This summer, Meta CEO Mark Zuckerberg extended a coveted invitation to Rishabh Agarwal: join the company’s ambitious new A.I. lab, complete with a multi-million dollar compensation package of stock and salary.
Zuckerberg’s vision for this lab was clear: to forge “superintelligence,” a technology capable of surpassing the human brain. Despite the unknown path to achieving such a feat, he encouraged Dr. Agarwal to embrace the bold leap, emphasizing that in a rapidly evolving world, inaction is the greatest risk.
However, Dr. Agarwal, already a Meta employee, declined the enticing offer.
Dr. Agarwal isn’t alone. Over 20 top-tier researchers have recently departed their high-profile positions at Meta, OpenAI, Google DeepMind, and other major A.I. players. They’ve chosen to join Periodic Labs, a burgeoning Silicon Valley startup, often forfeiting tens, if not hundreds, of millions of dollars in the process.
While established A.I. labs pursue abstract objectives like superintelligence and artificial general intelligence (AGI), Periodic Labs is charting a distinct course. Their mission is sharply focused on developing A.I. that can expedite novel scientific discoveries across disciplines like physics and chemistry.
“The primary aim of A.I. isn’t to merely automate white-collar jobs,” stated Liam Fedus, one of Periodic Labs’ co-founders. “Our core objective is to accelerate scientific progress.”
Fedus was a key member of the small OpenAI team responsible for creating the groundbreaking chatbot ChatGPT in 2022. He left OpenAI in March to establish Periodic Labs alongside Ekin Dogus Cubuk, a former researcher at Google DeepMind, Google’s premier A.I. division.
Already, several prominent A.I. companies are exploring technologies to quicken scientific breakthroughs. For instance, two Google DeepMind researchers recently earned a Nobel Prize for their work on AlphaFold, a project that offers significant, albeit incremental, advancements in drug discovery.
Many industry leaders contend that large language models (LLMs), the A.I. systems powering chatbots, are on the verge of monumental scientific breakthroughs. OpenAI and Meta assert their technologies are already advancing towards this goal in fields such as drug discovery, mathematics, and theoretical physics.
“We firmly believe that advanced A.I. can accelerate scientific discovery, and OpenAI is uniquely positioned to lead this charge,” commented Laurance Fauconnet, an OpenAI spokesperson.
Yet, Fedus challenges this perspective, suggesting that current A.I. giants aren’t truly on a path to genuine scientific discovery. He characterized Silicon Valley’s approach to large language models as “intellectually lazy.” Fedus and Cubuk aspire to rekindle an era when leading tech research institutions, akin to the historic Bell Labs and IBM Research, prioritized the physical sciences as central to their mission.
(It’s worth noting that The New York Times sued OpenAI and Microsoft in 2023 for alleged copyright infringement concerning news content used in A.I. systems. Both companies have denied these allegations.)
The A.I. systems underpinning chatbots like ChatGPT are known as neural networks, a name inspired by the intricate web of neurons in the human brain. By identifying complex patterns within immense datasets of text scraped from the internet, these systems master the art of human-like language generation. They can even be trained to write computer programs and solve mathematical problems.
However, Fedus and Cubuk argue that no matter how much textual data—from textbooks to academic papers—these systems analyze, they cannot inherently master the art of scientific discovery. To achieve this, they propose that A.I. must also engage with and learn from tangible physical experiments in the real world.
“A chatbot cannot simply ‘reason’ for days and suddenly unveil a monumental discovery,” Dr. Cubuk explained. “Humans can’t do that either. Significant breakthroughs typically emerge after numerous trial experiments, and even then, success isn’t guaranteed.”
Periodic Labs, having secured over $300 million in seed funding from prominent investors like a16z, has begun operations at a San Francisco research facility. Their ambitious plan involves constructing a dedicated lab in Menlo Park, California, where physical robots will conduct scientific experiments on an unprecedented scale.
Researchers at the company will design and oversee these experiments. Simultaneously, A.I. systems will rigorously analyze both the experimental procedures and their outcomes. The ultimate goal is for these systems to independently learn and autonomously drive similar experiments.
Just as neural networks glean insights by identifying patterns in vast textual datasets, they possess the capacity to learn from diverse forms of data, including images, sounds, and even physical movements. Critically, these systems can process and learn from multiple data types concurrently.
For example, by analyzing a collection of photographs alongside their descriptive captions, an A.I. system can establish the relationships between visual elements and linguistic descriptions. It can, for instance, infer that the word “apple” corresponds to a round, red fruit.
At Periodic Labs, A.I. systems will integrate knowledge from scientific literature, direct physical experimentation, and iterative efforts to refine and enhance these experimental processes.
Consider a scenario: one of the company’s robots might perform thousands of experiments, meticulously combining various powders and materials in pursuit of a novel superconductor. Such a material holds the potential to revolutionize numerous types of electrical equipment.
Under the guidance of human scientists, the robot could select promising powders based on existing research, mix them in a lab flask, heat them in a furnace, test the resulting material, and then repeat this entire process with different material combinations.
By thoroughly analyzing this extensive cycle of scientific trial and error — effectively identifying the patterns that lead to successful outcomes — an A.I. system could, in theory, learn to automate and significantly accelerate future experiments of a similar nature.
“The system won’t achieve a breakthrough on its very first attempt; rather, it will iterate relentlessly,” Dr. Cubuk affirmed. “Through countless repetitions, our aspiration is to arrive at discoveries much faster.”
A.I. researchers have explored these concepts for years. However, the necessary computing power and other critical resources required to propel such an ambitious endeavor have only recently become widely accessible.
Despite these advancements, developing this caliber of technology remains profoundly challenging and time-intensive. Building A.I. that interacts with the physical world is inherently more complex than developing A.I. solely within digital environments.
“Will this solve cancer in two years? Absolutely not,” remarked Oren Etzioni, the founding CEO of the Allen Institute for AI. “But is it a promising, visionary bet? Without a doubt, yes.”