Agentic Tasks: a Few Samples

What models are capable of and what they aren’t capable of continues to astound me. I’ve been pushing myself to use pi for a number of tasks these few weeks and wanted to share some learnings.

Image manipulation: Replacing Profile Picture
Link to heading

This was shockingly good. Pi one-shot cropping the new image and replacing it. No notes.

Image manipulation: Text Replacement
Link to heading

This was surprisingly a struggle (after how well the image manipulation worked). The model consistently struggled to follow instructions, and frequently went “above and beyond” - frequently ruminating significantly to get a subpart result. I haven’t found a great solution for this yet. It does seem like the thinking causes the model to ruminate, in a way where it talks itself out of doing what I want it to do.