IME, not long. ChatGPT (v3) did poorly with Zig when it first came out, despite some Zig in its training code, and despite equivalent Python tasks being handled

IME, not long. ChatGPT (v3) did poorly with Zig when it first came out, despite some Zig in its training code, and despite equivalent Python tasks being handled trivially. I worked around that for a bit by including a moderately lengthy prompt prefix with example Zig code and descriptions of interesting language features, which worked well for bulk boilerplate at the time. A year later, the model was updated (presumably with additional GH Zig source as training data) and works fine without any prompt prefix strategy, at least for the simple sorts of tasks I'd use an LLM for.