Using Opus 5.5 to discover a new eyewitness record of the dodo

Another day, another historical cipher broken by a frontier model.

又一天，又一个历史密码被前沿模型破解。

Yesterday, the security researcher Carter Church announced that he had used GPT-6 Astra (working on the problem for six hours) to break a Napoleonic-era cipher that had previously resisted all decryption attempts.

昨天，安全研究员卡特·丘奇 (Carter Church) 宣布，他已使用 GPT-6 Astra（研究了该问题 6 个小时）破解了拿破仑时代的密码，该密码此前曾抵制所有解密尝试。

Church’s post on the subject is an interesting example not just of a reproducible methodology, but also of how to vibe code an information-dense writeup that is notably different from any traditional academic aesthetic, but which actually does surface primary sources and meaningful information in a fairly deep way:

Church 关于这个主题的帖子是一个有趣的例子，不仅展示了可重复的方法论，而且展示了如何编写信息密集的文章，这种文章与任何传统的学术美学明显不同，但实际上确实以相当深入的方式呈现了主要来源和有意义的信息：

The aesthetic weirdness is just the tip of the iceberg here.

审美上的怪异只是这里的冰山一角。

What I notice most about these forays into historical sleuthing using AI (and my own attempts at same) is the epistemological weirdness of how current frontier models now operate when given historical research tasks.

关于这些使用人工智能进行历史调查的尝试（以及我自己的尝试），我最注意到的是当前前沿模型在执行历史研究任务时如何运作的认识论怪异。

The best way to demonstrate what I mean here is by sharing the step-by-step process by which I was able to find what appears to be a previously-unnoticed Dutch report of hunting dodos dating to 1615.

展示我在这里的意思的最好方法是分享一步一步的过程，通过这个过程，我能够找到一份看似以前未被注意到的荷兰关于 1615 年狩猎渡渡鸟的报告。

The core steps were:

核心步骤是：

Start from an actual base of specialist knowledge to define a specific research question (for example, I had previously researched the history of the exotic animal trade in the seventeenth century, and am planning to write a book on animal extinctions in the early modern period which will have a dodo chapter)

从实际的专业知识基础出发，定义一个具体的研究问题（例如，我之前研究过17世纪珍奇动物贸易的历史，并打算写一本关于近代早期动物灭绝的书，其中会有渡渡鸟一章）

Identify a large, freely-available, well-edited corpus of historical sources (in this case, the GLOBALISE archive of Dutch East India Company archives, which is an amazing resource)

确定一个大型、免费提供、经过精心编辑的历史资料库（在本例中，是荷兰东印度公司档案的 GLOBALIZE 档案库，这是一个令人惊叹的资源）

Download the sources and run them through an embedding model to allow semantic search to find passages that might help answer the research question

下载源代码并通过嵌入模型运行它们，以允许语义搜索找到可能有助于回答研究问题的段落

Use semantic search to surface candidate passages and then ask frontier models to read them and produce a ranked list of the best matches for human review

使用语义搜索来显示候选段落，然后要求前沿模型读取它们并生成最佳匹配的排名列表以供人工审核

Do any of the source passages help answer the question?

任何来源段落有助于回答这个问题吗？

If so, iterate on them.

如果是这样，请迭代它们。

If not, keep looking in new archives or with new search terms.

如果没有，请继续查找新的档案或使用新的搜索词。

This is not, on its face, all that different from how I do research on my own.

从表面上看，这与我自己进行研究的方式并没有什么不同。

I tend to search around in historical databases using various search terms that pop into my head, and then scan through the results until a passage catches my interest, then I read more carefully and iterate.

我倾向于使用突然出现在我脑海中的各种搜索词在历史数据库中进行搜索，然后浏览结果，直到一段话引起我的兴趣，然后我更仔细地阅读并迭代。

The difference is that an AI agent like Opus 5.5 can spawn dozens of copies of itself to read through sources in multiple languages.

不同之处在于，像 Opus 5.5 这样的 AI 代理可以生成数十个自身副本来读取多种语言的源代码。

If you give these agents an API key, they can also run their own embedding searches on new sources that emerge during their research, as Opus did here during the run that led it to the newly-discovered passage:

如果你给这些代理一个 API 密钥，他们还可以在研究过程中出现的新来源上运行自己的嵌入搜索，就像 Opus 在运行过程中所做的那样，从而找到新发现的段落：

Few historical creatures have been more widely studied than the dodo, the large Mauritius bird that famously went extinct due to overhunting around 1681.

很少有历史生物比渡渡鸟受到更广泛的研究，渡渡鸟是毛里求斯的一种大型鸟类，因 1681 年左右的过度捕猎而灭绝。

But this is actually a large part of why finding more information about the animal turned out to be “ tractable ” for an AI agent: it already had a very large identified source base to draw on, and it was able to scan the scholarly literature to find out where previous archival finds relating to dodos had been made (here is Opus 5.5 writing up its own process in a research dossier ).

但这实际上是为什么找到更多关于渡渡鸟的信息对于人工智能代理来说是“容易处理”的一个重要原因：它已经有一个非常大的已识别来源库可供利用，并且它能够扫描学术文献以找出以前与渡渡鸟相关的档案发现的来源（这里是 Opus 5.5 在研究档案中写下了自己的过程）。

Most of what Opus 5.5 found as it searched through the millions of records in the Dutch East India Company records was already known to the many historians and scientists who have studied the history of the dodo.

许多研究渡渡鸟历史的历史学家和科学家都已经知道 Opus 5.5 在搜索荷兰东印度公司记录中的数百万条记录时发现的大部分内容。

But one manuscript source from 1615, a ship’s log available here , seems not to have been noticed before.

但一份来自 1615 年的手稿来源，即此处提供的一艘航海日志，以前似乎没有被注意到。

The journal was probably written by Isbrant Cornelisz van Petten, who captained a Dutch East India Company merchant vessel called Wapen van Amsterdam .

该日记可能由 Isbrant Cornelisz van Petten 撰写，他是荷兰东印度公司一艘名为“瓦彭范阿姆斯特丹”号商船的船长。

The Wapen made landfall on the island of Mauritius in April, 1615, where the crew collected water and food to prepare for the rest of their voyage to the East Indies.

瓦彭号于 1615 年 4 月在毛里求斯岛登陆，船员们在那里收集水和食物，为前往东印度群岛的剩余航程做准备。

Among other things, they “caught many tortoises, dodos [ dodeersen ], and some geese and parrots.”

除其他外，他们“捕获了许多乌龟、渡渡鸟 [ dodeersen ]，以及一些鹅和鹦鹉。”

Here is the Dutch transcription and English translation of the relevant passages (note that the transcription and translation are not perfect — if you work on early modern Dutch please let me know of any corrections!):

以下是相关段落的荷兰语转录和英语翻译（请注意，转录和翻译并不完美 - 如果您研究早期现代荷兰语，请告诉我任何更正！）：

I read through the available secondary literature on the topic, like Parrish’s 2013 book The Dodo and the Solitaire , and other specialist articles.

我通读了有关该主题的现有二手文献，例如 Parrish 2013 年的书《渡渡鸟和纸牌》以及其他专业文章。

It genuinely looks like this is a new addition to the timeline of dodo, which previously had a gap in the 1611-16 period.

看起来这确实是渡渡鸟时间线的新补充，之前该时间线在 1611-16 年间有一段空白。

Now, as for whether this actually matters all that much: it not exactly earth-shattering.

现在，至于这是否真的那么重要：它并不完全是惊天动地的。

But I do think that this is a publishable result, especially when combined with another new finding from the same search, which found a probable new reference to another extinct bird from Mauritius, the red rail .

但我确实认为这是一个可以发表的结果，特别是当与同一搜索中的另一个新发现相结合时，该发现发现了可能对毛里求斯另一种灭绝鸟类“红轨”的新参考。

An account from 1638, the year the Dutch first colonized the island, turns out to describe “field-hens” using the Dutch word ( velthoenderen ).

1638 年，即荷兰人首次在该岛殖民的那一年，有一份记录使用荷兰语单词 (velthoenderen) 来描述“田野母鸡”。

Experts on the topic have previously identified this word as one that the Dutch used to describe red rails.

该主题的专家此前已确定该词是荷兰人用来描述红色铁轨的词。

But this particular reference was apparently missed because a French scholar in 1890 mistranslated the word as perdrix ( partridges).

但这个特定的参考文献显然被遗漏了，因为一位法国学者在 1890 年将这个词误译为 perdrix（鹧鸪）。

Opus 5.5 was able to go back to the original manuscript source and correct this.

Opus 5.5 能够返回原始手稿来源并纠正这一点。

For me, though, the most interesting possible finding is still very much in doubt.

但对我来说，最有趣的发现仍然存在很大疑问。

But if it can be nailed down, it really would be quite fascinating.

但如果能确定下来，那确实会很令人着迷。

This is because it might help explain one of the most mysterious paintings from the 17th century: the Mughal Emperor Jahangir apparently owned a living dodo.

这是因为它可能有助于解释 17 世纪最神秘的画作之一：莫卧儿皇帝贾汉吉尔显然拥有一只活渡渡鸟。

However, no one knows who gave it to him, when, or why, and Jahangir never writes anything about it in his memoirs .

然而，没有人知道是谁、何时、为何把它给了他，贾汉吉尔也从未在他的回忆录中写过任何有关它的内容。

Two years earlier Mansur had painted a Mauritian dodo that is still cited by biologists as the most accurate surviving representation of the bird.

两年前，曼苏尔画了一只毛里求斯渡渡鸟，至今仍被生物学家视为现存最准确的鸟类代表。

In other words, Jahangir was exactly who a canny merchant or courtier would go to if they came across a highly unusual-looking bird.

换句话说，精明的商人或朝臣如果遇到一只长相极不寻常的鸟，就会去找贾汉吉尔。

It turns out that no one really knows the exact date of this painting.

事实证明，没有人真正知道这幅画的确切日期。

Some scholars go with “circa 1625,” others with “circa 1615.” (We do know that an English merchant reported the existence of Mauritian dodos in India in 1628, but whether this included the dodo shown in the painting is unknowable).

一些学者认为“大约 1625 年”，另一些学者则认为“大约 1615 年”。 （我们确实知道，一位英国商人于 1628 年报告了印度存在毛里求斯渡渡鸟，但这是否包括画中所示的渡渡鸟不得而知）。

To my great surprise, Opus 5.5 dug around in a very wide range of manuscripts to surface the following theory: Jahangir’s dodo was quite possibly the same animal that a Portuguese Jesuit described on Mauritius in 1616.

令我大吃一惊的是，Opus 5.5 在广泛的手稿中进行了挖掘，得出了以下理论：贾汉吉尔的渡渡鸟很可能与葡萄牙耶稣会士 1616 年在毛里求斯描述的动物是同一种动物。

The Jesuit referred to this Mauritian bird as an “ostrich.” The thing is, Mauritius has no ostriches!

耶稣会士将这种毛里求斯鸟称为“鸵鸟”。问题是，毛里求斯没有鸵鸟！

Unfortunately, the original Portuguese text of this account is now thought to be lost, but a French translation survives, and that’s what Opus is citing here:

不幸的是，这篇报道的葡萄牙语原文现在被认为已经丢失，但法语翻译幸存下来，这就是 Opus 在这里引用的内容：

The theory that this “ostrich” was actually a dodo is already known to experts on the topic, and the French translator glosses it as such.

这只“鸵鸟”实际上是渡渡鸟的理论已经为该主题的专家所知，法语翻译者也这样解释它。

But it appears that the connection between this Portuguese dodo caught en route to Goa in 1616 and the dodo that ended up reaching Jahangir at some point between 1615 and 1625 has not yet been made.

但这只 1616 年在前往果阿途中捕获的葡萄牙渡渡鸟与最终在 1615 年至 1625 年间某个时间到达贾汉吉尔的渡渡鸟之间似乎尚未建立联系。

Granted, the chain of transmission has not been established.

当然，传播链尚未建立。

But my hunch is that this may in fact be the origin of Jahangir’s dodo.

但我的预感是，这实际上可能是贾汉吉尔渡渡鸟的起源。

This is because we happen to already have a very clear chain of transmission of another exotic bird from the Portuguese based in Goa to Jahangir’s court, dating to 1612: an American turkey!

这是因为我们碰巧已经掌握了另一种外来鸟类从果阿葡萄牙人到贾汉吉尔宫廷的非常清晰的传播链，其历史可追溯至 1612 年：美洲火鸡！

This is something I plan to dig into more.

这是我打算进一步深入研究的内容。

If it can be traced more reliably, I think the identification of Jahangir’s dodo would be a pretty big deal.

如果能够更可靠地追踪，我认为贾汉吉尔渡渡鸟的识别将是一件非常大的事情。

The Mughal court dodo is possibly the most famous and scientifically important individual dodo that ever lived, because Jahangir’s brilliant court painter, Ustad Mansur, left behind the most accurate surviving depiction of it.

莫卧儿宫廷渡渡鸟可能是有史以来最著名、最具科学意义的渡渡鸟个体，因为贾汉吉尔杰出的宫廷画家乌斯塔德·曼苏尔留下了现存最准确的渡渡鸟描绘。

At minimum, I can say that this deep dive into the Dutch records has made it clear to me that frontier AI models are capable of surfacing new historical finds based on independent archival sleuthing.

至少，我可以说，对荷兰记录的深入研究让我清楚地认识到，前沿人工智能模型能够根据独立档案调查得出新的历史发现。

That’s not something I could have said months ago.

这不是我几个月前就可以说的话。

What they can’t currently do is ask the right questions or determine the significance of finds.

他们目前不能做的是提出正确的问题或确定发现的意义。

The main thing that contemporary AI can do for historical research is, in effect, the digital equivalent of counting sheep.

实际上，当代人工智能在历史研究中所能做的主要事情就是数字化地数羊。

Because they never get bored, they can search through enormous datasets to find new evidence for existing claims (or, potentially, disprove them).

因为他们永远不会感到无聊，所以他们可以搜索大量数据集来寻找现有主张的新证据（或者可能反驳它们）。

They are worse at coming up with new ideas of their own.

他们更不善于提出自己的新想法。

What seems to work best is if they are placed on the boundary between two disciplines, given a source base, and told to work methodically toward answering a human-generated question .

似乎最有效的方法是将它们置于两个学科之间的边界，给定源基础，并告诉它们有条不紊地回答人类生成的问题。

They are also notably bad at judging the historical significance of what they find.

他们还特别不善于判断所发现的东西的历史意义。

Again and again, in my attempts to find something tractable, Opus 5.5 and GPT-6 ended up spawning up to a dozen independent agents that that drilled down into minutiae and got utterly lost in the weeds.

在我一次又一次尝试寻找易于处理的东西时，Opus 5.5 和 GPT-6 最终产生了十几个独立代理，它们深入细节并完全迷失在杂草中。

Sometimes, the weeds ended up being fun.

有时，杂草最终会变得有趣。

For instance, one agent discovered a very entertaining and novel account of an English ship captain, Jonathan Hide, who stole valuable ebony wood plus “two sea cows” (!) from the Dutch colony in Mauritius.

例如，一名特工发现了一个非常有趣和新颖的故事，讲述了一位英国船长乔纳森·海德（Jonathan Hide）从毛里求斯的荷兰殖民地偷走了宝贵的乌木和“两只海牛”（！）。

When a Dutch official confronted Hide and his crew, “their carpenter threatened to split my head with his axe.” The official adds:

当一名荷兰官员与海德和他的船员对峙时，“他们的木匠威胁要用斧头劈开我的头。”官方补充道：

The Captain also took about 20 land tortoises on the voyage, claiming he intended to put them ashore on St Helena, so as to breed the said tortoises there .

船长在航程中还带了大约20只陆龟，声称打算将它们放到圣赫勒拿岛的岸上，以便在那里繁殖这些陆龟。

This is historically significant as an environmental history finding.

作为一项环境历史发现，这具有重要的历史意义。

It turns out that the animal in question, the Mauritius giant tortoise , also ended up going extinct.

事实证明，这种动物毛里求斯巨龟也最终灭绝了。

Whether these twenty individuals ever did end up on St. Helena, many thousands of miles away, is unknown.

这二十个人是否最终到达了数千英里之外的圣赫勒拿岛，目前尚不清楚。

But we do know that Captain Hide’s ship sank near the Azores on the return voyage home.

但我们确实知道，海德船长的船在返航途中在亚速尔群岛附近沉没。

This is a pretty great story!

这是一个非常棒的故事！

Early modern people transplanting crops and animals is very much in my wheelhouse, and I love the carpenter brandishing an axe at the Dutch official.

早期现代人移植农作物和动物是我的驾驶室，我喜欢木匠向荷兰官员挥舞斧头。

I will likely use it in some of my published work at some point.

我可能会在某个时候在我发表的一些作品中使用它。

Most of the time, however, agents veered off into archives that are totally illegible to me and require very extensive specialist knowledge to fact check.

然而，大多数时候，特工们会转向我完全难以辨认的档案，并且需要非常广泛的专业知识来进行事实核查。

For instance, this was the product of a multi-hour search into Inca khipu :

例如，这是对 Inca khipu 进行数小时搜索的结果：

Is that decisive?

这是决定性的吗？

Does any of this mean anything?

这些有什么意义吗？

I genuinely have no idea.

我真的不知道。

I assume Gary Urton does.

我想加里·厄顿会的。

But I don’t really want to bother someone who has dedicated his life to the study of Inca khipu with a question generated by an AI agent which spent a few hours on the same topic.

但我真的不想用一个人工智能代理提出的问题来打扰一个毕生致力于印加奇普研究的人，而这个问题是由人工智能代理在同一主题上花了几个小时提出的。

This is really the crux of the current problem in a number of fields that AI research agents stand poised to transform: how should human experts interface with these things?

这确实是人工智能研究人员准备改变的许多领域当前问题的症结所在：人类专家应该如何与这些事物互动？

I feel comfortable working with the material relating to, say, English merchants stealing giant tortoises, or Jahangir’s dodo, because this is the sort of stuff I’ve been studying for well over a decade now.

我对处理与英国商人偷窃巨型乌龟或贾汉吉尔渡渡鸟有关的材料感到很舒服，因为这是我十多年来一直在研究的东西。

But what happens when amateurs running AI research agents veer off into genuine new scholarly territory, without the knowledge or network to double check it?

但是，当运行人工智能研究代理的业余爱好者在没有知识或网络进行双重检查的情况下转向真正的新学术领域时，会发生什么？

This is what Carter Church gets at in the post I linked to at the top:

这就是卡特·丘奇在我在顶部链接的帖子中得到的内容：

A patient specialist could have done this in 1970.

1970 年，一位患者专家就可以做到这一点。

But it would have meant weeks of a specialist’s time: period French, a steady eye for 1,300 hand-drawn signs, and the statistics, all spent on one plate in one army journal.

但这意味着专家要花费数周的时间：用法语、对 1,300 个手绘标志和统计数据进行密切关注，所有这些都花在一本军队杂志的一个盘子上。

People with those skills exist, and they have more important problems to solve.

拥有这些技能的人是存在的，他们有更重要的问题需要解决。

So the letter sat unread, because no one who could read it could justify the time.

所以这封信一直没有读，因为没有人能读懂这封信，无法证明时间合理。

I spent six hours of model time and a few evenings of my own.

我花了六个小时的模特时间和几个晚上的时间。

Tomokiyo says he’s receiving solutions faster than he can record them.

Tomokiyo 表示，他接收解决方案的速度比他记录解决方案的速度还要快。

I’m sure few of these are from cryptographers.

我确信其中很少有来自密码学家的。

A year ago that sentence probably wouldn’t have made sense.

一年前这句话可能没有意义。

I’d guess most “unsolved” lists, in most fields, hold more problems like this than anyone assumes.

我猜想，在大多数领域，大多数“未解决”的列表中存在的此类问题比任何人想象的都要多。

Those lists are about to get a lot shorter, and the people shortening them will increasingly be people who were simply curious.

这些列表将会变得越来越短，而缩短列表的人将越来越多地是出于好奇的人。

Personally, I would be thrilled if a bunch of curious amateurs running a small army of AI agents dug into the dodo material, or into the early modern drug trade, or the history of consciousness science in the 19th century, or any number of other things I’m actively researching.

就我个人而言，如果一群好奇的业余爱好者管理一支人工智能代理小军，深入研究渡渡鸟材料、早期现代毒品贸易、19 世纪意识科学的历史，或者我正在积极研究的任何其他事物，我会感到很兴奋。

But it is going to get really weird, really soon, when these things happen.

但当这些事情发生时，很快就会变得非常奇怪。

The bottleneck will soon become not research findings themselves, but the attention of experts in niche topics .

瓶颈很快将不再是研究成果本身，而是专家对利基主题的关注。

A guest post on mathematician Terrence Tao’s blog recently concluded: “We’re gonna need a lot more mathematicians.”

数学家陶哲轩博客上的一篇客座文章最近得出结论：“我们将需要更多的数学家。”

I agree, and I would add: we’re gonna need a lot more historians and humanists.

我同意，并且我想补充一点：我们将需要更多的历史学家和人文主义者。

• Who wrote Queen Elizabeth I’s most scathing letters?

• 谁写了伊丽莎白一世女王最严厉的信件？

( Smithsonian - this is a great example of original historical analysis based on cross-comparison between manuscript sources).

（史密森尼 - 这是基于手稿来源之间的交叉比较的原始历史分析的一个很好的例子）。

Res Obscura is a reader-supported publication.

Res Obscura 是一本由读者支持的出版物。

To receive new posts and support my work, consider becoming a free or paid subscriber.

要接收新帖子并支持我的工作，请考虑成为免费或付费订阅者。

Thanks for sharing this.

感谢您分享这个。

I'm a reporter on the Visual Investigations team, at The New York Times, and we often think about doing a historical Visual Investigation.

我是《纽约时报》视觉调查团队的一名记者，我们经常考虑进行历史视觉调查。

So far, we haven't yet.

到目前为止，我们还没有。

But these kind of things are inspiring – especially if you are getting more confidence in that Mauritius-to-India link.

但这类事情是鼓舞人心的——特别是如果您对毛里求斯与印度之间的联系更加有信心的话。

Hi Ben, I wasn't able to read your Opus 5.5 research dossier – could you please check if your link to that artifact is accessible to all?

你好，Ben，我无法阅读你的 Opus 5.5 研究档案 - 你能检查一下你指向该工件的链接是否可供所有人访问吗？

I'd previously created an MCP server to the Globalise VOC corpus ( https://github.com/kintopp/globalise-mcp ) and a quick search there for Dodos using several early modern Dutch name variants came up with this: https://claude.ai/share/156b605d-827e-40db-b808-efc8a31e2680

我之前为 Globalize VOC 语料库创建了一个 MCP 服务器（https://github.com/kintopp/globalise-mcp），并使用几个早期现代荷兰语名称变体快速搜索 Dodos 得出了以下结果：https://claude.ai/share/156b605d-827e-40db-b808-efc8a31e2680