Re:Language Models Are Injective and Hence Invertible
by sigmoid10 on Thursday October 30, 2025 @07:08AM
(#45758658)
It doesn't really matter which vector you are looking at, since they are using a tiny constraint in a high dimensional continuous space. There's gotta be an unfathomable amount of vectors you can fit in there. Certainly more than a few billion.
[ Parent
| on HN ]
Re:Language Models Are Injective and Hence Invertible
by sebastianmestre on Thursday October 30, 2025 @04:38PM
(#45765027)
No, yeah, totally. Even assuming binary vectors 2^768 is a ridiculously huge number. The probability of collision even assuming a bad sampling that discards 75% of dimensions is still vanishingly small.