The investment of $ 14.3 billion in Meta in the AI scale, the main actor of the AI data industry, was indeed a very strange case.
Meta acquired 49% of the company in the agreement announced last Thursday. Scale has announced that his CEO, Alexandr Wang, has ceased to become a framework in charge of a new “superintendent” unit inside the technology giant. (The agreement has not yet received regulatory approval.)
The agreement was good news for Meta, which was largely considered to be late in the AI breed and needed a new AI leadership, and for Wang, who, at 28, will become one of the most powerful AI players in the technological industry in the agreement.
But the agreement was less obviously beneficial for the scale itself, which is likely to lose lucrative affairs because of its new proximity to Meta. OPENAI and Google, two main customers of Scale and the main competitors of Meta, would have started to finish their work with Scale in the wake of the agreement.
“Laboratories do not want other laboratories to determine the data they use to improve their models,” said Garrett Lord, CEO of Handshake, a competitor on a scale, who says that demand for his business “tripled overnight” following the meta-affair. “If you are General Motors or Toyota, you don’t want your competitors to enter your manufacturing plant and see how you run your processes.”
Other competitors on a scale say they have seen a similar burst of disagreement. “Last week was completely crazy,” said Jonathan Siddharth, CEO of Turing, a company that helps all large AI companies connect with human experts to create proprietary training data. In the past two weeks, Turing has added potential contracts worth $ 50 million, according to Siddharth, “as Frontier Labs recognizes that the advancement of the act requires really neutral partners.”
“This is the equivalent of an exploding oil pipeline between Russia and Europe,” said Ryan Kolln, CEO of Appen, another AI training data company, describing the disruption of the industry’s supply chain. “Customers evaluate very quickly: how to get an alternative offer?”
Kolln adds: “Now Meta being such a large owner of scale, the ability to [Meta] To obtain information on what other foundation model laboratories do becomes much more difficult to manage. »»
Employees with several scales signed contracts to move on to two rival data companies last week, according to people with direct knowledge of the hiring processes.
A spokesperson for the AI scale had no comments, but pointed out to a report that quoted Openai’s financial director by saying that Optai would continue to work with the Meta investment. The spokesperson for Openai and Google refused to comment, but everyone has pointed out of the reports indicating that they have finished their work with Scale. Meta and Anthropic did not respond to requests for comments. (Time has a technological partnership with the AI scale.)
The amount of money which could ultimately change hands following the Meta agreement is immense. Each of the main companies of AI now spends about $ 1 billion on human data per year, according to Lord – and their data budgets increase and do not decrease. While Scale’s competitors are jostling to fill the vacuum left by the Meta agreement, the company’s drama indicates a fundamental remodeling of the construction of the most precious AI models in the world.
Change of tides in the data industry
Scale made its debut as a data labeling company, the armies of rallying human entrepreneurs from around the world – mainly in low -income countries like India, Venezuela and the Philippines – which would gain such as task to do things like labeling images or answer simple questions.
This type of work was useful in the first stages of AI development, when IA companies still had trouble teaching image models to make the difference between cats and dogs, or teach language models to chain coherent sentences.
But as AI models have improved, the type of data that AI companies are looking for have radically changed. This change has become even more pronounced after the industry has moved to so -called “reasoning” models: the AI that note a thought before setting up on an answer. These models are now better than most humans in code writing, carrying out research and answers to complex scientific questions.
This paradigm of “reasoning” led Openai, Google and Anthropic to seek mainly expert data. The most lucrative training data is now written by people with doctorate, who write the exact steps they take while solving problems, so that AI models can learn to imitate this behavior.
“The industry moves to the need for more intelligent and smarter humans,” said Siddharth, CEO of Turing. “For some areas, even a single expert human is not enough to move the needle. You need a team of expert humans.”
What each company of AI exactly asks its human experts to do is a closely kept secret. All AI laboratories tend to converge the same strategies over time, say the initiates, but the more each laboratory can keep its secret training processes, the more time they can spend time at the “border” of industry, with their model of AI better than their rivals.
This is why Meta’s great investment in the scale seems to have annoyed all the companies of the border AI. Meta can currently be late in the AI race – but if he can access some of the most precious secrets of her rivals, there is a chance that he can start to quickly fill the gap.
