Without it, there would be no Li Fei-Fei's ImageNet... and now it is gone.

A crowdsourcing task platform that paid just a few cents per assignment once underpinned the entire deep learning era.

Now, it is shutting down.

Amazon has just announced that its crowdsourcing platform Mechanical Turk will completely cease operations on September 30 this year.

At its peak, over 500,000 workers were on the platform doing odd jobs online: viewing images, classifying content, labeling data, and manually completing all the small tasks that computers could not learn to handle back then.

Later, this massive "human computing power" directly contributed to the creation of the ImageNet dataset —

Without MTurk, the manual labeling of tens of millions of images for ImageNet would have been nearly impossible to complete.

Without ImageNet, Li Fei-Fei might not have risen to fame overnight.

The team of Hinton would not have become legendary after that breakthrough, and deep learning might not have suddenly experienced its renaissance in 2012.

21 years later, real AI has become increasingly capable of completing these tasks on its own.

And the group of people who once helped AI appear intelligent, along with the platform itself, have now received the "Shutdown Notice".

500,000 People Doing Gig Work Online: MTurk Brought in Real Humans to "Cover Shifts"

Mechanical Turk (MTurk) was launched in 2005.

In fact, at the very beginning, Amazon never intended for this tool to rewrite the history of AI.

Back then, MTurk solved a very practical and dramatic problem: some tasks were extremely difficult for computers, but ridiculously easy for humans...

So Amazon came up with a very straightforward solution: If programs can't handle it, just break the problem down and assign it to real people~

As a result, these split small tasks were called HIT, short for Human Intelligence Task.

For example, if a company had 10,000 images to classify, MTurk's operating model was to split them directly into 10,000 independent tasks and release all of them on the platform at the same time.

100 people can work on it together, 1,000 people can do it too, and if tens of thousands of people go online at the same time, the work gets done even faster.

In other words, MTurk did not make a single person work faster, but turned a laborious task that could only be done in a queue into something that could be "globally parallelized".

Truly a manually operated data flywheel...

And the name Mechanical Turk itself perfectly matches this operating logic.

In the 18th century, there was a famous "Mechanical Turk" in Europe. From the outside, it was a machine that could play chess on its own, and even played against Napoleon.

Later people found out that there was a real chess player hidden inside the machine! (I totally underestimated this...)

Amazon moved this allusion to the Internet, and the fit was almost perfect: the program seemed to process tasks automatically on the surface, but the real people behind the screen were the ones doing the actual work.

Jeff Bezos gave it a description back then that turned out to be almost prophetic:

Artificial Artificial Intelligence.

Once this idea was proven to work, MTurk quickly grew from an internal tool of Amazon into a global crowdsourcing marketplace.

By 2011, more than 500,000 workers from 190 countries had gathered on MTurk.

As the number of workers grew, the types of tasks that could be accepted expanded rapidly: image screening, speech transcription, text classification, data cleaning, sentiment judgment, content moderation and more.

Any task that "is difficult for machines but can be judged by humans in a few seconds" can be posted on the platform.

This pattern lasted for 21 years.

Without MTurk, There Would Be No Modern ImageNet and the Deep Learning Revolution

Let's go back to around 2006.

Shortly after starting her teaching career at Princeton, Li Fei-Fei decided to do something that was quite counter-trend at the time —

To build a large enough visual encyclopedia for machines.

At that time, mainstream computer vision datasets usually only contained thousands to tens of thousands of images. With too little data, models were prone to overfitting: they could memorize the patterns in the training set, but could hardly truly understand the diverse objects in the real world.

So Li Fei-Fei's team decided to do something very crazy at that time: build a "large enough" visual world for machines.

The ImageNet dataset was born as a result.

ImageNet took WordNet as its framework, split the real world into individual concepts, and then added hundreds or thousands of real images under each concept.

Dogs, cars, apples, chairs, birds, fish... All definable objects were added to the dataset as much as possible.

But soon, a more realistic problem came up —

Images were not hard to find, they could be searched on Google, Yahoo and Flickr, but the trouble was that the retrieved images were not necessarily correct, and many of them were completely irrelevant!

So after the search engine retrieved the images, manual work was still required for the final step: verifying each image one by one.

To give you a staggering number: the total number of candidate images ImageNet had to process later exceeded 160 million! At this point, the manpower problem became a total deadlock...

When ImageNet was almost stuck by the massive amount of manual image screening work, Li Fei-Fei's team turned to the newly launched Amazon Mechanical Turk.

The ImageNet team began to split the massive image screening work into microtasks, and distribute them to global workers on the MTurk platform.

A grand AI infrastructure was thus broken down into countless ordinary "mouse clicks".

The ImageNet paper published in 2009 noted that the dataset already included 5,247 concepts and 3.2 million sorted images at that time.

The number given by the ImageNet team when reviewing the entire project later was: 49,000 MTurk workers from 167 countries.

In other words, the massive image screening project behind ImageNet was supported by a temporary global labeling army.

This is why ImageNet might never have been successfully built back then without MTurk...

After ImageNet was built, the subsequent story began to accelerate.

In 2010, the ImageNet Large Scale Visual Recognition Challenge was launched. Models from all over the world began to train on the same dataset and compete on the same leaderboard.

Two years later, the team that would later be repeatedly written into AI history made its debut —

Geoffrey Hinton from the University of Toronto, and his two students Alex Krizhevsky and Ilya Sutskever.

The three of them participated in the ImageNet Large Scale Visual Recognition Challenge with a deep convolutional neural network.

They directly pushed the Top-5 error rate of the 2012 ImageNet Challenge to about 15%, leaving traditional methods far behind.

This made the entire computer vision community see for the first time intuitively that the combination of massive data, GPU computing power and deep neural networks could bring a huge leap in performance.

From left to right: Ilya Sutskever, Geoffrey Hinton and Alex Krizhevsky

Since then, the name of Alex Krizhevsky has been permanently linked to AlexNet.

Ilya later joined Google Brain, co-founded OpenAI and served as its chief scientist, and founded SSI after leaving.

Professor Hinton later won the Turing Award and became one of the recognized godfathers of deep learning.

Li Fei-Fei later returned to Stanford to teach, co-led the Stanford Institute for Human-Centered AI (HAI), and established her position in the history of modern computer vision with ImageNet.

Later, VGG, GoogLeNet, ResNet came one after another.

The ImageNet leaderboard was refreshed year after year. Deep learning expanded from the computer vision field to speech processing, then to natural language processing, and finally arrived at the current large model era.

The world is really amazing —

Li Fei-Fei's team built ImageNet with the help of MTurk; ImageNet made Li Fei-Fei widely famous; Hinton and his students became legendary with the help of ImageNet; deep learning achieved its renaissance through this breakthrough.

Nearly 50,000 ordinary people from 167 countries sat in front of their computers, and "clicked" hundreds of millions of candidate images into the ImageNet dataset little by little.

AI Finally Learned to Do the Work, But MTurk Shut Down First

That's why the shutdown of Mechanical Turk today looks like a perfect closed loop of an era.

Amazon's official statement is very restrained, only saying that it continuously evaluates its products and services, so it decided to discontinue MTurk.

CNBC previously cited Krista Pawloski from the data workers' rights organization Turkopticon, saying that MTurk has been declining over the years, Amazon has invested less and less in it, and a new generation of data labeling platforms are constantly taking away workers and clients.

But the more direct reason, to put it bluntly, is that the most valuable tasks it used to do are increasingly being completed by AI itself...

In 2005, letting a machine judge whether there is a dog in an image was a real difficult problem, but now it has become a basic task that multimodal models can easily complete.

At the same time, the AI industry's demand for "human labor" has also changed.

Cutting-edge models need professionals such as programmers, doctors, and lawyers to evaluate codes, review responses, test reasoning performance and safety.

As a result, new platforms such as Scale AI, Mercor, and Prolific have emerged, starting to screen and manage experts to provide more professional training and evaluation data.

In contrast, MTurk's crowdsourcing model of "anyone can take tasks for a few cents per order" is increasingly out of step with the times...

What's more surreal is that even MTurk workers themselves have started to use AI!

A 2023 study by Swiss scholars found that among the surveyed MTurk workers, the proportion of those who used AI models to complete text tasks was as high as 46%.

Originally, clients paid for real human judgment, but some workers would accept the order and then assign the task to models like ChatGPT to answer.

This is more or less a bit of black humor...

The fate of MTurk is actually a microcosm of the development of AI over the past 20 years.

At the very beginning, humans hid behind the machines, labeling images one by one, and painstakingly built ImageNet, pushing open the door to the deep learning era.

Later, AI grew up along these data, and finally learned to view images, listen to sounds, and write texts on its own.

21 years later, the group of people who once helped AI appear intelligent have stepped down from the stage.

The "Mechanical Turk" that hid a real chess player inside can finally close its cabinet door.

Reference Links:

[1]https://www.cnbc.com/2026/08/25/amazon-service-that-jeff-bezos-called-artificial-ai-is-shutting-down.html

This article is from the WeChat official account "QbitAI", written by Meng Yao, and republished with authorization from 36Kr.