Savanna works in the same way as Nix: each training stage is defined fully in code, so it computes a hash over all inputs (training image, code version, environment, training configuration, etc.) and maps this to the eventual output. When the same inputs come along again, Savanna just reads the output from cache. As the post mentions, this is particularly helpful in sweeps that span multiple training stages.
I think it's cool that prompts are checked-in to version control, so the thought behind the action is captured. Also, there's a clear distinction between AI-generated and 'organic' code, which allows you to explicitly decouple core business logic from trivial grunt work. Although this particular implementation is likely not valuable, I feel there is value hiding somewhere in the concept. Certainly, the way humans currently collaborate with AI on code more often than not leads to unmaintainable software.
This sucks, though it’s formative. An experience that I value highly from my time studying was working on a group project in a team with misaligned goals: it teaches you how much it matters to find good people to work with in the real world!