Sorting interview prep by employer, not topic, is the right way to study for AI roles

The most useful thing about this repo is its organizing principle: not topics, but employers. Thirty-five companies, each with its own section covering the reported interview loop, the roles it targets, and the questions that have actually surfaced in those loops. That structure matters because the same underlying concept - KV cache memory, say - gets asked at very different depths depending on whether you're interviewing at NVIDIA (roofline arithmetic, derive the formula) versus a product company like Glean (here's a latency budget, where do you cut).

The common-questions section is genuinely well-constructed. Questions are listed once, with tags showing which companies ask them, so there's no repetition when you move into company-specific sections. The categories are also honest about what AI engineering interviews actually test: not just transformer internals, but inference economics, agent loop design, RAG pipeline architecture, evaluation methodology, and safety. The coding section includes problems that show up repeatedly across companies - async batch processors over rate-limited APIs, streaming SSE parsers, token-bucket rate limiters made distributed - which are more representative of what these roles actually require than standard algorithmic puzzles.

The company-specific depth varies in interesting ways. The Anthropic section describes a CodeSignal-style assessment that grows through four progressive levels against a black-box evaluator, and notes that some loops now include an AI-collaboration round where you're graded on how you direct and verify Claude. The xAI section flags that loops are fast and weight raw shipping ability heavily, with a four-hour timed product build. The Hugging Face section is unusual in that your public GitHub history is described as genuinely part of the evaluation. These aren't generic interview tips - they reflect the actual culture and priorities of each organization.

The answers are links to Outcome School blog posts and videos, which is the repo's commercial angle. That's fine; the question lists stand on their own, and the linked content covers the material at reasonable depth. What the repo cannot do is tell you what a given team is hiring for right now - the README says explicitly that interview loops change constantly and vary by team, level, and region, and to treat each company section as a map of what that company cares about rather than a script.

For someone preparing to interview at one of these companies, the most practical move is to start with the common questions tagged to your target company, then work through the company-specific section. The system design questions in particular - design the serving stack for a consumer chat assistant at hundreds of millions of users, design permission-aware retrieval, design an on-prem deployment for a European bank - are the kind of open-ended problems where knowing the shape of what the interviewer expects is most of the preparation.