Daniel Jeffries (Unknown)

Well, I suppose you could argue that they're not bugs, they're features.

So all of these features are actually just shortcuts.

They're not statistically salient.

They're just a form of overfitting.

But there's also this robustification in the extreme of training regime.

So maybe it's possible that it could both be true, that there are features which are bugs, but also at a certain scale, we can just allow the statistical process to use them or not use them.

Very cool.

Can you tell us about your black box attacks paper?

I actually remember, was that 2018?

Yeah.

I was running these paper review calls at Microsoft, and that was one of the papers that we did.

I actually remember now.

Yeah, you're right, because you don't get the logic.

So you have to do it in the kind of sparse domain where you don't have all of the labels.

Wow.

That's really cool, man.

Amazing.

I actually had read that paper.

And you did another paper as well, which was faster black box attacks.

So this was like an optimization, presumably, of that.