[object Object]May 27

I’m glad people are finally noticing LLM translators will just make plausible sentences the fuck up when they’re fed anything but a perfect source to translate, which makes them exhausting and damaging for language learning and a variety of other situations where you’re expected to actually, you know, use the fucking language for anything but a highly inaccurate skim

thank fuck all the language learning companies didn’t jump onto the LLM train, right?

…right?

Show thread[object Object]May 27

this post brought to you by the language learning app I’m using semi-consistently telling me that names, formatting artifacts, and words in languages other than the one being learned, all have fabricated definitions related to the sentence they appear next to

this is a very hard error to catch if you aren’t paying attention and it’s actively slowing me down

Show thread[object Object]May 27

“LLMs might be horrible at almost everything but at least they’re good at translating and accessibility”

are they though? or are they fucking terrible and you didn’t notice because the output looked good and you didn’t check it?

Show threadDaniël Franke May 27
@zzt LLMs: good at everything that you are not equipped to verify.
Show threadkabel42May 27
@ainmosni @zzt ai is great at things that are very cheap to validate, because it learns to get the person that gets 3¢ per validation to click the green button
Show threadDaniël Franke May 27
@kabel42 @zzt sadly, while the validation is cheap, the generation isn't, but that's not their problem.
Show threadkabel42May 27
@ainmosni @zzt Is it? I can't validate a translation is good that fast, I cant even decide which squares are bicyles and not mopeds
Show threadDaniël Franke 
@kabel42 @zzt I meant in cases where validation is cheap, like the insane tightloop in claude code that validates if claude generated valid json with an open jsonschema and then lets claude try again if it doesn't validate.
May 27 at 3:17pmWeb