SO is probably a very significant factor in the success of LLMs but it's decline will not affect LLM development. LLMs will simply be trained on the conversations people are having with them.
Unlikely to happen with conversations, yes, but very likely to happen implicitly as part of work with coding agents. Successful agent trajectories will be used to learn what works and what doesn't for a given API.
People go to the llm to request help. How is that conversation going to be a good source to increase future knowledge?
No one is going to ChatGPT to explain to it how they solved a problem.