Comment on Meta’s new AI image generator was trained on 1.1 billion Instagram and Facebook photos
Mahlzeit@feddit.de 11 months agoCan we get back to this? I am confused why you believe that AIs like ChatGPT spit out “exact copies”. That they spit out memorized training data is unusual in normal operation. Is there some misunderstanding here?
GiveMemes@jlai.lu 11 months ago
I don’t think we’re really talking to each other, but more past each other so I took a break.
To answer the question, it was an analogy and the ransomware part was to show the non-intelligence and creationary lack of AI more than be applied to the programming analogy. Sorry if that was confusing.
It was an ars technica (iirc) article I read in which the author made a working ransomware with GPT-4 by having it initially create a program to encrypt a file, then had it encrypt directories instead, then added flags and debugged it all of which he claims can be done by pretty much anyone malicious with access. Nowhere along the way did chat-gpt realize what it was doing though. A human would have.
Also ime at least I got completely copy and pasted paragphs from gpt 3.5 a few times dunno how much 4 has improved upon that. I think my disagreement with you about AI copyright infringement is that you think that AI can create new things whereas I don’t think that. I don’t think that because it can only ever rehash its training data. Our current AI systems can’t actually create new thoughts though. For example with your how to write a book author analogy, those people haven’t just read people’s advice and are now putting it on paper. Those people have also read tons and tons of novels. Taken classes on English and created and defended original ideas as part of that. If you trained an AI on English classes and novels it would have no idea how to write a “how to write a novel” type book while a person would. You have to have it copy something in order for it to perform, it’s just the way that it works.
Furthermore it really wouldn’t take a huge change to copyright law, just clear differences between the rules that apply to sentient vs non-sentient sources.
Mahlzeit@feddit.de 11 months ago
This touches several difficult topics.
I don’t think that matters to copyright law, as it exists.
Copyright law is all about substantial similarity in copyrightable elements. All portraits are similar by virtue of being portraits. Portraits are not copyrighted, nor can one copyright genres and such. A translation of a text has superficially no similarity with the original, but has to be authorized.
What you are saying would mean, that similarity is no longer a requirement for an infringement. That’s a big change. It is copyright, after all.
Non-sentient sources are not new. Take cameras, for example. Cameras have been improved over time so that less skill is necessary to operate one. It’s no longer necessary to manually focus, to set the exposure time, to develop the film, … This also means that photos today have less human creative input. In current smartphone cameras, neural AIs make many decisions and also “photoshop” the result.
It doesn’t really make sense to me to treat modern cameras differently to old ones. Or: Someone poses and renders a figure in Blender. What difference does it make if they use an old-fashioned physical based render or a genAI?
Nevertheless, the question whether AIs can create something new, can be answered. The formal definition of “information” is that it is a reduction in uncertainty. For example, take the sequence of letters: “creativit_”. You probably have a very clear idea what the last, missing letter is. So learning that it is “y” doesn’t give you much information.
But take the sequence: “juubfpvoi_”. The missing letter could be any lower-case letter. You may not feel very informed when you learn that it is “f”, but it does represent a much bigger reduction in uncertainty.
When we write texts, we use the same old words in the dictionary; just a few 10,000 at most. We string them together with the same old rules of grammar to tell the same old things. The sky is blue, things fall down, not up; people love and hate, and in the end the good guys win. You can probably think of exceptions to all these. They are exceptions. We create small variations on the same old themes. We rehash.
If a story does not cater to expectations, then it’s not believable. People should behave as we know people to behave. The laws of nature should be consistent and familiar. Most of all: The conventions of the genre should be followed. As a human, you are supposed to lift ideas from previous works. New ideas may be appreciated, but are not required.
The second string was, in fact, created by a machine; not an AI, but an RNG. Even with many GBs of output, it should be impossible to find any biases or patterns that allow one to guess at the next letter. I didn’t make one up myself because humans are not very random even when we try. And when we write, we do our best to reduce our randomness even further. We try not to invent new spellings; ie make spelling errors.
AIs receive input from a pRNG, which means that they create new things. What they are supposed to do is to strip away all that novel information and create something largely predictable. They often fail and, say, create images of humans with an innovative number of fingers. LLMs make continuity errors, or straight start to spout gibberish. The problem is that AIs create too many new things, not that they don’t.