Jeff Klune

And so we end up getting a lot.

What we don't do is publish all of the experiments that didn't work, and we don't document these negative results.

And we also end up kind of with things like overly hyped results.

Maybe somebody does a couple experiments and they seem exciting and they can think of a couple more experiments they should do to really make sure that's a real result.

And they might just say, oh, I'm too busy to do those.

But in the back of their head, they're thinking, and that could potentially disprove this result, but I don't want to take the risk because right now I have something that looks quite sexy.

So AI scientists do not have to get tenure.

They don't have to get grants.

They're not motivated by fame.

and impact.

And so they can be more honest.

They can be more rigorous.

They could spend a lot of time replicating the work of other labs to make sure that it's real, something that human scientists almost never have incentives to do.

And so we could create a world in which the science we get back is much more reliable and trustworthy, at least in those dimensions.

There is a world, however, in which we could do it completely wrong.

We could only publish the most

high impact, sexy AI scientific results.

We could incentivize the AI scientists to go out and get the most high impact results because we want our system to out-compete the systems of other people and show how good it is.

And then we would just kind of re-inject all of the problems we have in the current community, including the things like the replication crisis, if we don't do it right.

So there's an opportunity to get way better here, but it would be kind of comical if we ended up creating all of the same problems all over again.