I still have to spend a lot of resources on colloquialisation, but it's an ever-improving process. I'm standardizing the initial input ("ingestion") and getting better at distinguishing between when I should use colloquial "human" interaction and when I should skip formalities and make direct computational requests.
It still makes a lot of mistakes. I think I've accidentally unveiled some basic permutative operations the AI performs which drives these more esoteric decisions. Instead of giving me the permutations and descriptions I wanted, it was giving me the result of this code snippet:
for i, j in zip("103245", "103245"[1:]):
print(i, "->", j)
Whereas the solution would have been computable with:
for i, j in zip("012345", "103245"):
print(i, "->", j)
I'm struggling to communicate the concept of a bijection between the zero-based indices and the roles/functions themselves, and I don't want to use the term bijection because it diminishes the value of the terms role and function as they pertain to sociomathematical teams. The point is to incorporate certain intuitive mathematical notions which vibe with human quantitative reasoning, not to subsume sociology with mathematical terminology.
The full version of the most recent interview can be found here.