Arseny Khakhalin

Typical in academia and Twitter: writing a nice abstract or article and then realizing you're out of characters. I wondered if gptchat could help shorten texts by removing unnecessary words. Here's what I asked it to do with my oversized tweet: fit it in 280 characters.

(the text above is already a shortened version; see the attached pic for the original, which was a whole 2 sentences too long! Seems like a quite useful tool, potentially!)

Dec 4, 2022 at 9:25pmWeb
Show threadArseny KhakhalinDec 4, 2022

That said, #ChatGPT doesn't seem to be doing a particularly good job with #scientific #texts, and also it doesn't seem to be able to #count the characters in its own output. So typically if you ask to #shorten the text, it first overshoots by a bit, then undershoots by a bit.

Here I'm trying it on my own old abstract as an example (from 1340 chars it first went to 1200, then to 700, while I asked for 1000)

Show threadOliver SteinbockDec 4, 2022
@ampanmdagaba Noticed this too. Commands like "double the length" sometimes fail completely.
Show threadArseny KhakhalinDec 4, 2022
@SteinbockGroup It totally makes sense, given how the system is built, but seeing how it emulates Linux and whatnot, I still somehow thought that may be it can guesstimate it :) No, apparently, not to well.
Show threadJohannesDec 4, 2022
@ampanmdagaba
I spend a lot of time re-writing my grant proposals to make them easy-to-parse and concise because my innate writing style is rather baroque, so a text-generating language model might speed that up. When I prompted Galactica to write a review on a topic I just assigned a student, the answer had factual errors (or actually fabrications), but stylistically it mimicked what would conventionally be considered good writing in biomedical literature.
Show threadArseny KhakhalinDec 4, 2022
@johannes_lehmann I played with the chatbot a bit more, and it does seem to mimic the style, but it's not really helpful. (Which is strange, as I read on Twitter some people claiming that for really academic topics it actually works better, not worse). It doesn't seem to catch the nuanced importance of some words, compared to others, which ruins details. And as it cannot count its own output, it does not strategically prioritize cuts. So, still a long way to go!