Roadmap.

Phase 0 · ≈ 1 week

Orientation

Kill the "I don't know where to start" paralysis by shipping a trained model in your first few days, set up a research environment you'll actually keep, and start the two habits — a log and reading — that run through every later phase.

What to learn · in priority order

  1. 1

    Environment (half a day)

    Python + PyTorch. A single consumer GPU, or free Colab/Kaggle to start — don't spend money yet. Confirm torch sees your GPU/MPS. Don't over-build the setup; you'll redo it later.

  2. 2

    First vertical slice (1–2 days)

    Run the fast.ai Lesson 1 notebook, or train nanoGPT on tiny-shakespeare following the README (nanoGPT is deprecated but frozen — ideal for a first run: it will never change under you). Watch a real training loop run and a loss curve drop. You will not understand most of it. That's fine.

  3. 3

    Public scaffolding (half a day)

    A GitHub repo (e.g. learning-ml), a research log (LOG.md — one entry per session; the single highest-ROI habit in the roadmap), and a paper-notes folder.

  4. 4

    Join one community (1 hour)

    The EleutherAI Discord (#beginners / #research) is the highest-signal home for independent model researchers. Make an account, read, don't post yet.

  5. 5

    Set your target (1 hour)

    Re-read the Levels ladder. Write one line in your log: "My 6-month target is L2 — run a clean extension of a small-models result." Vague goals produce vague effort.


Primary path · use these, not ten others

fast.ai Lesson 1 or nanoGPT on tiny-shakespeare

Either gets a real loss curve dropping on day one. Pick one, run it today.

TinyStories (arXiv 2305.07759)

Your first figures-first paper skim — short and motivating.


The deliverable

A public repo with a loss curve screenshot and a started LOG.md. Trained something end-to-end, even without understanding it yet.


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

0/5 shipped

artifact: A public repo with a loss curve and a log.

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


Traps

Tooling rabbit holes

Don't spend the week on the perfect dev environment, dotfiles, or a GPU rig. Colab is fine for now.

"I need to understand it first"

No. Run it, then understand it. Inverting that instinct is the whole point of this week.

Skipping the log

The log is your external memory and the raw material of your first writeups. Start it now or you never will.