Fast Company
Meta shifts from metaverse dreams to chasing AI “superintelligence,” aiming to build AI better than humans at all tasks. Led by Alexandr Wang and packed with top talent, Meta’s new lab reflects a broader tech obsession. Yet definitions vary, practical uses remain limited, and real-world impact is still unclear.
Don’t take the race for ‘superintelligence’ too seriously
Author: Harry McCracken
The technology industry has always adored its improbably audacious goals and their associated buzzwords. Meta CEO Mark Zuckerberg is among the most enamored. After all, the name “Meta” is the residue of his 2021 rebranding of the company formerly known as Facebook Inc. That was supposed to emphasize an utter commitment to building the metaverse—an ambitious yet fuzzy concept blurring aspects of virtual reality, augmented reality, and social networking.
Except Zuckerberg’s goals have already drifted. Now he’s all-in on attaining an advanced form of AI called superintelligence. It’s the focus of a new organization called Meta Superintelligence Labs, which Meta is stocking with some of the planet’s top AI talent. Whether the company is really dangling $100 million offers to snag some potential recruits remains unclear. But MSL is led by Alexandr Wang, who joined Meta in June as part of a deal involving it investing $14 billion in his startup, Scale AI—an eye-watering sum all by itself.
The term superintelligence isn’t new (it was the title of a 2014 book by philosopher Nick Bostrom), but it may turn out to be 2025’s buzzword of the year. A few more data points:
- OpenAI CEO Sam Altman recently declared that his company “is a lot of things now, but before anything else, we are a superintelligence research company.”
- One of Altman’s OpenAI cofounders, famed AI scientist Ilya Sutskever, left a year ago to cofound a new startup called Safe Superintelligence, which just lost another of its founders, Daniel Gross, to Meta’s new lab.
- Last month, when Microsoft said it had trained AI to diagnose disease more accurately than doctors can, its AI CEO, Mustafa Suleyman, called the feat “a genuine step toward medical superintelligence.”
So what is superintelligence? In an excellent piece on its emergence as a Silicon Valley obsession, Bloomberg’s Shirin Ghaffary defines it as “AI that is not just at parity with most people, but even better than all humans at all tasks.” Ghaffary says the industry is gravitating toward superintelligence because it’s a more tangible goal than artificial general intelligence, or AGI. Google DeepMind chief AGI scientist Shane Legg, who popularized that term, defined it for me as “something that can at least match human capability in the sorts of cognitive tasks that people can typically do.”
“Even better than humans at all tasks” would seem to be a substantially loftier ambition than merely “matching human capability” in “tasks that people can typically do.” But over at IBM.com, an article says we definitely haven’t achieved AGI but arguably have achieved superintelligence. That’s because the IBM piece’s far less sweeping definition of superintelligence only involves it outperforming humans at certain jobs, not all of them. It cites four existing examples: IBM’s own Deep Blue (chess) and Watson (Jeopardy!) along with Google DeepMind’s AlphaGo (Go) and AlphaFold (protein structure predictions).
All of this leaves me with more questions than answers:
- If IBM’s four examples aren’t enough to confirm that superintelligence has been attained, how many would be?
- Who gets to decide what “better than humans” means?
- What happens if AI sails past humans on some fronts while remaining stubbornly behind on others?
- Don’t many of our skills involve the ability to interact with the physical world, making surpassing them as much about robotics as software? (Meta is working on that, too.)
This we do know: Like AGI before it, superintelligence is inherently aspirational. It exists as a notion in part so that groups of people have something to race toward. In that sense, it bears some resemblance to past moonshots such as, well, NASA’s Apollo program. The difference is that superintelligence offers no well-defined end point akin to landing a human being on the moon, which you’ve either accomplished or you haven’t. There will be no superintelligence equivalent of Neil Armstrong setting foot on the lunar surface on July 20, 1969.
Back in March 2024, I wrote about the pointlessness of fixating on what AGI is and when it might be achieved. Swapping in superintelligence as fodder for this debate doesn’t accomplish anything. Yes, we all need to gird ourselves for a world in which AI competes with people for jobs. I don’t discount the possibility of it presenting existential risks to humanity. It will unquestionably cause trouble of types yet to be identified. And—fingers crossed—it may help with some of our thorniest unsolved problems.
These points remain true no matter how you define superintelligence and regardless of whether it’s ever reached. Which means that fixating on the race for it isn’t a terribly productive way of readying ourselves for AI’s future impact on our lives.
Meanwhile, AI companies still have a spotty record at figuring out practical applications for AI in its present, less-than-superintelligent state. That’s especially true for Meta, which has larded my Facebook feed with clueless “Meta AI” interjections wholly incapable of grasping the conversations they’re trying to join. Meta getting better at using the AI it already has would be an encouraging sign that its quest for superintelligence isn’t just raw, unbridled ambition in search of actual useful purpose.
Credits: TCA, LLC.