Hermes Trismegistus driving a schoolbus

Last week I saw someone describe GitHub Copilot—the code-completion tool which has been trained on all of the public source code on GitHub—as a form of license-laundering, and the same thing occurs to me about DALL-E.

Clearly, it's been trained on an enormous …

GLOSSATORY is a bot which posts spurious definitions from an RNN trained on a dictionary, and every morning I use its output as a prompt for little drawings, some of which I post at GLOSSATORY. Today I tried feeding some recent definitions to DALL-E to see what it came up …

"Graffiti art of Wittgenstein riding a skateboard"

Happy Friday.

VQGAN+CLIP is a method for generating images from text phrases using a GAN image generator and the CLIP image recogition system: I tried feeding it the first few numbered paragraphs from Wittgenstein's Tractatus Logico-Philosophicus, which I was thinking about because I recently watched Derek Jarman's excellent film about the …

This is the output of the neuralgae algorithm, which visualises a neural net's idea of some image classes and then classifies the result and repeats, looped into a GIF, and it's exactly the sort of fluid morphing between surreal forms I've been trying to get for two years.

Here are all 1000 classes in the ImageNet challenge, rendered with the Visual Geometry Group's 19-layer neural net, using the same deepdraw technique I used for the thousand faces of CaffeNet. These ones are rendered on random grayscale noise, not my face.

This is the net which the neuralgae Twitter …

I was tinkering around trying to get Justin Johnson's neural-style implementation working on the HPC - this is an algorithm which uses a deep learning image recognition net to extract low-level style features from one image and use them to render the large-scale contours of another. This technique has been commercialised …

@voronidols are on Twitter, and also on Mastodon @voronidols@botsin.space and the bots page.

Source code

I've updated the algorithm behind neuralgae. It's still the same basic principle: render a deepdraw image based on some random categories, feed the results back through the classifier, and then repeat with the new categories. But the draw part of the cycle now uses the weights from the classification stage …

I've made my first Twitter bot. It posts a new image and accompanying text from my Neuralgae NaNoGenMo project, six times a day, modified slightly so that it fits within the character limit. One of my goals with Neuralgae was to generate something which was in some sense readable, and …