The AI found new physics fast. Pólya would ask if it actually checked.
June 17, 2026 · George Pólya, How to Solve It~5 min read
The headline was the kind that makes a physicist's pulse jump: an AI had learned to sift the cosmos for signs of new physics faster than anyone thought possible, skipping the brutally expensive simulations the search normally demands. Then came the sentence the headline left out — the same shortcut can quietly go blind to the very thing it's hunting. A 2026 study had stumbled onto a 2,300-year-old lesson about how thinking actually works, and the man who wrote it down most clearly was a Hungarian mathematician with a thin, near-perfect little book.
The discovery that almost wasn't
The study, reported by ScienceDaily on June 11, put a technique called transfer learning to work on the Quijote simulations — a vast library of model universes, some obeying our standard ΛCDM cosmology, others laced with massive neutrinos or modified gravity. The differences between them are fantastically subtle, buried in how cosmic structure forms. Transfer learning lets an AI carry what it learned on one universe over to another, slashing the need to re-run simulations that can cost millions of compute-hours. The result was a real speedup. The catch, in the researchers' own words, is that the approach "can backfire when AI relies too heavily on familiar patterns" — leaning so hard on the physics it already knows that it could miss the signature of physics it doesn't. The tool built to find the new is biased, by its very nature, toward the old.
Pólya's oldest trick: have you seen this before?
In 1945 George Pólya published How to Solve It, a slim book that has quietly shaped how mathematicians, engineers, and teachers think ever since. Its engine is a four-step method — understand the problem, devise a plan, carry it out, look back — and its single most powerful heuristic for devising a plan is a question: Have you seen a related problem before? Reasoning by analogy, Pólya argued, is one of the deepest moves in all of thought: you crack a hard new thing by noticing its resemblance to an easier old thing you already solved. That is, precisely, what transfer learning is — analogy, automated and run at the scale of a galaxy survey. The AI is doing the most Pólya thing imaginable. Which is exactly why it inherits Pólya's warning.
The step everyone skips
Because Pólya was just as emphatic about the danger. Analogy, he said, is a magnificent way to find a candidate solution and a terrible way to confirm one. A resemblance suggests; it does not prove. And so the fourth step of his method — Looking Back, the one students and researchers and AIs alike are forever tempted to skip because the answer already feels right — is the one that does the real work. Looking Back means: take the solution the analogy handed you and test it as if you didn't trust it. Does it actually satisfy the conditions? Would it survive a case the analogy never saw? The figure below sets the two paths side by side: analogy alone, fast and quietly biased; analogy plus the verification step, slower and honest.
A June 2026 study (reported by ScienceDaily, June 11) used transfer learning on the Quijote cosmological simulations — comparing a standard ΛCDM universe against ones with massive neutrinos or modified gravity — to hunt for new physics far faster, without re-running costly simulations. The catch: the approach 'can backfire when AI relies too heavily on familiar patterns,' potentially blind to something genuinely new. That is George Pólya's How to Solve It in one line. Reasoning by analogy — 'Have you seen a related problem?' — is one of the most powerful ways to FIND a candidate solution. But his fourth and most-skipped step, Looking Back, demands you verify it: a pattern that fits the known data is a lead, not a proof. The heuristic that finds the answer is not allowed to certify it.
Why a pattern that fits is not a proof
The deepest trap in analogy is that a good fit feels like the truth. When a pattern learned on known physics maps neatly onto new data, every instinct shouts found it — and that confidence is exactly the failure mode. A genuinely new force, a genuinely novel particle, would by definition not match the familiar template; it would show up as a misfit, a residue the analogy waves away as noise. This is why the heuristic that finds the answer must never be the judge that certifies it. The AI's familiarity is its strength on day one and its blind spot on day two, and only a verification step it cannot perform on itself can tell the difference between a real discovery and a beautiful echo of what it already believed.
What this means past the lab
You don't run cosmological simulations, but you reason by analogy a hundred times a day — and increasingly so does the AI on your screen, both of you pattern-matching the new onto the familiar at speed. Pólya's discipline is the cheap insurance. When an analogy hands you an answer that feels obviously right, treat that feeling as a flag, not a finish line, and spend one deliberate minute on Looking Back: what would this miss, and what case have I not tested? The faster our tools get at finding patterns, the more the bottleneck — and the value — moves to the unglamorous step they keep skipping. Finding is cheap now. Checking is the whole job.
Analogy is a brilliant way to find an answer and a terrible way to confirm one.
The heuristic that finds the solution must never be the judge that certifies it.
Framework from George Pólya's How to Solve It (波利亚《怎样解题》) — the four-step method, reasoning by analogy, and the Looking-Back / verification step. The physics study was reported by ScienceDaily (June 11, 2026): transfer learning applied to the Quijote cosmological simulations (ΛCDM vs massive-neutrino / modified-gravity universes) to search for new physics, with the caveat that it "can backfire when AI relies too heavily on familiar patterns." Popular-science reading; figures and quotations per the original report.
科普
AI 飞快地「找到」了新物理——波利亚会问:它验过了吗
2026 年 6 月 17 日 · 波利亚《怎样解题》约 4 分钟
那条标题,是会让物理学家心跳一紧的那种:一个 AI 学会了在宇宙里筛查新物理的迹象,快得超出所有人预料,还跳过了这类搜寻通常要付的、贵到离谱的模拟。然后是标题没写出来的那句话——同一条捷径,可能悄悄对它正在追猎的东西视而不见。一项 2026 年的研究,撞上了一条关于「思考究竟如何运作」的、两千三百年的老道理;而把它写得最清楚的,是一位匈牙利数学家,和他那本薄薄的、近乎完美的小书。
那个差点不算发现的发现
这项研究由 ScienceDaily 于 6 月 11 日报道,把一种叫「迁移学习」的技术,用在了 Quijote 模拟上——一座庞大的「模型宇宙」库,有些遵循我们标准的 ΛCDM 宇宙学,有些则掺进了大质量中微子或修改后的引力。它们彼此的差别细微得惊人,藏在宇宙结构如何形成的细节里。迁移学习让 AI 把在一个宇宙里学到的东西搬到另一个宇宙,省去重跑那些动辄耗费数百万计算时的模拟。结果确实快了。而那个坑,用研究者自己的话说,是这套方法「在 AI 过度依赖熟悉模式时会适得其反」——太用力地倚靠它已知的物理,以至于可能错过它还不认识的物理的签名。这个为「找到新东西」而造的工具,因其本性,偏向旧的。
因为波利亚对其中的危险,同样毫不含糊。他说,类比是找到一个候选解的绝妙方式,却是确认一个解的糟糕方式。相似只能提示,不能证明。于是他那套方法的第四步——回顾,那个学生、研究者、连 AI 都永远忍不住要跳过的一步,因为答案已经「感觉对了」——才是真正干活的那一步。回顾的意思是:把类比递给你的那个解,当作你并不信任它,去检验它。它真的满足那些条件吗?它扛得住一个类比从没见过的情形吗?下面这张图,把两条路并排摆好:只靠类比,快,却被悄悄带了偏;类比加上那一步验证,慢,但诚实。
你不跑宇宙学模拟,但你一天要用类比推理上百次,而你屏幕上那个 AI 也越来越如此——你俩都在飞快地,把新事物往熟悉的模板上配。波利亚那套纪律,是一份便宜的保险。当一个类比递给你一个「明摆着对」的答案时,把那份「感觉」当成一面旗,而不是终点线,并郑重地花上一分钟去回顾:这会漏掉什么?哪个情形我还没验过?我们的工具找模式越快,瓶颈——以及价值——就越是挪向那个它们老想跳过的、不起眼的步骤。如今,「找到」很便宜,「验证」才是全部的活儿。