I was working with GPT the other day on a new (not programming) problem for the first time, and I asked it to download a Google Drive spreadsheet and take a look at it. Which I was quite surprised to get a flat response that it couldn't do that, since just the other day, I had been working through a programming problem with GPT 3.5 (and I've never paid for it) when there was a point in the dialogue that I was asking it to generate a parser, and I suggested it would be really handy if there was some way I could provide it with example data for the parser to be tested against.

Wouldn't you know, it said this:

> I'm sorry, but as a text-based AI model, I cannot directly receive or process files or attachments. However, I can guide you on how to provide a sample PDF and assist you further with the script.

> If you'd like to share an example PDF, you can upload it to a file-sharing service (such as Dropbox, Google Drive, or any other file-hosting platform) and provide a link to the file. I can then guide you on how to modify the script based on the specific structure and content of the PDF.

Can it download the PDF, or is it faking? I've told it that it is a competent Ruby programmer as a first prompt, and this is the only reason that I can think of why I'd get such different results. Is it drawing a line between "reading the structure" and "processing a set of data files" or is it really just faking me out? The next prompt after I shared a link:

> After reviewing the PDF, I'll provide an updated version of the script that should work well with the given example. Please allow me some time to analyze the PDF and make the necessary adjustments.

... OK, should I just wait here then? I thought about it for a minute, and then I provided it with the header row for the output CSV, which was the only thing I really needed it to get from the PDF file. We moved on after that, and I do not know if it actually ever read the PDF, or if it was pulling my leg supportively, "yeah I can't really download a file, but sure send me the file, I am a great Ruby programmer as you said before, so surely I will know what to do with it friend-o!"

Is it really just the presence or absence of the reassuring "you are a programmer and you know what you're doing" prompt at the beginning of my conversation with it, that made the bot behave differently in both circumstances? I was told that it would be important to do things like this. What's really happening here, I wonder? I would love to understand a bit better!

We did get to the end of the problem, and no the parser did not work reliably (but it worked much better than I was honestly expecting, also I did not try with GPT4 so maybe that would have been the difference...)