If you’re watching for indicators of AI market froth, the meteoric rise of micro1 is about as clear as it gets. At the start of the year, it was an AI-powered recruitment service clearing roughly $7 million annually. Now, eight months after a pivot into data annotation for AI training, it has crossed $100 million in annualized revenue and fielded investment offers at a $2.5 billion valuation (just several months ago, it closed a funding round valuing the company at $500 million).
It’s been a wild ride for the company’s 24 year-old CEO Ali Ansari, who is now on the cusp of becoming one of the world’s youngest billionaires. Should micro1 lock in or exceed the $2.5 billion valuation currently being batted around by investors, Ansari’s stake in the company — approximately 42% — would be worth over $1 billion, according to Forbes’ estimates.
Ansari first became intrigued by AI training when a large data labeling firm came to them seeking micro1’s help with recruitment.
“It was a mind-blowing project for us,” Ansari said. “We were like, why is this company hiring hundreds of engineers in two weeks? We said, holy shit, we should really focus on this market.”
So like competitor Mercor, which also started out as an AI-powered recruiting service, Ansari tilted micro1 into data labeling.
AI training —human annotation of the information used to train the large language models that undergird artificial intelligence — is one of the most high-growth sectors in Silicon Valley right now. In order for AI models to become “smarter,” they need humans to add context and meaning to the information they’re trained on. Because AI’s performance grows in direct relation to the amount of high-quality data it is able to ingest when being ‘trained,’ the market for reliable data providers has grown rapidly. Ansari estimates that the major AI labs currently spend $15 billion a year on AI training.Two years from now he thinks that number will exceed $100 billion.
It’s a burgeoning market. Depending on who you ask, data labeling will either become a massive, sprawling industry or itself be replaced by AI advanced enough to train itself. But now, it’s big business, one that’s minted four new billionaires in the past few months alone: the Mercor founders and Surge founder Edwin Chen.
Ansari won’t disclose specific customers besides Microsoft, but said micro1 works with a number of frontier AI labs and most of the “Magnificent Seven” tech companies. All of them are looking for domain experts to annotate training data on everything from customer service to investment banking. Many of these experts are highly educated and get paid between $60 to $170 an hour to grade AI outputs. A handful, like medical or finance experts, can make as much as $500 an hour.
Ansari envisions a future where nearly everybody can be an AI trainer, not just white collar experts. He’s developing a service that pays people to videotape themselves going about their daily routine, such as folding laundry, to create training data for robotics AI models.
“AI training is fundamentally affecting the economy; it’s an entirely new job sector,” he said.
According to Adam Bain, co-founder of 01A Ventures and a micro1 investor, data labeling was until recently an ‘underloved, underappreciated’ space.
“The consensus was that it was not a good space to invest in,” Bain told Forbes. “And data labeling did start out very basic, but it’s now about literally trying to find people who are smarter than the models…it’s gotten very complex.”
Investors worried data labeling would become obsolete once AI achieved artificial general intelligence (AGI) and could match human-level cognitive abilities. They were also hesitant about touching the unsexy business of managing short-term projects with large groups of contractors.
“We were concerned about the size of the prize,” said Jamin Ball, a partner at Altimeter Capital, who has not invested in any data labeling companies. “Data labeling seemed like a commodity that would have low margins at maturity.”
Ball has since changed his mind and is now actively looking to invest in data companies. Data is the “oxygen” to AI models, he said.
Succeeding as a data labeling entrepreneur requires an understanding of the AI market’s whims, which is much more difficult than Accenture spinning up an offshore call center. Ansari is great at forecasting the future, said Bain, who describes his attitude as “360-degrees: he is equally passionate about building a great company, spending time with customers, and working incredibly hard.”
Ansari moved to Los Angeles from Iran just over a decade ago, and has been dabbling in entrepreneurship since he was a teen. During middle school he built a modest business reselling textbooks on eBay; in high school he created an online math tutoring platform he was able to sell for a profit (“low six figures”) when he graduated. While running a software consultancy as a Berkeley undergraduate, Ansari was frustrated with how difficult it was to find competent offshore engineers. So he used OpenAI’s GPT-3 model to build an AI-powered recruiter that would talk to candidates for him and assess their skills. When the AI-powered recruiter exceeded $1 million in annual revenue, he shut down the software consulting firm to focus on it. Then he pivoted into data labeling.
Micro1 differentiates itself through a philosophy Ansari calls “humans first,” which prioritizes the experience of human data labelers. The process begins with an AI interview, and then a realistic job simulation exercise, so the AI trainers know what to expect.
Once hired, experts are paired with “human data managers,” typically recent graduates from top universities, whose job is to help experts navigate the world of AI training. Project managers at micro1 are partially compensated based on micro1’s “Expert Happiness Index” and the experts’ performance is assessed with detailed analytics and stack ranked.
“If the experts are happy, they produce better quality work, and the labs get a better model,” Ansari said.
While some critics find it grim that workers are training the AI that might someday replace them, Ansari argues this next job sector will actually generate an infinite amount of work for people to do, and be a boon to out-of-work blue collar workers.
He’s most excited about the data market emerging around advanced robotics, which he predicts will someday dwarf the current AI training market. For companies building humanoids, there isn’t really a pre-made corpus of training data, there is no internet to mine in service of advancing a large language model. This isn’t the kind of data you scrape from publishers or Wikipedia, it is the kind of data you need to create and capture.
To do that micro1 is shipping kits of equipment, including Meta’s Rayban glasses, to people who will create foundational datasets for robots by recording themselves performing various tasks — making their beds, repairing a leaking faucet, or putting dishes away. That is a vast and sprawling task if humanoid robots are to someday do even half of what we do. And it’s the reason Ansari is so confident in the future of the market he’s playing in.
“The only way we get to the end state is when we are able to completely model the world out perfectly, and that will never happen,” he said.
