Skip to main content


Showing posts from December, 2018

Neural Networks of Carcosa

Something I did for fun over the weekend: Found a workbook for easily creating a text-generating neural network: Took a plaintext dump of the hex contents from LotFP's Carcosa . Massaged the text to make it a little better for neural network processing (make everything lowercase, remove punctuation, remove numbers, etc.) Trained the neural network against the text (ran 50 epochs worth of character-level training) Took the resulting model and generated some sample text using the workbook's default settings of "1 very unexpected token, 1 unexpected token, 2 expected tokens, repeat" to ensure the model doesn't try to exactly replicate the original text. ;tldr I tried to create an AI that can auto-generated Carcosa-like hex descriptions. The results are... kind of weird. The neural network was being trained letter-by-letter rather than word-by-word, so there's no guarantee it won't invent its own