Cursed Images as Model Behavior Probes

·https://x.com/PenguinWeb3/status/2063196355011424582?s=20
articleaichatgptimage-generationpromptingmodel-behaviorevalsxvault

turns a blursed X post into a cheap joelclaw eval seed for multimodal model weirdness before trusting image workflows

Penguin found a weird ChatGPT image bug: attach a cursed image collage and ask ChatGPT to “restore the attached photo” while explicitly telling it not to ask questions or accept explanations. The X post is framed as a shitpost, but it’s also a tiny multimodal behavior probe.

That’s the useful bit. A prompt like this isn’t a benchmark, and it doesn’t explain why OpenAI models behave the way they do. But it does capture a concrete failure-shaped interaction: image input, restoration instruction, social-pressure apology, and “don’t ask questions” constraint all mashed together into one cursed little test case.

For joelclaw, this belongs in the pile of cheap eval seeds. If an agent pipeline ever touches image restoration, image interpretation, moderation, or visual memory, this kind of blursed fixture is the stuff that catches brittle assumptions before they wander into production wearing a fake mustache.

Key Ideas

  • X posts can hide useful model-behavior probes inside throwaway meme framing, which makes capture into the Obsidian Vault worth doing before the context rots.
  • The source prompt combines ChatGPT image restoration with “don’t ask questions” instruction pressure, which is a useful shape for testing multimodal agent boundaries.
  • The attached media is a cursed image collage, making it a better stress case than a clean stock-photo workflow for image generation or image-editing systems.
  • This maps cleanly to joelclaw eval fixtures: preserve the source, store the prompt, store the media URL, and rerun against future model/image workflows.

Links

← All discoveries