#916 2005 · Amazon (Mechanical Turk, Venky Harinarayan) · E-commerce / crowdwork platforms
Amazon's catalog defeated its own code, so engineers wired human judgment in as a function
问题
Duplicate product listings on Amazon's site kept defeating automated matching software built to catch them
背景
As Amazon's catalog grew far beyond books, it accumulated large numbers of duplicate product pages — the same item listed multiple times with slightly different titles, images, or descriptions. Engineers tried automating the cleanup with matching algorithms, but the task kept defeating software: a human could glance at two listings and tell in seconds whether they were the same product, while writing rules general enough to catch every variation without producing false matches proved impractical at catalog scale.
Amazon manager Venky Harinarayan's team built what he called a hybrid machine-human computing arrangement: a system that broke the deduplication task into small, independent units, each answerable by a human in seconds, and distributed those units automatically over the internet to a pool of remote workers who could complete them whenever and wherever they chose. Jeff Bezos saw the internal tool's broader potential and had it turned into a public marketplace in 2005, named Mechanical Turk after the 18th-century chess-playing automaton that had secretly concealed a human operator inside.
换别人会怎么做
The available options were to keep refining automated matching algorithms in hopes of eventually catching every duplicate variant, or hire a dedicated internal team to manually review the catalog, an approach that scaled only as fast as headcount and cost accordingly.
他们看到了什么
Harinarayan's team saw deduplication wasn't a hard computational problem, it was a fast human judgment repeated at scale, scaling with recruits instead of engineering effort spent automating intuition away.
那一手
Amazon built and then opened publicly a marketplace where any business could post small 'human intelligence tasks' — like identifying duplicate product listings, transcription, or image labeling — that software could call programmatically, and any worker anywhere could complete for a small per-task fee, paid the moment their own catalog-matching problem needed the same kind of judgment.
为什么管用
Breaking the catalog problem into small, independent 'human intelligence tasks' meant no single worker needed context on the whole catalog, only enough to answer one quick comparison, which kept the labor cheap, parallelizable, and callable programmatically like any other API. Because workers could complete tasks whenever and wherever suited them rather than being centrally employed, the system could scale its available labor supply up or down with demand instead of being bounded by a fixed internal headcount.
值了多少
The internal tool solved Amazon's own catalog problem, then launched publicly in 2005 as a marketplace still running today.
什么时候会失灵
The mechanism depends on the task decomposing into small units answerable correctly by an anonymous worker with no broader context, and on being able to verify or aggregate results cheaply enough to catch bad-faith or low-quality submissions; tasks requiring deep specialized expertise, sustained context across many related items, or judgment calls with no clear right answer don't hold up well when split across disconnected, interchangeable workers.
后来呢
Mechanical Turk became the template for the broader 'crowdwork' and human-in-the-loop AI industry, later underpinning much of the labeled training data behind modern machine learning systems, all descended from a tool built to solve a strictly internal catalog problem.
资料来源
- [1]Untold History of AI: How Amazon's Mechanical Turkers Got Squeezed Inside the MachineIEEE Spectrum, 2021spectrum.ieee.org