Startup types used to say that its not enough to be better than the current solution - you have to be 100 times better to justify the switching cost, and with LLMs it seems like that is like 1000x. It sucks.
Still, even assuming this is correct (which is not yet anywhere close to being certain), as long as there will be humans deciding what goes into production, decision making will be the bottleneck to address. If people rely on reading, it's too slow. Way too slow. If people only look at the system from outside, they will be making uninformed decisions.
Moldable Development offers a different option :)
I find myself wanting to talk about this in more detail, but this account is anonymous - could I email you?
Or contact me on social media.
How you create contextual tools is secondary. Emacs' ability to create extensions is certainly interesting. But here are some questions to consider: - how many contextual tools do you have for your system? 10s, 100s, 1000s? - or, for how many and what kind of questions don't you have contextual tools? and why is that?
At the extreme, when practicing Moldable Development, we tend to address dozens of questions per day per developer through contextual tools :)
The GUI possibilities of Emacs are a bit limited though. Glamorous Toolkit looks better in that respect. I need to actually try it!
Something possibly similar is Studio: https://github.com/studio/studio - also Smalltalk. But seemingly dead (or at least sleeping). I encountered this via its author's RaptorJIT project, a fork of LuaJIT apparently intended to turn it into something maintainable by mere mortals: https://github.com/lukego/blog/issues/19 - some videos here: https://www.youtube.com/@lukego/streams