AI Combed Through 400 Years of Archives and Found a Forgotten Meteorite and Lost Rhinos

Jesse Waites's post "I pointed AI at 400 years of archives" earned 95 points and 47 comments on Hacker News. The author writes that he ran four centuries of material through AI and found a forgotten meteorite, lost rhinos and several other stories. The announcement does not show which archives he used, which model he ran or how much it cost. Below: what is claimed, how such projects are usually built and where they are weak.

What we know from the claim

The promise itself is simple: a large corpus of old documents that nobody has read in full, and a model that can read it in a reasonable amount of time. Four centuries of material means manuscripts, printed editions, correspondence, catalogs and possibly newspapers. Such collections are usually poorly indexed. A person finds what they search for by keyword, and everything else sits untouched.

The findings in the headline look like exactly the result of this kind of reading without a predefined query. A meteorite that was once described and then forgotten. Rhinos whose traces vanished from the records. If the author has not exaggerated, the model worked as a reader who does not need to know the question in advance.

How these projects are usually built

Projects like this are most often built in several steps. This is a general pattern, not a description of this particular post.

• Recognition: scans become text through OCR or handwriting recognition, and this is where the most quality is lost on old typefaces and handwriting.
• Tagging: an LLM extracts people, places, objects, dates and events from each document.
• Matching: records from different decades are linked together, and stories surface that no single document shows on its own.
• Verification: a human checks the findings against the primary source.

The last step determines whether you end up with a real find or a nice-sounding fabrication. Models tend to fill in missing context, especially when the text was poorly recognized. A "forgotten meteorite" is valuable only if a link to an archive page lets you see the actual record.

Where the pitfalls are

The first is recognition errors. If "rhinoceros" in the original turned out to be a different word, and the OCR or the model bent it to fit a striking story, the find disappears. The second is selection. The post reports what was found. The headline says nothing about how many false positives had to be thrown out. The third is that "forgotten" does not mean "unknown": historians of local collections may have known about the meteorite all along, and the information simply never reached wide access.

Given their number, the Hacker News comments most likely revolve around the same questions: how it was verified, how much it cost and which of the findings are genuinely new. The comments themselves are not quoted in the source, so this is a guess, not a summary.

Key parameters

• Material horizon: about 400 years.
• Claimed findings: a forgotten meteorite, lost rhinos, "and more".
• Community response: 95 points, 47 comments on Hacker News.
• Unknown from the announcement: model, corpus size, cost, share of false positives.

What to do with this

If you have an archive, whether corporate correspondence, ten years of technical documentation or a digitized library, the scheme from this post works as a recipe: first have the model read everything without a predefined question, then pick out the anomalies and check them by hand. Every finding should link to a specific page or file. Without that link, the result can be neither used nor defended.

Count the cost of recognition and the cost of tagging separately: on poor scans, the former is often more expensive than the latter. Before citing the post as proof of what AI can do, read the original and look for the methodology and examples of verification. The Hacker News announcement does not contain these details.