← 返回 Avalaches

MIT的 Mert Demirer 及其共同作者进行的一项最新研究指出,人工智慧辅助软体开发的初始速度与其最终生产力收益之间存在显著差距。虽然使用人工智慧的开发人员所创建或编辑的档案增加了将近300%(约3倍),但此增幅在提交审查的工作件数中减半至150%(约1.5倍)。最终,在已完成的软体发行版本中,此提升幅度缩小了5倍,仅呈现30%(约0.3倍)的温和成长。此差异说明了审查和发布生产级软体中的人类瓶颈如何减少了人工智慧所带来的直接产出收益。

此外,软体产量的激增并未带动同等的消费。尽管过去一年中行动应用程式的发行量显著增加,但应用程式的下载量并未成长,大多数新应用程式都未能获得受众。因此,各家公司正在重新评估其人工智慧支出。例如,Uber 执行长 Dara Khosrowshahi 透露,该公司在单一季度内就耗尽了其整个2026年的人工智慧预算,促使其转向使用成本较低的模型,仅将前沿模型保留用于关键案例。此外,研究表明,将廉价的开源模型与前沿人工智慧顾问相结合,能以更低的成本提供更优异的结果。

这些温和的生产力收益反映出传统组织架构在利用新科技时所面临的挣扎。历史上,当工厂仅用电动马达取代蒸汽机而未重组工作流程时,早期的电气化带来的收益微乎其微;真正的繁荣是在数十年后,随著分散式、针对特定工作站的马达出现才到来。虽然现有企业将人工智慧嫁接到现有工作流程中的进展有限,但像 Anthropic 和 OpenAI 这样以人工智慧为核心的本土公司正经历爆发性的营收和生产力成长。当前的效率低下反映了先进工具与不合适架构之间的相互作用,随著这些瓶颈的缓解,这种情况应该会得到解决。

304ce6606eb8.png


cc0971fe5ba8.png


a2a4868c27d1.png



A recent study co-authored by Mert Demirer from MIT highlights a significant gap between the initial speed of AI-assisted software development and its ultimate productivity gains. While developers using AI created or edited nearly 300 percent more files, this increase was halved to 150 percent for pieces of work submitted for review. Ultimately, the boost contracted fivefold to a modest 30 percent uplift in completed software releases. This discrepancy illustrates how human bottlenecks in reviewing and releasing production-grade software diminish the direct output gains from AI.

Moreover, the surge in software production has not driven equivalent consumption. Despite a marked increase in mobile app releases over the past year, app downloads have failed to grow, with most new apps failing to secure an audience. Consequently, companies are reconsidering their AI expenditures. For instance, Uber CEO Dara Khosrowshahi revealed that the company exhausted its entire 2026 AI budget in a single quarter, prompting a shift to cheaper models, while reserving frontier models only for critical cases. Additionally, research suggests that pairing cheap open-source models with frontier AI advisors provides superior results at lower costs.

These modest productivity gains reflect how traditional organizational structures struggle to exploit new technology. Historically, early electrification yielded minimal gains when factories merely replaced steam engines with electric motors without reorganizing workflows; the real boom occurred decades later with decentralized, workstation-specific motors. While incumbent firms see limited progress grafting AI onto existing workflows, AI-native companies like Anthropic and OpenAI are experiencing explosive revenue and productivity growth. Current inefficiencies reflect the interaction of advanced tools with poorly suited structures, which should resolve as these bottlenecks ease.
2026-06-07 (Sunday) · ec03962642cd4f2c4cdf93597512582dd0293ae4