A student stopped me after a talk last month and said: “I know Python. I've built a web app. How do I start doing research?” Good question. Here is the short version I would give you.

Month 1–3: read ten papers, read them twice

Pick one topic. Ten recent, highly-cited papers in that topic. Read each one twice. First pass for scope; second pass with a pen. For each paper, write a one-page summary with: contribution, method, limits, what you would cite it for.

Bad: reading thirty papers once. Good: reading ten papers twice.

Month 4–6: reproduce something

Pick the paper whose code is available (most are, on GitHub). Reproduce a main result. You will hit undocumented preprocessing. You will hit a hyperparameter that isn't in the paper. Debug it. This is research.

Month 7–9: small original experiment

From the reproduction, spot one question the paper didn't answer. Run a focused experiment. Write it up, even if it's negative. Especially if it's negative.

Month 10–12: first workshop paper

Target a workshop, not a top-tier conference. Your first paper will likely not be brilliant. It will teach you how to write, review, and respond to reviewers.

What to ignore

  • Twitter threads that promise “AGI by Friday”.
  • The compulsion to have an opinion on every new model release.
  • Any advice, including this post, that someone gives you without knowing your specific topic.

When in doubt, come talk to your advisor. They have seen this before.