Unlock Editor’s Digest Lock for Free
The dystopian thriller Soylent Green ends with the hero discovering that the mysterious food that supports New York’s citizens and maintains the film’s name is actually made of humans. The Meta Platform reports that it could acquire 49% of the artificial intelligence company scale AI for $15 billion. AI is also a person.
So far, building AI has been an exercise to manage rare, inanimate resources, mostly inanimate. For example, there is a shortage of chips on the top end. The weight of the power constraint. The majority of Meta’s capital expenditure, estimated to be up to $720 billion this year, is allocated to servers and data centers, the technology equivalent to brick and mortar.
However, other ingredients are also scarce. One is decent data and is essential for effectively training AI models. Openai co-founder Ilya Sutskever warns that the planet will reach “peak data” just as it is heading towards “peak oil.”
Scale AI is a company that seeks to postpone the moment when data is gone. Label, clean up, classify and modify the so-called “tokens” and the models that use them. Founder Alexandr Wang says that freely available data has reached its limits and that what comes next is difficult to convert these algorithms into a form that can be used.
This is also a problem for meta boss Mark Zuckerberg. Privacy is one of the issues that companies are working on. For example, European users can deny the company the right to train models using public Instagram posts, even though they can’t make US friends. Online forum Reddit is suing humanity, an AI model maker, to “scrape” user posts without permission.
As Meta and his peers know, part of the solution is more people. Scale AI is a proponent of “reinforcement learning from human feedback” or RLHF, where real employees are course correction and coaching models. Wang believes that algorithms work best when they come with people.
That might explain why Meta considers not only buying half of the scale AI, but also leading the founder a new “super intelligence” division. The labels fit the process as well as the product. Leading AI to commander involves combining pairs with not only humans but also with very clever scientists and great thinkers, not just what the Brainiacs know, but how they reason.
Certainly, $15 billion is a lot of money to seduce wise people. Scale AI was expected to generate $2 billion in revenue this year, information reports. As reported, Meta offers are 15 times more than that to value the entire company.
john.foley@ft.com
