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
for producing this kind of recursion.
The the reason the language it contains with it this with sequence of language that they learn
are it's it's built to have this recursive cup c uh capability w within it.
That this word is going to produce the next word, which is going to produce the next word premised on the entire sequence before it.
And that's that's the that's the crazy thing.
There was also this re an interesting result that Anthropic put out a paper a little while ago.
I think it's called the biology of the biology of large language models.
The next token, even though you're only producing the very next token from the sequence, but the language models have learned that f because they've learned sequences to next token, they've learned that they're within any point along the sequence that that p that point in the sequence is pregnant with
the potentiality for uh uh uh the the not just the next token, but many other tokens going uh uh uh moving forward.
It's the whole trajectory that are uh that uh is is sort of encapsulated.
in that matrix that you're talking about earlier, right?
The matrix is just a matrix for taking a sequence, produce the next token.
But no, no.
The matrix is customized so that it's gonna run recursively.
Right.
And and so it's built in such a way, it's it's tuned in such a way that it's not it's gonna produce the next word the.
Well that's a s that's a that's not useful.
No, the is the next piece of the gr the auto regressive chain that's gonna produce the man when to the store, right?
And so
It's it's not just any old matrix.
Showing 921–940 of 2,007 · page 47 of 101
← Previous
Next →