Roadmap.

Phase 5 · ongoing→ Levels 3–4

Specialization & research

Stop following a curriculum and start doing research — pick a niche, produce your first original result (however small), engage a community, and ship it publicly. This phase has no end; it's the transition from learning the field to contributing to it.

What to learn · in priority order

  1. 1

    Pick a niche — deep, not wide

    One sub-area where you can build a real result on one GPU: efficiency (quantization/distillation/pruning), mechanistic interpretability, data curation/synthetic data, or small reasoning models/post-training.

  2. 2

    The research loop

    Reproduce something in your niche → find the open thread (the unrun ablation, the unswept variable) → scope it (prior art, feasibility, confounds) → run it cleanly → ship the artifact.

  3. 3

    Get visible

    Ship in public — open weights + reproducible eval + writeup beats a clever-sounding paper with no artifact. EleutherAI (#research, SOAR) solves collaborators, compute, and the arXiv-endorsement gate simultaneously.

  4. 4

    Realistic venues

    ICLR Blog Posts track, NeurIPS ENLSP workshop, the ML Reproducibility Challenge. Not the main-track lottery.

Primary path · use these, not ten others

The small-models research on-ramp

Your applied specialization guide — scoped projects, the field map, the how-to-get-visible-and-publish playbook. Read it in full now.

ARENA (if interpretability) ↗

The mech-interp / alignment-engineering curriculum, with TransformerLens.

The research-buddy

The tool for scoping and pressure-testing project ideas — prior-art check, feasibility, confound check. Essential from here on.

The deliverable

Your first original result: fix one model family, one dataset, vary one axis, ≥3 seeds, same-size baseline. Ship open weights/code + reproducible eval harness + honest baselines + a clear writeup. This is your L3 credential.

Milestones · check only when you can do it from a blank file

0/8 shipped

artifact: A public original result. → Level 3 reached. (Level 4 = recognized, self-directed work over years.)

rule: every phase must end with a public artifact (repo + writeup). no artifact → phase not done, regardless of what you've watched.

Traps

✗Niche-hopping

Depth compounds; breadth doesn't. Stay in one sub-area long enough to develop taste before switching.

✗Reaching for novelty too early

Your first "original" result should be a clean extension of existing work, not an invented direction. Earn novelty.

✗Building tools instead of doing research

Ship results, not infrastructure. (Over-investing in the research-buddy counts.)

✗Isolation

The single biggest accelerator available to you is a community that grants compute, gives feedback, and co-authors. Use it.