A Field Guide to AI Fellowships

Research Scientist

3 min read

Research Scientist

In one line

You generate and answer new research questions — design the experiment, run it, figure out what it means, write the paper. You shape what the lab works on, not just how.

What it actually is

Research scientists are the people who decide what’s worth studying and then study it rigorously. At a frontier lab or a research org, a scientist might notice that nobody really understands some behavior of large models, form a hypothesis, design experiments to test it, and publish the result so the field moves. The output is knowledge — usually a paper, sometimes a new method everyone adopts.

It’s the most “academic” of the tracks, and that’s not an accident: most research scientists either have a PhD or built an equivalent track record of original research in industry. The core skill being hired for is taste in research questions — knowing which questions are both important and tractable — which is genuinely hard and takes reps to develop.

What you actually do day to day

  • Read a lot of papers and form opinions about which directions matter.
  • Design experiments that could actually distinguish between hypotheses (harder than it sounds).
  • Pair with research engineers to run those experiments at scale.
  • Argue about results in front of a whiteboard.
  • Write — papers, blog posts, internal reports. A scientist who can’t communicate is half a scientist.

What you have to do to get in

The path

The honest mainline is: do research, show you can do research. That usually means a PhD, but the real requirement underneath is a track record of original work — a paper, a strong replication-plus-extension, a result someone cared about. Fellowships like the MATS scientist-style streams and the Anthropic Fellows program exist partly to let people build that track record without a traditional PhD.

For a beginner aiming here, the realistic 12-month plan is: build foundations → do one supervised research project (SPAR, a MATS-style program) → turn it into a public output → use that to apply to the next thing.

Skills required

See Research Skills and Technical Skills for the deep versions. The short list:

  • Math: linear algebra, probability, multivariable calculus, some optimization. You need to read ML papers fluently, which means the math can’t be a wall.
  • ML depth: how transformers work, training dynamics, evaluation. (Deep learning and transformers)
  • Research taste & rigor: forming hypotheses, designing clean experiments, not fooling yourself. (Research Skills)
  • Writing: papers and clear technical prose. (Career and Communication Skills)

Is this you?

Signs you lean scientist

  • You’re the person who asks “but why does that work?” and isn’t satisfied by “it just does.”
  • You enjoyed (or suspect you’d enjoy) the part of a project where you don’t know the answer yet.
  • You like writing and arguing about ideas, not just building.
  • You’re willing to play a longer game for a deeper kind of contribution.

Reasons to think twice

  • It’s the slowest cold-start without a PhD. If you want to be in a lab within a year, Research Engineer is usually faster.
  • A lot of the day is reading and writing, not building. Some builders hate this.

Pointers & extra resources

Research Engineer · Interpretability · Alignment and AI Safety · Skills Map · Tracks Overview

Graph View

Backlinks