Ask HN: What questions do LMMs fail with?

Character-level manipulation fails. I suspect this is because GPTs are trained on tokens rather than sequences of characters, means they're fundamentally dyslexic and can't really spell.

https://imgur.com/if92ZIQ

https://imgur.com/sYXFPRu

Nice. Thank you for this insight.

LLM can't do linguistic nuances like irony, humor and sarcasm. It also have difficulty with ambiguous statements and multiple negation. It may also have problem with technical jargon (though it work quite well with medical terms for some reason).

It keeps inventing features and command-line flags. A dozen times it presented me with code (Postgresql/PostGIS related) where it turns out a method cannot be called with those parameters.

https://blog.opencagedata.com/post/dont-believe-chatgpt "TL;DR ChatGPT claims we offer an API to turn a mobile phone number into the location of the phone. We do not."

encoding="utf-8" for creating files. it always gets this wrong. I have to write it and when I asked it to make another change it will forget it.

New ideas