AI labs need to start funding historical research

gist

Benjamin Breen argues frontier models can now help solve tractable historical problems and urges AI labs to fund historian collaborations. Using GPT-6 and Claude Opus 5.5, he reports progress on cryptography, cross-language text tracing, and niche findings such as a Hartlib–Newton alchemical anagram link. He asks labs to fund digitization, free API access for historians, and shared lists of open historical problems.

ideas

• Tractable historical problems look like math problems. They are expert-identified, digitized, matched to model strengths, codeable, and falsifiable, which is why reasoning models surged in mathematics before the humanities.
• Three problem classes fit those constraints. Cryptography and codebreaking, tracing texts across translations and adaptations, and linking findings stuck in niche subfields are the openings Breen thinks frontier models can help with now.
• Case studies mix null results with publishable leads. Astra treats John Dee’s Liber Loagaeth as mostly nonsense syllables; Opus links Hartlib and Newton via vitriol anagrams; Charles V cipher letters validate capability after prior solutions are found.
• Labs should fund infrastructure, not hobby compute. Breen wants collaborative digitization, free historian API access, and millennium-style open-problem lists instead of personal Codex and Claude Code allowances.
• Archival access is the binding constraint. Digitized manuscripts behind privilege walls, and the much larger undigitized premodern corpus, block work more often than model skill alone.

quotes