The Billionaire, Age 27, Powering the Boom in AI --- Alexandr Wang's Scale AI deploys a global army of gig workers to shape how the big AI models behave

"Alexandr Wang became one of the world's youngest self-made billionaires by building a sprawling army of more than 100,000 contractors, who perform the grunt work that powers the modern AI boom.

Sitting behind computers in cities across the world, his startup Scale AI's workers type out the stories, label the images, and craft the sentences that furnish chatbots with the text they need to better understand human speech patterns. Dubbed data labeling, their tasks range from composing haikus and summarizing news articles to writing stories in languages like Xhosa or Urdu.

The labor-intensive operation has become so in demand by businesses eager to enter the AI race that Scale's revenue pace tripled last year, boosting its valuation to $14 billion. Wang's stake is estimated at about $1.7 billion. The 27-year-old founder likens his company's importance in the AI revolution to the computing chips developed by Nvidia -- currently the hottest tech company in the world.

Inside Scale's 180,000-square-foot office in San Francisco, employees recruit labelers by posting advertisements on websites like Reddit and LinkedIn, touting the benefits of flexible remote work. Others review batches of data, or monitor issues around onboarding workers and paying them.

The startup's contractors often don't know they are working for Scale. They complete tasks through two websites, called Remotasks and Outlier, neither of which publicize their affiliation with the startup. Projects are given code-names so workers don't know which clients they are labeling for. Scale typically uses animals as code names, such as Ostrich for OpenAI and Bee for Apple, though it has recently gotten more creative. Google's code name, for example, is boba, after the Asian bubble tea drink.

Managing this empire is difficult. Some workers quit, saying they are frustrated with delayed payments and gig-work wages, which can run as low as $8 an hour. Others have found ways to cheat at their jobs to increase their productivity and make more money. 

At times, the data they produce is so low-quality that Scale's own employees -- even executives and top engineers -- have had to redo it themselves.

Wang, who is Scale's chief executive, says he's trying to move beyond it. He called the business "mundane and unsexy," and says he's developing a new suite of software products, such as tools to help businesses create AI applications, that typically command higher margins and are more favored by the venture-capital crowd.

Meta's code name is Flamingo -- a stuffed version of which sat atop an employee's desk on a recent visit to the startup's headquarters. After Scale AI bungled a project last year for the tech giant, Wang declared a company emergency and launched an all-hands-on-deck effort to fix the job, called Flamingo Revival, according to former Scale employees.

Early last year, Meta Platforms asked the startup to create 27,000 question-and-answer pairs to help train its AI chatbots on Instagram and Facebook.

When Meta researchers received the data, they spotted something odd. Many answers sounded the same, or began with the phrase "as an AI language model . . ." It turns out the contractors had used ChatGPT to write-up their responses -- a complete violation of Scale's raison d'etre.

The researchers communicated the disappointing results to Scale, prompting Wang to rally the entire company to try and save the contract.

He asked employees to drop everything and create new writing samples to send to Meta. An internal leaderboard showed who had completed the most labeling tasks. The prize for the winner: a paid vacation.

Later, Scale discovered that much of the bad data sent to Meta had come from Kenyans who had become experts in making a quick buck off the Remotasks platform, the former employees said. 

Scale restricted several new labeling projects to workers based in the U.S. and other wealthy, English-speaking countries.

The change didn't stem the fraud entirely: Some foreign workers found ways to skirt the new rules by buying labeling accounts registered to U.S. residents that they found for sale in group chats on WhatsApp and Facebook.

A Scale spokeswoman said that Scale has cracked down on such activity and the percentage of its freelancers exhibiting fraud fell to under 0.1% in July.

"We have a strong relationship with Scale AI and look forward to continuing to build on it in the future," a Meta spokeswoman said.

Wang chased success on an accelerated timeline. He was born in Los Alamos, New Mexico, to Chinese immigrant scientists working at the Los Alamos National Laboratory. He has said they began teaching him advanced physics starting in kindergarten.

Wang started making plans to create a company in the ninth grade when he and a friend made a Google Doc of startup ideas, he wrote in a blog post titled, "What I learned in 2016," the year he started Scale.

As a teenager, Wang debated, played violin, and traveled the country for math and physics competitions. He left high school a year early to work in Silicon Valley, where he became an engineer at the question-and-answer site Quora. Later, he wrote in the blog post about how his 12-hour days there taught him the importance of hard work.

Wang said he first began to see the groundbreaking potential of AI after attending a summer camp in San Francisco called SPARC, the Summer Program for Applied Rationality and Cognition. It was designed to bring together talented math and science students. His time there as a teen introduced him to early AI researchers like OpenAI co-founder Greg Brockman and Dario Amodei, the CEO of Anthropic, and inspired him to learn about the technology, he said.

In 2015, Wang went to MIT. As a freshman he juggled five graduate computer science courses and spent his spring break working on an iPhone app called Ava to help users book doctor's appointments. Soon after, he dropped out.

Wang made the decision after a conversation with Eric Wu, the CEO of the online house-flipper Opendoor. Wu tried to convince Wang to accept a job at the infant startup instead of working at established tech companies, encouraging him to take risks while he was young. The advice convinced Wang he needed to start a company himself. "I knew I would regret it if I never took the risk to be an entrepreneur at the perfect time," he wrote.

In the summer of 2016, Wang enrolled Ava in a startup training program run by the prestigious venture firm Y Combinator, then run by Sam Altman, who is now the CEO of OpenAI. Soon after, he and his co-founder Lucy Guo broadened the idea to Scale AI. Wang and Altman lived together during the pandemic in San Francisco, and Altman ended up indirectly owning a stake in Scale through his deal with Y Combinator, which retains equity in the startups it nurtures.

Scale AI got going amid a boom in funding for self-driving cars. Developers of the technology needed labeled images to help autonomous vehicles recognize objects like stop signs and pedestrians, and turned to the startup for help.

A few months after its founding, Scale signed up Cruise and Tesla as early customers, while the startup set out to build its network of contractors. In 2017, Wang created Remotasks, a subsidiary that focused on recruiting cheap labor abroad. At one point, Scale set up facilities in Africa and Asia to train data labelers.

To recruit some of Scale's first contractors, Guo joined Filipino remote work groups on Facebook, sharing a quiz she created.

Scale soon recruited hundreds of contractors through online chat groups. Many came from the Philippines, where groups of labelers worked in internet cafes, playing video games while completing assignments on Remotasks.

Back in San Francisco, Wang ran Scale like a Silicon Valley startup. He created company mantras like "Why Not Faster" and "Run Through Walls," and courted venture-capital backers including early Facebook investor Accel and Peter Thiel's Founders Fund. In 2019, Scale hit a $1 billion valuation.

When the self-driving car boom slowed, Wang sniffed out other revenue sources. In 2019, he signed Scale's first generative-AI contract, with OpenAI, to label data for an early version of the language model behind ChatGPT. The following year, he signed a deal with the U.S. Army to help it build data sets for its AI efforts.

Wang "is very good at seeing around the corner," said Mike Volpi, a retired partner at Index Ventures who sits on Scale's board. "He has multiple entrepreneurial and visionary skills that are blended into a single person."

During the height of the pandemic, Scale negotiated a roughly $40 million contract with Meta to label data for new shopping features on Facebook and Instagram. Employees working on the project were so burned out that the company sponsored a paid vacation to Cancun, Mexico, in mid-2022 to help them unwind.

A few months later, Meta canceled the deal after struggling to grow its e-commerce business. In January 2023, Wang laid off 20% of Scale's staff, saying that many industries Scale served were hurt by the tech downturn.

The company bounced back as tech giants poured billions of dollars into generative-AI projects to catch up with OpenAI after its viral launch of ChatGPT.

Last year, the startup negotiated a roughly $120 million contract with Google to assist on its Gemini language model. Its annual pace of revenue shot up from $227 million to $680 million, according to investor documents.

Fresh off a record $1 billion fundraising round, Wang opened a new office in San Francisco this summer -- taking over a space formerly occupied by Airbnb. It has the look and feel of a big tech company: floors built around a plant-filled atrium with modern-style furniture and a coffee bar staffed with a full-time barista.

Earlier this year, Wang made an appearance at the Met Gala in New York City, before attending the invitation-only Allen & Co. conference in Sun Valley, Idaho, where he schmoozed with executives like media mogul Barry Diller.

"Truly no one flies in and out to more parties than you, it looks like a full-time job," OpenAI's Altman posted on X in February responding to Wang's post, which said that after traveling for the past few years, he'd decided the most engaging people are at home in the Bay Area.

Competitors ranging from legacy data-labeling outfits to hungry startups are racing to steal Scale's business, offering the same formula for supplying contractors on-demand. Customers like OpenAI are also recruiting their own data labelers to reduce their dependence on the startup.

"If they want to be a $50 billion company, they've got to find another act, another arc to their story," said Vince Hankes, a partner at Thrive Capital, a long time Scale investor.

---

How Scale AI works

One example of a task handled by a contract employee

1. A worker in the Philippines applies for a job to become an AI writing evaluator at Outlier, a unit of Scale AI.

2. The worker is hired and assigned to write text that the model can use to learn how humans think. Sample assignments could include creating and answering such questions as: "Explain the moon landing to a 6 year old," or "Write a story about frogs," or "Give me some books about humanity's first contact with aliens."

3. Scale AI provides the writing samples to customers, like Google and Meta, to train their AI models, with the goal of teaching them how to converse in more helpful and natural ways.

4. The worker is paid, averaging $8/hour in the Philippines, according to Outlier." [1]

It all sounds like a primitive scam. It is amazing that no one in Lithuania thought of doing this. We also have boba, but without bubbles.

1. EXCHANGE --- The Billionaire, Age 27, Powering the Boom in AI --- Alexandr Wang's Scale AI deploys a global army of gig workers to shape how the big AI models behave. Berber, Jin.  Wall Street Journal, Eastern edition; New York, N.Y.. 21 Sep 2024: B.1.