It’s kind of wild to criticize people for not providing the conversation prompt logs they had with LLMs and then publish code on GitHub generated by an LLM without publishing their own LLM conversation log.
Is there a slide deck or transcript of this? I don’t watch videos this way.
I’m still reading the machine generated transcript of the video. But to keep it short:
The author was messing with ISBNs (international standard book numbers), and noticed invalid ones fell into three categories.
- Typos and similar.
- Publishers assigning an invalid ISBN to the book, because they didn’t get how ISBNs work.
- References "hallucinated"¹ by ChatGPT, that do not match any actual ISBN.
He then uses this to highlight that Wikipedia is already infested by bullshit from large “language” models², and this creates a bunch of vicious cycles that go against the spirit of Wikipedia of reliability, factuality, etc.
Then, if I got this right, he lays out four hypotheses (“theories”) on why people do this³:
- People who ignore the limitations of those models
- People seeking external help to contribute with Wikipedia
- People using chatbots to circumvent frustrating parts of doing something
- People with an agenda.
Notes (all from my/Lvxferre’s part; none of those is said by the author himself)
- “Hallucination”: misleading label used to refer to output that has been generated the exact same way as the rest of the output, but when interpreted by humans it leads to bullshit.
- I have a rant about calling those models “language” models, but to keep it short: I think “large token models” would be more accurate.
- In my opinion, the author is going the wrong way here. Disregard intentions, focus on effect — don’t assume good faith, don’t assume any faith at all. Instead focus on the user behaviour; if they violate Wikipedia policies once warn them, if they keep doing it remove them as dead weight fighting against the spirit of the project.
