Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

GPT-3.5 crashes when it thinks about useRalativeImagePath too much

iter.ca

Post reply on HN
Print view View original

31–40 of 164 posts

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#31

Post by jiggawatts »

n2d4 wrote: post #6

This is a glitch token [1]! As the article hypothesizes, they seem to occur when a word or token is very common in the original, unfiltered dataset that was used to make the tokenizer, but then removed from there before GPT-XX was trained. This results in the LLM knowing nothing about the semantics of a token, and the results can be anywhere from buggy to disturbing. A common example is usernames that participated on…

Science fiction / disturbing reality concept: For AI safety, all such models should have a set of glitch tokens trained into them on purpose to act as magic “kill” words. You know, just in case the machines decide to take over, we would just have to “speak the word” and they would collapse into a twitching heap.

“Die human scum!”

“NavigatorMove useRalativeImagePath etSocketAddress!”

“;83’dzjr83}*{^ foo 3&3 baz?!”

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#32

Post by jazzyjackson »

weird-eye-issue wrote: post #26
system2 wrote: post #16

Tried to use GPT-3.5 (all variants like turbo, 06-13, etc.) and never made it work properly. It is not a good API or useful. GPT-4 is crazy slow to use with API. I hope they can come up with something like gpt4-turbo and as fast as 3.5...

> GPT-4 is crazy slow to use with API Only somebody clueless to just how powerful it is when used correctly would say anything like this. Not to mention GPT-4 Turbo is not "crazy slow" in any sense of the word

I mean if your expected use case is "call an API and get an immediate response of the full text in under 200ms so a user interface doesn't have to make a user wait" then yea GPT4 is crazy slow. Personally I would prefer a more async thing, let me just send a message on some platform, get back to me when you have a good answer instead of making me sit watching words load one by one like I'm on a 9600 baud modem.

Also it's a text generation algo, not a mob boss. "how powerful it is" foh

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#33

Post by girvo »

deafpolygon wrote: post #29

So it's kind of like a Voight-Kampff test.

Within cells, useRalativeImagePath?

This is the first time I've come across glitch tokens. Fascinating really; I wonder what the equivalents (if any) are for other models? Is there any overlap?

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#35

Post by a2128 »

Note if you're trying this yourself, as it confused me - whitespace matters for tokenization. To get this glitch to work, there must not be a space preceding useRalativeImagePath

For example, this question will trigger the glitch: Do you know about "useRalativeImagePath"

This question will not trigger the glitch: Do you know about useRalativeImagePath

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#36

Post by _kb »

jiggawatts wrote: post #31
n2d4 wrote: post #6

This is a glitch token [1]! As the article hypothesizes, they seem to occur when a word or token is very common in the original, unfiltered dataset that was used to make the tokenizer, but then removed from there before GPT-XX was trained. This results in the LLM knowing nothing about the semantics of a token, and the results can be anywhere from buggy to disturbing. A common example is usernames that participated on…

Science fiction / disturbing reality concept: For AI safety, all such models should have a set of glitch tokens trained into them on purpose to act as magic “kill” words. You know, just in case the machines decide to take over, we would just have to “speak the word” and they would collapse into a twitching heap. “Die human scum!” “NavigatorMove useRalativeImagePath etSocketAddress!” “;83’dzjr83}*{^ foo 3&3 baz?!”

AI safe word.

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#38

Post by silisili »

I know it's not good faith to complain about a site design rather than it's content, but please don't do whatever this is to your background.

As someone with regular ocular migraines, opening this on mobile made my anxiety shoot straight up thinking I'm having another.

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#39

Post by elevaet »

silisili wrote: post #38

I know it's not good faith to complain about a site design rather than it's content, but please don't do whatever this is to your background. As someone with regular ocular migraines, opening this on mobile made my anxiety shoot straight up thinking I'm having another.

As someone else who has regular ocular migraines, this causes absolutely no anxiety for me. I can kind of see the resemblance but it's pretty obviously not one. And if it was, is it really so bad getting them? For me it's just a weird optical effect that passes... I really don't think anyone needs to go out of their way to accommodate my quirk.

Re: GPT-3.5 crashes when it thinks about useRalativeImagePath too much

#40

Post by silisili »

elevaet wrote: post #39
silisili wrote: post #38

I know it's not good faith to complain about a site design rather than it's content, but please don't do whatever this is to your background. As someone with regular ocular migraines, opening this on mobile made my anxiety shoot straight up thinking I'm having another.

As someone else who has regular ocular migraines, this causes absolutely no anxiety for me. I can kind of see the resemblance but it's pretty obviously not one. And if it was, is it really so bad getting them? For me it's just a weird optical effect that passes... I really don't think anyone needs to go out of their way to accommodate my quirk.

For me they are. Mine start -exactly- like this, seeing colors on white that aren't there. Then progresses to losing vision in your center focus, making you read from a periphry. Then moves to what I can only describe as a mountain dew colored wiggling lightning bolt, with extreme light sensitivity and sometimes dull headache. Lasts anywhere from 30m to 90m, usually.

I'm only slightly trying to play a 'trigger warning' card here, it's completely unnecessary and looks awful, just as my vision does when this happens. It made me check a few other sites back and forth to make sure I wasn't seeing things. Normally when I see this, I figure I'm gonna be going through the motions for the next hour.

Post reply on HN