Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Dystopian thriller “Soylent Green” ends with the hero discovering that the mystery food that sustains New York’s citizens, and whose name the movie bears, is actually made of human beings. Reports that Meta Platforms may acquire 49 per cent of artificial intelligence company Scale AI for $15bn illustrate a similar plot twist. AI, too, is people.
So far, building AI has mostly been an exercise in managing scarce resources of the inanimate kind. Top-end chips, for example, are in short supply; power constraints weigh. Most of Meta’s capital expenditure, estimated at up to $72bn this year, has been earmarked for servers and data centres, the tech equivalent of bricks and mortar.
But other ingredients are becoming scarce too. One is decent data, essential for training AI models effectively. OpenAI co-founder Ilya Sutskever has warned of reaching “peak data” in the same way the planet is heading towards “peak oil”.
Scale AI is a company that tries to defer the moment when the data runs out. It labels, cleans up, categorises and corrects so-called “tokens” and the models that use them. Founder Alexandr Wang says freely available data has reached its limits, and what comes next will be trickier to translate into a form these algorithms can use.
This is a problem for Meta boss Mark Zuckerberg too. Privacy is one issue the company is tackling: European users, for example, can deny the company the right to use their public Instagram posts to train models, though their friends in the US cannot. Online forum Reddit is suing AI model-maker Anthropic for “scraping” user posts without permission.
Part of the solution, as Meta and its peers know, is more people. Scale AI is an advocate of “reinforcement learning from human feedback”, or RLHF, where actual employees course-correct and coach models. Wang reckons algorithms work best when they come with a person, and will for the foreseeable future — a hybrid approach sometimes known as “centaur AI”.
That may explain why Meta would consider not just buying half of Scale AI, but also plucking out its founder to head a new “superintelligence” division. The label would fit not just the product, but the process. Getting AI to the commanding heights involves not pairing it with not just humans, but very smart ones — leading scientists and great thinkers — to learn not just what brainiacs know but how they reason.
Granted, $15bn is a lot of money to lure some clever people. Scale AI was expected to make $2bn of revenue this year, The Information reports. As reported, Meta’s offer would value the whole company at 15 times that.
Then again, this is a trifle for a company expected to dole out $44bn in dividends and share buybacks this year, according to LSEG. If Zuckerberg is the first to reach AI supremacy, unlocking a market potentially worth trillions of dollars, $15bn would seem a pittance. AI was already a bidding war for chips, power and data; human neurons now join that list.
