So is that ;) But yeah, the link between context length and performance reduction is well researched. Automatic context management has been getting better, but it’s not a panacea, especially in very large codebases and long sessions.

> Maybe we are ok accepting some "inefficiencies" in the code for the LLM but we get better CPU utilization or cache/RAM usage

If by ‘inefficiencies’ you mean choosing hardware-friendly architectures over those that correspond to a human-friendly world model, then I would agree with you (this was already required in my domain pre-LLMs). Even still that does not stop one from adhering to DRY, KISS, etc as the abstractions required are often either free or close to free at runtime.