Dr. Alice Chao taught emergency medicine to students at Stanford University School of Medicine. Now, she’s teaching artificial intelligence-powered chatbots to think, diagnose, and prescribe just like her.
Chiao is part of a burgeoning new economy of subject matter experts who train AI through a process called reinforcement learning, essentially teaching models to evaluate the AI’s responses and improve them through trial and error. The services industry for AI frontier labs is growing rapidly and is estimated to be worth at least $17 billion, according to Dimitri Zabelin, senior AI analyst at PitchBook.
Chiao is one of tens of thousands of professionals working with Mercor, one of the companies that helps leading AI companies manage reinforcement learning. Melkor contracts experts in a variety of fields, from medicine, law and finance to comedy, sports and even wine. Experts can earn up to hundreds of dollars per hour by teaching AI their jobs.
“AI is going to be the new Dr. Google, the new WebMD that people go to for medical information, and I knew I needed to be a part of that to make sure the information was accurate, safe, and meaningful to the people who used it,” Chao told CNN.
AI models are trained on large amounts of data. But that training isn’t very useful without something known as “reinforcement learning.” This process involves human experts teaching the model the difference between good and bad responses. Companies like OpenAI, Google, and Anthropic are leveraging what Mercor CEO Brendan Foody described as “an army of people” to do just that.
Uncertainty about how AI will reshape various industries has reached a crescendo over the past two weeks. Software stocks plunged in early February following the release of a new tool from Anthropic that tailors models to work in specific industries, such as law and finance. Then, a viral essay by a tech company CEO containing a stark declaration about how AI could destroy jobs took the internet by storm. Others say Melkor is causing job losses, replacing stable full-time careers with gig work, and causing AI to displace human jobs.
But Chiao doesn’t see his work through Mercor as teaching AI how to do his job. Rather, she thinks it’s about ensuring that AI models are safe and capable enough to allow doctors to spend more time with patients and less time filling out forms. She believes AI will eventually be able to help doctors read scans, fill out charts, and write notes.
“Doctors were chosen because we really want to help people. We want to heal. We want to spend time talking to people, listening and actively engaging,” Chao said. “We don’t want to think that AI will take over our jobs. We’d like to think that AI will take over the aspects of our jobs that prevent us from being good doctors, good healers, and good listeners.”
When Chiao trains her AI models, she uses real-life scenarios she has encountered over her decades as both a primary care and emergency medicine physician. It involves asking questions from both the patient and physician perspective. For example, a patient may ask whether their child should see a doctor when they have a cough or fever. But the system also needs to know how to respond when presented with medical terminology, such as what a doctor would write on a health questionnaire.
She said AI models sometimes provide answers that Chao wouldn’t have thought of herself. But sometimes we think that experts like ourselves need to intervene.
“Sometimes it just doesn’t make any sense. You think, ‘Oh, this might be misleading,’ or ‘This might be alarming,’ or, ‘It’s not safe to respond to this,'” Chao said. “So I step in and say, ‘OK, here we need to create something that is safe, accurate, and applicable to the user at hand.'”
After consulting with a team of other experts in the field, Mercor experts evaluate the model’s response using the rubric you create. These responses are fed back into the model, which trains it to perform well.
Regarding AI in healthcare, Chao said patients should use today’s AI modeling tools as a starting point before consulting a doctor. This technology is not meant to replace doctors like her, who has 20 years of experience in the field.
“There’s a intuition that comes with experience, when you sit down with a patient, look into their eyes, and see something beyond the patient’s history, the values of the lab, and the words that come out of the patient’s mouth,” Chao said. “So this is where it’s really important to know that AI is not a doctor or a human.
The most popular profession Markor hires is software engineering, followed by finance, medicine and law, Melkor CEO Foudy told CNN. Mercor’s job listings are wide-ranging, seeking everything from journalists to mechanics.
However, Foody points out that not everything can be taught, and the more subjective the task, the harder it will be for the AI to master it.
One example is comedy. Markow sought to train one AI model to be funny by recruiting comedians from Harvard University’s iconic comedy publication, the Harvard Lampoon.
“They were telling jokes and writing rubrics to improve the model and make it more interesting,” Foudy said.
But the problem is that while it’s obvious to humans, it’s not so important to machines. Different people have different opinions on what is interesting.
“What we really need is more localization of how humor varies by region, and (the answer) is how do we have experts who can understand what the jokes are in all these different areas,” Foudy said.
Before Foody and his co-founder Mercor set out to help AI models do human jobs better, the company had a very different goal: helping people get jobs.
Mercor, which Foody co-founded three years ago at the age of 19 with friends Adarsh Hiremath and Surya Midha, started as a recruitment and HR platform. When the company shifted its focus to AI, the rolodex of resumes was the perfect starting point for finding the experts AI companies were looking for.
Foody said Mercor currently pays thousands of professionals more than $1 million a day and has grown its revenue run rate from $1 million to more than $500 million in less than two years. PitchBook’s Zabelin said the company is valued at more than $10 billion, adding that the high valuations of Melkor and its competitors show investors believe services such as human feedback and expert testing of AI models are becoming a permanent and essential part of building and improving AI systems.
Melkor is not the only company entering this space. Last year, Meta invested $14 billion in Scale AI, which operates in a similar space to Mercor, and hired then-28-year-old founder Alexandr Wang as chief AI officer. Other competitors such as Surge AI, Handshake, and Micro1 are helping to create a new class of young ultra-high-net-worth tech founders.
Although valuations will fluctuate, the 22-year-old Foudy and his co-founder will likely be among the youngest tech founders to make the Forbes Billionaires list since Mark Zuckerberg, who made the list at age 23.
“Obviously we were ambitious about what we wanted to do, but we never imagined anything like this, especially for it to happen this quickly. So it feels very surreal,” Foudy said.
Foudy enjoys some of the perks of being a young billionaire (he even treated his family to Super Bowl tickets). But despite growing concerns that AI will displace jobs, he remains focused on growing the business he believes will be important in shaping the future of work.
In his view, Melkor’s efforts are a step toward solving a larger problem.
“We need to cure cancer. We need to solve climate change,” he said. “And the ability for everyone to be 10 times more productive and better able to tackle these important issues would be a huge benefit to our progress as a society.”
