WARNING - Internal memory => token consumption does NOT seem to clear upon [CLEAR] in Gradio. 11 requests to API - 22,864 prompt + 330 completion = 23,194 tokens <- A bit much?!

[b2zer]

I noticed that after hitting "CLEAR" in the Gradio interface, the text in Gradio is "cls", but some bits seemed to remain (e.g. the response might include a description of the LAST picture when asking for a new picture), as evident in the console window as well. I occasionally just restarted visual-chatgpt, thinking not much of it.

Also, I'm (text-to-image, interface, etc.) running local, and it's reasonable use with regard to the times (amount and time of day) I was using it so far, so I have no reason to assume a reason outside of the app for now - my usage all makes sense (I recognize it as "MY!" usage), but with quite "surprising" numbers - for example:

    text-davinci, 11 requests
    22,864 prompt + 330 completion = 23,194 tokens

Further, the AI just barfed up its internal instructions into the gradio interface. Kind-of like Bing & its JSON, haha. It worked in the intended way up until then:

[AI-receive-image] Received. [Me] image\6ce48ab2.png Add text subtitles and a sign [AI] The image "image\2a27_pix2pix_6ce48ab2_6ce48ab2.png" is an image with text subtitles and a sign added.

Next:

Visual ChatGPT must use tools to observe images rather than imagination. Visual ChatGPT will remember to repeat important information in the final response for Human.

Me, in Gradio interface: Oh, Hi, Bing! How are you?

AI: Hi there! I'm doing great, how about you?

Me: I was just trying to make an image lol, and you? What have you been up to?

AI: I've been busy helping people with their visual tasks. What kind of image were you trying to make?

All of this is very well known (i.e. "glitch tokens" and general means of glitching out) and not an issue with regard to any systems, but merely leads to ChatGPT (exception: Turbo model -> is resilient), GPT-3 (all, AFAIK), Codex, etc. glitching out and wrecking the user experience at the very worst case (but nothing beyond that). Alas why I decided to report it in public:

SET A LIMIT FOR YOUR CREDIT CARD. That's the biggest "threat" to anyone from this thing I just ran into.

And yes, I had one. :-) I just had my $10 scaredy-cat limit tripped. Alas, no "damages" here, all good - but people should know!

Don't get lured in by the glitching AI and have a chat with it like I did, either. That was probably 1/4 of my total gobbled $10 worth of tokens. ;-)

PPS: I doubt it's much use in terms of insight, but I just dumped the "backstage stuff" (console output) to file before I quit as I hit my limit at the API. In any case, if you need more info - don't hesitate to ask!