It’s a rote observation at this point, but: there’s a clear discrepancy between the demonstrated value of LLMs (which is, to be clear, significant!) and the aggressive manner in which Microsoft has introduced them into their products. The latter is what you do when you can’t demonstrate value, and it produces worse outcomes by design (because it starts from the assumption that users need to be given the stick instead of the carrot).
There have been a lot of recent changes to VS Code that feel like this: the Copilot pane has been refactored to take up more space and behave less like other composed panes in the window; the integrated terminal now does overly clever and brittle things to introduce suggestions in REPLs like Python’s. Those kinds of changes have pushed me more to Zed recently, which has all of the same AI features but without the user hostility.