Frontier AI / foundation models Interview Loops
Frontier AI labs don't share one interview structure. They share a subject. The number of stages and the formats vary a lot. OpenAI's engineering screen is two interviews, Mistral's process has six stages, and Perplexity's onsite lasts four and a half hours. But the content comes together: practical multi-level coding, ML coding for LLMs, debugging code that runs but learns the wrong thing, design for serving and training, and a project or paper discussion. No research or engineering loop here asks SQL. Data science is the exception, and it's the one role where SQL and experiments still lead the loop. Rules on AI tools differ more here than anywhere. OpenAI encourages them in its data science take-home, and Anthropic bans them in live rounds.