My word choice is more driven by exposure to the dataset that I'm working with. There's been some recent successes with autoencoders trained on virtual sensory input (i.e., video games) with surprising results, e.g., neural networks that can simulate the dynamics of these environments with surprising fidelity.

Of course, learning to play video games at a high level is trivially easy, as everyone in the field now knows. The next challenge is, naturally, to make money doing this. But how? After the traditional thirty seconds of research before undertaking a major project, I determined that the only way to make money from video games is to become a popular streamer.

So now I am training an agent to generate video of it playing and reacting to an imaginary game and equally fictitious Twitch viewers, with a dataset drawn from the top Fortnite streamers. The reward function is comprised of a blend of subscribers, donations, and (logarithmically scaled) misogyny in the chat. Thus far, I've only managed to create some sort of window into hell, where the "game" consists of unceasing violence, murder after murder after murder as towers of mismatched material swell and fall in ever transforming locations on the isle while the chat endlessly subscribes, spams, and emotes in cackling glee and the superimposed webcam video features a... thing with too many eyes and hands screaming incoherently.

At first I thought it was a problem with my dataset, so I started watching some of the streams myself. This has not yielded insight into the whole "nightmare vision" output of my model, but it has expanded my vocabulary on the twin subjects of combustibles and comestibles, which I feel is a reasonable trade-off for the sanity battering associated with this whole endeavour.