484
this post was submitted on 17 Aug 2023
484 points (96.0% liked)
Technology
59086 readers
3617 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
But it's not just converting them into a different format. It's not even storing that information at all. It can't actually reproduce anything from the dataset unless it is really small or completely overfitted, neither of which apply to GPT with how massive it is.
Each neuron, which represents a word or a phrase, is a set of weights. One source makes a neuron go up by 0.000001% and then another source makes it go down by 0.000001%. And then you repeat that millions and millions of times. The model has absolutely zero knowledge of any specific source in its training data, it only knows how often different words and phrases occur next to each other. Or for images it only knows that certain pixels are weighted to be certain colors. Etc.
This is a misunderstanding on your part. While some neurons are trained this way, word2vec and doc2vec are not these mechanisms. The llms are extensions of these models and while there are certainly some aspects of what you are describing, there is a transcription into vector formats.
This is the power of vectorization of language (among other things). The one to one mapping between vectors and words / sentences to documents and so forth allows models to describe the distance between words or phrases using euclidian geometry.
I was trying to make it as simple as possible. The format is irrelevant. The model is still storing nothing but weights at the end of the day. Storing the relationships between words and sentences is not the same thing as storing works in a different format which is what your original comment implied.
I'm sorry you failed to grasp how it works in this context.
You made me really interested in this concept so I asked GPT-4 what the furthest word away from the word “vectorization” would be.
Interesting game! If we're aiming for a word that's conceptually, contextually, and semantically distant from "vectorization," I'd pick "marshmallow." While "vectorization" pertains to complex computational processes and mathematics, "marshmallow" is a soft, sweet confectionery. They're quite far apart in terms of their typical contexts and meanings.
It honestly never ceases to surprise me. I’m gonna play around with some more. I do really like the idea that it’s essentially a word calculator.
Try asking it how the vectorization of king and queen are related.