Sholto Douglas

How is that encoding for like- Superposition.

You're just encoding a ton of features in these high-dimensional vectors.

In a brain, is there, like, an exonifiering or however you think about it, like...

I don't know how you think about how much superposition is there in the human brain.

Oh, okay.

So crucially, then superposition is not an artifact of a neuron.

It is an artifact of like the space that is created.

Okay, cool.

Yeah, thanks.

we kind of talked about this but like I think it's just like kind of wild that it seems to the best of our knowledge the way intelligence works in these models and then presumably also in brains it's just like

There's a stream of information going through that has quote-unquote features that are infinitely, or at least to a large extent, just like splittable.

And you can expand out a tree of like what this feature is.

And what's really happening is a stream, like that feature is getting turned into this other feature or this other feature is added.

I don't know.

It's like, that's not something I would have just like thought.

That's what intelligence is.

You know what I mean?

It's like a surprising thing.

It's not what I would have expected necessarily.

What did you think it was?