Frontier LLMs move from transcribing history to actually solving it

A historian of science argues that the latest frontier models — GPT-6 (Sol/Astra) and Opus 5.5 — have crossed a threshold that didn’t exist a year ago: they can now attack genuine open problems in historical research rather than just transcribe documents. The models perform best on ‘tractable’ questions, which in the humanities means problems where experts have already defined the target, the source data is digitized, the work rewards multilingual reasoning or bespoke code, and — critically — a proposed answer can be proven or disproven. That last constraint is why reasoning models have flourished in mathematics but rarely in the humanities, and it limits their historical utility to a few areas: codebreaking, tracing texts across translations, and connecting findings scattered across niche subfields.

The connective work may prove the most valuable. When Astra cracked a July 1941 Enigma message that had resisted human analysts, the real breakthrough wasn’t the cryptanalysis but the model surfacing a collection of related German radio messages held in the Bundesarchiv — including file references that human experts couldn’t locate through their usual channels and can’t fully account for. The author frames this as a recurring pattern: the models are ‘maniacally determined’ on problems they judge solvable, pushing searches in directions specialists struggle to reconstruct after the fact. Other results are more modest. Turned loose on John Dee’s ‘angelic’ manuscript Liber Loagaeth, Astra concluded the text is mostly nonsense syllables produced by a charlatan scryer, though it cross-referenced character-repetition statistics against Dee’s diary and flagged one passage that genuinely encodes a name from Dee’s invented mythology.

The broader pitch is a call to action: AI labs, historians, and funding agencies should build deliberate collaborations, since expert domain knowledge plus frontier models plus heavy compute is already yielding unexpected findings — such as identifying a French alchemical source that Isaac Newton had loosely translated into Latin, an attribution apparently never made before. Not every experiment lands (fresh WWI/WWII ciphers to break have largely been picked over), but in one ongoing project the model itself proposed mining Darwin’s writings to uncover undocumented links between Darwin and his informants — a research idea the author judged as legitimate as a real dissertation topic.