Raw LLM Responses

I think part of the difference between the human inspiration and the AIs training comes down to this rather nebulous concept of style. A human can develop a style of painting for example. And that style may draw on other artists for inspiration, it may be down mechanical reasons like the joints in your hand, visual processing but a style for humans is a way of painting, a particular semi-structured algorithmic approach to applying paint to canvas. An AI isn't applying paint to a canvas, it's taken text in input, producing coloured pixels as output, everything in between exists as a numerical vector. It is not learning a style to produce a painting, rather it skips over that step in production, it never builds up from a base or learns to apply oil to canvas. Instead it relates features of a finished painting to text. Keyword being features. It learns what arrangement of pixels relates to what piece of text that's broadly it. So features could include something like large groups of similarly coloured pixels and these groups could relate to the concepts of van Gogh's thick brushwork. But equally features could also include someone's signature or a particular motif they paint with. Or even whole subjects from someone else's painting or work. Features are just groups of pixels in an image. Features can include stylistic elements but they can also include elements we'd consider plagiarism if a human just transposed them into a new picture. Because ultimately it's not learning to paint, or draw or write. It's learning to produce sets of pixels, it's learning to produce an image of a painting, it's learning how text relates to sets of pixels.