// This is really well done. The narrator breaks the conceit around 4:30
in the video to explain what he's doing, but the first four minutes are surprisingly coherent and basically indistinguishable from any other well produced popsci youtube video. It sounds more like a dialect of English from a non-native speaker than something utterly without comprehension.
Of course, the effect works because the machine was trained on the specific style of the artist, and he's selling it well. But the important point is that it works. AI isn't quite a fluent speaker of natural language, but it can talk and we can understand what it's saying. This is remarkable in ways that are hard to overstate.
The really interesting thing is that the dialect is systematic, across all media; the script sounds in language the way deepdream looks in images. These networks tend to wear their training on their sleeves, so you can see that DeepDream has a penchant for dogs, or that this script was trained on Big Idea scripts, without too much difficulty. The complexity of the behavior feels rudimentary in the same systematic way, and once you have a feel for how these networks function it's not too hard to spot the behavior at work.
I'm tempted to recognize all the results of this latest generation of ANNs, from the first DeepDream images to the AlphaGo wins, as history's first genuine AI culture. It has a unique and systematic and recognizable style, coherent across genre and form. The style is generated and curated and supported largely by the human culture it operates in, but we're so taken with it partly because it is so new and interesting and unlike the cultures we've built before.
I'm not saying AI is yet equivalent to human culture, at least not yet. Right now, AI is still mostly just the latest corporate toy (in fairness, we're not so different in this regard either). But future AI will trace its ancestry right through to the cultures we're living through today. https://www.youtube.com/watch?v=3p10knivMRg
via Max Adler