Linkposting, tagging and excerpting in accord with 'Should pretty much all content that's EA-relevant and/or created by EAs be (link)posted to the Forum?'.
Will's distillation of Alexander's ~3,500 words into ~500:
A very careless plagiarist takes someone else’s work and copies it verbatim: “The mitochondria is the powerhouse of the cell”. A more careful plagiarist takes the work and changes a few words around: “The mitochondria is the energy dynamo of the cell”. A plagiarist who is more careful still changes the entire sentence structure: “In cells, mitochondria are the energy dynamos”. The most careful plagiarists change everything except the underlying concept, which they grasp at so deep a level that they can put it in whatever words they want – at which point it is no longer called plagiarism.
“But don’t humans also have genuinely original ideas?” Come on, read a fantasy book. It’s either a Tolkien clone, or it’s A Song Of Ice And Fire. [...] Dullards blend Tolkien into a slurry and shape it into another Tolkien-clone. Tolkien-level artistic geniuses blend human experience, history, and the artistic corpus into a slurry and form it into an entirely new genre. [...]
“But don’t scientists have geniunely original ideas?” Scientists are just finding patterns in reality nobody has ever seen before. You say “just a pattern-matcher”, I say “fine, but you need to recognize patterns in order to copy them, so it’s necessarily a pattern-recognizer too”. And Einstein was just a very good pattern-recognizer.
“But don’t humans have some kind of deep understanding that pattern-recognition AIs don’t?”
The point is, GPT-2 has faculties. It has specific skills, that require a certain precision of thought, like counting from one to five, or mapping a word to its acronym, or writing poetry. These faculties are untaught; they arise naturally from its pattern-recognition and word-prediction ability. All these deep understanding things that humans have, like Reason and so on, those are faculties. AIs don’t have them yet. But they can learn.
Now imagine you prompted it [a pattern-recognition AI] with “P != NP”. This time give it near-infinite training data and computational resources. Its near-infinite training data will contain many proofs; using its near-infinite computational resources it will come up with a model that is very very good at predicting the next step in any proof you give it. The simplest model that can do this is probably the one isomorphic to the structure of mathematics itself (or to the brains of the sorts of mathematicians who write proofs, which themselves contain a model of mathematics). Then you give it the prompt P != NP and it uses the model to “predict” what the next step in the proof will be until it has a proof, the same way GPT-2 predicts the next word in the LotR fanfiction until it has a fanfiction.
The version that proves P != NP will still just be a brute-force pattern-matcher blending things it’s seen and regurgitating them in a different pattern. The proof won’t reveal that the AI’s not doing that; it will just reveal that once you reach a rarefied enough level of that kind of thing, that’s what intelligence is. I’m not trying to play up GPT-2 or say it’s doing anything more than anyone else thinks it’s doing. I’m trying to play down humans. We’re not that great. GPT-2-like processes are closer to the sorts of things we do than we would like to think.