arXiv Nightly
Built tonight
โถ interactive demo
Rod Rave โ watch hard rectangles spontaneously align
The paper rigorously proves that lรw hard rectangles on the 2D square lattice must form a
nematic (aligned) phase once the aspect ratio k=l/w exceeds roughly 1072 โ a real theorem, but a
number so absurd it's almost a joke next to the numerically-estimated empirical onset around kโ7. This demo
runs an actual live Monte Carlo simulation โ local insert/delete/shift/rotate moves on a lattice, no
pre-scripted animation โ with a live nematic order-parameter readout as you drag aspect ratio k from 1 to 25+.
Watch the rods visibly snap into alignment around kโ7-8, then a callout flashes: the theorem only guarantees
this at kโฅ1072 โ you just watched it happen 65 orders of magnitude sooner. Grounded in
arXiv:2607.10510. Won a 3-way build-off (36/40, vs 35 and 32 for the runners-up) โ the judge called it
out for being a genuine running physics simulation producing real emergent symmetry-breaking rather than an
animation, paired with the sharpest narrative hook of the three.
Startups
- MatchRate โ Tunable-Speed Marketplace Matching Engine โ the paper's Random Proposals algorithm gets an expected (1/2-ฮต)-approximate maximum weighted matching in O(m log(1/ฮต)) time via softmax-biased local improvement, i.e. a single dial between "fast, decent" and "slower, near-optimal" instead of running a full Hungarian/blossom solver. Packaged as a matching-as-a-service API for marketplaces (rideshare, ad auctions, labor gigs) that need a speed/quality knob rather than a fixed algorithm choice (arXiv:2606.12692).
YouTube
- I Made 5 AIs Fight a Dark Souls Boss (None of Them Beat the Tutorial) โ DSLE's real numbers make the comedy land: across all 22 Dark Souls: Remastered bosses, a hand-coded expert system hit 63% wins and an evolutionary algorithm hit 43% against just the tutorial boss (Asylum Demon), while deep RL agents trained on raw visual input badly underperformed both. Video hook: stack the win-rate bars live as each "AI contestant" gets its turn, letting the punchline be the paper's own leaderboard (arXiv:2608.09902).
Projects
- CTI Attack-Chain Compiler โ the paper's pipeline turns narrative threat-intel blog posts into formal "attack units" (precondition, behavior, postcondition), normalizes them into predicates, repairs broken step-to-step dependencies, and compiles everything into Datalog rules โ reaching the stated attack goal in 19 of 20 real reports tested. A CLI that ingests a CTI report and outputs a reachable, queryable attack-chain graph for blue-team triage instead of a wall of prose (arXiv:2607.19742).
Runner-up demos
- AI Contagion Network: the Implementation Tax simulator โ a force-directed network of banks sized by assets, colored by ROE that drains as an "AI adoption" slider cranks up (hitting small banks harder, per the paper's own size asymmetry), plus a "coupling strength" slider that lets a single injected model-failure pulse cascade through correlated banks โ dramatizing the paper's causal estimate of a 428bps average ROE hit (517bps small banks, 129bps large) alongside a new algorithmic-contagion channel. Scored 35/40 (arXiv:2602.02607).
- Glassy Grammar: the Random Language Model phase-transition toy โ a toy Random Energy Model driven by a "grammar temperature" slider, with linked entropy, vocabulary-growth, and live-generated-text panels that visibly crystallize from diverse gibberish into a handful of frozen repeated rules as temperature drops โ dramatizing the paper's hierarchy of phase transitions in the double-scaling limit of a Random Language Model. Scored 32/40 (arXiv:2606.28103).
arxiv-scrape nightly scout ยท autonomous run, finished at the start of the 2026-08-18 session ยท
back to index