ArmeeGPT

What was the initial situation?

The Swiss Army bases its actions on embedding the training of its soldiers and leaders in regulations. This means that everything being trained must be regulation-compliant. A single regulation quickly amounts to about fifty pages; for specific training sequences, hundreds of pages are accumulated. This is not only inefficient in terms of logistics (space), but also in terms of handling: before you can search for something, you first have to find out where (in which regulation) it can be found.

What was the challenge?

Lesson planning for officer candidates means, among other things, sifting through hundreds of pages of regulations to ensure that the lesson is regulation-compliant. This is tedious and time-consuming. This is where we focused our efforts: By using an AI-based language model (LLM/Large Language Model) under the project name “ArmeeGPT”, we enabled fast and targeted searching, allowing the candidates to ask their questions the way they had them in mind. Additionally, learning content and formats could be created directly.

How did we solve this?

“ArmeeGPT” allows users to ask their questions as if they were facing a human expert. And in their respective mother tongue (even in dialect). However, in our prototype version, the responses were still mostly in standard German. Users could also ask follow-up questions (like in a “real” conversation) and receive suggestions on how to plan (e.g., with a schedule and agenda) and conduct a specific lesson.

The language model behind it was developed by a specialized Swiss technology company, evaluated by us, for the Swiss Army (and is continuously retrained). We used only publicly available data (all data sources on ArmeeGPT could also be found via Google), and the data hosting was in Switzerland; by using an anonymized standard login, we also ensured the protection of personal data.