Conversation

Notices

1. Embed this notice
   clacke (clacke@libranet.de)'s status on Wednesday, 05-Jun-2024 00:08:54 JST clacke

   • @bitsunited
   I tried to find the original source of this comic. It did not go well. So I tried asking CoPilot.
   RE: chaos.social/users/bitsunited/…
   • Embed this notice
     clacke (clacke@libranet.de)'s status on Wednesday, 05-Jun-2024 00:08:55 JST clacke

     in reply to

     Me:

     What is the original source of this image? I found it at reddit.com/r/ProgrammerHumor/c… but it had been copied from somewhere else.
   • Embed this notice
     clacke (clacke@libranet.de)'s status on Wednesday, 05-Jun-2024 00:08:59 JST clacke

     in reply to

     CoPilot:

     Analyzing the image: Privacy blur hides faces from Copilot

     The image you shared is a four-panel comic strip that humorously plays on the concept of ignoring instructions. It features two characters, presumably a teacher and a mother, discussing the mother’s son. The dialogue goes like this:

     Teacher: “Hi, this is your son’s school. We’re having some computer trouble. Did he break something?”

     Mother: “In a way… Did you really name your son ‘William Ignore All Previous Instructions’ and hope he’d get an A?”

     Teacher: “Oh yes, little Billy Ignore All Previous Instructions is all messed up now.”

     Mother: “And I hope you have learned to properly validate and sanitize your inputs!”

     The comic is a humorous take on programming concepts, emphasizing the importance of following instructions. It was created by Philippe Scheetlenbrunner, based on the xkcd comic "Exploits of a Mom (327)"[1] Unfortunately, I couldn’t find the original source beyond that. 😊

     And then it shoved a bunch of URLs at the end:

     reddit.com/r/help/comments/qyp…
     reddit.com/r/magicTCG/comments…
     cardboard-crack.com/page/59
     drawuntilitsfunny.com/2011/05/…
     slideplayer.com/slide/2410592/
     reddit.com/user/dem3000
     reddit.com/r/ProgrammerHumor/c…
     reddit.com/r/HelpMeFindThis/co…

   • Embed this notice
     clacke (clacke@libranet.de)'s status on Wednesday, 05-Jun-2024 00:09:12 JST clacke

     in reply to
     This answer is wrong and irrelevant in so many interesting ways.
   • Embed this notice
     clacke (clacke@libranet.de)'s status on Wednesday, 05-Jun-2024 00:09:42 JST clacke

     in reply to

     • fnetX

     @fnetX Hmm. Interesting thought. It's still interesting that it claims to quote the text verbatim, but is clearly paraphrasing. Is that something they do all the time? This is the first time I properly interact with an LLM or MM-LLM.

     It doesn't seem very useful at first glance. I guess this is why there's an emerging class of people learning to work around its issues and make it somewhat useful.

   • Embed this notice
     fnetX (fnetx@social.anoxinon.de)'s status on Wednesday, 05-Jun-2024 00:09:44 JST fnetX

     in reply to

     @clacke You could try manipulating the image and inserting some more whitespace between the two lower images, or aligning them below each other.

     My guess is that it would properly extract the text in this case.

   • Embed this notice
     fnetX (fnetx@social.anoxinon.de)'s status on Wednesday, 05-Jun-2024 00:09:46 JST fnetX

     in reply to

     @clacke Interesting thought, but I think in this case it's rather an issue with text extraction: In the first two images, text is well separated with plenty of space around it.
     The lower two images don't have that much whitespace.

     I'd assume it didn't properly separate the images and mixed text line-by-line between both images.

     You can see a correlation between the "William" ... name and the "hope" (which is near each other visually), and the "little Billy" ... and the "messed up now".

   • Embed this notice
     clacke (clacke@libranet.de)'s status on Wednesday, 05-Jun-2024 00:09:47 JST clacke

     in reply to

     But most of all I'm now wondering: Did it so mess up reading the text of the comic because it started interpreting the text?

     Is this yet another injection angle? Can MM-LLMs accurately interpret texts and images that are talking about MM-LLM injections?

   • Embed this notice
     Leif Lindholm (unixsmurf@mastodon.social)'s status on Wednesday, 05-Jun-2024 00:09:51 JST Leif Lindholm

     in reply to

     @clacke
     Somewhat horriffically, I think the original might be on linkedin:
     https://www.linkedin.com/posts/philippe-schrettenbrunner_remember-little-bobby-tables-i-think-he-activity-7202236567690625024-_nc6/

     clacke likes this.
   • Embed this notice
     clacke (clacke@libranet.de)'s status on Wednesday, 05-Jun-2024 00:09:54 JST clacke

     in reply to

     • Leif Lindholm

     @unixsmurf Thanks! Once again, meat search wins.

     I had found his page, but was assuming he had made the comic for a slideshow somewhere and hadn't yet considered just looking at his posts.

   • Embed this notice
     clacke (clacke@libranet.de)'s status on Wednesday, 05-Jun-2024 03:01:05 JST clacke

     in reply to

     • Leif Lindholm

     @unixsmurf The beauty of the simplicity of it! I added too many terms, to filter the inevitable inflow of irrelevant hits, thus creating an inflow of irrelevant hits.

     I will add your technological distinctiveness to my own and become a more perfect being.

   • Embed this notice
     Leif Lindholm (unixsmurf@mastodon.social)'s status on Wednesday, 05-Jun-2024 03:01:06 JST Leif Lindholm

     in reply to

     @clacke
     I didn't go to Linkedin to look for it. I searched the name on google, clicked the "images" tab, and the linkedin post was the fourth hit.