Having an AI convergence day, so I married my tinkering with VQGAN+CLIP with GPT-J with the following experiment. Prompt is in standard font, response is in i

Having an AI convergence day, so I married my tinkering with VQGAN+CLIP with GPT-J with the following experiment. Prompt is in standard font, response is in italics. ---- I started messing around with image generation with VQGAN and CLIP. It is really cool, but I ran out of things to prompt it with, you know just short phrases that describe a scene that could be rendered. I shortly thereafter saw an article about GPT-J and thought it would be cool to use AI to generate the prompts. So I go the demo up and running and asked it to generate a prompt by typing this into the notebook. The computer responded with the following: Three cats chasing a balloon with laser beams shot at them from the ground. I did not type in "three", so it really was just a blue-print of the problem. But this got me thinking. I had seen in different contexts that we humans will generally enter some rules when generating a creative response and act as that request for the request. For example, when I am writing a resume I am usually very mindful that I do not include any errors in the use of personal pronouns. I know that when I write "I will go to the fair", I know that I could come up with the following but not the former. I will go to the fair I then noticed that the prompt actually contains two main requests. Three cats chasing a balloon And this is where things got really interesting. I have noticed that sometimes I will be able to play the two requests against each other. The sun shines on three cats chasing a balloon In this example I was able to generate a response that contained two unrelated requests that were both fulfilled from the prompt. This isn't a requirement, if the AI or its creators want to keep a loop from being possible, the loop will likely be re-rolled if it is present. ---- I then took the phrase 'Three cats chasing a balloon with laser beams shot at them from the ground.' and generated an image with VQGAN+CLIP and it came up with the following: https://imgur.com/KpAdZ3W https://imgur.com/KpAdZ3W