Using LLMs to trace alchemical knowledge and decode 17th century letters

The piece argues that frontier language models like GPT-6 Sol/Astra and Opus 5.5 should be actively funded for collaborative historical research because they can solve a class of well-defined, digitized, provable problems that benefit from multilingual reasoning, advanced math, and bespoke code. It lays out criteria for tractability and gives concrete examples: cryptographic codebreaking (GPT-6 Astra decrypted a stubborn 1941 Enigma message by assembling archival clues), tracing texts across translations (GPT-6 identified a Latin passage Newton had translated from a French alchemical source), and synthesizing dispersed findings across niche subfields. The key advantage is the models’ ability to notice and recombine disparate documentary evidence in ways human specialists had not.

Three case studies show how this works in practice. Astra found that John Dee’s Liber Loagaeth is largely nonsense syllables but detected repetition patterns and a genuine encoded reference to an angelic name. GPT-6 is being used to trace Darwin’s informants and references across multilingual corpora. Opus 5.5 downloaded thousands of Samuel Hartlib documents and cross-checked sources, uncovering different anagrammatic encodings for “vitriolum” used by Newton and Hartlib and marginalia confirming “vitriolum angaricum.” The conclusion advocates funding collaborations between AI labs and historians: these tools can surface meaningful, verifiable discoveries, even as their internal search strategies remain often opaque.