On Thursday, Meta unveiled a $14.3 billion investment and partnership that will be the core of a new artificial intelligence research lab dedicated to the pursuit of “superintelligence,” a swing-for-the-fences effort in the global technology race.
The new lab will be led by Alexandr Wang, a co-founder and the chief executive of Scale AI. Meta invested in his start-up as part of an agreement to bring Mr. Wang to the company.
Meta, which owns Facebook, Instagram and WhatsApp, has been offering compensation packages as high as $100 million to leading researchers across the field in an effort to staff the new lab, according to four people familiar with the effort.
News of the lab has left many people asking what exactly a superintelligence lab is meant to do.
Here is a guide.
What is superintelligence?
Companies like OpenAI and Google want to build “artificial general intelligence,” or A.G.I., which is a way to describe a machine that can do anything the human brain can do. It is an ambitious goal with no clear path to success. Many tech leaders say a breakthrough is not far away, but it’s hard to know what that means because A.G.I. is so loosely defined.
As a result, Silicon Valley leaders are now talking about building superintelligence, which would be even more powerful than A.G.I. It, too, is a loosely defined term, but generally refers to a machine that is more powerful than the human brain.
Has anyone else built a superintelligence lab?
Ilya Sutskever, the former chief scientist at OpenAI, has co-founded a company called Safe Superintelligence. It has no plans to release products. Its stated mission is to privately build superintelligence and release the technology when it has been proved to be completely safe. Dr. Sutskever is among those who believe that powerful A.I. could harm humanity.
Are we on the verge of A.G.I.?
No one really knows when or even if A.G.I. will happen, but many are saying they are getting close to it. OpenAI’s chief executive, Sam Altman, told President Trump that it would happen before the end of his term. Dario Amodei, the chief executive of Anthropic, an OpenAI rival, said it would happen within one to two years. But there is no hard evidence.
Measuring intelligence is also notoriously difficult. We can’t even agree on how to measure the intelligence of humans — much less compare humans and machines.
Then why do people say A.G.I. is close?
As companies like OpenAI fed more and more data into their systems, they improved at a steady rate. Some technologists believe the progress will continue at much the same rate — to A.G.I. and beyond.
But is that true?
Maybe. In recent months, companies used up just about all the English language text on the internet as they improved these systems. So they turned to a new method where their systems learn by trial and error — a process called reinforcement learning.
Technologists are divided on whether this technique will lead to A.G.I. A recent paper from Apple — which followed a similar paper from researchers at Arizona State University — shows that the latest A.I. systems could become less accurate as they work through the multiple steps of problem solving. It was an indication that the metronomic improvement that the tech industry has grown accustomed to may be slowing.
“These are great tools,” said Subbarao Kambhampati, a professor and A.I. researcher at Arizona State, who was an author of the report. “But I don’t believe we are anywhere near A.G.I. or superintelligence. With these tools, you always need a human in the loop.”
Didn’t Meta already have an A.I. lab?
Yes. Meta built its first artificial intelligence in 2013, when the A.I. race was in its infancy. That fall, both Google and Meta (then called Facebook) were fighting to acquire an important A.I. start-up, DeepMind. Google prevailed, and DeepMind now forms the core of Google’s A.I. efforts. So Meta’s chief executive, Mark Zuckerberg, decided to build his own lab.
To run the lab, he tapped Yann LeCun, a pioneer in neural networks who was a professor at New York University.
What is a neural network?
A neural network, the fundamental building block of technology such as OpenAI’s ChatGPT, is a mathematical system that can learn skills by analyzing enormous amounts of digital data. By identifying patterns in thousands of cat photos, for instance, it can learn to identify a cat.
In 2013, neural networks were getting very good at recognizing both sounds and images, allowing a digital assistant like Siri to recognize words and driverless cars to recognize pedestrians and street signs. That technology eventually led to today’s chatbots.
Does Meta have its own chatbot?
Yes. It’s called Meta AI. Because of its long history in A.I., the company responded quickly to the arrival of ChatGPT. It created a new generative A.I. that pushed similar technologies across many of its online services.
Meta accelerated the development of A.I. across the industry by open sourcing its key technologies, meaning it freely shared the underlying computer code with outside software developers and businesses.
Though this is a common practice in the tech industry, many of the leading A.I. companies were reluctant to open source their most powerful technologies. Meta was not.
So why is Meta building a new A.I. lab?
The company’s latest technology, Llama 4, did not live up to expectations. It was not as powerful as rival systems from the likes of OpenAI and Google.
Meta has also been grappling with employee churn and management issues related to the technology, according to two of people familiar with the company’s struggles.
Does that really warrant a new lab?
Mr. Zuckerberg is an ambitious person. And he is worried about falling behind the other industry giants.
The new lab is an effort to attract a stable of talented researchers. Meta has been making stunning offers to researchers, in some cases promising compensation packages as high as $100 million.
But Meta may need more than just money to woo top talent. The leading researchers already make millions of dollars a year and want to work on the most ambitious projects.
That is why Meta is using the superintelligence label.
“Where these terminologies come from is anyone’s guess,” Dr. Kambhampati said. “These terms have become a commercial thing, a branding thing, an advertising thing, rather than a technical thing.”
Cade Metz is a Times reporter who writes about artificial intelligence, driverless cars, robotics, virtual reality and other emerging areas of technology.
The post Meta Is Building a Superintelligence Lab. What Is That? appeared first on New York Times.
