Peter McCrory

And productivity and the labor market.

It's not just capabilities, it's capabilities plus what determines firm adoption and what is the consequence of that firm adoption.

So in that respect, we looked at some of our API traffic.

How are businesses embedding cloud capabilities to automate existing capabilities?

or new workflows.

And one of the things that we find is that for the most complex tasks that businesses bring to Claude, you can think here like automated biological research or automating some sophisticated machine learning model to predict weather patterns.

When businesses use Claude in this very sophisticated way, they rely on disproportionately more input context than for simple tasks.

Now, why is that interesting?

I think it's interesting because it points toward the source of tasks we don't see in our data.

Tasks that the model might be very good at, but the context information is not readily available.

So concretely, if you're a large complex organization,

The amount of tokens, the input tokens.

Yeah, exactly.

So you need to provide.

Yeah.

That's actually an interesting and subtle point here is that like by looking at enterprise adoption, businesses are incentivized to do this in the most efficient way.

And so the correlation that we trace out between output complexity and input tokens

likely captures some fundamental constraint on what is needed by businesses to tackle very hard problems.

And so I think this points in the direction of a standard lesson of new technologies, which is businesses need to make complementary investments.

to unlock the benefits of new technologies.