AI singer-songwriter ‘Anna Indiana’ debuted her first single ‘Betrayed by this Town’ on X, formerly Twitter—and listeners were not too impressed.
Experience is ongoing learning through the subjective self. When you experience the color red you do not just record it with your photoreceptors, and your experience of the color red is different from mine because we don’t just record wavelengths of light. We don’t just continue to learn from continual exposure to new data, we also continue to learn from generating our own data. In this way our subjective experience is qualitative, not simply quantitative. I don’t just see the specific light wavelengths, I experience the “redness” of red.
When LLM is trained on that kind of data it just starts to hallucinate. This is promising! I think the hallucination phenomenon is actually a precursor to creativity and gives us great insights into the nature of subjective experience. In a sense, my phenomenal experience of the color red is actually much like a hallucination where I am also able to experience the color’s “warmth” and “boldness”. Subjectivity.
it’s only qualitative because we don’t understand it
when an LLM “experiences” new data via training, that’s subjective too: it works its way through the network in a manner that’s different depending on what came before it… if different training data came before it, the network would look differently and the data would change the network as a whole in a different way
When an LLM feeds on its own outputs, though, it quickly starts to hallucinate. I think this is actually closer to creativity, but it betrays the fundamental flaw behind the technology - it does not think about its own thoughts and requires a curator to help it create.
I’ll believe something is an AI when it can be its own curator and not drive itself insane.
that’s a lack of understanding of concepts though, rather than a lack of creativity… curation requires that you understand the concept that you’re trying to curate: this looks more like a dog than this; this is a more attractive sunset than this
current LLMs and ML don’t understand concepts, which is their main issue
id argue that it kind of does “think about its own thoughts” to some degree: modern ML is layered, and each layer of the net feeds into the next… one layer of the net “thinks about” the “thoughts” of the previous layer. now, it doesn’t do this as a whole but neither do we: memories and neural connections are lossy; heck even creating a creative work isn’t going to turn out exactly like you thought it in your head (your muscle memory and skill level will effect the translation from brain to paper/canvas/screen)
but even we hallucinate in the same way. don’t look at a bike, and then try and draw a bike… you’ll get general things like pedals, wheels, seat, handlebars, but it’ll be all connected wrong. this is a common example people use to show how our brains aren’t as precise and we might like to think… drawing a bike requires a lot of very specific things to be in very specific places and that’s not how our brain remembers the concept of “bike”