Illia Polosukhin

It's a very peer-to-peer network, as you can see.

And again, this is what blockchain is, is a peer-to-peer network of, you know, kind of facilitating value transfer and kind of trust in a permissionless way.

And so that's really what we're going after.

Yeah, I think... So important component is if we know what training data went in, if we know kind of how it was trained, there's a lot more tools to do explainability, why it made this decision.

Right?

To give you an example that I kind of... I had this experience at Google where...

we were training a model for question answering.

You ask it, you know, we had like a test set where Barack Obama was born and its reply was Kenya.

And you're like, huh?

And so you go in, like you look at training data and it's because there's a bunch of websites that talk about Barack Obama was born in Kenya.

Right.

And so, you know, being an engineer, you're like, hey, you know, I need to improve my model.

So you remove the websites.

Yeah.

And, you know, naturally the model becomes more left leaning because you just removed a bunch of right leaning websites.

We talked about that.

And so all of those decisions, they're not malicious in any way, but you don't know what's happening.

Unless you see the training data, you don't know what the outcome model, what biases it has and how it arrived to this.

Exactly, yeah.

And so it's really important to have that visibility and kind of traceability for this, because that will give you indeed better explainability on why those decisions are made.