What LLM labs actually hire for
Research engineer roles at the big labs filter for production ML skills first, paper count second.
Max Mynter documented his job hunt into research engineering at the big LLM labs and the pattern he found surprised me. The role is not junior research scientist. It is senior ML engineer with a research slant. The filter starts at production. Can you train models at scale. Can you debug distributed PyTorch jobs that die at epoch seven. Can you write efficient data loaders. Publications help but they are not the core signal. The labs want people who can ship model improvements into inference, not just write the paper. Mynter mentions the interview format: technical screens on ML fundamentals, system design for training infrastructure, coding in actual PyTorch. No whiteboard proofs. No reciting Transformer architecture from memory. They want to see if you know why gradient accumulation breaks with certain optimiser states, or how to shard a 70 billion parameter model across eight nodes without deadlocking. The cybersecurity parallel is obvious. We hire detection engineers who can write Sigma rules and tune SIEM pipelines, not penetration testers who collect CVEs. Both roles touch security but the day-to-day is infrastructure, not exploration. Research engineering at an LLM lab is the same split. You implement the next quantisation technique, you do not invent it. What I would adjust if I were applying: build one end-to-end training repo with real monitoring, checkpointing, and multi-GPU support. Make it public. That is the portfolio piece that gets you through the screen. Papers are good. A repo that shows you can actually run a 200 hour training job without losing state is better.