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
They learn an embedding.
In order to do next token, they learn embedding that gives this this where the space, uh, you know, and we could think of almost like two dimensional, three dimensional space.
Oh, it's it's very high dimensional, but you know, for for uh for our purpose, we think about it that that where uh cat and dog live, you want those things to live closer together than uh cat and desk.
Um and of course it's much richer than that, right?
It's not just semantic like this
this very kind of superficial level of semantic uh similarity.
In fact, what it is is capture the somehow the semantics, so to speak.
are captured by the space.
Like the space, the shape of the space itself is what allows the model to understand sort of the the relation between words so that I can do the next next token generation.
But it's a very, very different space, right?
It's has to do with really with with relations between words and in terms of generation.
In terms of uh you know next token generation so that it's useful for that purpose.
What does the perceptual space look like?
Like, well, this perceptual space is gonna have a very different uh ki th the the the the axes there um almost certainly aren't gonna have the same kind of meaning as in the linguistic space.
There'll be something closer, maybe color features, shape features, something like that.
And where this microphone lives is within that space is gonna have radically different meaning.
Uh than saying you know it's you know
It's not apples and oranges, right?
That's those aren't different enough, right?
It's apples and math or something.
Showing 421–440 of 2,007 · page 22 of 101
← Previous
Next →