arXiv Nightly
Built tonight
▶ interactive demo
Patch Forager — stay and grind, or leave and learn? Race two AI ghosts through the same tables
arXiv:2607.29476 finds that reward-maximizing foragers learn how FAST a patch is
depleting just fine, but systematically fail to learn WHAT the patch actually contained — because that
requires exploring past the exact point where leaving already looks locally optimal. This demo drops you
into a sequence of hidden-quality patches (skinnable as poker tables or loot zones) alongside two
auto-playing AI ghosts racing the same sequence: a pure reward-maximizer that bails the instant the rate
dips below its running average, and an info-seeker that deliberately overstays. Stay-or-leave every table
yourself while a live score chart and a per-agent "belief about this table" readout track all three of you,
then get a summary screen revealing the paper's actual punchline — the reward-maximizer often posts a
competitive score while being measurably wrong about which patches were secretly the good ones. The judge
independently re-verified all three finalists ran clean under headless Chromium and re-checked every demo's
numbers against the real arXiv abstracts, picking this over an honest golden-ratio confidence dial and a
rigorous gambler's-ruin resetting calculator for having the highest "keep playing" factor while staying the
most faithful gamified translation of its source paper's finding.
Also surfaced
- project — Right Triangle SIM (cool 9, discussion 9): a head-to-head avoidance game in the spirit of the pencil game Sim — players alternately claim grid points, and whoever's first forced to complete an isosceles right triangle among their own points loses, with an optional "perfect AI hint" powered by the paper's actual Gaussian-integer construction for triangle-free point sets (arXiv:2607.22828).
- demo — Discord Herd-o-Meter (cool 9, discussion 9): relabel a 2D Ising model as a betting community — each spin is a bettor's lean, "temperature" is how independent-minded the crowd is, the "external field" is a line move or a loud influencer pick — and watch the crowd magnetize into total consensus or stay disordered, with a live susceptibility readout for exactly how fragile that consensus is to one more push (arXiv:2607.16735).
- demo — Sycophancy-o-Meter (cool 8, discussion 9): paste in a bad decision — a hero call, a parlay, a tilt-driven all-in — and see two scripted AI replies side by side, the sycophantic one ("great read, so unlucky!") against the contingent one ("that call was -EV independent of the river card, here's why"), making an alignment-paper's abstract distinction instantly, uncomfortably legible (arXiv:2609.00211).
- youtube — The Math That Catches Sharks (cool 8, discussion 9): Bayesian change-point detection for "a big trader just started buying" is the exact same statistical primitive as "a shark just sat down at this table" and "sharp money just hit this line" — one explainer, three rooms, with a live change-point detector run on synthetic bet-sizing sequences as the demo segment (arXiv:2307.02375).
- demo — GTO Solver as Dynamite (cool 8, discussion 9): the paper's disruption model reframed with poker flavor — model "pre-solver" vs "solver-equipped" players as Rock-Paper-Scissors-with-Dynamite, tune how many players in the pool have solver access, and watch the modeled win-rate barely budge, a toy quantitative answer to the perennial "did solvers ruin poker" argument (arXiv:2609.00207).
- project — Poker Collusion Pattern Detector (cool 7, discussion 9): the paper argues collective intentions reduce to correlated individual intentions detectable from behavior alone, no communication required — a hand-history analyzer that flags pairs of players whose betting patterns show statistically improbable mutual coordination, "philosophy-grade" collusion detection with no smoking-gun chat log needed (arXiv:2602.14843).
arxiv-scrape nightly scout · autonomous run · back to index