Skip to content
Stable Learning

How to Do Research

Written by Zhongming Yu

Nobody warns you that raw talent is the least scarce input to research. Every year, brilliant students — students who can derive anything, implement anything — walk into a PhD and walk out years later with a thin CV and a quiet sense of having been cheated. Almost none of them were beaten by the math. They were beaten by two decisions that felt too mundane to deserve their attention: what to work on, and how fast to move. Problem selection and working pace are where research outcomes are actually decided; execution is merely where they are collected. And beneath both decisions sits a prior question most students never ask out loud — why you are here at all — because no answer to what and how fast can rescue the wrong answer to why.

This track is the advice I give every new member of the lab before their first project, and it runs the full arc: first the why — the audit of your reasons that everything else depends on — then the three principles of the how, then the whole person who has to carry it all. The why and each principle compress to a single test you can run on yourself in one honest minute — four tests in all. This page is the map; the chapters are the territory, and each principle chapter carries a live simulator you can push until it fails.

Why Do Research

Section titled “Why Do Research”

Before what to work on and how fast to move, settle what you are working for — and in an era when LLM agent pipelines already draft papers end to end, that question is urgent, not philosophical. If your honest answer is “the degree,” my advice is to quit early: you will be miserable, and degree-driven work converges on the mass-produced incremental paper — exactly the artifact automation now produces at near-zero cost, so hand-crafting it is a waste of your best years. A PhD is a bet — enormous fixed cost, uncertain marginal return, several of your most important years staked on an extremely high-risk domain, with nobody guaranteeing a better job on the other side — and only two reasons are strong enough to place it: wanting the research experience and skills themselves, or pure curiosity and love for the science. So before anything else, ask:

The Why Test. If the degree disappeared tomorrow, would you still choose this work?

Read the full chapter: Why Do Research

Work on the Frontier

Section titled “Work on the Frontier”

The most common failure in research is not the problem that defeats you — it is the problem that quietly was never hard. It is dangerously easy to settle into a mature, comfortable corner of a field and spend a year stacking clever tricks onto a question nobody needed answered: motion that feels like research and produces nothing. Before you write a line of code, you owe yourself two answers — why is this worth doing, and why is it hard — and if you cannot produce both, the honest diagnosis is that you have not read enough of the current frontier to know where it is. So stop and ask:

The Frontier Test. Why is this worth doing, and why is it hard? Am I at the boundary of what is known, or decorating the interior?

Read the full chapter: Work on the Frontier

Bet on What Scales

Section titled “Bet on What Scales”

Some fields are traps with excellent lighting. They look glamorous for a season — GNNs had theirs, sparse-attention codesign had one too — but their life cycles are short because they carry a fatal, structural flaw: they fundamentally cannot scale up. In an environment where compute, data, and models grow relentlessly, an approach that fights scaling is an approach the future will delete — taking your years of work with it. You can do everything right inside a doomed field and still walk away with nothing — a long, disciplined effort spent building an elaborate empty room. So before you commit, ask:

The Scale Test. If compute, data, and models grow 100x, does my approach matter more — or less?

Read the full chapter: Bet on What Scales

Research Is a Race

Section titled “Research Is a Race”

Every open problem has an expiration date, and no one will wait for you. The result you are circling is being circled right now by other groups, and if you slack, your idea ships as someone else’s paper and someone else’s codebase — in agents research today, the gap between a promising idea and a published one is measured in weeks. This is why research demands 100% motivation and 200% effort: your thinking and your energy go into the paper every day, and rest and vacations are what you do after it ships. The alternative is well documented — a recycled manuscript, a new state of the art landing mid-revision, and months of self-blame that no amount of late regret can refund. So keep asking:

The Clock Test. If I paused this project for a month, would the result still be mine to publish?

Read the full chapter: Research Is a Race

The Complete Researcher

Section titled “The Complete Researcher”

The tests pick the problem and set the pace; you still have to be the instrument that executes, and research is a contest of the whole person. Your taste — trained on literature, philosophy, and art — decides which problems and which solutions are beautiful enough to matter, and your body decides whether you can sustain the effort at all: the best researchers in the world put in 80-plus-hour weeks, and nobody does that on an unmaintained body. Method compounds only on top of the person who carries it.

Read the full chapter: The Complete Researcher

How to Use This Track

Section titled “How to Use This Track”

Start with the why: read Why Do Research first, and do not skip it because it looks like the soft chapter — it decides whether the others are worth your time. Then read the three principle chapters in order — each ends where the next begins. Play with every simulator, and deliberately drive each one into its failure regime; the failures are the lesson. End with The Complete Researcher, the chapter about the instrument itself. Then come back: rerun all four tests at the start of every new project, because your reasons drift, the frontier moves, the scaling landscape shifts, and the clock never stops.

References

Section titled “References”