Great point. I regret not being more systematic about this. I have tried on most popular models since gpt 3.5 launched, but it’s all been very ad hoc with the same general approach of building up a language feature by feature.