Research Skills

The skills that separate “I can run experiments someone designed” from “I can figure out what’s worth studying and study it well.” These are what scientist-track fellowships are really selecting for, and they matter a lot for Interpretability and Alignment and AI Safety too. They’re harder to build than technical skills because they take reps on real problems — which is exactly why the supervised-research fellowships exist.

Reading papers

The entry skill. You can’t contribute to a conversation you can’t follow.

What "good enough" looks like

• You can read an ML paper and extract: what question, what method, what they found, what’s shaky.
• You don’t need to understand every equation to get the contribution.
• You’re building a mental map of who’s working on what.

How to get unstuck

First pass: abstract, intro, figures, conclusion. Don’t start at equation 1. Read the shape before the details. Speed comes from volume — your tenth paper is far easier than your first.

Forming good questions (research taste)

The hardest and most valuable. Knowing which questions are both important (worth answering) and tractable (answerable by you, soon). This is what a PhD slowly builds and what mentored fellowships try to accelerate.

You build taste by

• Reading a lot and noticing what’s missing or unexplained.
• Working under someone with good taste and stealing their instincts. (This is the whole point of MATS / SPAR.)
• Doing small projects and feeling which questions led somewhere vs. nowhere.

Experiment design & rigor

Designing experiments that actually distinguish between hypotheses — and not fooling yourself about the results.

What "good enough" looks like

• Your experiment could actually come out the other way (it’s a real test, not a demo).
• You control the obvious confounds.
• You’re suspicious of your own results, especially the ones you wanted.
• You know the difference between “this is suggestive” and “this is established.”

The cardinal sin

Fooling yourself. Feynman’s line — “you are the easiest person to fool” — is the whole job. The researchers reviewers want are the ones who try hardest to break their own exciting result.

Turning work into output

Research that nobody can read or reproduce barely counts. The ability to package a result — a paper, a clean repo, a public writeup — is part of the skill, not an afterthought. (More in Career and Communication Skills.)

---

Related: Skills Map · Technical Skills · Research Scientist · Interpretability