文 | 斯凯有话说
1000 片纯白地狱拼图，没有图案、没有色差、没有参考，人类玩家耗时 12 天完成。本以为擅长识别推理的 AI 可以轻松通关，没想到仅五小时就撑不住崩溃，纯白地狱为什么难住了人工智能？
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。到货后我开始体验，断断续续用了12天拼完。据玩家社群流传，人类最快纪录大约是160个小时，我属于中不溜的水平。
拼完的那晚，我盯着桌上那块纯白的板子，脑子里冒出的第一个念头不是成就感，而是一个问题：
这种活儿，不正是AI最擅长的吗？
拼图本质上是一个匹配问题：1000片，每片的形状独一无二（官方说的），规则封闭，答案唯一。不用理解画面，不用发挥创意，只需要看形状、做比对。这不就是把AI喂饱、让它支棱起来的标准场景吗？
于是我做了一个决定：用AI重新拼一遍这副已经拼好的纯白地狱，看看它到底要花多久。
我的方案
我的设计其实很朴素，也是任何一个理性人能想到的最优解：
在拼图桌上画一个20×50的格子，一共一千格，每格标上编号，1000片散片一片一格放好，拍成照片上传给AI。它要做的事只有一件——判断哪两片的边缘可以咬合。
照这个方案，原本需要160小时的盲目试错，理论上可以压缩成“照单抓药”：AI看图、提特征、算匹配、输出配对表，我负责动手验证。
我把照片传了上去。
第一个小时：它聊得头头是道
起初的对话相当愉快。AI准确地说出了纯白地狱的难点：“没有颜色线索，只能依赖边缘形状匹配。”它还主动帮我分析了“先拼边框片”的策略，语气像个经验丰富的拼图老师。
我让它判断照片里A12格和B07格这两片能不能拼。它回答：“从图片看，两片的边缘曲率呈现互补特征，大概率可以咬合。”
听起来很专业。我拿实物一试——拼不上。
第三个小时：我开始追问
我又试了几对，准确率惨不忍睹。于是我换了个问法，把两片的特写放大到极限再传上去。它依然在输出那种正确的废话：“边缘轮廓存在细微差异，需要更精确的测量。”
一句话点醒了问题所在：它在描述，不在测量。
它告诉我“这片凸得比较陡”，但拼图需要的是“凸点位于边缘37.2%处”。它可以每一次都言之凿凿，但两次描述同一片，措辞还会有出入。而对拼图来说，差之毫厘，就是拼不上。
更要命的是，当我让它连续推30片的拼接顺序时，它给出了一个“A接B、B接C、C又接A”的闭环。它做的是逐对的局部判断，没有能力维护一个全局自洽的拼图状态——每一步都像第一次看图。
第五个小时：我查到了一份“判决书”
不甘心的我去查了文献，然后发现了一个哭笑不得的事实：学术界早就拿拼图处刑过AI，而且用的题目比我的实验干净得多。
2026年8月，一个叫JigShape的研究测试了五个前沿视觉大模型。题目是计算机合成的拼图：形状数字化生成、每片严格唯一、图像完美清晰——没有我方案里照片模糊、光线不均这些干扰。结果：16片的拼图，五个模型里只有一个超过瞎猜水平；64片，集体回到随机；微调过的模型到144片，正确率不足5%。
研究者给这个现象起了个名字：规模悬崖。片数一涨，约束一多，模型全线崩塌。
我的五小时实验，本质上是在悬崖边上散了个步。
查到这篇论文的时候，我有点泄气——原来学术界早就把这件事研究透了，我的五小时实验，不过是给一个已知结论又添了一个 anecdote。但转念一想，知道一个结论，和用自己的手撞上这个结论，是两回事。前者是信息，后者是认知。而后来我才知道，这件事在 AI 研究圈更是早不是新闻——三十年前就有人总结过一个“莫拉维克悖论”：对机器来说，下棋容易，钉钉子难；推理简单，感知难。学界为此搞了几十年的神经符号结合研究。只是这些讨论，从来没有走出过论文的围墙——直到我用自己的五小时，替自己撞上了它。
为什么新AI做不到：卡住它的不是眼睛，是两样东西
实验失败后我一直在想：照片够清楚了，每片也确实不同，它到底卡在哪？
想明白之后发现，大模型输掉这场比赛，不是输在“看”，而是输在两个更底层的地方。
先说“测量”。大模型处理图像的方式，本质上是生成对图像的描述——这是它的出厂设定。它可以说“这条边有个凸起”，但说不出“凸点中心位于边缘37.2%处、曲率半径1.8毫米”这样的数字。而拼图的判定标准是物理咬合，容差以零点几毫米计。描述无法拼图，测量才能。更麻烦的是，它对同一片拼块的两次描述还可能措辞不一——一套输出无法复现的系统，做不了精密度要求为零误差的活。
再说“工作台”，这是我那个“A接B、B接C、C又接A”闭环的真正成因。人类拼图时，桌上已经拼好的部分是锚点：每拿起一片新拼块，你不是从一千片里盲选，而是对着一块已经成形的区域找位置，错了就换，拼对了就变成新的锚点。这是一个不断积累确定性、随时回溯修正的过程。而大模型没有桌子——它每回答一次，都是从零开始重新看图，已判断过的配对不会沉淀为“确定的事实”。上千条约束要在脑子里同时成立且互不打架，这种全局一致性，恰恰是它的架构里没有的东西。
“规模悬崖”说的就是这件事：16片时，几十条约束还能勉强糊弄过去；到64片、144片，约束数量指数增长，每一步的微小错误互相传染，最后满盘皆输。
人类赢的不是聪明，是手感和时间
想明白这些之后，我又反过来想了一个问题：那我12天是怎么拼完的？
论计算速度，人眼比对两条曲线的吻合度，远不如计算机在误差矩阵里搜一遍。但我有一样大模型完全没有的东西：触觉。
拼到最后几百片，我根本不怎么看。手指捏着拼块在已经成形的区域上滑动，凹是凹、凸是凸，卡进去的瞬间有一种轻微的“咬合感”——对了就是对了，手感会告诉你。这种反馈是实时的、物理的、零延迟的，大模型隔着一张照片，永远拿不到它。
除了手，还有直觉。拼了一百片之后，我对这套刀模的“性格”产生了肌肉记忆：这个厂的凸起偏圆、那类凹口偏深，拿到一片新拼块，手指一捻大概知道该去桌子的哪个区域找。这不是计算，是几千次试错压缩出来的模式识别。
以及最重要的——耐心。纯白地狱没有视觉反馈，拼一小时可能一无所获，支撑你继续的是“总会有进展”的信念。这不是能力，是意愿，而这恰恰是机器从来不需要具备、人类却随时可能丢失的东西。
所以这场较量真正的格局是：大模型有千里眼，但没有手指，也没有工作台；人类的手指是精密的测量仪，桌子是天然的存贮器，唯一的短板是耐心有上限。
反转：其实AI能拼，用的还是40年前的技术
写到这里，本该以“AI不行”收尾。但我多查了一步，发现了一个更有意思的事实。
自动拼图求解，是人工智能领域一个有几十年历史的老课题，论文就发表在IJCAI、AAAI这类正经AI会议上。经典解法一点也不玄：扫描每片，提取边缘轮廓，把每条边变成一串精确到像素的坐标；用欧氏距离计算两条边的吻合误差；得到一张1000×1000的误差矩阵后，交给搜索优化算法找全局最优解。这条路线上，数百片规模的无图案拼图早有自动求解的先例。
按AI教科书的定义——感知、知识表示、搜索、规划——这套系统是标准的AI。
所以准确的结论不是“AI拼不出纯白地狱”，而是：
AI能拼出纯白地狱，用“老AI”；拼不出它的，是“新AI”。
老AI把问题变成数学，然后算——它不需要看懂图片，它量。新AI从数据里学模式，然后生成——它不量，它谈。学出来的智能赢在开放和泛化，造出来的智能赢在精确和可靠。而纯白地狱这种“规则封闭、答案唯一、容错为零”的任务，恰好是造出来的智能的主场。
我的五小时崩溃，错不在AI，在于我把题目给错了对象——就像让一位博学的翻译去操作车床，不是他笨，是这活儿不归他。
真正值得警惕的事
这只是一篇人机分工的科普吗？其实我最想说的在后面。
麻省理工学院媒体实验室2025年做过一个脑电实验：54名参与者分三组写文章，一组用ChatGPT，一组用搜索引擎，一组纯靠大脑。四个月里，用AI的那组在神经、语言、行为三个层面的表现全面落后，大脑神经连接最弱，甚至难以准确引用自己几分钟前刚“写”完的文章。OECD在《2026年数字教育展望》里给这类现象起了个名字：元认知懒惰——把思考外包出去，制造一种“知识掌握的假象”。另一项覆盖666人的研究则发现，AI工具使用频率与批判性思维得分显著负相关，年轻人中最严重。
把这些研究和拼图放在一起，我看到的是一个方向性的错位：
我们正把本来会做的题外包出去，却指望AI去解它根本解不了的题。
写作、思考、判断——这些人类本来擅长的事，我们迫不及待地交给了大模型，然后认知能力一路滑坡；而纯白地狱这种该交给“老AI”的精确匹配问题，我们却在让大模型看图说话，然后失望于它猜不准。两头都搞反了。
而拼图给我的最大提醒是：在“外包思考”的名单上，手上的活儿其实排在最后。人类拼图靠的触感、直觉和耐心，是长在身体里的能力，AI夺不走它——真正在悄悄流失的，是我们坐下来动手做一件事的意愿。
结语
那块12天拼出来的纯白板子，现在挂在我的书房。
有人问我，拼它图什么？又不好看。以前我也说不清，现在我想明白了：它的价值就是那160个小时本身——在没有颜色、没有反馈、没有捷径的纯白里，你能依靠的只有自己的手、眼和那点不肯放弃的耐心。
AI时代，这种“看起来毫无必要的苦功”会显得越来越可笑，也会越来越珍贵。
真正该问的问题从来不是“AI能不能替我做这件事”，而是：这件事，本来该不该是我做的？
对了，那副拼图我最终没有让AI重拼。倒不是心疼它，而是我发现：用AI拼拼图这件事最大的悖论在于——如果你真的需要那块拼图，你就不会需要拼图。
本文事实依据：纯白地狱（Beverly「純白地獄」）片数版本与难度体验据玩家社群及媒体报道；JigShape论文arXiv:2607.27670（2026年8月）；MIT媒体实验室脑电研究arXiv:2506.08872（2025年，未经同行评审，结论为初步结果）；OECD《2026年数字教育展望》；Gerlich批判性思维研究（Societies，2025）。文中“160小时最快纪录”为玩家社群流传说法，未经官方证实；莫拉维克悖论、神经符号AI研究为领域经典共识。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Wen, Zi, Skye.
1000 pure white hell puzzles, no pattern, no color difference, no reference, human player time 12 days to finish. I thought AI was good at identifying reasoning, but I didn't think it'd last five hours.
Seeds
It started with a boring phone brush.
Late that night, I accidentally drew an open-box video from Douyin: The blogger opened a box of puzzles and fell out, and the image was a white one — not most, but all. 1000 pieces, no patterns, no colours, no border reference maps, each of which looks exactly the same. It has its name on it: Pure white hell.
As a heavy-handed enthusiast, this boring game hit me instantly. And the next day, I put in a list. I started to experience it, and it took 12 days to finish. According to the player community, the fastest human record is about 160 hours, and I'm in the middle.
The night I finished, I looked at that white board on the table, the first thought in my head was not a sense of achievement, but a problem:
Isn't that what AI's best at?
The puzzle is essentially a matching question: 1,000 pieces, each of unique shapes (officially), closed rules, the only answer. You don't have to understand the picture, you don't have to be creative, you just have to look at the shape and compare. Isn't this the standard scene where AI is fed to make it lucid?
So I made a decision: to spell this pure white hell again with al, and see how long it takes.
My plan.
My design is simple, and it's the best that any rational person can think of:
Draws a 20 x 50 grid on the jigsaw table, a total of 1,000 cells, each with a number, and 1,000 pieces of fragments, one by one, to be photographed and sent to AI. There's only one thing it has to do — to determine which edges can bite.
Under this scheme, it would have taken 160 hours to make a blind error, which could theoretically have been reduced to a “single-cap”: Ai, characterization, calculator, output pair, I'm in charge of proofing.
I'll pass the photos on.
The first hour: It's the headline.
The initial conversation was quite pleasant. AI accurately says the hard point of pure white hell: "No color clue, only edge shapes." It's also helping me with the strategy of "scrambling the border" and speaking like an experienced puzzle teacher.
I let it judge whether the A12s and B07s in the picture can be spelled. And it said, "Looking at the picture, the edge curvature of the two pieces is complementary, and the rate is likely to match."
Sounds professional. I took a physical -- I couldn't spell it.
3rd hour: I started asking.
I've tried a few more, and the accuracy rate is overwhelming. So I changed the question, and I zoomed the two pieces to the limit and passed them on. It still produces the right kind of crap: “There are fine differences in the contours of the edge that require more precise measurements.”
One sentence wakes up the problem: it describes, not measures.
It tells me that “this scene is steeper”, but the puzzle requires “a 37.2 per cent location on the edge”. It can be said every time, but twice in the same way and with different wording. And for the puzzle, it's not working.
And even worse, when I put it in the order of 30 pieces in a row, it gave it a closed circle of "A to B, B to C, C to A." It does so by way of local judgement and does not have the capacity to maintain a whole-of-the-art puzzle — every step is like the first.
Five hours: I found a verdict.
I went to the literature and found a sad fact: The academic community had already tortured AI with a puzzle, and the subject was much cleaner than my experiment.
In August 2026, a study by Jig Shape tested five front-line visual LLM. The title is a computer-synthetic puzzle: shape digitized, each strictly unique, and the image is perfectly clear — without the blurry of photographs and the unevenness of light in my programme. Outcomes: 16 puzzles, only one of the five models exceeds the level of blind guessing; 64, group back to random; and slightly adjusted models to 144, with a correct rate of less than 5 per cent.
The researchers gave a name to this phenomenon: a cliff of scale. The number rises, more binding, and the model collapses.
My five-hour experiment essentially took a step on a cliff.
When I got to the paper, I was a little upset -- the academic community had already studied it, and my five-hour experiment just added anecdote to a known conclusion. But, in retrospect, it is different to know a conclusion and to hit it with your own hand. The former is information and the latter is cognitive. And then I realized that this was not news in AI -- 30 years ago someone summed up the Moravik paradox: chess is easy for machines, nails are hard; reasoning is simple and difficult. The academic community has conducted several decades of neuronym synthesis studies for this purpose. It's just that these discussions never got out of the wall of the paper -- until I hit it myself for five hours.
Why can't the new AI do it? It's not the eyes. It's two things.
After the experiment failed, I kept thinking, "Where is it stuck?"
To understand, LLM lost the game, not in “see”, but in two lower places.
First, "Measurements". The way LLM processes images is essentially to generate a description of the images — it's a plant setup. It can be said that “there is a bulge on this side”, but it cannot be said that “the centre of the focal point is located at 37.2 per cent on the edge with a curvature radius of 1.8 mm”. The puzzle is defined by a physical bite, with a margin of zero millimeters. The description cannot be puzzled, so the measure can be. Even more troublesome, its two descriptions of the same piece of piece might be worded differently — a system whose output could not be repeated, and could not do what precision required zero error.
And this is the true cause of my "A to B, B to C, C to A" loop. The part of the human puzzle that has been spelled out on the table is the anchor: for each piece of new piece, you don't pick out a thousand pieces blindly, but you look for a place in an area that's already formed. This is a process of cumulative certainty and retroactive revision at any time. LLM does not have a table - each time it answers, it re-examines the map from scratch, and the judged pair does not sink into “established facts”. Thousands of constraints have to be built in their minds at the same time and not fight each other, and this global consistency is precisely what is not in its structure.
This is the story of the “crawl of scale”: in 16, dozens of restraints can barely be overcome; in 64, 144, the index of bound quantities increases, each with a small error that spreads over the other, and then is lost.
It's not smart, it's hand and time.
And then, after all this, I thought of a question: How did I finish 12 days?
On the speed of calculation, the human eye compares to the match of two curves, much less than the computer searching in the error matrix. But I have one thing that LLM doesn't have: touch.
I don't see it that way. Finger squeezing on a plattered area, dents, concussions, and a slight "bite feeling" at the moment when it's stuck -- that's right, that's right, that's right. This feedback is real-time, physical, zero-delayed, and LLM is separated by a photograph and never gets it.
Besides the hand, there's a hunch. After a hundred pieces, I had a muscle memory of the scissor's character: the rims, the dents, the dents, the pieces, the fingers, the area where the table should be. It's not a calculation. It's a model recognition that's been compressed thousands of times.
And above all — patience. Pure white hell, with no visual feedback, could get nothing for an hour, supporting your continued belief that there will be progress. This is not capacity, it is will, and this is precisely what the machine never needs to have, but humans can lose at any time.
So the real pattern of this contest is that LLM has a thousand-mile eye, but no fingers, no workstations; human fingers are sophisticated gauges, tables are natural receptacles, and the only short plate is patience with ceilings.
Inverted: Actually, AI can spell, using 40 years old technology.
Here, it should have ended with "Ai can't". But I looked over it and found a more interesting fact.
Automatic puzzle resolution is an old, decades-old subject in the field of artificial intelligence, and the papers are published at official AI meetings such as IJCAI, AAAI. The classic solution is not at all complete: scan each piece, extract the contours of the edge, and turn each side into a line of precise pixels coordinates; calculate the difference between the two sides using the Oxylene distance; and, after obtaining a 1,000 x 1,000 error matrix, hand it over to the search optimization algorithm to find the best solution in the whole picture. On this route, hundreds of large-scale, uncharted puzzles have a precedent for automatic resolution.
This system is standard AI, as defined in the AI textbook - perception, knowledge expression, search, planning.
So the exact conclusion is not "Ai cannot spell pure white hell," but:
AI can spell pure white hell with "old AI"; it's "new AI".
Old AI turns the problem into math, and then it counts -- it doesn't need to read pictures, it does. New AI learns the pattern from the data and then generates -- it doesn't matter, it talks. The smarts learned are open and broad, and the smarts created are accurate and reliable. And the task of “rules closed, only answers, zero tolerance” in pure white hell happens to be the home of created intelligence.
My five-hour breakdown, not AI, was because I gave the wrong subject -- like having a learned translator run the car bed, not being stupid, not being his job.
Something that really deserves attention.
Is this just a pop-up job? Actually, the last thing I want to say is back there.
The Media Laboratory of the Massachusetts Institute of Technology (MIT) carried out an EMP in 2025: 54 participants wrote in three groups, one with ChatGPT, one with a search engine and one with a pure brain. Over the past four months, the performance of the group using AI has been completely backward at the neurons, language and behavioral levels, and the brain has the weakest neural connection and it is even difficult to quote exactly what it “written” just a few minutes ago. In the Digital Education Vision 2026, the OECD gives a name for this phenomenon: a meager cognitive laziness — outsourcing of thinking to create a “knowledge-held illusion”. Another study, which covered 666, found that the frequency of use of the AI tool was significantly negative for critical thinking, the worst among young people.
To put these studies and puzzles together, I see a directional error:
We're outsourcing what we would have done, and we're counting on AI to solve what it can't solve.
Writing, thinking, judging — these things that humans were good at, we couldn't wait to hand over LLM, and then the cognitive ability slipped all the way down; and pure white hell, the exact match that should be given to “Old AI”, we're giving LLM a speech and then disappointed in it. Both sides are upside down.
And the biggest reminder of the puzzle for me is that on the list of "outsourcing thinking" the work is actually at the end. The touch, intuition and patience on which the human puzzle is based are the ability to grow in the body, and AI cannot take it away — what is really passing quietly is our will to sit down and do one thing.
Concluding remarks
The 12-day white board is in my study.
I was asked what to spell. It's not good-looking. I didn't know before, and now I think it's clear: it's worth the 160 hours themselves — in pure white without color, no feedback, no shortcuts — all you can rely on is your own hands, eyes and that patience that you won't give up.
In the AI era, such “unnecessary hard work” will become increasingly ridiculous and valuable.
The real question is never “Can AI do this for me” but: Should I have done this?
By the way, I didn't finally let AI do it again. It's not about hurting it, it's about finding out that the biggest paradox about the AI puzzle is that if you really need it, you don't need it.
The facts of the paper are based on: several editions of the Beverly "Pholy White Hell" film and difficult experiences from the player community and media coverage; JigShape paper arXiv:2607.27670 (August 2026); arXiv:2506.08872 (in 2025, without peer review), conclusions as preliminary results; OECD 2026 Digital Education Outlook; Gerlich critical thinking study (Societies, 2025). The “160 Hours Best Record” in the text is an unverified version of the player community; Moravik paradox, Neuristic AI research is a classic consensus in the field.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.