I realize what follows is quite a digression and probably not about what you intended, but after reading a few of your posts and typing this out, I figured I'd share anyway in case it resonates.
> generate new responses to text or image inputs by the user
I'd pick some fault with the choice of name "process generative". The cited artifacts seem to be neither of those things. They can be embedded into generative processes for sure, but it'd be the people who interact with them that make them processal and generative, in my view, not the models themselves. Below is the tl;dr:
I'd say the use of "new" in the quote above is load bearing. Technically and historically, the latent diffusion model underlying many image generators was developed to represent complicated probability distributions in a concise set of parameters. To use this model as an image generator, one must collect a set of samples of the image distribution one hopes to represent, and then apply a training procedure to develop the parametric representation. There's already several layers of representation happening here, each with a corresponding fidelity loss but also a kind of "reality loss" if you want to call it that: the subjects of images -> the images themselves -> vector representations of images -> a probability distribution over the vector representations -> parameters representing distributions over vector representations.
Once you finally have the parameters in hand, you can then sample from the represented distribution. This is the "generation" step, and what you are referring to as "new". I'd argue both words are inappropriate here in any but their jargon senses.
Something like ChatGPT has a similar flavor, though it arose from sequence-to-sequence translation research, which is not explicitly about representing complicated probability distributions. However, implicitly that's what it's doing (it's a descendant of conditional random fields, which were more explicit about this aim). At base when you enter a prompt, you're drawing a sample from a conditional probability distribution over sequence space, conditioned on the prompt sequence (I'm ignoring the guardrails and other wrappers around the core LLM for brevity).
So, what exactly is "new" in a sample from a probability distribution? Arguably nothing. Users might be surprised because they were not previously aware that some particular sample was "in there". But it's the kind of surprise a street magician trades in with a two-headed coin, the kind of surprise that happens in a board game. What we generally think of as "new", "novel", "creative", usually happens in the realm before this stack of representations, not several layers deep in it. Or it plays with the representations themselves, rather than keeping them fixed and sealed. Or it knocks the board game over entirely. Or it comes up with some other thing I haven't listed.
What exactly is "generative" about a sample from a probability distribution? Also arguably nothing. Yes, "generative" is a piece of jargon used to mean roughly "draw a sample from". But if we imbue "generative" with a sense of open endedness, a quality we think human language and creativity, biological evolution and ecosystems, and political and social systems have, among others,then a probability distribution cannot be generative. It encapsulates what Leonard Savage called a "small world", and even he acknowledged there's such a thing as a "large world" and that it's inappropriate to apply these small world methods and concepts to the large world.
To me, words like "generate", "new", and "process" refer to the large world. There might be small world analogs, but those will always be missing something important.
