Trenton Bricken

Machine learning is empirical.

We need to do this.

I think it's going to be pretty important for certain aspects of scaling dictionary learning.

Interesting.

So I think images are also in some ways just easier to interpret than text.

Yeah, exactly.

And so Chris Ola's interpretability work on AlexNet and these other models, like in the original AlexNet paper, they actually split the model into two GPUs just because they couldn't, like GPUs were so bad back then.

relatively speaking, right?

Like, still great at the time.

That was one of the big innovations of the paper.

But they find branch specialization, and there's a Distilled Pub article on this where, like, colors go to one GPU and, like, Gabor filters and, like, line detectors go to the other.

And then, like, all of the other... Yeah, yeah, yeah.

And then like all of the other interpretability work that was done, like the floppy ear detector, right?

Like that just was a neuron in the model that you can make sense of.

You didn't need to disentangle superposition, right?

So just different dataset, different modality.

And given Dworkesh's success with the Vesuvius Challenge, we should be pitching more projects because they will be solved if we talk about them on the podcast.

Yeah, so Bruno Olshausen, who I think of as the leading expert on this, thinks that all the brain regions you don't hear about are doing a ton of computation and superposition.

So everyone talks about V1 as having Gabor filters and detecting lines of various sorts.

And no one talks about V2.