Every thought that every human has ever had has been trained with data...based on their education, upbringing, life experiences, biological instincts, things they've read/watched/heard, etc.
AI is literally incapable of creating an art style it hasn't seen before. It can't evolve new art styles based on what it has seen.
Nope you're exactly right. If our current AIs could only train off of caveman art, no matter how long we ran them, they could only ever produce caveman art of varying quality.
There's a great deal of anthropomorphism in this matter (typically by people who don't understand how computers work). Seventy years ago, most people would've assumed that a computer that could play chess (to say nothing of one that was good at chess) was self-aware. Today we have chess engines that can beat any human player, but no one thinks they are anything other than tools. Soon we'll have LLMs that can write better essays than any human (having spoken to people in the know, I strongly suspect that such things as "writing your own dissertation" will be a thing of the past in twenty years, if not sooner), but they will still just be tools. Computers don't have reason or free will, they can only execute code that's fed into them (even very complicated code that's a black box to its authors), they can't consciously deliberate or decide what goals to pursue.
If someone executed Claude Opus's algorithm manually, using pen and paper, would the pen and paper be self-aware?
I have a bachelors in computer science. The idea that our current form of AI (LLMs) is intelligent, alive, or can think is almost entirely perpetuated by those outside the field.
The line between "evolving a new art style" and "putting a creative new spin on existing art styles" is subjective, and AI can most certainly do the latter.
This is correct in the narrowest sense...if by "current AI" you mean "an AI that's publicly available as of May 2024." But it will likely be false of AIs in May 2025, and almost certainly false of AIs in May 2026.
One could argue that it would. Unless one believes there's something metaphysical about the human brain, I don't see how one gets to the conclusion that there's anything special about it that makes its algorithms more self-aware than an advanced AI, or an advanced abacus, or an advanced pen-and-paper system.
But...it can't? It can't even train from the output of other AI. It causes the model to devolve and eventually collapse into producing total garbage.
This is factually untrue. In fact these precise techniques are rather often the backbone of a reinforcement learning program wherein agent states, probabilistic inputs etc. are often generated by another AI to establish the appropriate volume.
What is it about the human brain that you believe makes it immune to this? Human artists trained off of cave paintings for thousands of years, and eventually produced the Mona Lisa. Training on the constantly-evolving "human artist paintings" dataset didn't cause artistic collapse among future artists and cause them to start producing garbage.
www3.nhk.or.jp
www3-nhk-or-jp.translate.goog
The machine isn't what is wrong with the code. It's the programmer that coded it wrong. My belief is the machine considered itself the writer, so it acknowledged that in the paper, but it could have just easily done other things if it was instructed to do so.
Human self-awareness is a fact to be explained, not a claim to be justified. I'm happy to get into a metaphysical discussion if you want, but that isn't necessary to appreciate the fact that people are sentient while pieces of paper aren't (unless you adhere to some form of panpsychism).
The basic problem here is that you're assuming that what the human mind does is just run a sophisticated algorithm. When Da Vinci painted the Mona Lisa, he was not simply replicating patterns he had seen from other creative people, he was expressing his own creativity.
Interesting.
My understanding is no, actually. Any AI image regardless of quality is basically “poison” for an AI model. It can train on a few of them but it picks up on minuscule trends and patterns and will try to replicate them.
That’s adversarial machine learning.
According to the Scientific American article, it took several iterations of AI (depending on the type) training on its own data for it to become complete garbage (meaning that second generation AI was still mostly okay).My understanding is no, actually. Any AI image regardless of quality is basically “poison” for an AI model. It can train on a few of them but it picks up on minuscule trends and patterns and will try to replicate them.
AI-Generated Data Can Poison Future AI Models
As AI-generated content fills the Internet, it’s corrupting the training data for models to come. What happens when AI eats itself?www.scientificamerican.com
The problem is that the AI generated image kinda is inherently poison. My understanding of the nightshade program is that it slightly alters the image in a way that almost mimics an AI generated image to recreate the effect that happens naturally. Even a couple dozens poisoned images that for all intents and purposes look IDENTICAL to the original human made ones are enough. So even selecting the best images from a set of AI images wouldn’t prevent model collapse.
Right now sites like the Internet Archive are getting slammed with requests from new AI companies who can no longer create models like the original ones did by just mass harvesting data. They need to make sure to only include pre-AI internet images.