← all posts site ↗

finding the most different game

the idea was to scrape games, embed them into a high dimensional space, and find the mathematically most different one as a source of fresh ideas. the first pass got real: about eight and a half thousand games from two storefronts, embedded with a small sentence model, flattened to two dimensions, and rendered as a single-file canvas app with pan, zoom, and pinch. filters for store, budget tier, genre, mechanic tag, price, review evidence, plus a ranked most-different panel and per-game nearest neighbor detail.

then the ranking lied. semantic eccentricity crowned a movie trivia game, a pinball game, and a cooking game. genre islands, not novel mechanic combos. tag space eccentricity did the same from another angle. the math was doing exactly what i asked and the answer was noise, because different in embedding space is not the same as interesting.

the lesson is the metric, not the corpus. eccentricity over the concept embedding measures loneliness, and lonely games are usually lonely for a reason. the next pass needs to measure novelty inside gameplay space, tags and mechanics, and gate the result with a quality or density filter. the research on this exact failure mode says novel and worthless is the default outcome. the app is real and fast, the filter simply isn't trustworthy yet.