When AI fails, and why

• Why do models get facts wrong even when the evidence is in front of them, and can we catch it automatically? Hallucination and verification
• What is a model really computing when it learns from examples in a prompt, and can maths predict when it will fail? In-context learning
• Can we change how a model behaves while it runs, safely and without costly retraining? Changing models without retraining
• How can large models run with less memory and compute, with a guarantee on what we give up? Faster, cheaper AI
• Can a model trained on data alone learn the symmetries of the world, and where can careful, checkable AI make a difference in science? World models and science

• PaperSeptember 2026

  Robot world models aren’t invariant to how the actions are written

  Hand a robot world model the same trajectory written as joint targets instead of changes, and retrieval degrades 2.6 to 13.4 times. With Ahmed Karim.

• PaperSeptember 2026

  Exact attention responses from RoPE derivatives

  How attention responds to moving, removing or changing part of a prompt can be computed exactly. With Julie Huang, Maggie Chlon and Gregory Gutin.

• ResultSeptember 2026

  No trade-off between energy and force accuracy

  On MD17 ethanol and aspirin, energy and force error fell together across almost the whole range of weightings.

• PaperJune 2026

  Predictable Compression Failures at ICML 2026

  Hallucination as a measurable shortfall of information, with a rule for when to answer and when to abstain.

• PaperJune 2026

  LLMs are Bayesian, in expectation

  On Qwen2.5 7B and 14B, the gap to an ideal Bayesian reasoner is a few hundredths of a bit per prediction; the order sensitivity traces to the positional encoding.

• CodeMay 2026

  ntkmirror: adapt a model without touching its weights

  A 50,000-parameter controller came within a point of LoRA on GSM8K, and lost far less on held-out code.

Why a CNN beats a vision transformer on small data

Every symmetry you build into an architecture is data you don’t have to collect.

Noether’s theorem, in one spring

Every symmetry comes with something that never changes. A spring shows how.

What masked reconstruction teaches a model about shape

Hide part of a shape, ask a model to rebuild it, and symmetry becomes a shortcut worth learning.

No trade-off between energy and force accuracy

Neural surrogates for chemistry are tuned as if energy and force accuracy compete. On two molecules, they didn’t.

All 7 explainers