🤔 if i’m rewriting code in a high level language using an LLM, what’s the crossover point where the resulting program runs as many instructions as the model ran in creating it? perhaps this can be called the LLM code break when point. how many programs never reach this point in all the times that they are run, and might better have been computed on the model directly.

this leads to the thought of dynamic single shot models as an efficiency improvement.