This summer, the Meta chief executive, Mark Zuckerberg, invited Rishabh Agarwal to join the company’s new A.I. lab, offering him millions of dollars in stock and salary.
With the new lab, Mr. Zuckerberg said he wanted to build “superintelligence,” a technology that could eclipse the powers of the human brain. Though no one knew how to create superintelligence, he urged Dr. Agarwal to make a leap of faith.
In a world that is changing fast, Mr. Zuckerberg told him, the biggest risk you can take is not taking any risk.
But although Dr. Agarwal was already a Meta employee, he turned down the offer to join another company.
Dr. Agarwal is among more than 20 researchers who have left their work at Meta, OpenAI, Google DeepMind and other big A.I. projects in recent weeks to join a new Silicon Valley start-up called Periodic Labs. Many of them have given up tens of millions of dollars — if not hundreds of millions — to make the move.
As the A.I. labs chase amorphous goals like superintelligence and a similar concept called artificial general intelligence, Periodic is focused on building A.I technology that can accelerate new scientific discoveries in areas like physics and chemistry.
“The main objective of A.I. is not to automate white-collar work,” said Liam Fedus, one of the start-up’s founders. “The main objective is to accelerate science.”
Mr. Fedus was among the small team of OpenAI researchers who created the online chatbot ChatGPT in 2022. He left OpenAI in March to found Periodic Labs with Ekin Dogus Cubuk, who previously worked at Google DeepMind, the tech giant’s primary A.I. lab.
A number of leading A.I. companies are already working on technologies meant to accelerate scientific discovery. Two researchers at Google DeepMind recently won a Nobel Prize for their work on a project called AlphaFold, which can help accelerate drug discovery in small but important ways.
Leaders in the field often argue that large language models — the technologies that power chatbots — will soon achieve significant scientific breakthroughs. OpenAI and Meta say their technologies are already pushing toward this goal in areas like drug discovery, math and theoretical physics.
“We believe advanced A.I. can move scientific discovery forward faster, and that OpenAI is uniquely positioned to help lead the way,” OpenAI spokesman Laurance Fauconnet said in a statement.
But Mr. Fedus said that such companies are not on a path to true scientific discovery. “Silicon Valley is intellectually lazy” when describing the future of large language models, he said. He and Dr. Cubuk are hearkening back to a time when the tech industry’s leading research operations, including Bell Labs and IBM Research, saw the physical sciences as a vital part of their mission.
(The New York Times sued OpenAI and Microsoft in 2023 for copyright infringement of news content related to A.I. systems. The two companies have denied those claims.)
The A.I. systems that drive chatbots like ChatGPT are called neural networks, named for the web of neurons in the brain. By pinpointing patterns in vast amounts of text culled from across the internet, these systems learn to mimic the way people put words together. They can even learn to write computer programs and do math problems.
But Mr. Fedus and Dr. Cubuk believe that no matter how many textbooks and academic papers these systems analyze, they cannot master the art of scientific discovery. To reach that, they say, A.I. technologies must also learn from physical experiments in the real world.
A chatbot “can’t just reason for days and come up with an incredible discovery,” Dr. Cubuk said. “Humans can’t do that either. They run many trial experiments before they find something incredible — if they even do.”
Periodic Labs, which has secured over $300 million in seed funding from the venture capital firm a16z and others, has started its work at a research lab in San Francisco. But it plans to build its own lab in Menlo Park, Calif., where robots — physical robots — will run scientific experiments on a massive scale.
The company’s researchers will organize and guide these experiments. As they do, A.I. systems will analyze both the experimentation and the results. The hope is that these systems will learn to drive similar experiments on their own.
Just as neural networks can learn skills by pinpointing patterns in massive amounts of text, they can learn from other kinds of data, too, including images, sounds and movements. They can even learn from different kinds of data at the same time.
By analyzing both a collection of photos and the captions that describe those photos, for example, a system can grasp the relationships between the two. It can learn that the word “apple” describes a round red fruit.
At Periodic Labs, A.I. systems will learn from scientific literature, physical experimentation and repeated efforts to modify and improve these experiments.
For instance, one of the company’s robots might run thousands of experiments in which it combines various powders and other materials in an effort to create a new kind of superconductor, which could be used to build all sorts of new electrical equipment.
Guided by the company’s staff, the robot might choose several promising powders based on existing scientific literature, mix them in a laboratory flask, heat them in a furnace, test the material and repeat the whole process with different powders.
After analyzing enough of this scientific trial and error — pinpointing the patterns that lead to success — an A.I. system could, in theory, learn to automate and accelerate similar experiments.
“It will not make the discovery on the first try, but it will iterate,” Dr. Cubuk said, meaning it will repeat the process over and over again. “After lots of iteration, we hope to get there faster.”
A.I. researchers have explored similar ideas for years. But the computing power and other resources needed to drive this kind of effort have only recently become available.
Even so, developing this kind of technology is enormously difficult and time consuming. Developing A.I. in the digital world is far easier than in the physical world.
“Is this going to solve cancer in two years? No,” said Oren Etzioni, the founding chief executive of the Allen Institute for AI. “But is this a good, visionary bet? Yes.”
Cade Metz is a Times reporter who writes about artificial intelligence, driverless cars, robotics, virtual reality and other emerging areas of technology.
The post Top A.I. Researchers Leave OpenAI, Google and Meta for New Start-Up appeared first on New York Times.
