Research Engineer

In one line

You build the code, infrastructure, and experiments that turn research ideas into real results. If the scientist asks “what if we tried X?”, you’re the one who makes X runnable, fast, and trustworthy.

What it actually is

Research engineers are the backbone of modern AI labs. There are usually two to four times more RE roles than research scientist roles, because research at scale is mostly engineering: training runs, eval harnesses, data pipelines, tooling, reproducibility. An RE takes a fuzzy research idea and turns it into clean, fast, correct code that produces a result you can believe.

This is, for most beginners, the single best entry point into the field. It rewards exactly the thing a motivated learner can build on their own — demonstrable engineering skill — rather than credentials. And it converts naturally: plenty of REs grow into research scientists over time by gradually owning more of the “what should we study” question.

What you actually do day to day

• Implement a method from a paper and get it actually working (deceptively hard).
• Build evaluation harnesses so the team can measure whether a change helped.
• Optimize training/inference so experiments run in hours instead of days.
• Debug the weird stuff — why is loss spiking, why does this only fail on 8 GPUs.
• Pair closely with scientists; the best REs shape the research, not just serve it.

What you have to do to get in

The path

This is the most “show, don’t tell” track. You need code reviewers can look at. A clean repo that reimplements a known result, a small interpretability or RL project, a contribution to an open-source ML tool — any of these beats a paragraph about how passionate you are.

The fellowship on-ramp is excellent here: ARENA is literally designed to take someone with basic coding and turn them into an AI-safety research engineer in ~5 weeks. MATS engineering-flavored streams, LASR Labs, and the OpenAI Residency all take strong engineers without research pedigrees.

Skills required

Deep versions in Technical Skills. The short list:

• Strong Python, and comfort with PyTorch. This is non-negotiable and the most learnable. (Python and PyTorch)
• Deep learning fundamentals: you can implement a transformer block from scratch and explain every line. (Deep learning and transformers)
• Engineering hygiene: version control, testing, reading other people’s code, debugging methodically. (Software engineering hygiene)
• Enough math to not be scared of the papers. Less than a scientist needs.

Is this you?

Signs you lean engineer

• You’d rather have something running tonight than a perfect plan for next week.
• You like fast feedback loops — write code, run it, see if it worked.
• You’re the person friends ask to debug things.
• You learn by building, not by reading first.

Why I steer a lot of beginners here first

It’s the track where one weekend of focused work produces visible evidence. You can’t “show” research taste in a weekend, but you can ship a clean transformer-from-scratch notebook. Evidence is what gets you in. (The portfolio mindset)

Pointers & extra resources

• ARENA curriculum — even if you don’t get into the program, the materials are public and excellent. Work through them.
• Jacob Hilton’s Deep Learning Curriculum for self-study.
• Research Engineer vs Research Scientist for the role distinction.
• More in Reading and Courses.

Related

Research Scientist · Interpretability · Applied and Product ML · Skills Map · Tracks Overview