Did the xAI onsite in March. Full day, four rounds back to back. Virtual for me (I'm not SF-based) but I've heard in-person is similar structure.
Round breakdown:
Round 1: Coding. Two problems, 60 minutes. LeetCode medium to hard. I got a graph problem and a dynamic programming problem. Definitely harder than the OA. They expect you to think out loud and they will probe your complexity analysis.
Round 2: System design. 50 minutes. Prompt was infrastructure-flavored: design a distributed training job scheduler for large GPU clusters. Right in my lane but still challenging. They care about failure modes, recovery, priority queuing, and cost efficiency. The interviewer knew their stuff and pushed back on my proposals meaningfully.
Round 3: Behavioral. 40 minutes. Two or three questions, pretty standard. Ownership, disagreement, speed of decision-making. Nothing surprising here.
Round 4: Domain deep-dive / culture. 40 minutes. This one was less structured. They asked me about specific technologies I'd worked with, why I made certain architectural choices in past roles, and what I thought about current trends in distributed ML infrastructure. Also a few "why xAI" type questions.
Lunch break in the middle. No coding during lunch.
Feedback loop: I heard back in 5 business days. Passed to offer stage. The debrief apparently takes 2-3 days internally before they decide.
General vibe: the interviewers were sharp and engaged. Not the kind of loop where they're reading from a script. They riff off your answers. Good if you like that kind of conversation, probably stressful if you want predictability.