Excellent job learning a new DSL

We gave Gemini a language spec for a DSL that was designed specifically for LLMs to generate plus a textual description of what we wanted it to write. The DSL is specific to our app, and therefore no model has it in training data. While a little slow (maybe ~30s per request), Gemini did a great job at learning the new DSL in-context and producing valid outputs that fit the description.