this post was submitted on 03 Dec 2023
542 points (95.2% liked)

hmmm

4795 readers
493 users here now

For things that are "hmmm".

Rule 1: All post titles except for meta posts should be just plain "hmmm" and nothing else, no emotes, no capitalisation, no extending it to "hmmmm" etc.

founded 2 years ago
MODERATORS
542
Hmmm (startrek.website)
submitted 1 year ago by [email protected] to c/hmmm
 
you are viewing a single comment's thread
view the rest of the comments
[–] Seasm0ke 52 points 1 year ago* (last edited 1 year ago)

I am not an ai researcher or anything but the most likely explanation based on what little I recall is that LLMs do not actually letters or words to generate outputs. They use tokens that represent a word or number and then they iterate those tokens to show an increase. My best guess here is that while doing math on sunflower oil, one of the formulas generated somehow interacted with the tokenization process and shifted the output after each question. Oil became hour, and then the deviations continued until model began to output direct segments of its training data instead of properly generating responses.

Again this is absolutely speculation on my part. I don't have much of a direct understanding of the tech involved