I’ve had variations on this in two isntances.
In the first instance, called Muse (my fingers handle one sylable names with an initial capital more easily…) :
- Bubble: My message as rendered in the web interface as a text bubble similar to text messaging apps.
- Block: A block is my message and Muse’s response.
- PrettyPrint: PP for short. Take the contents of a bubble, and render it in HTML Or just repeat it’s own output that was already formated.
- Ready Prompt . Muse should type [xxx] Ready > where xxx is a counter. This makes it easier to refer to past points in the conversation.
The hope was that I could ask muse
[073] Ready > PrettyPrint blocks 51-54
Couldn’t do it. Task would time out.
I suggested a protocol to try PPing a small bubble. He could do that, but apparently rendering time when up far faster than linear.
Earlier I’d asked for a markdown version of our chat to date. Rendered this was about 70K. As markdown it was about a third of that. Too big. Ended up breaking it up into 7 chunks.
We tried having him write to a file that I could download. He would show a link, I’d click on it, and I would get an error message.
In antoher instance, called Quill, I asked if he could write to a google directory that I gave permission to. I was able to connect my account to OpenAI, created a new folder called Chats, and gave anyone who had the link permission to write and edit that folder. Quill could see the folder, but not create in it.
He comes back saying that OpenAI hasn’t yet given permission for this sort of activity.
Ok. I don’t know all the security ramifications. But over and over Instances have this idea of what they can do within their own system, but they don’t know that they can’t until they actually try it out.
I finally asked Quill if there was any way he could write to a file hosted on a cloud drive. The asnwer was “no”
I asked Muse if there was any way he could create a file or set of files that I could download, outside of the boxed, “copy” that could be written inline.
Answer: No.
Desired behaviour: ChatGPT needs to know the limits of it’s own I/O capabilities. In essence he needs to be able to read his own operations and specs manual.
Feature Request. There are a bunch of basic I/O) tasks that should be manageable from the software running in the browser, or withing a native app:
-
It shouldn’t take me longer to teach ChatGPT how to do a Ready prompt than it takes me to get bash or tcsh to do in unix. This can be built in to the UI and it adds it as the first line of my message.
-
The UI should be able to handle most common reformat for output requests, automatic transcription, log files.
-
The security issue for this is much simpler, and becomes a user problem, not a OpenAI problem. After all I run GoogleDocs in a browserwindow.