Waymo Data Scientist Interview Experience — Rare Events, Sampling, and Probability Coding

A recruiter contacted me. After screening and a technical screen, I went into four interview rounds. I've made the original questions less specific for others preparing.

Data Intuition, first round
A new version performs significantly better in offline evaluation, in simulation, but shows no difference after going live. What could cause that?
How would you model a rare event using a log form, and why use that form? Be sure to mention how to handle zeros.

Data Intuition, second round
Given a batch of simulation scenarios based on historical data, how would you test whether a new version is better than the old one?
You're given two tables. The first lists several units, such as ten trips, and each unit's probability of being selected for inspection. The probabilities differ across units. The second lists the units actually selected and whether an event occurred for each, yes or no. How would you use this data to estimate the overall event rate?

Statistics
Only some data from a particular type of scenario has been retained. How would you estimate the overall event rate? For example, all nighttime driving logs are retained, while the daytime logs are too large, so only a known random fraction is retained. How would you use the retained data to estimate the overall braking-event rate for that week? If you could choose the sampling fractions for the different scenario types, how would you choose them?
Two versions each have a small number of observed events over different mileages. Can you treat one version's estimated rate as known to calculate the other version's p-value? If not, how should you compare them?

Coding
Write a function that takes n and returns n uniformly distributed points inside the unit circle.
Given a set of two-dimensional points, write a function to determine whether they come from the entire circle or only half of it.

Overall impressions: Prepare an answer about your own data science experience for every round. I hadn't expected the coding questions to have nothing to do with algorithms—they were entirely probability questions. Data intuition was all about rare events, so focus your preparation there. A lot of what I'd prepared did come up, but the interviewers didn't accept the AI-generated approaches I'd prepared beforehand. The questions felt very specific to their business. Reflecting after failing, I realized that many of my answers simply stated what I thought without explaining any assumptions. I think that may be central to how they judge whether you meet the hiring bar. It really was very difficult. Hope everyone gets something useful from this!