The Magic Was a Person
In 2011 an app told you the calories in a photo of your dinner and reviewers called it jaw-dropping. It took a couple of minutes to answer because a human being was looking at your lunch.
I went back and read a review from April 2011 this week. CNET, Rick Broida, an iPhone app called Meal Snap. You photographed the food on your plate and it told you how many calories were on it. Two dollars ninety-nine.
He tested it on a small pile of strawberries. It identified them correctly and estimated 38 to 57 calories, which is about right. He called the app jaw-dropping and put it in the same bracket as Google Translate, Shazam and Word Lens, which in 2011 was the highest compliment available to software. His one real complaint was speed. The strawberries took a couple of minutes to come back.
That couple of minutes was a person.
Calorie Counting Magic
Meal Snap was made by DailyBurn, and the way it worked was this. You took the photo, the photo went to Amazon Mechanical Turk, and somebody, somewhere, opened it on their screen and typed “strawberries”. That word got matched against a calorie database, and the number came back to your phone. The app’s full name on the App Store was Meal Snap - Calorie Counting Magic. The CEO used the word magic in interviews too. He acknowledged there were humans involved somewhere in it, but he never got specific, and nobody made him.
The funny part is that the whole thing is right there in the review if you know what you are looking at.
It took minutes for strawberries but seconds when Broida typed the food names into the caption field himself, because when you tell the worker what it is, the worker does not have to think. The estimates came back as wide ranges. And a serving of lentils came back as peas and gravy.
Peas and gravy is not a machine vision failure. A model that gets lentils wrong gets them wrong by confidence, it tells you they are beans, or chickpeas, or something else in the legume-shaped part of its weight space. Peas and gravy is a person squinting at a photo of brown mush at some hour of the night, making a call, and moving on to the next one because the next one is worth another four cents.
TechCrunch worked it out within forty-eight hours of launch and published the guess. The company calls it magic, they wrote, but we assume they have a handful of people, maybe through Mechanical Turk, breaking the meal down item by item. They were exactly right. It changed nothing. Everyone kept calling it magic, including the people who had just been told it wasn’t.
Artificial Artificial Intelligence
Here is the thing I keep enjoying about this. Nobody was hiding it.
Amazon launched Mechanical Turk in 2005, and Jeff Bezos described it in public as artificial artificial intelligence. Not artificial intelligence. Artificial artificial intelligence, with the second artificial doing an enormous amount of work in that sentence.
And they named it after the Turk, the chess machine built in 1770 that toured Europe beating aristocrats and generals, and which contained, in the cabinet underneath the board, a man. That is the most famous fake machine in history. Its entire cultural function for two hundred and fifty years has been to serve as the go-to example of a thing that looks automated and is not.
So the label on the box was accurate, the name of the box was a warning, and the founder of the company said the quiet part out loud at launch. We just did not read any of it. For fifteen years, “take a photo of your dinner and get the calories” sat there as the canonical example of impossible magic, the joke app in Silicon Valley, the thing that obviously could not be done. It could not be done. It was being done by people, at scale, in public, under a sign that said so.
Why It Actually Died
Meal Snap was not a scam, and I want to be fair to it, because what DailyBurn did is a legitimate and well-understood technique. You fake the capability with humans, you find out whether anybody actually wants it, and you automate it later once you know the answer. It has a name, Wizard of Oz prototyping, and it is genuinely good practice. Building the hard thing first and then discovering nobody wanted it is a much more expensive mistake.
What killed the app was not accuracy. It was the unit economics.
Every single photo cost money. Not amortised money, not fixed cost you pay once and spread over a million users, actual cash out the door on every individual snap, paid to a person. The fees piled up per interaction and never came down, because there was no curve to ride. A human looking at a photo of a burrito costs the same on the ten millionth burrito as it did on the first.
That is the failure mode nobody warns you about with a Wizard of Oz build: every happy user makes the business worse. Growth is the thing that kills you. Your best week is your most expensive week. You cannot get out of it by getting bigger, which is the only escape hatch software normally has, and the only way out is for the technology to arrive before your runway ends. For Meal Snap it did not. DailyBurn pivoted, and the app went away.
The version of this that matters now is that the trade has flipped. In 2011 the fake was cheap to build and ruinous to run. Today the real thing is a few lines against a vision model, and your marginal cost is tokens, which is small and falling. It is the first time I can think of where doing it properly is cheaper than faking it. Which does not mean skip the Wizard of Oz step, it means run it for a week to find out if anyone wants the thing, not for two years to find out if you can build it. That question is answered now.
The Bit Where It Gets Funny
Fifteen years on, the impossible party trick is a commodity.
Cal AI launched in May 2024, built by two teenagers. You photograph your dinner, it tells you the calories and the macros, and it uses the phone’s depth sensor to work out portion size, which is the part Meal Snap could never have done no matter how many people you put behind it. Fifteen million downloads, somewhere around thirty million a year in revenue, bought by MyFitnessPal. SnapCalorie, from the bloke who built Google Lens, does the same thing for free. It is instant.
Broida’s complaint in 2011 was that it took a couple of minutes. Nobody would ship that latency today. And the reason it is instant now is precisely that nobody is looking at your lunch.
Which gets to the thing I actually think is interesting here, because Amazon announced on the 25th of August that Mechanical Turk closes on the 30th of September. Twenty-one years, five hundred thousand registered workers across a hundred and ninety countries, thirty-five days notice.
It is tempting to file that as AI taking jobs, and I do not think that is quite it. Those were not jobs that AI came for. They were jobs that only existed because AI had not arrived yet. The task description was, in effect, be the machine until the machine gets here. Somebody was holding a seat warm for a computer, and nobody told them that was what the role was.
The Man In The Box Had A Chatbot
The ending is better than the shutdown notice, though, and it happened three years earlier.
In 2023 a group of Swiss researchers checked what Mechanical Turk workers were actually doing and found that up to forty-six per cent of them were using AI to complete the tasks. Which is perfect. The machine that pretended to have a mind was operated by a man pretending to be a machine, who by the end was quietly using a machine.
You cannot really call that cheating either. The task was to produce the output a computer could not produce. Once a computer could produce it, the fastest way to do the task was to ask the computer, and the workers, who were being paid by volume, figured that out well before the platform did. The premise collapsed from the inside, at the level of the individual worker, task by task, years before anyone wrote the press release.
What It Was Worth
I do not want to write the eulogy here, because the numbers do not support one.
The best study we have, out of CHI in 2018, tracked 2,676 workers across 3.8 million tasks and found a median wage of about two dollars an hour, with only four per cent of workers clearing the US federal minimum. The ILO got much the same figure across five platforms, and that is before you count the unpaid time, hunting for tasks that are worth taking and sitting qualification tests that lead nowhere. Requesters could reject your work and keep it, the rejection stayed on your record permanently with no appeal, and a bad requester could mass-reject everything you had ever done for them. Amazon’s position throughout was that it ran a marketplace and was not an employer, so there was nobody to complain to.
That is not a good job disappearing. That is a bad job disappearing, and it should be said plainly.
The honest complication is that it was still the only work some people could get to. The organisers who have spent years pushing Amazon on exactly the pay problems above are the same ones pointing out that it was reachable from a phone, at any hour, with no commute and no fixed shift, which made it available to people with disabilities, caregiving at home, or nothing local worth having. Both of those are true at once, and thirty-five days is not much notice for the ones who were living on it.
So the complaint worth making is not that AI took good jobs. It is that the work was underpaid for twenty-one years while it was pretending to be a computer, and the pretending was the entire product. The value was in the illusion, and the illusion was the one part the people producing it were never paid for.
Where That Leaves It
Broida wanted two things from Meal Snap in 2011. He wanted it faster, and he wanted it more accurate.
He got both, fifteen years late. The speed came from removing the person, and the accuracy came from a model trained on roughly the kind of work that person was doing.
John Croucher builds practical software and AI systems for Australian businesses, focused on solving real problems with measurable outcomes.