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 months

No recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.

Appearances

newest first · ▶ plays the moment

Theories of Everything with Curt Jaimungal · Elan Barenholtz: Language Is an Autonomous Organism, Not a Tool · 13 Jun 2025 podcast

to what it receives at the at the far end, uh at the f at the sort of the endpo of that qualitative processing.

Uh that is essentially the idea.

Um so this is this is the exact problem I am working on right now.

Um there was a fantastic paper uh that I just uh came across uh about a week ago.

There was a paper uh that uh it was published in Archive recently.

It's called Harnessing the Universal Geometry of Embeddings.

And what this paper showed is that you could have uh completely different models.

solving different linguistic tasks.

Uh for example, you could have GPT, then you could have BERT, which solves a somewhat different task.

So there's masked tokens as opposed to uh auto-regressive uh next token generation.

And what they found was that you could learn what is it what's this latent space.

What you could do is uh hand off, uh take take the embedding.

The embedding is basically you can think of that as numerical representation.

It's it's it's it's a high dimensional numerical representation um of your tokens.

Uh so here's a token.

Uh that's gonna this token's gonna represent the word dog, and then we're gonna we're we're gonna take that token and uh embed it in a much higher

higher dimensional space.

And what they found is uh that if you take uh the embedding, the higher higher dimensional representation from one model, uh see ChatGPT, and then uh take uh representation from a different model.

uh that you could actually get the uh you you could then you could you could take the embedding, send it to this latent space.

Uh if you cycle it through, get the uh big you have to I know it's starting to get in the weeds a little bit.

Showing 221–240 of 2,007 · page 12 of 101