Becoming an AI Researcher
Topic 19 · Part 4 · Your paper (capstone) · ⏱ about 120 h

Capstone: Write Your Own Paper

Pick a feasible first project in the field you choose, then work through a stage-by-stage paper workbook that saves as you go, exports a Markdown or LaTeX draft, and ends with a reviewer-style self-assessment.

Show:
In this topic
Beginner

What this capstone is for

Everything in Parts 1–3 comes together here: you will produce a real first paper draft in your chosen field. A completed, honest, well-evaluated study (posted on arXiv or submitted to a workshop) with released code is the strongest single piece of evidence in a PhD application, stronger than grades alone.

Be realistic: a first project usually takes 3–6 months part-time, and most first papers are rejected at least once. The workbook keeps you moving: each stage has a concrete output, and none of them requires a big GPU budget.

Beginner

Feasible first projects

Good first projects answer a narrow question well with modest compute. Common, respected types:

TypeExampleWhy it works
Careful reproductionreproduce a recent paper's main table; test whether results hold with more seeds or a new datasetteaches the whole pipeline; reproducibility tracks and challenges publish these
New setting or populationevaluate known methods in a low-resource language, a new medical domain, or a different country's dataa clear gap; your local knowledge is an advantage
Analysis or evaluation studywhen and why does method X fail? robustness, fairness or calibration across groupsno new method needed; rigorous analysis is valued
Benchmark or dataseta carefully documented dataset for an under-served task (with ethics approval and consent)high impact if done well; plan for data governance
Small method improvementa simple change with a clear rationale, tested fairly with ablationsharder to show convincingly; needs strong baselines
Try it

Try it: project ideas for your field

Choose your field and resources to see project patterns that suit a first paper. Then make one specific to a dataset or setting you know.

Interactive: this part runs in your browser.

Try it

The paper workbook

Work through the stages in order; your text is saved in this browser as you type (export regularly as a backup). Each stage shows the questions a reviewer will ask. Export a Markdown draft or a LaTeX skeleton for Overleaf at any time.

Interactive: this part runs in your browser.

Beginner

A realistic timeline

Weeks (part-time)StageOutput
1–3question, reading, gapone-sentence question; literature matrix of 15–30 papers
4–5planwritten experiment plan (metrics, baselines, seeds, splits) reviewed by someone
6–12experimentsbaselines reproduced first, then your runs; results files and scripts in Git
13–15analysistables and figures from code; ablations; error analysis
16–19writingfull draft; feedback from two readers; revision
20–24submissionarXiv preprint, code release, workshop or conference submission

If you have no supervisor yet, ask a lecturer, a researcher whose paper you build on, or a community such as ML Collective for feedback on your plan and draft. Short, specific requests ("could you look at my two-page plan?") get replies.

Check yourself

Quiz

1. Which first project is most feasible with a laptop and a free GPU?

2. Why reproduce baselines before running your method?

3. The strongest single piece of evidence in a PhD application is usually…

Go deeper

📚 Further reading

Trusted, mostly free sources: official documentation, well-known textbooks and courses. When you want to confirm or go deeper on anything here, start with these.

Ready