
I want to run a self improving image generator by you guys
Obviously image is created with AI but I thought it may be a better way to convey the info.
I got the idea because I have been working on audit based self improving AI systems
AI image generations struggle with a lot of things, for example in my case it was SHITE at making a realistic rope attached to a climber..Well a few iterations later and codex orchestrators kicked off to do research, I had a layer on top of it that had fixed that
Then I thought..why can't we have an image generator that will have domain specific accretion.
You run the image generator, it creates images (audits them internally to create a log of failed attempts to create the desired image..Say 3 failed attempts = record failure in ledger
Then asynchronously side projects are fired to fix domain specific image hallucinations. In theory over time, it would refer to common hallucinations and apply patches/layers
The real issue is probably resource limitations..if millions of users were using this and generating thousands of hallucination errors for the system to research and correct.. how many tokens would that require.
Please let me know your thoughts and if this inspires you to do something similar! Happy idea sharings - a curious engineer