LLMs have three major components: a massive database of “relatedness” (how closely related the meaning of turns are), a transformer (figuring out which of the previous words have the most contextual meaning), and statistical modeling (the likelihood of the next word, like what your cell phone does.)
LLMs don’t have any capability to understand spelling, unless it’s something it’s been specifically trained on, like “color” vs “colour” which is discussed in many training texts.
"Fruits ending in ‘um’ " or "Australian towns beginning with ‘T’ " aren’t talked about in the training data enough to build a strong enough relatedness database for, so it’s incapable of answering those sorts of questions.
It can’t see what tokens it puts out, you would need additional passes on the output for it to get it right. It’s computationally expensive, so I’m pretty sure that didn’t happen here.
I’m not an expert, but it has something to do with full words vs partial words. It also can’t play wordle because it doesn’t have a proper concept of individual letters in that way, its trained to only handle full words
they don’t even handle full words, it’s just arbitrary groups of characters (including space and other stuff like apostrophe afaik) that is represented to the software as indexes on a list, it literally has no clue what language even is, it’s a glorified calculator that happens to work on words.
It’s crazy how bad d AI gets of you make it list names ending with a certain pattern. I wonder why that is.
LLMs aren’t really capable of understanding spelling. They’re token prediction machines.
LLMs have three major components: a massive database of “relatedness” (how closely related the meaning of turns are), a transformer (figuring out which of the previous words have the most contextual meaning), and statistical modeling (the likelihood of the next word, like what your cell phone does.)
LLMs don’t have any capability to understand spelling, unless it’s something it’s been specifically trained on, like “color” vs “colour” which is discussed in many training texts.
"Fruits ending in ‘um’ " or "Australian towns beginning with ‘T’ " aren’t talked about in the training data enough to build a strong enough relatedness database for, so it’s incapable of answering those sorts of questions.
It can’t see what tokens it puts out, you would need additional passes on the output for it to get it right. It’s computationally expensive, so I’m pretty sure that didn’t happen here.
doesn’t it work literally by passing in everything it said to determine what the next word is?
I’m not an expert, but it has something to do with full words vs partial words. It also can’t play wordle because it doesn’t have a proper concept of individual letters in that way, its trained to only handle full words
That’s interesting, didn’t know
they don’t even handle full words, it’s just arbitrary groups of characters (including space and other stuff like apostrophe afaik) that is represented to the software as indexes on a list, it literally has no clue what language even is, it’s a glorified calculator that happens to work on words.
I mean, isn’t any program essentially a glorified calculator?