AI and openness at CERN: FirstPrinciples demos the AI Physicist at the Open Science Fair

FirstPrinciples presented an early demo of its AI Physicist at the Open Science Fair, held this year at CERN. The event sparked conversations on trust, openness, and the role of AI in research, underscoring how collaboration will shape the future of discovery.

A couple of weeks ago at CERN, FirstPrinciples presented an early demo of the AI Physicist at the Open Science Fair, a meeting ground for researchers, policymakers, and open-data advocates reshaping how science is done. Our modular platform for domain-specialized reasoning in physics became our own experiment in open science.

If the IAIFI Summer School was about training hybrid scientists, the Open Science Fair was about something broader: rethinking the conditions of science itself. Who can access it? How is it built? What does openness mean in an age of billion-parameter models?

This year’s theme, “Fusing Forces – Accelerating Open Science through Collaboration,” set the tone. In an era when global challenges demand collective intelligence, the Fair spotlighted how partnerships across borders, disciplines, and institutions can accelerate discovery, improve knowledge accessibility, and build a more transparent and equitable research ecosystem.

Open Science Fair 2025 Banner (Credit: Open Science Fair)

The new gatekeepers of knowledge in the Age of AI

For Matthias Le Dall, FirstPrinciples’ Lead Data Scientist, that theme resonated with both urgency and history. “Historically, science has never been open,” he explains. “In antiquity, knowledge was held by the few select people who were educated enough to know how to read and write, who also usually happened to be the wealthy. In more modern times, as science and literacy spread, knowledge has been concentrated in labs and research institutions, which require funding to run. Now AI introduces a new kind of gatekeeping: not knowledge itself, but one of large-scale and highly costly compute infrastructure.”

FirstPrinciples' Matthias Le Dall and Nawar Ismail at “Wandering the immeasurable" – a 15-tonne sculpture next to the Globe of Science and Innovation at CERN

Building and training powerful AI systems now requires financial investment on a scale that rivals major experimental facilities. “The question becomes: how do you make it open when such upfront investment is required?”

Another tension also emerged. In essence, machine learning models are data compressors transforming complex datasets into simple predictions.

Fundamentally, that sits uneasily alongside the deeply rooted (and deeply necessary) culture of attribution throughout science. “Machine learning models blur the lines of the origin of the data they learn from", Le Dall notes, “which is quite at odds with how scientists cite each other's works."

The AI Physicist meets healthy skepticism

Against this backdrop, FirstPrinciples presented the AI Physicist as an experiment in openness:  a system designed not for engagement or profit, but for physics itself.

Healthy skepticism met the demo at first. Physics is hard, after all, and the idea of automating parts of it pushes at long-held boundaries. But once the conversation turned to concrete methods — symbolic regression, equation testing at scale, transparent reasoning steps — skepticism shifted to curiosity.

“We explained how you can, in principle, extract mathematical intuition from LLMs,” says Nawar Ismail, who co-presented the demo. “Imagine asking a thousand models to propose a theory, then testing each against real data. Even without knowing it, the model can converge on something that describes the data.” That’s the kind of rigorous experimentation we’re building into the AI Physicist itself.

“Many people were talking about ‘bad agents.’ We wanted to show what a ‘good agent’ might look like:  one built solely for studying physics.” - Nawar Ismail

The framing of AI not as a chatbot but as a scientific agent helped set the demo apart and sparked conversations well beyond physics. Attendees floated applications in geoscience, remote sensing, and design optimization, hinting at a wider appetite for AI tools grounded in domain expertise.

Not all feedback was enthusiastic and some attendees raised fears about AI replacing scientists. “A few years ago, people thought AI would only replace mechanical jobs, not creative or scientific ones,” Le Dall notes. “This event showed that fear is now moving into research fields too.” But the prevailing tone at CERN was constructive rather than defensive, and FirstPrinciples engaged in many passionate debates about quality control, attribution, and funding alongside a clear willingness to share ideas.

Matthias Le Dall presenting the early AI Physicist at Open Science Fair 2025, hosted at CERN.

Where skepticism meets collaboration

Those conversations mirrored the AI Physicist’s own architecture: modular, transparent, and built for integration rather than lock-in. One promising lead came from an astrophysicist working on similar model fine-tuning; the team is now planning a follow-up collaboration.

Knowledge graphs emerged as a common bottleneck across disciplines, underscoring that the challenges of AI-driven science aren’t unique to physics.

“We’re going to face headwinds twice: first from scientists whose job is to be skeptics, and then from the unconventionality of doing science with AI.” - Matthias Le Dall

The Fair also left the team with harder questions. “The consensus from conversations is that papers written by AI would be considered as spam, given the flooding of potentially thousands of papers.” Le Dall recalls. “But one of our key goals is to create an AI that can automatically write papers.” 

Yet the optimism held. If the future of science depends on openness and collaboration, then the challenge is to embed those values into the tools we’re building today.

For FirstPrinciples, the Open Science Fair affirmed why openness isn’t just an ideal but a necessity. We operate across borders and disciplines because curiosity and exploration benefit everyone, not just a select few. By building tools like the AI Physicist, we’re working toward a research ecosystem that is transparent, collaborative, and globally accessible. In our view, openness is not simply an ethical stance; it’s an accelerant for discovery itself.

Building openness from the start

At CERN, that vision felt within reach. The Open Science Fair showed how much common ground exists among communities that don’t often meet — AI researchers, open-data advocates, physicists, and policymakers. Each sees different problems but the same opportunity: to build openness in from the start, rather than bolt it on later.

Artistic rendition of "Wandering the immeasurable" sculpture next to the Globe of Science and Innovation at CERN.

For FirstPrinciples, the AI Physicist isn’t just a physics engine. It’s a test case for how domain-specialized AI can be developed transparently, in collaboration with the people it’s meant to serve.

As science becomes more complex, its future may hinge not only on how powerful our models are, but how open and interoperable we make them. The conversations at CERN revealed something encouraging: the scientific community isn't just ready for this challenge, it's hungry for it. The skepticism is there, but so is the recognition that the biggest questions ahead can't be solved in isolation. The tools we build today will determine whether AI accelerates discovery for everyone, or just for those who can afford the compute. At CERN, the choice felt clear.

This article was created with the assistance of artificial intelligence and thoroughly edited by FirstPrinciples staff and scientific advisors.