Essay: a multimodal AI gave up on a 1,000-piece all-white jigsaw in five hours
A first-person essay republished by Huxiu AI describes an attempt to have a multimodal AI chatbot solve a 1,000-piece all-white "Pure White Hell" jigsaw by photographing the pieces laid out in a numbered 20×50 grid and asking which edges interlock. The author says accuracy was poor from the first hours, that the model described edge curvature rather than measuring it, and that when asked to order 30 consecutive pieces it returned a closed loop (A joins B, B joins C, C joins A); he abandoned the attempt after about five hours. The essay cites the JigShape study, which it says tested five frontier vision models on synthetic puzzles and found only one beat guessing at 16 pieces while all dropped to chance at 64, and contrasts this with classical contour-matching solvers that it says have handled puzzles of several hundred pieces.
The author argues that LLMs fail at such closed-rule, zero-tolerance matching tasks because they describe images instead of measuring them and lack a persistent global working state, a limitation he links to the "scale cliff" reported in the cited JigShape research.
引子事情始于一次无聊的刷手机。那天深夜我无意间在抖音刷到一个开箱视频：博主拆开一盒拼图，倒出来，画面里是一片纯白——不是大部分白，是全部白。1000片，没有图案，没有色块，没有边框参考图，每一片都长得一模一样。盒子上印着它的名字：纯白地狱。作为重度手玩爱好者，这种无聊至极的游戏瞬间击中了我。第二天就下了单。到......
本文来自微信公众号： 斯凯碎碎侃 ，作者：斯凯，原文标题：《心锚｜让AI拼我用了12天完成的1000片“纯白地狱”，它五小时就崩溃了》
引子
事情始于一次无聊的刷手机。
那天深夜我无意间在抖音刷到一个开箱视频：博主拆开一盒拼图，倒出来，画面里是一片纯白——不是大部分白，是全部白。1000片，没有图案，没有色块，没有边框参考图，每一片都长得一模一样。盒子上印着它的名字：纯白地狱。