Oz Amram

I’m Oz — an ML + physics researcher at Fermilab, moving into AI safety

I’m a Wilson Fellow and Associate Scientist at Fermilab and a member of the CMS experiment at the Large Hadron Collider. For the past several years my research has been on machine learning for particle physics: anomaly detection, generative models, and foundation models, mostly aimed at finding things in complex collider data that nobody knew to look for.

I am now transitioning to AI safety research, and am actively looking for roles in the field.

The rapid increase in AI capabilities over the last year, and recent public misalignment incidents have convinced me AI safety is an urgent issue, and worth leaving my current research and tenure-track position behind for. I believe I have the technical skills, research history and motivation to be effective in alignment or interpretability research roles.

What carries over

The problems are not the same, but a lot of the machinery is.

Finding behaviour nobody specified in advance. Anomaly detection has been my biggest research focus of the past few years: searching for new particles in complex, massive datasets from the LHC, with no labelled examples. I have developed and used weakly supervised and unsupervised methods ML methods for this task, and established best practices of how such methods should be evaluated and validated. I led the first such search at CMS, running five complementary methods over 30 million collision events.

Knowing when a model can be trusted off-distribution. Classifiers in particle physics are often trained on simulation, but deployed on real data. And the gap between the two is hugely important. I developed a method CMS now uses as standard for calibrating that gap and putting defensible uncertainties on it.

Saying why a model flagged something. For that search I also built the interpretability framework — characterising what made a flagged event anomalous and mapping it back to physical detector signatures. It was the first for an anomaly search at the LHC.

Evaluating generative models honestly. I wrote CaloDiffusion from scratch, and now lead the effort benchmarking generative models for the CMS calorimeter upgrade. Most of that work is the evaluation framework, building quantitative metrics that capture how closely the model is matching physics-based simulations, and highlighting which features are mismodeled.

My résumé is a two-page summary aimed at AI safety roles. My full CV has the complete record.

The physics

CMS studies the fundamental particles and forces that make up all matter in the universe. We collide protons at the highest energies we can reach and sift through the millions of collisions produced every second for signs of new interactions. My research has focused on applying novel machine learning methods to the analysis of this data.

I completed my PhD in physics at Johns Hopkins University in 2022, joined Fermilab as a postdoc, and in 2026 became a Wilson Fellow (tenure-track assistant professor equivalent).

A few of the things I’ve worked on:

Anomaly Detection

Model-agnostic searches that let the data itself flag unexpected new particles, instead of testing one theory at a time. I led the first such search at CMS.

Applications of Generative Models

Harnessing modern generative AI --- diffusion models and normalizing flows --- to accelerate detector simulation and to perform high-dimensional, data-driven statistical inference.

Foundation Models

Training large models on real LHC collision data, and mapping out how they scale, as a foundation for many downstream physics tasks.

You can find more on my projects page.

I have a blog on substack where you can read my marginally-filtered thoughts. I also used to write for ParticleBites, summarizing recent particle physics papers for a broad audience.