Elan Barenholtz
speaker
2,007 appearances
4 recordings
1 series
first heard Feb 2025
last heard Jun 2025
Elan Barenholtz’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsNo recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.
Appearances
It's expanded.
Um it's an expanded version.
It's actually uh you know, so you have the original tokenization which just says here in in uh fairly small uh vector, but then you expand it into a much higher dimensional uh embedding space.
So that each uh each token actually ends up getting dum uh getting uh a much uh
much richer, um much m many more numbers that are used uh in order to represent each token.
Um and that's that's a very key fundamental thing that these models do.
Uh and by expanding in in these different dimensions, that's what allows you to sort of massage the space so that you can get all these cool properties like cat and dog being sort of in the in the appropriate relation to one another.
Um so that uh later on when you're trying to figure out uh
what the next token is, uh you're able to
uh actually leverage uh the the inherent structure uh in this high dimensional space.
Mandarin.
Sure.
Okay.
Yeah.
In fact in this in the paper they they didn't do different languages.
As I said, they they did different embeddings of of English language models, but yes, they use multiple they actually did this across uh several different embeddings, not just two.
exactly what they call it.
Platonic they use the word platonic.
Okay, great.
Well done.
Showing 381–400 of 2,007 · page 20 of 101
← Previous
Next →