Featured Conversation
Calculation, Not Cognition: A Particle Physicist on the Limits of AI's Next Decade
A conversation with Tommaso Dorigo on why particle physics holds AI to a far stricter standard than medicine does, what 'deployment' means when a model is just a calculation tool, and why he doesn't expect AI to solve AGI in the next decade.
Editorial
This month, we stepped outside medicine to hear from a discipline that has quietly been using AI for far longer, and under far stricter rules: particle physics. Tommaso Dorigo has spent his career holding models to a standard of statistical proof that clinical medicine rarely approaches, and that contrast runs through everything below. For him, ‘deployment’ means using a model as a calculation tool to optimize an experiment, not a system making judgment calls about a patient. That distinction is precisely why his perspective belongs in these pages: it draws a clear line between AI as instrument and AI as decision-maker, and it’s a useful mirror for a field like medicine, where that line is still being drawn.
— Hiba Hamdar, MD, Editor
Featured Guest
Tommaso Dorigo, Ph.D.
First Researcher, INFN-Padova · Editor-in-Chief, AI and Brain
Tommaso Dorigo is an experimental particle physicist and First Researcher at Italy's National Institute for Nuclear Physics (INFN), where he has been a member of the CMS experiment at CERN since 2001, following more than a decade on the CDF experiment at Fermilab. His research spans particle physics, statistical inference, and machine learning, with a current focus on the AI-assisted end-to-end optimization of physics measurements and detector design. He founded and coordinated the MODE Collaboration, a multi-institution group developing differentiable programming tools for particle physics, and serves as President of the USERN network for interdisciplinary research and education. He is Editor-in-Chief of the journal AI and Brain, and the author of the book Anomaly! Collider Physics and the Quest for New Phenomena at Fermilab.
The Conversation
How many trials, validation rounds, or iterations does an AI model typically need to undergo before it can be considered reliable enough for real-world acceptance — and does this differ meaningfully between fields such as particle physics and medicine?
There is no number I can quote. In general, though, particle physics has much more strict standards than medicine when it comes to hypothesis testing, e.g. a 5% effect will typically allow you to write about a correlation (not causation, but it will be branded as such) between a drug and some effect. While for subnuclear phenomena, we require a 3*10^-7 effect to accept that we have established a true effect about particle physics. You can draw your conclusions for what this kind of difference implies on the practice of use of AI models in the two disciplines.
What criteria do you personally consider most important when judging whether an AI tool is ready for deployment, versus still experimental?
Deployment in my field just means using a model as a calculation tool for optimization or for statistical inference. As such, we do not have implications that straggle into ethics or transparency etcetera, and we typically use what works. Optimization tasks in particular are "safe," as an AI model that approximates the behavior of a physical system does not even need to be precise in predicting the behavior of the system, in order to find the absolute maximum of a utility function: it just needs to be able to estimate with some accuracy the derivative of the utility function with respect to the optimization parameters, which is a much easier task.
In your experience, what is the biggest gap between an AI model that performs well on paper (in benchmarks or test datasets) and one that performs reliably in real-world, high-stakes conditions?
Unfortunately I do not work with such kind of AI, I work with AI for modeling experiments, which are still far from real-world conditions.
Given the rise of no-code and AI-assisted development tools, do you believe someone with little to no formal programming background could realistically build a meaningful AI tool today — and if so, what would be the biggest risks or blind spots in what they build?
Anybody today can go very far with LLMs use. The risk is that they do not understand well what they are developing. In my perspective and in my field this translates merely in bogus results, untrustable predictions, fallacious inference.
How do you see the role of domain experts (such as physicians) changing in AI development — from being purely end-users to becoming active builders of their own tools?
I think the expert needs to be always in some way interacting with the models in order to avert the problem that developed models do not work as intended. Human-in-the-loop schemes are today an important standard in this sense.
What do you see as the most significant current limitation of AI systems — whether in physics, medicine, or general applications — that the field has not yet solved?
Real-time processing of sensory data and integration in a continuous learning and refining of models is still far to be achieved. Since a large LLM today needs enormous CPU and energy resources to be trained, the shift to a continuous learning setup is difficult because of these limitations.
Where do you personally see AI heading over the next 5 to 10 years, particularly in terms of its reliability and trustworthiness for critical decision-making?
I don't know if a real new revolution awaits us in 5-10 years time - we have broken ground pushing the SoTA in LLMs, but it is difficult to push further in that direction, and while these models will be around for good, they do not "solve" the AGI problem. What I think will happen will be a partial burst of the AI bubble in the financial market, with many small companies dying, but a consolidation of the bigger players. In terms of reliability and trustworthiness, I think there will be a consolidation of what we have already seen.
What, in your view, is actually required for a person to be considered a researcher? Does publishing many papers make someone a researcher, or is that a separate question from the substance of the work itself?
Nice question. A researcher is a curious person who is an expert in some specific domain, and is capable of using SoTA tools to push the boundary of human knowledge in some direction. This usually results in publications as the natural outcome and documentation of the acquired new knowledge, but publications are a means, not a goal. Usually a good researcher is capable of raising the interest of the community and acquire funding, but this also is not necessarily what happens especially with people who think diagonally and out of the box at the problems they consider.
Is there a personal experience or turning point in your career that shaped how you think about the responsible development and validation of AI systems?
In physics we don't risk to break things, and AI models remain calculational tools, so "responsible development" means little in my field, unfortunately, so I do not have first hand information to share about this.