Noam Shazeer

You know, we're...

training a little bit on video data, but we're definitely not anywhere close to thinking about training on all the visual inputs you could get.

So you have visual data that we haven't really begun to train on.

And then I think we could extract a lot more information from every bit of data we do see.

You know, I think one of the ways people are so sample efficient is they explore the world and take actions in the world and observe what happens.

Yeah.

Right?

Like you see it with very small infants, like picking things up and dropping them.

They learn about gravity from that.

And that's a much harder thing to learn when you're not initiating the action.

And I think having a model that can take actions as part of its learning process would be just a lot better than just sort of passively observing a giant data set.

Is God toward the future then?

Something where the model can observe and take actions and observe the corresponding results seems pretty useful.

Yeah, I mean I think there's always this tradeoff of should we publish exactly what we're doing right away?

Should we put it in the next stages of research and then –

roll it out into like production gemini models uh and not publish it at all or is there some intermediate point and for example in our computational photography work in pixel cameras you know we've often uh taken the decision to develop interesting new techniques like the ability to do um

super good night sight vision for low light situations or whatever, put that into the product and then published a real research paper about the system that does that after the product is released.

And I think

you know, different techniques and developments have different treatments, right?

Like, so some things we think are super critical, we might not publish.