This approach has been around for a while and there are a number of applications/systems that were using the approach. The thing is that it’s not a different model, it’s just a different use case.
Its the same way OpenAI handle math, they recognize it’s asking for a math solution and actually have it produce a python solution and run it. You can’t integrate it into the model because they’re engineering solutions to make up for the models limitations.
All the evolution in AI right now is just trying different model designs and/or data. It’s not one model that is being continuous refined or modified. Each iteration is just a new set of static weights/numbers that defines it’s calculations.
If the models were changing/updating through experience maybe what you’re writing would make sense, but that’s not the state of AI/ML development.