Hacker News new | past | comments | ask | show | jobs | submit login

"buns glazing"

Lol typo or intentional?




I'm guessing intentional spoonerism.

https://en.wikipedia.org/wiki/Spoonerism


I love how autocorrect mistakes sometimes just come out sounding like rhyming slang… a less obtuse version than most real dialects of rhyming slang, more understandable by most people.


There was an AI thread a few days ago about how Open AI could not come up with a catchy slogan for selling hot cross buns in the UK. Could be a reference to the thread, or the actual slogan used in the UK.


It’s an interesting shortcoming of the large language models that they struggle to “play” with the rules of the language very much. I’ve tried to get GPT 3, 3.5, and 4 to generate business names and other names for things and it’s not particularly good at it. It frequently “runs out of creativity, and starts to repeat the same combination of basing building blocks. The worst example I came up with had moderately specific business names with a particular theme like an “1890s” or “1970s” or “late 1990s computer startup” style of name.. it would frequently run out and start repeating the same building blocks in under 50 examples, the absolute worst started repeating itself at about item 20 and the list gets increasingly repetitive from that point.

I built a modified name generator based on some existing open source code that’s been around for several years now based on GPT2 and sure I couldn’t ask to go give me something with a specific theme… but damn it was way more creative at generating new words that could be names and it was t hard to filter the output to be useful since I could get it locally in batches of thousands overnight.


On the contrary, I find that GPT - especially GPT-4 - is very good at playing with the rules of the language: https://gist.github.com/int19h/6fa34a86923cd681396393b21b9ab...


You’ve been much more proscriptive, and given it a larger task than I was trying to highlight, and I can see GPT4 doing well at the kind of task I’ve personally experienced it doing well at.

What I was trying to highlight is the sort of “small task” not, very short essay and feel free to make up specific words using specific linguistics as some guidance.

I’ve seen it make up words in this way too, even without prompting it to allow it to make up a word. I was more trying to highlight that its not particularly good at prompts like “please give me a list of 50 names for a hypersonic aircraft company for my science fiction novel”, counterintuitively it requires significant additional prompt text to expand its creativity in response to prompts like this, and not repeat itself if you ask it for a list of another 50 names.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: