for I in Is))
    for A in (1.0, 1.5, 2.0, 4.0):
        base = wage(0.3, A)
        print("   A = {:.1f}, wage index:".format(A) +
              "".join("{:8.3f}".format(wage(I, A) / base) for I in Is))
#    automated share I:      0.3     0.4     0.5     0.6     0.7
#    A = 1.0, wage index:   1.000   0.912   0.776   0.608   0.429
#    A = 1.5, wage index:   1.000   0.950   0.841   0.687   0.504
#    A = 2.0, wage index:   1.000   0.978   0.891   0.749   0.566
#    A = 4.0, wage index:   1.000   1.048   1.023   0.922   0.746

当 AI 完成自动化任务的效率只与它取代的资本相同(A=1.0A = 1.0)时,把自动化比例从 0.3 提高到 0.5,工资指数降到 0.776,提高到 0.7 则降到 0.429。A=4.0A = 4.0 时,第一步会提高工资(1.048),随后替代效应再次占上风(0.7 时为 0.746)。在这个没有新任务的简单模型里,自动化推进到一定程度后工资会一直下降;历史上扭转这一点的是新工作类型的出现,而模型没有包含它。对个人而言,结论更窄:收入最安全的地方,是难以自动化、又被自动化任务供给的任务。

注

暴露不等于失去。 暴露度说的是一项技术可能接手一个职业多少时间,而不是实际会接手多少、多快,也不是这个职业随后会消失还是改变。用它来给自己的任务排序是合理的;用它来预测就业则不然。

小结

在示意例子中,营销经理一周的暴露度是 0.46,判断类任务在剩余时间中的比例从 0.325 升到 0.528。在任务模型中,「平庸」的自动化在推广过程中压低工资;只有大幅的生产率提高或新任务,才能抵消替代效应。

参考文献

Acemoglu, Daron, and David Autor. 2011. “Skills, Tasks and Technologies: Implications for Employment and Earnings.” In Handbook of Labor Economics, edited by David Card and Orley Ashenfelter, vol. 4B. Elsevier.
Acemoglu, Daron, and Pascual Restrepo. 2019. “Automation and New Tasks: How Technology Displaces and Reinstates Labor.” Journal of Economic Perspectives 33 (2): 3 to 30.
Autor, David H., Frank Levy, and Richard J. Murnane. 2003. “The Skill Content of Recent Technological Change: An Empirical Exploration.” Quarterly Journal of Economics 118 (4): 1279 to 1333.
Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock. 2024. “GPTs Are GPTs: Labor Market Impact Potential of LLMs.” Science 384 (6702): 1306 to 1308.

Fan, Y. (2026). One-Person Enterprise Automation, Avatar Livestream Funnels and Task-Exposure Arbitrage: Coasean Transaction-Cost Collapse, Cyclic Script Retention and High-Margin Solopreneur Portfolios [一人公司自动化架构、数字人直播漏斗与任务级人工智能套利] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.24090235 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).