Okay

... to create the agent output in the end.

Yeah.

Today, what, what we'll cover first, um, why synthetic data in the first place, right? Um, a bit of a background about pre-training data in 2026.

Then, um, how we structured the, the experiments.

So we looked at three axes, mainly the prompt, the generator model, and the source data.

Then we, we ran three hundred and thirty-three experiments in total to figure out what moves the needle, what, um, leads to good pre-training data.

[chuckles]