The $14.3 billion investment in SAIE, a leading player in the AI data industry, was actually a very strange deal.
Meta acquired 49% of the company in a deal announced last Thursday. Scale has announced that CEO Alexandr Wang will be leaving to become an executive in charge of the new “Super Intelligence” unit within the tech giant. (The transaction has not yet been approved by the regulatory authorities.)
The deal is good news for Meta, who has been widely seen as lagging behind in the AI race and in need of new AI leadership, and at age 28 he will become one of the most powerful AI players in the tech industry as part of the contract.
However, this transaction is clearly not beneficial to the scale itself and could result in a loss of favorable business as a result of new proximity to the meta. Two of Scale's major clients and Meta's major rivals, Openai and Google, reportedly began finishing their work with Scale after the contract.
“The lab doesn't want to know what data other labs are using to improve their models,” says Garrett Lord, CEO of Handshake, who says that demand for his company's services has “tripled overnight” as a result of the meta transaction. “For General Motors or Toyota, we don't want to see how our competitors enter the manufacturing plant and carry out the process.”
Competitors of other scales say they have seen similar treatments. “Last week was totally sane,” says Jonathan Siddharth, CEO of Turing, who helps all major AI companies connect with human experts and create their own training data. Over the past two weeks, Turing has added potential contracts worth $50 million, Siddharth said, “Like Frontier Rabo recognizes that to advance AGI, we need a true neutral partner.”
“This is equivalent to an oil pipeline exploding between Russia and Europe,” explains Ryan Kolln, CEO of Appen, another AI training data company, explaining the disruption in the industry's data supply chain. “Customers are really quickly assessing it. How do you get an alternative supply?”
Kolln adds: [Meta] Getting information about what other Foundation Model Labs are doing is much more difficult to manage. ”
According to people with direct knowledge of the employment process, employees of multiple sizes signed an agreement last week to move to two rival data companies.
A Scale AI spokesman had no comment, but pointed out his comments towards a report citing Openai's Chief Financial Officer that Openai will continue to work with Scale following its meta investment. A spokesperson for Openai and Google declined to comment, but they turned their time on reports that each said they had involved work on scale. Meta and humanity did not respond to requests for comment. (Time has a technology partnership with Scale AI.)
The amounts that could ultimately change hands as a result of meta transactions are immeasurable. The leading AI companies currently spend around $1 billion a year on human data. Mainly, the data budget has not declined. Scale competitors, as Jostoll, to fill the void left by meta transactions, point out a fundamental reshaping of how the world's most valuable AI models are built.
The data industry's tide change
Scale was launched as a data labeling company marshaling the army of human contractors around the world. This is primarily in low-income countries such as India, Venezuela and the Philippines, where you get pennies for doing things like labeling images and answering simple questions.
This type of work was useful in the early stages of AI development when AI companies struggled to teach image models to communicate the differences between cats and dogs, or to teach language models to connect consistent sentences.
However, as AI models improved, the type of data AI companies wanted fundamentally changed. This shift became even more pronounced after the industry shifted to the so-called “inference” model. AIS is about writing down a line of thought before you settle for the answer. These models are superior to most people when writing code, conducting research, and answering complex science questions.
This “inference” paradigm has led us to seek primarily Openai, Google, humanity and others. Experts data. The most profitable training data is currently written by people with PhD degrees. The PhD writes down the exact steps to take while solving problems, so AI models can learn to mimic this behavior.
“The industry is changing to needing smarter, smarter people,” says Siddhath, CEO of Turing. “In some areas, even a single expert is not enough to move the needle. You need a team of experts.”
What AI companies are looking for precisely from experts is a closely guarded secret. According to insiders, all AI labs tend to go around the same strategy over time, but the more labs can keep their training process secret, the more they can spend on the industry's “frontiers” and the AI models are better than their rivals.
So, the large investment in meta scale seems to be making all frontier AI companies uneasy. Meta may be late in the AI race at the moment, but if they have access to some of their rivals' most valuable secrets, they could begin to close the gap quickly.
