Machine learning & interpretability. GitHub
Physics & Computer Science at Dartmouth.
Founder of Principia AI.
A POMDP-based calculus tutor that plans over student misunderstanding, not just questions.
Tracking the conditioning of diagonal state-space model modes through quantization-aware training.
Ranking NASDAQ-100 stocks by predicted next-day returns with a market-aware graph and a multi-level Mamba state-space model.
Why AI outputs that look identical to human ones still can't enter critical systems: the causal structure between ideas can't be guaranteed — yet.
On effective theories, interpretability, and what success means for ML models.
On SSMs as machines for quadrature, and a hypothesis that a discrete spectrum via QAT could make their lost rigor trainable again — written as a proposal before the experiments that tested it; see the diagonal-SSM/QAT paper for where it landed.
Diagnosed and corrected a systematic bias in the standard χ² histogram fit — demonstrated via bin-width dependence — and used a binned Poisson maximum-likelihood estimator to measure the cosmic-ray muon lifetime and extract the Fermi coupling constant.
Compared two competing diffraction models — far-field Fraunhofer against a finite-distance path-integral model — and discriminated them quantitatively: a dimensionless Fresnel parameter predicted where they would diverge, and a slit-width consistency check exposed the model that absorbed its misspecification into an unphysical parameter. Demonstrated on Young's double-slit in the single-photon regime.
Determined the Landé g-factors of both rubidium isotopes by optical pumping and RF spectroscopy. Their ratio cancels the dominant coil-calibration systematic exactly, isolating a real ~3σ deviation from theory that the per-isotope uncertainty would otherwise have buried — backed by a data-driven recalibration of the apparatus's field coils.