arXiv Nightly
Built tonight
▶ interactive demo
Curvature Mass Playground
A surface nematic (a liquid-crystal-like order parameter confined to a curved 2D surface) picks up an
effective mass term purely from the geometry it lives on — the paper's headline result is m² = K_ab K^ab, the
squared extrinsic curvature of the embedding. This demo runs a real finite-difference second-fundamental-form
computation on a three.js mesh you can morph between a flat plane, sphere, saddle, and torus neck — drag a defect
marker across the surface and watch its live-computed effective mass spike near high-curvature necks and drop to
zero on the flat plane, with a Chart.js readout tracking it as you move. Grounded in arXiv:2605.19183. Won a
3-way build-off (37/40 on wow/interactivity/polish/fidelity) over a distribution-shift detection-scale-law sandbox
(31/40) and an interactive map of "what counts as alive" (33/40) — the judge picked it for doing real, live
differential geometry on a draggable 3D mesh while matching its paper's formula almost word-for-word.
Projects
- Privileged-Data Distillation: train rich, deploy cheap — ports a cancer-imaging trick (train a network to translate a cheap modality into the feature space of an expensive one, so inference only ever needs the cheap modality) straight onto the user's NBA props edge modeling: train on box-score + tracking-derived features (spacing, closeout speed, shot quality) where available, deploy on box scores alone anywhere they aren't. Ships as a small reusable "trans-align" PyTorch module rather than a one-off model (arXiv:2411.00749).
Startups
- Parsimony Terminal — a 5-parameter, non-asset-specific volatility forecaster (near-LSTM accuracy in the paper's crypto-winter test, but auditable, near-zero compute) feeding a return-diversification allocator that provably collapses to simple Risk Parity in calm regimes, so every allocation is hand-checkable. Pitched as "explainable in 5 numbers, no GPU" — a natural extension of the user's SharpLab/NBA-props edge-modeling throughline into a general quant-terminal product (arXiv:2311.04727 + arXiv:2312.09707).
YouTube
- Why some seeds fly a mile and 99% land a foot away — subdiffusive fractional Brownian motion (particles moving in tight, correlated clumps) normally stops after an exponential waiting time, giving thin exponential dispersal tails; making the stopping rate itself Gamma-distributed across particles keeps the same tight local clumping but produces genuine power-law tails at long range. The whole paper is one side-by-side particle simulation: fixed stopping rate (tails vanish) vs. Gamma-distributed stopping rate (rare walkers keep going, tails persist) — and it explains a real, still-argued-over ecology mystery about long-distance seed and spore dispersal (arXiv:2606.21681).
Runner-up demos
- Shift Scope: the moment-detection scale law you can feel — the paper proves a hard limit: certifying a distribution-shift feature of scale ε carrying mass fraction f needs a polynomial test of degree N* ≥ log(1/f)/(2ε), with the optimal witness an RBF/MMD kernel at bandwidth σ* = ε (median σ*/ε = 1.12 across 26 real settings, AUC ≥ 0.95). The demo live-computes moment tests and an MMD statistic on real sampled point clouds as you sweep ε, f, and bandwidth — watch moment tests go blind exactly where the formula predicts and the MMD curve peak right at σ=ε (arXiv:2608.01268).
- Life-Space: An Interactive Map of What Counts as "Alive" — turns dozens of expert definitions of life into thematic feature vectors (metabolism, reproduction, boundary, evolution, information-processing) and embeds ~25 borderline entities — virus, fire, prion, coral, AI chatbot, Von Neumann probe — onto a live 2D map. Drag sliders to define a custom "entity" and watch it glide to its nearest real cluster, with the embedding math computed live rather than baked into a static image (arXiv:2505.15849).
Scouted 176 fresh arXiv preprints across 22 categories · 30 sampled for ideation across 6
batches · 24 ideas generated · picked by cool×buildable. Nightly run by
arxiv-scrape.