Ada Palmer

you know, gears for watches.

And so he asks, why didn't Italy have the Industrial Revolution?

And

I wonder if you stand by the answer you just gave.

But I would have also thought that the competitiveness between different Italian cities

city-states would have made it so that, like, hey, if they get better textile machines and whatever before you, it's kind of a disaster because they're right there.

This, by the way, was another interesting thing from the book, which was the first printed books, there was like not, as you just mentioned, there's not this market of commodity things that are produced cheaply that like the average person is going to be like, oh, if I can get this for $10.99, I'll go buy it.

And so they're trying to make this thing look like it was produced by artisanal luxury.

Existing benchmarks tend to seriously over-specify problems.

Take something like Web Arena.

Tasks often spell out every single step.

For example, I will arrive at the Pittsburgh airport soon.

Provide the nearest Hyatt hotel in the vicinity and then give me the minimal driving distance to the supermarket.

But real users don't talk like that.

They'll say, I'm landing in Pittsburgh.

Where do I get groceries?

Where should I stay?

If we want agents to generalize, they can't rely on spoon-fed instructions.

They have to handle underspecified prompts and infer missing constraints from the environment.

LabelBox built the data set specifically for this.