Study plans
AI researcher preparation
A 12-month sequence for candidates building toward a research-oriented role at a frontier lab or applied research team. Heavy on mathematics, foundational ML, and publishable work; light on vendor credentials.
This plan is for candidates building toward research-track roles — frontier lab research positions, applied research teams at AI-native companies, or research-adjacent positions at universities and institutes. The signal that closes these hiring loops is substantially different from the signal that closes applied-engineering loops. Vendor credentials carry little weight; what carries weight is mathematical depth, foundational ML competence, and demonstrably-publishable work.
Twelve months is a working window. For most candidates this plan extends to 18-24 months when realistic time constraints are factored in. The plan is dense.
Plan at a glance
| Phase | Window | Goal | Workload |
|---|---|---|---|
| 1. Mathematical foundations | Months 1-3 | Calculus, linear algebra, probability rigor | ~12 hrs/week |
| 2. ML foundations | Months 4-6 | The full classical-to-deep ML stack | ~12 hrs/week |
| 3. Research-track depth | Months 7-9 | One subfield deep + paper-reading discipline | ~12 hrs/week |
| 4. Publishable output | Months 10-12 | A workshop paper or comparable artifact | ~15 hrs/week |
Phase 1 — Mathematical foundations (months 1-3)
Goal: actual mathematical competence. Not “passing-familiarity” but the ability to read a paper’s methods section and reproduce its derivations.
Recommended programs:
- MIT MicroMasters in Statistics and Data Science (subset) — pick the relevant probability and linear-algebra modules. The MicroMasters is rigorous in a way most adjacent material is not.
- A standard mathematics-for-ML reference — the standard textbooks here are well-known and we will not name a specific edition; the field has multiple solid options.
Milestones:
- Comfortable derivation of gradient descent, backpropagation, common optimization algorithms.
- Linear-algebra fluency at the level of confidently working with covariance matrices, eigendecomposition, SVD.
- Probability theory at the level of being able to read a Bayesian inference paper.
Prerequisites: undergraduate-calculus comfort. If you don’t yet have that, add a four-week sprint up front on calculus refresher (Khan Academy + a problem book; the problem book matters more than the videos).
Phase 2 — ML foundations (months 4-6)
Goal: the full classical-to-deep ML stack, with mathematical depth.
Recommended programs:
- Stanford Online — XCS AI courses — the XCS-prefixed Stanford courses are research-calibrated. The machine learning and deep learning tracks are the most relevant.
- Harvard CS50’s AI track — useful as a complement, particularly the search and constraint-satisfaction material that many deep-learning-first programs skim past.
- DeepLearning.AI Deep Learning Specialization — for the applied-implementation layer.
Milestones:
- Implemented at least one classical ML algorithm from scratch (linear regression with regularization, a decision tree, a small SVM).
- Implemented at least one deep-learning model from scratch (a small transformer, a CNN, a recurrent model — pick one and go deep).
- Comfort with PyTorch at the level of reading other people’s research repos.
Phase 3 — Research-track depth (months 7-9)
Goal: one subfield deep, plus the paper-reading discipline that research-track candidates use to navigate the field.
The subfield. Pick one. The most-tractable subfields for self-directed research-track preparation in 2026 are: alignment / safety, mechanistic interpretability, agent design, reinforcement learning from human feedback, multi-modal architectures, sample-efficient learning. There are others; these are the ones with the most-accessible literature and the most-active independent research communities.
Paper-reading discipline:
- Read at least three papers per week in your subfield. Take notes on each.
- Implement the toy version of at least one paper per month.
- Write a short reading note on each paper. The notes do not need to be published; they need to be written.
Recommended programs:
- None at the credential level. This phase is research-reading work, not coursework. Some credentialing programs in adjacent areas (Cornell’s ML certificate, Imperial College’s offerings) can supplement, but they are not the load-bearing material.
Phase 4 — Publishable output (months 10-12)
Goal: a workshop paper or comparable artifact that demonstrates publishable-quality research work.
The work:
- Identify a small but defensible research question in your subfield — something narrow enough to make tractable in 8-12 weeks, something not already solved.
- Run the experiment. Write it up to publishable standard.
- Submit to a workshop at one of the major conferences (NeurIPS, ICML, ICLR workshops; ACL or EMNLP workshops if your subfield is NLP-leaning).
A workshop paper is not a top-conference paper. It does not need to be. What it demonstrates is that you can identify a question, run an experiment, and write a result. Hiring committees for research-track roles read workshop papers as a meaningful signal.
Alternative artifacts:
- A substantial open-source research-grade implementation (e.g., a faithful reproduction of a recent paper, with notes on what was hard).
- A long-form analysis published independently (a comprehensive analysis of a specific architectural pattern, an empirical study of a specific phenomenon).
The workshop paper is the gold-standard option; the alternatives are real but slightly lower-signal.
What this plan is not for
- Engineers transitioning into applied AI roles. This plan is research-oriented and substantially over-invests in mathematics for that purpose. Use AI engineer in 6 months instead.
- Founders and operators. This plan is wildly over-invested for founder-track literacy. Use Founder-track AI literacy instead.
- Candidates without time. Twelve months is the minimum, and most candidates spend longer. If you cannot commit ~12-15 hours per week sustained, this plan will not produce its intended outcome at its intended pace.
A representative completion pattern
Research-track candidates are different from the operator-track and founder-track candidates Edge Curriculum profiles most often. We have not, to date, published a profile of a candidate following this specific plan; the canonical pattern in the broader field is the post-undergraduate self-directed researcher who anchors on a workshop paper and applies to PhD programs or to applied research roles. The pattern is well-documented in field literature; the credentials matter less than the research output.
Update log
- 2026-05-12: Initial publication.
Study plans on Edge Curriculum are working recommendations, not prescriptions. For corrections or suggested improvements, corrections@edgecurriculum.com.