Not just totally aligned, but also "just-in-time" and ephemeral: an LLM goes and does/builds it on request, and then destroys all artifacts when the request has been fulfilled?
If the persistence of something incurs a cost in storage and security, and the traditional penalties of efficiency and learning-curves are removed, why not recreate it every time its needed?
Maybe not "evolve" but certainly would be different every time, right? If it already knows every conivable way to generate something, it stands to reason it will be different every time, as every request will be a little different, even if it's just the temperature of the room.
As a thought exercise: if I could automatically generate, at no cost in money or time, a "disposable" vehicle for every single trip I took, wouldn't it make sense that it would be different every time, even though the basic requirement (transportation) is the same? Why bother generating a car with A/C if it's cold out? Why have a passenger seat if I'm just going to the corner store for milk?
Ah, yes, preference! How well will the AI know what we prefer without us having to articulate it every time. Very good, I like this.
Let's turn that on it's head: maybe we engage with AI/AI products that seem to just provide outcomes we prefer. Let's say all AI will probably be able to generate all manner of things, but we might like Z AI because when I ask for Y kind of widget it builds ones I seem to just...prefer more. It seems to "just get me".
It's interesting to think there's still a market for this type of small scale enthusiast application.
One scenarios I've been thinking about recently is the death of the generic app.
If LLMs end up being able to create software easily, why wouldn't we choose to have software that's totally aligned our specific needs?