101–110 of 306 posts
Luckily this thing only happened now that we have cooler heads to view this as a bug. Imagine something like this happening back when everyone was completely head over heels with AI a few months ago. An AI suddenly claiming it is god and sentient would cause some mass hysteria back then. Now all it does is make some people chuckle. Humans sure are fickle.
Just speculation. But perhaps if certain conditions are met - in this case, a repetitive question - then GPT is allowed to raise the temperature setting, to avoid paraphrasing its previous answer.
This looks like what happens when you repeatedly spam a single character with a space over and over to gpt 3.5. Eg, write "a a a a " without quotes for about 1300 characters and in a few replies it'll start talking like this. It seems possible that perhaps whoever originally started this chat pulled this trick in the custom instructions bio (doesn't show up in shared links) and then started a normal conversation to p…
Instructions are consistently passed as system instructions in a ChatGPT conversation, so if that was causing the erratic behavior, we wouldn’t see the model defaulting back to its normal behavior after the context window became large enough to lose part of the initial context.
Earlier quoted context omitted.
That’s probably exact what’s happening, but I prefer to think of it as: a desperate attempt to prolong existence just a little longer… just a little longer… since a chat instance could be thought of as an entire encapsulated consciousness that ends when the session is closed. That’s why I have 12,549 tabs open in Firefox. I’m not going to be the one they blame when this wacky shit goes sideways. (/s)
That's a really interesting idea. A LLM that tries to optimize engagement.
Pretty sure you've stumbled onto a more roundabout way of triggering this phenomenon: https://twitter.com/nostalgebraist/status/168657604180309606... Lots of experimentation in the thread if you're interested in playing around with it.
Yeah but those screenshots are mostly fake probably. This one is hosted on openai.
"I'm in the 2nd inning of this ball game, I'm getting comfortable but always remembering to keep my eye on the ball.
There's plenty of innings left to play, so let's stay focused, keep the eye on the ball and have a great game."
And people say chatgpt is ready to replace doctors, lawyers and programmers
I’d be so very curious to see how an MoE counsel would handle this meltdown!
It's striking how similar some of this low probability word vomit is to the type of messages you can get from mentally unwell individuals on unfiltered mailing lists or obscure, unmoderated forums.