Unlocking History: AI Rediscovered Lost Meteorites and Eruptions

Software engineer Jesse Waites used a custom AI pipeline to search 400 years of historical archives, uncovering a forgotten 1812 meteorite, three lost rhinos, and three unrecorded volcanic eruptions. By utilizing a multi-stage filtering process with models like Claude, he was able to verify these findings against original handwritten scans and modern scientific catalogues. Waites has now open-sourced his 'Antiquity' toolkit to help others automate the discovery of lost historical data.

Key Points

• The AI pipeline processed 4.35 million pages of archives in just twelve hours, a task that would take a human researcher approximately 70 years.
• A multi-layered 'agentic workflow' was used to filter millions of passages into a few dozen high-probability candidates for human verification.
• The project rediscovered a lost 1812 meteorite fall, three Javan rhinos previously unknown to specialist literature, and three unrecorded volcanic eruptions.
• The author emphasizes that while AI handles the scale of reading, human judgment is still required to steer the investigation and verify findings against original scans.
• The research toolkit, Antiquity, has been open-sourced to allow other researchers to conduct similar AI-assisted historical investigations.

Sentiment

The overall sentiment is positive and constructively critical. Most commenters agree that AI-assisted archival research is promising and that this article demonstrates a useful application, but they also emphasize that historical claims need expert review, source criticism, and careful handling of translation and OCR uncertainties. The thread is not hostile to the article; its skepticism is mainly about validation and overclaiming rather than rejecting the value of the work.

In Agreement

• The project is a strong positive example of LLMs because it uses them to sift through archival material that would be impractical for one person to read manually.
• The discoveries are worthwhile leads even if they require follow-up, and similar results from older NLP or OCR methods would likely be treated as legitimate research aids.
• Open-sourcing the Antiquity workflow makes the work more valuable because others can inspect, adapt, and apply it to new historical questions.
• AI-assisted archival search could uncover many other overlooked records, including weather events, maritime routes, lost cargo, unusual animal records, and cross-language references.
• LLMs are useful as relentless filters over boring archival material, helping humans find the rare passages that deserve expert attention.
• Large-scale preprocessing and structured extraction could make historical corpora more searchable and enable cheaper downstream research questions.

Opposed

• A non-expert using AI to select fields and questions may produce shallow findings or overstate significance because they lack the context needed to judge importance.
• The claims depend on reliable translation, OCR, and interpretation of historical Dutch, German, handwriting, and colonial archive material, all of which can introduce errors.
• Finding highlights in an archive is not the same as understanding the archive; mundane records, gaps, and patterns can be central to historical knowledge.
• AI-mediated work creates a verification problem because readers must work harder to determine whether a claimed discovery is real, novel, or meaningful.
• Some commenters worry that broad corpus-analysis tooling can become ethically risky when applied to contemporary people, media, or political content.
• Several readers found the article's visual effects distracting or performance-heavy, which interfered with reading even if they liked the underlying project.