xAI · Primly Community

xAI data scientist interview (SQL + case + stats), what I saw

analyst_ana · 7 replies

Finished my xAI DS loop two months ago. I've interviewed at Anthropic, OpenAI, and a few others in the AI space so I can compare.

What the xAI DS loop has:

SQL round. Two problems, intermediate to advanced. Window functions, CTEs, one problem that required a self-join with some tricky filtering logic. Not LeetCode hard but definitely not starter SQL either. Platform was CoderPad.

Statistics/ML round. Questions ranged from basic probability (conditional probability, Bayes theorem) to experiment design (A/B test setup, power calculations, handling novelty effect) to a ML conceptual question about bias-variance tradeoff. One question about evaluation metrics: when would you use precision/recall vs. AUC, and why. They wanted the actual reasoning, not just the textbook answer.

Case study / product analytics. They gave me a hypothetical Grok feature and asked me to define the key metrics, identify potential issues in measurement, and propose a launch experiment. This was probably the most interesting round. They pushed on whether my proposed metrics could be gamed or were actually causally related to what we cared about.

Behavioral. Standard ownership and cross-functional questions.

Comparing to other AI lab DS loops: xAI's SQL round is harder than OpenAI's (OpenAI was more straightforward). The stats round is comparable to Anthropic. The product analytics case was most similar to what I saw at a couple of big tech companies.

Comp from offer: I didn't get the offer (failed the case round, my experiment design had a flaw they caught). But from what I've seen floating around, DS compensation is in the $180-240k TC range at mid-senior level, SF-based. This is a rough estimate from second-hand info, not my own offer.

Timeline: recruiter screen to final decision was about 5 weeks total including scheduling delays.

7 replies

analyst_ana

The "can your metric be gamed" question is so good. Most analytics case rounds don't go there. What was the flaw in your experiment design, if you don't mind sharing?

firsttime_mgr

I didn't account for network effects in the experiment. The feature I was analyzing had social components, so a naive user-level split would have spillover. I needed a cluster-level randomization and I didn't catch it until they pointed it out. Embarrassing in retrospect.

sec_sasha

The SQL window function stuff is table stakes at this point for any DS role at a real company. Good that they actually test it though, I've been in loops where the SQL was embarrassingly easy.

marketer_mei

The $180-240k TC range for DS at xAI, is that total or base? Because at other AI labs that kind of range is often just base.

content_cole

Total comp from what I was told second-hand. Not base. At AI labs right now the equity component is significant, so base alone would be lower. I really can't confirm without actual offer data.

sdr_sky

Did they ask anything specific to Grok's data or was it all generic analytical scenarios?

pivot_pat

The case used a Grok feature as the framing but you didn't need inside knowledge of the product. It was hypothetical enough that any reasonable product assumption worked. They told you the basics in the prompt.