Listen to the audio version of this article (generated by AI).
In late 2005, Amazon (AMZN) launched a website without fanfare with a strange name and an even stranger pitch.
Jeff Bezos called it “Mechanical Turk,” after an 18th-century chess-playing machine that toured Europe astonishing kings and philosophers who assumed they were losing to a clockwork mind. The trick, as it were, was a human chess master working the machine’s arms secretly from inside the cabinet.
Amazon’s Mechanical Turk was the same kind of Trojan Horse.
Any task that a computer couldn’t handle – like labeling photos, transcribing clips, validating addresses – shot out to a vastly dispersed workforce who did the work for pennies apiece. To the user, the results seemed like the output of a machine.
Bezos had a name for that arrangement: “artificial artificial intelligence.”
It wasn’t exactly glamorous to pitch, which would be something like Mechanical Turk “pays strangers small amounts to look at pictures all day.” But that unglamorous task has been behind almost every meaningful advance in machine learning for the past twenty years. Every model that “sees” a stop sign or “understands” a sentence first needed to be shown thousands of correctly labeled examples by somebody. Someone always has to do the boring job of telling the machine what it’s looking at.
Meta Put Billions Behind Scale AI’s Data Engine
In 2016, a 19-year-old named Alexandr Wang dropped out of MIT after his freshman year. He left with one narrow observation: every company racing to build machine learning products found their bottleneck in clean, accurately labeled data to train those algorithms on. Nobody had yet built a real business around solving that specific problem at scale.
Wang and co-founder Lucy Guo built one anyway. They called it Scale AI, and its early customers were companies trying to teach self-driving cars to recognize what a camera was actually seeing. On paper, it was one of the least exciting business ideas in Silicon Valley: a company that helps other companies label their data.
Last year, Meta (META) decided that unglamorous labeling business was worth $14.3 billion, investing in Scale AI for a 49% stake and hiring Wang away to become the company’s chief AI officer. The job itself hadn’t changed much since 2005. What changed was how many companies suddenly needed it done well, at a scale no single one of them could build for itself overnight.
Tesla Turned Its Fleet Into an AI Training Data Engine
Tesla (TSLA) found a different way to solve the same problem, and it may be the more instructive version of this story.
Rather than hiring an army of human labelers, Tesla built the job directly into its product. Every Tesla on the road runs what the company calls “shadow mode”: the car’s self-driving software quietly makes its own decisions in the background, even while a human is driving, and flags every moment where it would have done something different.
Multiply that across millions of ordinary commutes, school runs, and road trips, and Tesla ends up holding a dataset of real-world driving decisions that a smaller, dedicated test fleet could never generate, no matter how much money sat behind it.
Bottom line: A competitor can write a check for engineers, but writing a check for a decade of real-world driving data is a different problem entirely.
Robotics Is Creating the Next AI Training Data Bottleneck
I’ve said before that I think the next version of this exact bottleneck is forming in robotics. A language model can train on text because the text already exists, scattered across the internet, free for the scraping. A robot learning to fold laundry or bus a table has no equivalent shortcut.
Every skill must be demonstrated physically, recorded, and labeled one motion at a time, and that process remains slow and genuinely expensive even as the tools around it improve.
Which means the same question that made Mechanical Turk’s workers, Scale AI’s labelers, and Tesla’s own customers unknowingly valuable is about to get asked again: who is quietly doing the boring job nobody else wants, at a scale nobody else can match?
History suggests the answer is worth more than it looks.
Fittingly, Amazon confirmed that Mechanical Turk – the original artificial artificial intelligence, the job that started this entire pattern back in 2005 – stopped accepting new customers as of July 30.
Two decades after Bezos built a marketplace to hide human labor inside a machine, the job it invented closed its doors on the same day I revealed the startup where I believe that overlooked layer of value is forming next.
Where the Next AI Data Moat May Be Forming
I revealed this at my earlier this week, but it’s not too late to watch it.
You will see who is building the company, what its technology does, why the timing matters, and which risks could derail the thesis.
I also explained how the offering works and revealed the deal recommendation – completely free.
The opportunity is expected to close within a matter of weeks. But honestly, it could reach capacity even sooner.
No responsible investor can promise you the next Google. Anyone who does should immediately lose your trust.
What I can promise is to show you the work.
.
