Using LLMs to trace alchemical knowledge and decode 17th century letters
3 hours ago
- Frontier AI models like GPT-6 and Opus 5.5 have advanced enough to help solve actual historical problems, not just perform research assistant tasks.
- Pairing historians working collaboratively with current frontier models could produce numerous meaningful advances in historical knowledge and interpretation.
- Tractable historical problems for AI require digitized data, multidisciplinary reach, provable solutions, and often involve cryptography, text tracing, or connecting niche findings across fields.
- Concrete examples include Astra breaking a 1941 Enigma message by discovering new archival sources, and Opus identifying Newton's use of anagram for Hungarian vitriol in Hartlib's papers.
- Astra's analysis of John Dee's coded Liber Loagaeth found it to be mostly nonsense syllables, but detected occasional encoded references, demonstrating both potential and limitations.
- Systematic efforts are needed: digitizing manuscripts, providing free API access and compute to historians, and collaboratively identifying 'millennium problems' in history similar to mathematics.
- AI can expand research questions beyond any single person's knowledge, though it also risks cognitive offloading if used merely for writing or replacing original thought.
- The author plans to survey historians for open problems suitable for this approach and invites collaboration.