| > Like many, I was viscerally shocked that the results were possible. Why? There were news about bots writing news ~5 years ago. Given a few simple facts the AI generated the regular info-scarce but fluffy news-piece. Now OpenAI added better everything (better language models, more data, better "long-term memory" for overall text coherence), and we got better fluff. It seems like a GAN and a simple Markov chain generator. (Even if it's not that simple of course.) And maybe it's the equivalent of the "modern art meme" style transferred to AI/ML research. ( https://i.pinimg.com/236x/71/e1/21/71e12151f4b59d8433d32c126... ) What I'm trying to convey is that wrecking the net with auto-trolls was already possible, but for some reason Mechanical Turk was cheaper. > OpenAI warned everyone of an “exploit” in which text humans can trust to be human-generated Sokal already did that, and so did http://thatsmathematics.com/mathgen/ ... but of course this might be qualitatively different, because it can be targeted. (Weaponized, if you will.) But the defense/antidote is the same, but it takes a lot more than 6 months to make people better at critical thinking, but maybe you already heard about the difficulties of that :) |
