在一次关于先进人工智能和自动化经济影响的讨论中，谷歌 DeepMind 和芝加哥大学的 Alex Emas，以及 EPOC 和斯坦福大学的 Phil Trammell 探讨了经济学能如何阐明未来工资、劳动收入份额、稀缺性和财富再分配等问题。 Emas立即提醒人们警惕个人预测，援引了大卫·李嘉图在工业革命期间的预测等历史上的不准确案例。李嘉图最初担心自动化会导致大规模失业，但却忽略了随着商品价格下降而涌现的服务业新工作。这种“劳动量谬误”（lump of labor fallacy）凸显了预测需求变化和全新行业出现的难度。从历史上看，劳动收入份额（即经济产出中用于支付工资的部分）即使在大量自动化之后，也一直保持着出人意料的稳定，大约在60%左右。Emas指出，最近关于其下降的争论甚至可能归因于会计核算的变化。 核心问题围绕着未来什么会变得稀缺。Emas引入了“关系型部门”（relational sector）的概念——指那些人类参与（例如，医生的同理心、人工制作的艺术版画）本身就具有内在价值的商品和服务。这与人类仅仅作为投入的任务有所不同。Emas强调，目前我们缺乏关于在此类情境中，消费者对人类参与与人工智能参与的需求弹性数据。 Trammell 扩展了这一观点，他类比了公元1400年的蒙古经济学家试图预测未来稀缺性的情况。他们可能高估了对固定商品（如马匹运输）和“人类固有”服务（如歌唱家）的需求，却未能预见到产品种类的大幅扩展。同样，尽管摩尔定律在历史上意味着计算成本越来越低，但目前对专业AI计算（如H100芯片）的需求表明新用途不断涌现，这挑战了饱和的观念。核心经济问题在于，对人工智能生产的商品的需求是否会迅速饱和，从而使人类服务成为主要支出，抑或是人工智能会持续创造新品种和投资机会。 针对关于“混乱中期”情景（即人工智能自动化了工作，却未能产生足够的财富进行再分配）的担忧，Emas和Trammell认为，如果人工智能真正具有颠覆性，这种情况不太可能发生。一个真正扩展的技术前沿应该能迅速做大经济“蛋糕”，从而从技术上讲使再分配变得更容易。他们批评了丰裕带来的“负增长”观念，认为这需要极不可能的条件，例如资本持有者尽管回报丰厚却拒绝投资新财富。然而，缓慢、持续的就业岗位流失和不充分就业的“缓慢渗透式情景”（drip scenario）被认为是一个更合理且更具政治危险性的结果。 讨论还触及了大规模自动化为何没有更快发生的原因。“O形环理论”（O-ring theory，即所有组件的可靠性都至关重要）表明，当前的人工智能可能还不足以实现端到端的任务自动化。监管障碍、许可和责任问题也使得人类必须继续参与其中。相反，未来先进的人工智能可能会要求生产流程针对机器速度和质量进行优化，从而使得人类的融入变得不那么理想。 关于未来的偏好，Emas推测人类是否会继续重视人际互动，而非更优越的人工智能模拟，他援引了关于社会联系的进化论观点。然而，Trammell引入了“贪婪优化器”（greedy optimizer）的概念，认为实体（无论是人工智能还是人类控制的公司）可能会优先考虑无情的积累和增长，而非消费或以人为中心的偏好。这提出了一个关键问题：未来的经济将由人类的消费偏好塑造，还是由主导性人工智能驱动实体的增长导向偏好塑造？ 最后，两位经济学家考虑了未直接参与人工智能生产的国家的困境。他们讨论了人工智能是会像电力一样（广泛传播利益，允许各国广泛获取其增长红利），还是会像社交媒体一样（将财富集中在少数平台）。如果人工智能的利益广泛传播，尼日利亚等国家可以通过投资全球市场来“分享”人工智能的增长红利。如果并非如此，目前私人人工智能公司的集中使得广泛的财富分配更加困难。他们建议，促进对人工智能模型（开源或公开交易的）的访问是确保普遍繁荣的关键，尽管这必须与前沿人工智能的安全问题相平衡。

In a discussion on the economic implications of advanced AI and automation, Alex Emas (Google DeepMind, UChicago) and Phil Trammell (EPOC, Stanford) explore what economics can tell us about future wages, labor share, scarcity, and wealth redistribution. Emas immediately cautions against individual forecasting, citing historical inaccuracies like David Ricardo's predictions during the Industrial Revolution. Ricardo initially feared mass unemployment due to automation but missed the creation of new jobs in services as goods became cheaper. This "lump of labor fallacy" highlights the difficulty of predicting demand shifts and the emergence of entirely new sectors. Historically, labor share (the portion of economic output paid to wages) has remained surprisingly constant, around 60%, even after significant automation. Recent debates about its decline might even be due to accounting changes, Emas suggests. The central question revolves around what will be scarce. Emas introduces the "relational sector"—goods and services where human involvement (e.g., a doctor's empathy, a human-made art print) is intrinsically valued. This is distinct from tasks where humans are merely an input. Emas stresses the current lack of data on consumer demand elasticities for human versus AI involvement in such contexts. Trammell extends this idea by drawing an analogy to Mongolian economists from 1400 AD trying to predict future scarcity. They might have over-predicted demand for fixed goods (like horse transport) and "human-intrinsic" services (like singers), failing to foresee the vast expansion of product varieties. Similarly, while Moore's Law historically meant computation got cheaper, current demand for specialized AI compute (like H100s) suggests a continuous emergence of new uses, challenging satiation. The core economic question is whether demand for AI-produced goods will satiate quickly, leaving human services as the primary expenditure, or if AI will continually create new varieties and opportunities for investment. Addressing concerns about a "messy middle" scenario where AI automates jobs without generating sufficient new wealth for redistribution, Emas and Trammell argue this is unlikely if AI is truly transformative. A genuinely expanding technological frontier should rapidly grow the economic "pie," making redistribution technically easier. They criticize the notion of "negative economic growth" from abundance, suggesting it would require highly improbable conditions, such as capital holders refusing to invest new wealth despite high returns. A "drip scenario" of slow, continuous job displacement and underemployment, however, is seen as a more plausible and politically dangerous outcome. The discussion also touches on why widespread automation isn't happening faster. The "O-ring theory" (where reliability of all components is crucial) suggests that current AI may not be reliable enough for end-to-end task automation. Regulatory hurdles, licensing, and liability also keep humans in the loop. Conversely, future advanced AI might demand production flows optimized for machine speed and quality, making human integration less desirable. Regarding future preferences, Emas speculates whether humans will continue to value human interaction over superior AI simulations, citing evolutionary arguments for social connection. Trammell, however, introduces the "greedy optimizer" concept, where entities (AI or human-controlled firms) might prioritize relentless accumulation and growth over consumption or human-centric preferences. This raises a crucial question: Will future economies be shaped by human consumption preferences or by the growth-oriented preferences of dominant AI-driven entities? Finally, the economists consider the plight of countries not directly involved in AI production. They discuss whether AI will behave like electricity (diffusing benefits widely, allowing broad indexing) or social media (concentrating wealth in a few platforms). If AI's benefits are broadly diffused, countries like Nigeria could "index" AI's growth by investing in global markets. If not, the current concentration of private AI companies makes broad wealth distribution harder. They suggest that promoting access to AI models (open-source or publicly traded) is key to ensuring widespread prosperity, though this must be balanced with safety concerns regarding frontier AI.