Pozorvlak

RE: https://nomenloony.com/@nomenloony/116019280517956996

I was surprised by this, so I tried it myself on the free version of ChatGPT. The textual instructions it gave me were AFAICT completely correct and focused on safety, but when I asked it for a diagram it produced something almost as bad as this. I've seen it produce diagrams that don't match the text before.

Transcript here: https://chatgpt.com/share/6985005b-d5dc-8007-af30-711d912feab7

Feb 5 at 8:45pmWeb
Show threadDmitry TantsurFeb 6
@pozorvlak in my experience, common LLMs are terrible at diagram generation. I've tried with a much simpler flow diagram and only got some success when I told it to generate *code* to generate a diagram.
Show threadBugsFeb 6

@pozorvlak It matches my experience that chatgpt's image generation is bad at diagrams. I've played with juggling patterns and hobby circuit design: clearly explained in text, nonsense diagrams.

Uninformed guess: text is produced linearly, every token chosen in context of what's already written, creating coherent ideas. Vs images generated all at once, so lots of plausible-looking initial ideas all presented together with no checking whether they make sense in context?

Show threadBugsFeb 6
@pozorvlak ...for the limited meaning of "checking" and "makes sense" that can be applied to how these models work, I guess. I'm very ignorant here.