Re: Stunt
Anyhow, feeding it ridiculous amounts of examples (check, they train on insanely large datasets) and trying to improve training algorythms [sic] to extract more information per chunck [sic] of data fed (check, they do that too) is going to create additional information beyond the correlation between mere facts.
Information theory states the exact opposite; you cannot create additional "information" in a data set by extrapolation, all you can do is add distortion.
Those generated papers showing structure and logic show structure and logic because there is common structure and logic in the training data; that information (in the formal sense of the word) already exists in its inputs. The fact that, by statistical chance, the output appears to be novel, does not make it so, and the very real, and obvious risk here is that it outputs something that looks superficially reasonable, but is 100% false. This may appear identical to the real output of a human author writing a paper, and without detailed fact-checking and analysis it may be indistinguishable, but it will not be based on actual facts and reasoning, it will merely hold the appearance of something this is. If, by chance, its output is actually factually true, it isn't because it has formulated a new logical argument, it is entirely through random chance, or worse still, it is just regurgitating its training data verbatim. Even a stopped clock is right twice a day. What this sort of thing really does, is waste people's time, because now nothing can be taken on trust any more (whether it should have been is another matter), and peer review now means checking that sources exist, and that arguments based on them are logically sound as well as checking the methodology and results of the paper in question. All we have done here is automate academic fraud, which I think most people would agree is not a step forward in the goal of advancing human knowledge.
