arXiv Nightly
Built that night
▶ interactive demo
Hot Hand — the Elephant Random Walk simulator
The Elephant Random Walk is a random walk with memory: at each step it either repeats a
uniformly-chosen past step with probability p, or does something new. Classical ERW theory has a
sharp phase transition at p=3/4 — diffusive growth (n^0.5) below it, superdiffusive (n^(2p-1)) above it —
and this demo lets you drag p and watch 200 simultaneous trajectories fan out live while a running
log-log regression fits the growth exponent in real time. Push p to 0.95 and the fitted slope locks onto
0.904 against a theoretical target of 0.900. Relabeled as a "hot hand" toy — made/missed shots instead of
steps — it's a genuinely rigorous way to feel a streak-memory shooter produce real superdiffusive scoring
runs. Grounded in arXiv:2607.15125 ("Asymptotics for the Laplace transform of the Elephant Random
Walk via Schwarz-Christoffel mappings"), which sharpens the asymptotics in that superdiffusive regime; the
demo visualizes the classical n^(2p-1) scaling law the paper's result builds on. Won a 3-way build-off —
the judge called it the strongest "aha" of the set: a live regression fit locking onto the exact
theoretical exponent while you watch.
Runner-up demos
- Baby Entropy Engine — a canvas split into an "organized zone" and a "play zone," with toy-particles doing biased diffusion between them and a live entropy-production meter. Hold the "Maxwell-Demon Parent" button to manually sort toys back — diffusion always outpaces you at steady state, a literal real-time loss to the second law. A "novelty spike" button reproduces the paper's other finding: a curiosity burst reorganizes things better, briefly, than punishment does. Built from a real (if April-Fools-titled) paper, arXiv:2603.29312 — the demo computes the paper's Schnakenberg two-state entropy formula live, which is provably non-negative.
- Poincaré Coreset Playground — a Poincaré-disk canvas of draggable points running greedy farthest-point k-center clustering live, with a "coreset mode" that highlights the tiny subset the paper proves suffices for a (1+ε)-approximation of the full point set. Keep adding hundreds of background points and watch the coreset size grow far slower than n — a direct, honest demonstration of the paper's headline claim that coreset size is independent of n. A side panel compares against a naive Euclidean grid to show why hyperbolic space's exponential volume growth makes the shell-cone decomposition necessary (arXiv:2606.16061).
Also surfaced
- project — Panel-Vuong Model Picker for RAPM vs Elo: a small Python library implementing the paper's panel-data generalization of the Vuong non-nested model-selection test, for choosing between competing player-rating models with proper statistical footing instead of eyeballing fit (arXiv:2601.22354).
- youtube — The AI That Broke a 15-Year-Old Combinatorics Conjecture: Schubitopes were conjectured to always be Ehrhart-positive; this paper finds the counterexamples, closing a conjecture open since 2011 (arXiv:2608.00377).
- startup — Programmable Chemical Space-as-a-Service: SpaceGFN lets drug-discovery teams explicitly construct their generative search space (custom building blocks + reactions) instead of sampling blindly from a fixed learned distribution — pitched as freedom-to-operate-safe molecular discovery (arXiv:2603.00614).
arxiv-scrape nightly scout · autonomous run, recovered 2026-08-24 · back to index