Trenton Bricken

And I think it's because we just haven't been able to make sense of it.

What is V2?

It's like the next part of the visual processing stream.

And yeah, so I think it's very likely.

And fundamentally, superposition seems to emerge when you have high dimensional data that is sparse.

And to the extent that you think the real world is that, which I would argue it is, we should expect the brain to also be underparameterized in trying to build a model of the world and also use superposition.

It's a combinatorial code.

Yeah, exactly.

Well, actually, that's a great segue because all of this feels like GoFi.

Like you're using distributed representations.

But you have features, and you're applying these operations to the features.

I mean, the whole field of vector symbolic architectures, which is this computational neuroscience thing, all you do is you put vectors in superposition, which is literally a summation of two high-dimensional vectors.

um and you create some interference but but if it's higher dimensional enough then you can you can represent them uh and you have variable binding where you connect one by another and like if you're dealing with binary vectors it's just the x or operation so you have a b you bind them together and then if you query with a or b again you get out the other one

And this is basically the like key value pairs from attention.

And with these two operations, you have a Turing complete system, which you can if you have enough nested hierarchy, you can represent any data structure you want, etc, etc.

we try and get it to do as much interpretability work and other safety work as possible.

I mean, we have our responsible scaling policy, which has been really exciting to see other labs adopt.

I mean, I think we need to make a lot more interest.

If it's as capable as GPT-7 implies here, I think we need to make a lot more interpretability progress to be able to comfortably give the green light to deploy it.

I would be like, definitely not.