Point72 Gift Funds AI Research Program at Caltech
A $3 million gift from Point72 Asset Management establishes a five-year research program at Caltech aimed at building AI tools for scientific discovery.
When Jerrell Watts (MS '96, PhD '98) arrived at Caltech as a graduate student, neural networks were more promise than practice. As an undergraduate at UT Austin, he had spent a summer trying to train a neural network to sort plastic for recycling. The work ran headlong into the hardware limits of the era: the computers were simply too slow, and their memories too small, to train networks capable of solving interesting problems.
Three decades later, that constraint has evaporated. Watts, who earned his MS and PhD in computer science and is now head of algorithmic trading at Point72 Asset Management, has committed $3 million through Point72 to establish the Point72 Artificial Intelligence Research Program at Caltech.
"Now is the time to push AI and machine learning models into more areas, make them more transparent and interpretable, and enable human-AI collaboration in problem solving," says Watts, who has spent the last two decades watching machine learning reshape his own field.
The five-year gift will fund graduate students, computing resources, and large-scale data production. It also advances the Caltech Initiative on the Computing and Information Sciences and Computing in Science (CIS²) which seeks to leverage AI across every discipline and within the Division of Engineering and Applied Science (EAS).
Creating a Runway for AI
"The EAS division embarked on CIS² to accelerate research by encouraging scientists and engineers to collaborate on new AI modalities," says Harry Atwater, the Otis Booth Leadership Chair for the division. "The Point72 fund creates a transformative opportunity at the heart of that effort to develop new AI tools that will catalyze scientific discovery at Caltech and throughout the greater scientific community."
Pietro Perona, the Allen E. Puckett Professor of Electrical Engineering and director of Information Science and Technology, grounds that bet in a broader light. "AI cannot be successful as a standalone discipline," he says. "It has to be successful by impacting the world in some practical way. Caltech is very much about science, and the marriage of AI with the sciences is where we're looking."
Helping to shape the program's research agenda is Yisong Yue, a professor of computing and mathematical sciences. He frames the opportunity in terms of scale. Scientists across Caltech and JPL are collecting enormous amounts of data, he says, and most of it goes unused or is not utilized to its fullest extent. "We really need computational tools to help scientists manage that," Yue says.
Pursuing Three Lines of Research
The Point72 program will support three lines of research. The first focuses on training AI models to work with the kinds of data scientists actually collect—not text, which powers tools like ChatGPT, but video, sensor readings, spectral data, and molecular structures. The goal is to make these "multimodal models" more efficient, requiring less data and computing resources to train.
The second project may be the most immediately practical. Yue's group is building AI agents that scientists can teach their domain-specific expertise by conversing with them. In one experiment, one such agent cut a data-cleaning process from weeks to about an hour. The scientist spent 30 minutes to an hour showing the AI examples of neurological data and explaining the nuances. The AI then stored that knowledge, wrote a program consistent with it, and ran the program over the entire dataset.
Watts discussed this work with Yue during a recent visit to Caltech and recognized the parallel to his own field. In finance, he says, large language models can process decades of text data that would take teams of people weeks or months to get through. "It's that multiplier effect," he says, "and then to do so with a similar accuracy as the human would do, or even maybe greater accuracy."
These "data technician" agents also carry a subtler benefit, according to Yue. Because each scientist's AI retains a trace of how they were taught to process data, those traces can be compared across labs. "I can actually compare what each AI system has learned from each scientist," Yue says. He can surface cases where researchers unknowingly disagree on how to handle the same data—a potential tool for addressing the reproducibility challenges that have dogged fields from psychology to biomedicine.
The third project ventures into less charted territory: designing systems where multiple AI agents work together. For example, Caltech researchers are studying how to build collaboration protocols and contracts for communities of AI agents that draw on the same principles that govern how humans coordinate in markets and institutions.
Room to Explore
Together, these projects point toward what Watts and Yue describe as the automated discovery of scientific laws. In finance, Watts says, machine learning models are often black boxes. They produce correct outputs, but no one can explain how they got from inputs to answers. That opacity becomes a serious problem when every day presents circumstances the model has never encountered before. What Point72 is looking for, he says, is AI that can deliver interpretable models of reality, something closer to a formula that can be understood, analyzed, and applied beyond the data it was trained on.
Perona sees the same aspiration from the research side. A deep network trained on observations of falling objects can approximate the relationship between height and velocity, he explains, but it doesn't know the relationship is simple, and it doesn't know why. If you've only dropped tennis balls from a two-story building, the network has no idea what happens from a hundred stories up. "If you know the laws of physics," Perona says, "we can predict events that we've never seen before."
Perona credits Watts with a willingness to fund research without specifying predetermined outcomes and giving Caltech researchers room to explore. It's a model of philanthropy, Perona says, that he hopes other alumni and donors will follow: pick an interesting area and then let faculty and students pursue their ideas.
For Watts, the bet traces back to his graduate school days. The lab he worked in at Caltech was extremely collaborative, he says, and it taught him the value of being patient with research. Even in finance, which is fast-paced, research can take months or years to reach fruition.
"I've tried to exhibit the same patience," he says, "in allowing research time to develop."