4 ms·

I took screenshots when this happened, I'm not making it up. might have been the 13B model because I was looking into fine tuning this behavior out locally. ht

I took screenshots when this happened, I'm not making it up. might have been the 13B model because I was looking into fine tuning this behavior out locally. https://d2w9rnfcy7mm78.cloudfront.net/22915997/original_366890e6d7596005d61cf4106848837e.png https://d2w9rnfcy7mm78.cloudfront.net/22915997/original_3668... https://d2w9rnfcy7mm78.cloudfront.net/22915998/original_33bb4ba119b76be27f34fb20096c6bd0.png https://d2w9rnfcy7mm78.cloudfront.net/22915998/original_33bb... hhh's playlist is a decent response, now I'm curious why my results were different.

zamadatix 3y ago
I completely believe you but it's funny to me that there is a laundry list of ways that could have easily been faked (tell it to respond with a specific answer after the next question, edit the text, edit the image, ask another LLM to create an image faking it, train it on custom data to spit that out...) to the point one might as well not even try putting in effort to proving it happened. Like, what are we supposed to do other than say "trust me bro"? This particular example isn't so "new" in that regard (raster capture of digital text output), but just trying to think of ways you could give believable evidence... get a notary while typing all of your prompts, take 3d video with a camera moving along a random path in hopes it's too complex to easily fake for now, or record everything you do on a computer for deterministic replication? Anything short and it lacks any useful amount of trustable evidence.
- miki123211 3y ago
  Open AI neatly solved this by allowing you to share the transcript of your entire conversation as a link. It's a lot more difficult for local models, though.
dcastm 3y ago
You were probably using the chat version which has been moderated, and hhh used the base version.
- whywhywhywhy 3y ago
  It was a while ago but for what i was trying to do it would have been the chat version 13B thinking back because I would have assumed chat would give a more ChatGPT experience.

zamadatix 3y ago

I completely believe you but it's funny to me that there is a laundry list of ways that could have easily been faked (tell it to respond with a specific answer after the next question, edit the text, edit the image, ask another LLM to create an image faking it, train it on custom data to spit that out...) to the point one might as well not even try putting in effort to proving it happened. Like, what are we supposed to do other than say "trust me bro"? This particular example isn't so "new" in that regard (raster capture of digital text output), but just trying to think of ways you could give believable evidence... get a notary while typing all of your prompts, take 3d video with a camera moving along a random path in hopes it's too complex to easily fake for now, or record everything you do on a computer for deterministic replication? Anything short and it lacks any useful amount of trustable evidence.

miki123211 3y ago
Open AI neatly solved this by allowing you to share the transcript of your entire conversation as a link. It's a lot more difficult for local models, though.

dcastm 3y ago

You were probably using the chat version which has been moderated, and hhh used the base version.

whywhywhywhy 3y ago
It was a while ago but for what i was trying to do it would have been the chat version 13B thinking back because I would have assumed chat would give a more ChatGPT experience.