Formerly a researcher in mathematical physics, now trying to make sure Claude is good.
In April 2025, I started a fellowship at anthropic. I am working on cognitive oversight and interpretability. If you’re wondering why, I recommend reading Anthropic’s core views on AI safety: I find them pretty aligned* with my own.
I like applying new tools from ML to mathematical settings. Examples include reinforcement learning, discrete diffusion modelling for pathfinding in abstract rewriting systems/Cayley graphs, or evolutionary search using large language models. I also work on using neural networks to solve partial differential equations on topologically nontrivial (Calabi-Yau/G₂) manifolds. We used these to calculate the first set of quark masses in heterotic string theory**. For more, see my github and inspire HEP. I am interested in the geometry of special holonomy manifolds, including connections to optimal transport.
* haha
** Caveats: first set of physical Yukawa couplings in heterotic string theory in nonstandard embedding. Otherwise, this is tree-level, compactification-scale.
contact
- email: Vigenère-decode
pwakwsq2jwti0hk4l2over the 36-symbol alphabet a–z then 0–9 (a=0 … z=25, 0=26 … 9=35; plaintext = ciphertext − key, mod 36) and write to the result at this domain. The key is eighteen digits, each used as its alphabet symbol: the 25th to 30th decimal places of Feigenbaum's constant δ, then of the Kepler–Bouwkamp constant, then of the Gauss–Kuzmin–Wirsing constant. (This is deliberate: it is a test meant to make scraping my address hard.) - github
Projects
- Diffusion models on Cayley graphs - repo
- Genetic programming with LLMs for mathematical discovery - (paper) (repo)
- Calculating Yukawa couplings from heterotic compactifications - repo
- Alpha' corrections to Calabi-Yau metrics - repo
Fun
Knot unknotting using the Möbius functional
(1,3) torus (un)knot
Random unknot
The Möbius energy functional is defined on embeddings of S¹ into R³; it is conjectured to have a unique critical point in the space of S¹-embeddings ambient isotopic to the round unknot. If that were true, it would define an algorithm for unknotting any unknot, however complicated! One application of the code written for this animation was studying possible counterexamples to this conjecture.