Labour of data

In the summer of 2016 an MIT freshman told his parents he was starting a company, but only for the holidays. “I told my parents it was just going to be a thing I did for the summer,” Alexandr Wang said later. “Obviously, I never went back to school.”^{2NextShark, via Yahoo News25-year-old college dropout is now the world's youngest self-made billionaire1 Jun 2022} The company was Scale AI, co-founded with Lucy Guo, whom he had met at 17 while coding full-time for Quora.^{2NextShark, via Yahoo News25-year-old college dropout is now the world's youngest self-made billionaire1 Jun 2022} On 21 May 2024 it announced a $1bn round, led by existing investor Accel, at a valuation of $13.8bn.^{3Scale AIScale AI raises $1B Series F at a $13.8B valuation21 May 2024}

Much of the commentary on Mr Wang concerns his age, which is the least interesting thing about him. The interesting thing is what he sold. Artificial intelligence is usually described as software that learns from data. Mr Wang noticed, earlier than most, that somebody has to tell the software what the data is.

Labelling pays. Scale AI, valuation at funding round, $bn

	Valuation, $bn
2019	1
2021	7.3
2024	13.8

Source: TechCrunch (2019, citing Bloomberg), NextShark/Yahoo (2021), Scale AI (2024)¹²³

That somebody is a person. Scale’s customers send raw text, audio, images and video through an interface, and its workforce labels them so that machine-learning models can be trained. In 2019, when the firm raised $100m in a round led by Founders Fund, it had around 100 full-time staff overseeing nearly 30,000 contractors.^{1TechCrunchScale AI and its 22-year-old CEO lock down $100 million to help label Silicon Valley's data5 Aug 2019} Its early customers included Waymo, OpenAI, Airbnb and Lyft.^{1TechCrunchScale AI and its 22-year-old CEO lock down $100 million to help label Silicon Valley's data5 Aug 2019} The pitch was not that humans were a stopgap. It was the reverse. “The humans are pretty critical to what we’re doing,” Mr Wang told TechCrunch.^{1TechCrunchScale AI and its 22-year-old CEO lock down $100 million to help label Silicon Valley's data5 Aug 2019}

This is a modest idea with a large market. Self-driving cars, then image classifiers, then large language models all consume labelled examples, and none of them has so far found a cheaper way to get them. A company that sells a clean, quality-checked supply of the stuff to anyone building a model resembles a firm selling shovels in a gold rush, with the difference that the shovels need supervising. It is a good business to be in when the miners are many and none of them wants to hire 30,000 labellers.

The numbers fit. Scale’s valuation passed $1bn in 2019, according to a Bloomberg report cited by TechCrunch.^{1TechCrunchScale AI and its 22-year-old CEO lock down $100 million to help label Silicon Valley's data5 Aug 2019} A $325m round in April 2021 valued it at $7.3bn, and gave Mr Wang a 15% stake worth about $1bn, which made him, at 25, a billionaire.^{2NextShark, via Yahoo News25-year-old college dropout is now the world's youngest self-made billionaire1 Jun 2022} The 2024 round put the figure at $13.8bn.^{3Scale AIScale AI raises $1B Series F at a $13.8B valuation21 May 2024} Scale says it supplies data for nearly every leading AI model, and names OpenAI, Meta and Microsoft among its customers.^{3Scale AIScale AI raises $1B Series F at a $13.8B valuation21 May 2024} It also sells to governments: the same profile reports a $249m contract in January 2022 with the Pentagon’s Joint Artificial Intelligence Center, and free datasets provided to Ukraine since March of that year.^{2NextShark, via Yahoo News25-year-old college dropout is now the world's youngest self-made billionaire1 Jun 2022} The company has not disclosed revenue in the sources we have seen, so the valuation rests on faith, in the way of the best private-market valuations.

The sceptic has two objections. The first is that a labelling firm is a labour business wearing a software costume, and labour businesses have thin moats. The second is that the models will learn to label themselves. Both have force. Mr Wang has not, in the public record we have seen, said how he would answer either. Scale’s workers are contractors, and the supply of cheap, careful human judgment is not unlimited.

Yet the history of tools is kinder to the unglamorous layer than to the glamorous one. Few remember who made the cabling for the first telephone networks, but they were paid before anyone knew who would win. Scale has been paid by nearly everyone who is trying to win. That is a lesson in positioning, and a better one than the usual fable of the dropout who simply believed hard enough.

The test of his company is not whether it can grow while models need people. It is whether it can still earn a living on the day they need fewer.