Comment by snickerbockers
1 day ago
Generally novel either refers to something that is new, or a certain type of literature. If the AI is generating something functionally equivalent to a program in its training set (in this case, dozens or even hundreds of such programs) then it by definition cannot be novel.
This is quite a narrow view of how the generation works. AI can extrapolate from the training set and explore new directions. It's not just cutting pieces and gluing together.
Calling it “exploring” is anthropomorphising. The machine has weights that yield meaningful programs given specification-like language. It’s a useful phenomenon but it may be nothing like what we do.
Or it may be remarkably similar to what we do
In practice, I find the ability for this new wave of AI to extrapolate very limited.
Do you have any concrete examples you'd care to share? While this new wave of AI doesn't have unlimited powers of extrapolation, the post we're commenting on is asserting that this latest AI from Google was able to extrapolate solutions to two of AI's oldest problems, which would seem to contradict an assertion of "very limited".
Positively not. It is pure interpolation and not extrapolation. The training set is vast and supports an even vaster set of possible traversal paths; but they are all interpolative.
Same with diffusion and everything else. It is not extrapolation that you can transfer the style of Van Gogh onto a photographl it is interpolation.
Extrapolation might be something like inventing a style: how did Van Gogh do that?
And, sure, the thing can invent a new style---as a mashup of existing styles. Give me a Picasso-like take on Van Gogh and apply it to this image ...
Maybe the original thing there is the idea of doing that; but that came from me! The execution of it is just interpolation.
This is how people do things as well imo. LLM does the same thing on some level but it is just not good enough for majority of use cases
This is knock against you at all, but in a naive attempt to spare someone else some time: remember that based on this definition it is impossible for an LLM to do novel things and more importantly, you're not going to change how this person defines a concept as integral to one's being as novelty.
I personally think this is a bit tautological of a definition, but if you hold it, then yes LLMs are not capable of anything novel.
8 replies →
uhhh can it? I've certainly not seen any evidence of an AI generating something not based on its training set. It's certainly smart enough to shuffle code around and make superficial changes, and that's pretty impressive in its own way but not particularly useful unless your only goal is to just launder somebody else's code to get around a licensing problem (and even then it's questionable if that's a derived work or not).
Honest question: if AI is actually capable of exploring new directions why does it have to train on what is effectively the sum total of all human knowledge? Shouldn't it be able to take in some basic concepts (language parsing, logic, etc) and bootstrap its way into new discoveries (not necessarily completely new but independently derived) from there? Nobody learns the way an LLM does.
ChatGPT, to the extent that it is comparable to human cognition, is undoubtedly the most well-read person in all of history. When I want to learn something I look it up online or in the public library but I don't have to read the entire library to understand a concept.
You have to realize AI is trained the same way one would train an auto-completer.
Theres no cognition. It’s not taught language, grammar, etc. none of that!
It’s only seen a huge amount of text that allows it to recognize answers to questions. Unfortunately, it appears to work so people see it as the equivalent to sci-fi movie AI.
It’s really just a search engine.
9 replies →
>I've certainly not seen any evidence of an AI generating something not based on its training set.
There is plenty of evidence for this. You have to be blind not to realize this. Just ask the AI to generate something not in it's training set.
1 reply →
Isn't that what's going on with synthetic data? The LLM is trained, then is used to generate data that gets put into the training set, and then gets further trained on that generated data?
You didn’t have to read the whole library because your brain has been absorbing knowledge from multiple inputs your entire life. AI systems are trying to temporally compress a lifetime into the time of training. Then, given that these systems have effectively a single input method of streams of bits, they need immense amounts of it to be knowledgeable at all.
OK, but by that definition, how many human software developers ever develop something "novel"? Of course, the "functionally equivalent" term is doing a lot of heavy lifting here: How equivalent? How many differences are required to qualify as different? How many similarities are required to qualify as similar? Which one overrules the other? If I write an app that's identical to Excel in every single aspect except that instead of a Microsoft Flight Simulator easter egg, there's a different, unique, fully playable game that can't be summed up with any combination of genre lables, is that 'novel'?
I think the importance is the ability. Not every human have produced (or even can) something novel in their life, but there are humans who have time after time.
Meanwhile, depending on how you rate LLM's capabilities, no matter how many trials you give it, it may not be considered capable of that.
That's a very important distinction.