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图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Man, Yan, Skye has something to say.
1,000 pure white hell puzzles, no pattern, no color difference, no reference, human player time is 12 days to finish. I thought AI was good at identifying reasoning, so it was easy to get through customs, and I didn't think it would last five hours.
Rats
It started with a boring brush.
Late that night, I accidentally drew an open-box video from Douyin: The owner opened up a box of puzzles and fell out, and it was white -- not mostly white, but all. 1000 pieces, no drawings, no colours, no border reference diagrams, 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. The next day, the bill was issued. 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 the white board on the table, and the first thought came out not of achievement, but of 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 images, you don't have to be creative. You just have to look at shapes and compare them. Isn't this the standard scenario where AI is fed and elasticized?
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 thing any rational person can think of:
Draws a 20 x 50 grid on the puzzle table, a total of 1,000 cells, each numbered, and 1,000 pieces of fragments in one cell, which is taken and sent to AI. There's only one thing it's going to do — to judge which edges can bite.
Under this scheme, it would have taken 160 hours to test blindly, and theoretically it would have been compressed into a “single-cracker”: the Ai map, the characterization, the calibration, the output matching table, and I'm in charge of proofing it.
I passed the picture on.
First hour: It's the first thing he talks about.
The initial conversation was quite pleasant. AI accurately says the hard point of pure white hell: "No colour clues, only edge shape matching." It also helped me to analyze the strategy of the Scripts first, like an experienced puzzle teacher.
I let it judge whether the A12 and B07 pieces in the picture are fit. And it said, "Looking at the picture, the edge curvature of the two pieces is complementary, and the probability is likely to match."
Sounds professional. I'm going to take a physical -- it won't work.
Hour 3: I started asking questions
I tried a few more, and the accuracy rate was overwhelming. So I changed the question, zoom in on the two pieces and spread them over. It's still exporting the right kind of crap: "There are fine differences in the contours of the edge that require more precise measurements."
One sentence raises the question: it describes, not measures.
It tells me that “this scene is steeper”, but what the puzzle requires is that “the focal point is 37.2 per cent on the edge”. It can speak every time, but twice it is the same and the wording will differ. And for the puzzle, the difference is not.
And even worse, when I put it in the order of 30 pieces in a row, it gave a closed circle of "A to B, B to C, C to A." It makes local judgements on a case-by-case basis and does not have the capacity to maintain a whole-of-the-art puzzle — every step is like the first.
Five hours: I got a verdict.
I went to the literature and found a sad fact: Academies have tortured AI with puzzles long ago and use much cleaner topics than my experiments.
In August 2026, a study by Jig Shape tested five front-line visual LLM. The title is a computer-composed puzzle: shape digitized, strictly unique, perfect image — without the vagueness of the photographs and the uneven light in my program. Results: 16 pieces of puzzle, only one of the five models exceeds the level of blind guessing; 64 pieces return to random groups; and slightly adjusted models reach 144, with a correct rate of less than 5 per cent.
The researchers gave the phenomenon a name: a cliff of scale. The number rises, more binding, and the entire model collapses.
My five-hour experiment essentially took a step on a cliff.
When I found out about 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 it turns out that there is a difference between knowing a conclusion and hitting 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 been conducting several decades of neuronyms combined studies. 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.
I've been thinking since the experiment failed: where is it stuck?
To figure out, LLM lost the game, not in “seeing”, but in two lower places.
First, the word “measure”. The way LLM processes images is essentially to generate a description of the images — it is 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 warp radius of 1.8 mm”. The puzzle is defined by a physical bite, with a difference of a few millimeters. The description cannot be puzzled, so the measure can be. Even more problematic is the fact that its two descriptions of the same piece of pieces may be worded differently — a system whose output cannot be repeated and which does not perform the work of precision that requires zero error.
And this is the real cause of my "A to B, B to C, C to A" loop. The part of the human puzzle that has been put together on the table is the anchor: for each piece of new piece, you're not blindly selected from a thousand pieces, but you're looking for a place in an area that's already formed, and you change it to a new anchor. This is a process of cumulative certainty and retroactive amendment at any time. And LLM does not have a table - every time it answers, it re-examines it from scratch, and the judged pairs do not sink into “established facts”. The thousands of constraints that need to be put together in their minds without fighting each other are precisely what is not in their structure.
This is what the “crawl of scale” says: in 16, dozens of constraints can barely be overcome; in 64, 144, the binding quantity index increases, and every small error spreads across one another and ends up losing.
It's not smart, it's hand and time.
After trying to figure this out, I thought back to the question: How did I finish 12 days?
The speed of calculation is much less consistent than the match of the two curves. But I have something that LLM doesn't have: touch.
I don't even see it. The finger stomps on the pellets, the dents and the bumps, and the moment it gets in has a slight "bite feeling" -- that's right, that's right, that's right. This feedback is real-time, physical, zero delay, and LLM is separated by a photograph and never gets it.
In addition to hands, there's intuition. After a hundred pieces, I had a muscle memory of the shape of the knife: it had a vertebrae, a dent, a new piece, a finger that probably knew where to look. It's not a calculation. It's a mode recognition that's been tested 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 is progress. This is not the capacity, it is the will, and this is precisely what the machine never needs to have, and human beings can lose at any time.
So the real pattern of this contest is that LLM has a thousand-mile eye, but no fingers or workstations; human fingers are sophisticated meters, tables are natural receptacles, and the only short sheet is patient upper limits.
Inverted: Actually AI can spell, using 40 years old technology.
Here, it should have ended with "Ai." But I looked over it and found a more interesting fact.
Automatic puzzle resolution is an old, decades-old topic in the field of artificial intelligence, and the papers are published at official AI meetings such as IJCAI and AAAI. The classic solution is not at all complete: scanning each piece, extracting the contours of the edge, turning each side into a string of precise pixels coordinates; calculating the difference between the two sides using the Oxylene distance; and having obtained a 1,000 x 1000 error matrix, handing it over to the search optimization algorithm for the best solution in the world. On this route, hundreds of large, uncharted puzzles have long been automatically solved.
This system is standard AI as defined in AI textbooks - perception, knowledge expression, search, planning.
So the exact conclusion is not "AI can't spell pure white hell," but:
AI can spell pure white hell with "old AI" and it's "new AI".
Old AI turns the problem into math, and then counts -- it doesn't have to read pictures, it's volume. New AI learns patterns from data, and then produces -- it doesn't matter, it talks. The smarts learned are open and inclusive, and the smarts created are accurate and reliable. And the task of “rules closed, only answer, zero tolerance” in pure white hell happens to be the home of built intelligence.
My five-hour breakdown was not AI, but I gave it to the wrong subject -- like having a learned translator run the car bed, not being stupid, but not the job.
Something that really deserves to be warned.
Is this just a pop-up job? Actually, I'd like to say the most.
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. In the last four months, the performance of the group with AI has been completely backward at the neuro-linguistic and behavioural levels, with the weakest neural connection, and it has been difficult to quote exactly what it “written” just a few minutes ago. In the Digital Education Vision 2026, the OECD gives a name for such phenomena: a measurability laziness — outsourcing thinking to create a “knowledge-held illusion”. Another study, covering 666, found that the frequency of the use of AI tools was significantly negative in terms of critical thinking, the worst among young people.
Put these studies and puzzles together, I see a directional error:
We are outsourcing what we would have done, and we expect AI to solve what it can't solve.
Writing, thinking, judging — these things that humans are good at, we can't wait to hand over LLM, and then the cognitive ability goes down; and pure white hell, the exact match that should be given to the “old AI”, is a problem that we are giving LLM to talk, and then disappointed in it. Both of them are upside down.
And the biggest reminder of the puzzle is that on the list of "outsourcing thinking" the work is actually at the end. The touch, intuition and patience on which human puzzles are based are the ability to grow inside the body, and AI cannot take it away — what is really passing away is our willingness to sit down and do something.
Concluding remarks
That 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 per se — 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 do this?
By the way, I didn't finally let AI do it again. It's not that it hurts, it's what I found: the biggest paradox about the AI puzzle is that if you really needed it, you wouldn't need it.
This paper is based on the following facts: several editions of the Beverly White Hell film and difficult experiences from players' communities and media reports; JigShape paper arXiv: 2607.27670 (August 2026); arXiv: 2506.08872 (in 2025, without peer review and with preliminary results); OECD Digital Education Outlook 2026; Gerlich critical thinking study (Societies, 2025). The “160 Hours Best Record” in the text is an unverified version of the player community; the Moravik paradox, the neural symbol AI study is a classic consensus in the field.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.