Saturday, April 29, 2023

ChatGPT does not get it yet - Learning is not about repeating, but instead about [critical] thinking.


I translated the two books that I have published so far from Spanish to English using #ChatGPT, and here is what I learned:

1. Agreed with https://arxiv.org/pdf/2301.08745.pdf that ChatGPT is a good translator for spoken language.
2. 37 problems in a total of 31,459 words. Not bad at all.
3. 9 of the problems were also present in DeepL. The rest were not.
4. ChatGPT can be instructed to keep the format, and with that convert dialogs in Spanish (which use quotation dashes) to dialogs in English (which use quotes).
5. However, DeepL autosuggestion feature is a productivity boost and with just 9 errors shown I think it is still a more effective way to translate in cases where respecting format is not a must-have.
6. If text is heavily formatted like for example XML variants like HTML, then ChatGPT is unable to keep all escaped XML entities like '&lt;'. Instead, it randomly keeps them or shows them as literals, like in this case '<'. Prompting more gets the engine in repeating mode, demonstrating that there is no creativity and no critical thinking in ChatGPT. This to me is a demonstration that the ultimate question scientists have to answer is *how* these models actually "learn" so that we can predict how they will "behave". Learning is not about repeating, but instead about [critical] thinking. We know why and how an atomic bomb works, but we don't fully understand why and how AI works.

I will report back when I finish my upcoming poetry book, which I will only publish once I have the final material in the two languages because ChatGPT has taught me a lesson: In reviewing multilingual content, an author finds better ways to say the same even in their own native/mother tongue.

No comments:

Followers