For years, A.I. has been considered a “black box” technology, whose inner workings remain largely mysterious even to its creators because the system is supposed to “learn” itself. But a group of Harvard researchers are on a quest to find out exactly how A.I. learns. At the Upgrade 2025 conference in San Francisco last week, Japanese IT services company NTT Group launched the Physics of Artificial Intelligence (PAI) Group, based at Harvard University’s Center for Brain Science. Just as physics reveals the laws that govern motion and energy, the group aims to identify the fundamental principles that drive A.I.’s learning and reasoning.
“To truly evaluate and solve A.I.’s black-box paradox, we need to understand it on a psychological and architectural level—how it perceives, decides and why,” Hidenori Tanaka, PAI’s group leader, told Observer at last week’s event. Tanaka is also a lead A.I. researcher at Harvard’s Center for Brain Science.
Currently, the capabilities of an A.I. are measured through an approach called benchmarking, which typically involves testing models against a set of standardized tasks or problems, such as answering scientific questions, recognizing images or playing games. Tanaka believes this method is limited. “We need to go beyond benchmarking. It’s an insult to judge A.I. models based on mere computational power and how well they solve a couple of tough problems. It completely misses the cognitive depth models may have,” he said.
More importantly, understanding how an A.I. model works better helps humans improve it. “Uncovering the root causes of an A.I. model’s initial reasoning behavior can play a key role in minimizing bias and hallucinations in upcoming systems,” Kazu Gomi, president and CEO of NTT Research, told Observer at last week’s event.
To achieve this, PAI is building “model experimental systems,” or controlled digital experiment environments through which developers can observe how an A.I. model’s learning and reasoning curve evolves over time. By crafting numerous multimodal datasets consisting of images and text across physics, chemistry, biology, math and language topics, the team is probing how A.I. interprets previously unseen concepts and retains information. PAI plans to partner with A.I. developers worldwide to improve these datasets through insights gleaned from real-world experiments.
“The goal is to give A.I. systems a structured playground. Unlike internet-scraped data, these are intentionally crafted datasets, with each data point entry serving a distinct, predefined function,” explained Tanaka. “Just like medications act on specific neurons to treat a physical condition in humans, we’re looking at how information triggers responses within A.I. models at the neural or node level.”
Tanaka has a Ph.D. in theoretical physics from Harvard. His research focuses on studying the brain’s computational principles and aligning them with physical laws.
PAI is a spinoff from NTT Research’s Physics and Informatics Lab, established in 2019, where Tanaka previously led a group researching machine learning and sustainable data center solutions. The group has delivered over 150 papers in A.I. research. PAI’s previous research on neural network pruning algorithms has been cited over 750 times.
PAI’s core team includes multiple Harvard researchers. It also collaborates with Harvard neuroscientist Venkatesh Murthy, Princeton professor and former NTT Research scientist Gautam Reddy, and Stanford’s Surya Ganguli. Ganguli is a senior researcher at Stanford’s Institute for Human-Centered Artificial Intelligence (HAI), led by the A.I. pioneer Fei-Fei Li, and has co-authored several papers with Tanaka.
