A new approach to tuning generative AI models
feat. hallucination as nudge
The Pierre that I have summoned here and now has, as his very nature, the fact that he is not here and now - in other words, his unreality.
This is what we have called 'nothingness,' the material proper to the imaginative consciousness. Earlier, we characterized the imaginative consciousness, as distinct from the perceptual consciousness, as nothingness. Imagination passes through a 'double nihilation.' What happens when I imagine Pierre? First, I grasp the world as a situation in which Pierre is absent. I look around and check whether Pierre is here. In my visual field as I search for Pierre, the other things are given as mere background, with their outlines blurred and faded. I become certain of the fact that 'Pierre is not here right now.' At the level of perception, the unreality of Pierre stands out clearly. The imaginative consciousness begins by positing precisely this perceptual situation in which Pierre is absent. My consciousness rejects the situation in which Pierre is absent and, going against the real, tries to make Pierre present here and now.
'Pierre is sitting in the chair across from this table right now ······'
This is the process of double nihilation. In short, to imagine Pierre is to posit a reality in which Pierre is absent, and at the same time to summon the non-existent Pierre into the here and now.
From among the points where the first generation begins,
find the point where the hallucination begins.
An LLM that distinguishes perception from image
and memory from imagination
will come to possess a momentary consciousness.
