My conversation with ChatGPT:
I could swear someone here was looking forward to using this as a grammar checker. “2 R's”?
Right. The LLM tokenises before actually processing a request. Tokens are numbers, and there are no R’s in those numbers, so the LLM guesses the amount of R’s.
When you ask this with spaces in btween the word, the LLM will answer correctly.
Hope this helps
But… how do you put spaces between a single word?
Then I guess the tool is fundamentally flawed.
Ask various LLMs a typical question a writer would like to know:
- Character 1 is the daughter of A, the granddaughter of B and the great-granddaughter of X.
- Character 2 is the daughter of C, the granddaughter of D, the great-granddaughter of E and the great-great-granddaughter of X.
- X is the descendent of both Character 1 and 2.
- What is the family relationship between Characters 1 and 2.
You’ll get a variety of “logical” responses between LLM’s.
Helped me big time with a teenager now labelling her distant relative as Twice Removed in her contact list, aside from the extent they go to prove their own argument.
So yeah, AI does help. And it’s not like I can’t write those responses in my own words. It gives rise to unexpected comedy among all the other tensions at play.
Well, AI - in this case ChatGPT - still has weaknesses and does not solve a task in every case the way a human would, but the LLMs are capable of learning, and if they specialize in text editing (DeepL or LT), then they are also helpful tools for text control.
That’s not deep learning it’s Sub-A learning, one letter at a time.