Post

shambibble
@shambibble.com

it's true, specifically the special kind of stupid that comes with not having a backspace key. (if anything, i find these sort of errors endearing. the LLM knows immediately that it fucked up and is trying to save it the only way it knows how)

Luminiferous Ether/Or(wellian)@cira.bsky.social

ChatGPT is such a special kind of stupid

1 repost4 likes

shambibble
@shambibble.com

i bet bronze age scribes whose errors could fuck up several dozen man-hours' worth of quarrying would feel even more kinship with autoregressive models than i do

Luminiferous Ether/Or(wellian)
@cira.bsky.social

Is that what’s going on? I’m curious what the “tells” are for this… because it reaffirmed the date at the end? I’d be more inclined to find it endearing if less had access and took what it said as gospel. I can get it to tell me complete falsehoods about 25% of the time with a single question

shambibble
@shambibble.com

can't say for certain without looking at the logprobs which the closedAI dweebs don't give you anymore, but it matches a pattern i see in open models a lot. my guess is chernobyl burbled up in the low % chance (the same way you might momentarily think of and discard it)

shambibble
@shambibble.com

but because LLMs don't do reanalysis, or "system 2" intelligence, or even have the ability to erase their own context, once it blurts the word "chernobyl" there is nothing left for to do but invent some new context, however silly, that is consistent with its knowledge of the actual year