Multimodal Is Great Until You Need It to Be Exact

I love that one chat can now take a screenshot, a PDF, and a question and answer all at once. Then I asked it to read a single serial number off an invoice and it gave me one digit wrong, and I almost sent a payment to the wrong place.

Multimodal models are brilliant at the gist and sloppy at the detail. They'll summarize a document perfectly and transpose a number without noticing. For what does this say they're magic. For copy this exactly they're a liability.

The fix is to split the job. Let the model do the understanding, then make the human do the one verification that matters. I now read the critical digits myself and let the model handle everything around them.

This isn't a complaint. It's a boundary. The tool is a reader, not a clerk, and treating it like a clerk is how money moves to the wrong account.

Use multimodal for the sense, not the specifics. The gist is free. The exact line is still on you, and pretending otherwise is how small errors become real ones.