A SentinelOne engineer, Carter Church, has utilized the GPT-6 Astra AI model to decipher a letter from Napoleon’s era that had remained undeciphered for 217 years. This breakthrough, detailed in a blog post, demonstrates the growing capability of artificial intelligence in solving complex historical ciphers.
Broader Context
Source materials place the factual news in this context: SentinelOne engineer Carter Church used GPT-6 Astra for about 6 hours of model time to read a letter written under Napoleon’s orders, which no one had deciphered for 217 years, from a 1969 journal illustration.
The AI Breakthrough
A letter believed to be from 1809, discovered on a scanned image from a 1969 French military journal, the Revue historique des Armées, had remained inaccessible for over two centuries. The journal itself had mislabeled the year as 1807. According to Carter Church, the GPT-6 Astra model processed the text for approximately six hours to achieve the decipherment, a task that would have previously required extensive human expertise and time.
Historical Context and the Letter’s Content
The letter was originally intended for General Marmont and was written under Napoleon’s orders, though issued from Eugène de Beauharnais’ headquarters. The first line, in clear French, reads: “General Marmont, you should have received my letters of the 8th, 14th, and 20th.” The remainder of the letter consists of 24 lines of coded text, employing a homophonic cipher with approximately 1,300 cipher units and 155 different symbols, making traditional frequency analysis insufficient.
French cryptohistorian Daniel Tant had previously identified 33 letter pairs, deciphering about one-third of the text. The remaining two-thirds required further inference. Astra’s approach involved techniques like simulated annealing combined with French letter combination frequencies. A crucial step involved cross-referencing symbol interpretations for consistency, and a counter-proof reported the decipherment was not derived from the training data itself.
The deciphered content reveals troop deployments: 40,000 Bavarian, 30,000 Polish, Saxon, Davout 80,000, Masséna 60,000, and Oudinot 40,000 troops, along with information about a Russian advance into Austria. These details precisely match Napoleon’s correspondence, Volume 14908, dated March 16, 1809, confirming the letter’s authenticity.
Implications for Historical Research
Following the decipherment, Satoshi Tomokiyo, who maintains the Cryptiana list of unsolved historical ciphers, updated the entry to ‘solved’ and corrected the date to 1809. This event highlights how AI is breaking down barriers in historical research. While human experts could have potentially deciphered the letter with significant effort, the cost and time investment were substantial obstacles.
According to Carter Church, the primary bottleneck for solving such historical codes has historically been cost, a barrier now being lowered by AI. This development suggests that many previously unsolved historical mysteries may now be within reach, with human curiosity being the main remaining requirement for further discoveries. Uncertainties remain regarding the exact key used for the cipher and whether other historical codes are similarly solvable with advanced AI tools.
