[unknownhad]

I think this assumes software is a static target (Which it is not) . We are not just using LLMs to scan old code developers are using LLMs (like Copilot and others) to write new code and they are doing it by the shovel-load. The pace of shipping has gone up which means the pace of introducing new bugs has gone up right alongside it. The bug pool does not empty out because we keep refilling it every sprint.

Plus, the definition of the "easily found stuff" is a moving target. The AI models aren't static either. What takes a human reverse-engineer a week of deep insight today might just be a standard automated API call by 2027.

So while I would love for the dust to settle in a year, I think we are just looking at the new normal.

Thanks for reading the post and for the great counter-point!

[IcePic]

If we get new code by the "shovel", then it is likely trained on old bad code, so it might just (re)introduce old types of bugs by the shovel until all new fixed code overshadows the old code by a margin, which in turn will take a long while.