A field manual · not a contract
From strong engineer to independent researcher.
A step-by-step path through ML, LLMs, and the craft of research itself. Mastery is years, not months — but useful competence is months, and you climb by shipping artifacts, not finishing courses. The bottleneck is reps and taste, not information.
fig. 1 — your progress, plotted the only honest way: 0/37 checklist artifacts shipped. check boxes on the phase pages to descend.
§ 0 · In plain words
If the jargon loses you
An AI model is a guess-the-next-word machine — a program with billions of tiny dials that reads your words and guesses which word comes next, over and over.
Training means showing it mountains of text and, every time it guesses wrong, nudging all the dials a tiny bit so it's slightly less wrong next time. Billions of times.
A researcher runs fair experiments on these machines: change exactly one thing, keep everything else identical, run it several times, report honestly, publish so strangers can check.
This roadmap is the study plan from "can build software" to "can discover new, true things about how these machines work."
every term in this site is decoded in the glossary — keep it open in a tab.
§ 1 · Core philosophy
Five principles — more important than the curriculum
- 01
Build first, theory just-in-time
Do NOT spend 6 months on linear algebra before touching a model. Build a working thing, hit a wall, learn exactly the theory that wall requires, continue. The math matters — you pull it in when a model forces you to, not preemptively.
- 02
Spiral, don't sequence
The phases are a competence map, not a strict order. You'll build a tiny transformer in Phase 2 before you "finish" Phase 1's math — that's correct. A vertical slice early beats a perfect horizontal foundation.
- 03
Artifacts over courses
A finished course is worth nothing; a reproduced result with a writeup is worth a lot. Every phase ends in a shippable artifact. If you "completed" a phase but have nothing to show, you didn't complete it.
- 04
Reproduce before you innovate
You earn the right to have ideas by first reproducing other people's. Reproduction is where 80% of the real learning lives and where taste is built.
- 05
Learn in public
A private learner is an invisible learner with no feedback loop. Push code to GitHub, write short notes, post them, join a community. It's the feedback mechanism that makes you improve and the network that gets you collaborators and compute.
§ 2 · The levels
What "researcher" actually means
A ladder of capabilities, each reached by a deliverable — not a credential. Times assume ~10–15 focused hrs/week. Your near-term target is L2.
§ 3 · "You are here"
Find your entry point
Don't assume you're at zero. Start at the first thing you can't confidently do — by implementing, not "I've seen it." Most strong engineers land at Phase 1, move through it fast, and slow down at Phases 3–4 where the field-specific depth lives.
1.Derive backprop for a 2-layer MLP by hand and implement it from a blank file (no autograd).
2.Implement multi-head self-attention from scratch and explain why it's permutation-equivariant without positional encodings.
3.Explain the Chinchilla scaling law, what MFU is, and roughly what it costs to train a 1B model on one GPU.
4.Explain the difference between SFT, DPO, and GRPO, and name two ways an eval can lie to you.
5.Take a recent small-models paper, identify its baseline + the one confound that would invalidate it, and design a controlled extension.
§ 4 · The phase map
Six phases, one spiral
Orientation
Kill the paralysis. Ship a trained model in your first few days.
≈ 1 week
0/5 artifacts
Foundations
Backprop, gradient descent, and a neural net — from a blank file.
→ Level 1 (Reproducer) begins
≈ 4–8 weeks
0/5 artifacts
Transformers & LLMs
Build and train a GPT from a blank file. No black boxes left.
≈ 4–8 weeks
0/6 artifacts
Training & systems
How real models are trained efficiently. Reproduce a result, run one clean ablation.
→ Level 2 (Extender) begins
≈ 8–12 weeks
0/7 artifacts
Post-training & eval
SFT, DPO, GRPO — and evaluation as a discipline nearly everyone does badly.
≈ 4–8 weeks
0/6 artifacts
Specialization & research
Stop following a curriculum. Pick a niche, ship your first original result.
→ Levels 3–4
ongoing
0/8 artifacts
the labs · interactive explainers
one idea per page, every number draggable. read less, poke more.
Woven through all phases: research-method — the craft of being a researcher. Start it in week 1, not at the end. Taking the reinforcement-learning route instead? The RL branch grows from the same trunk.
§ 5 · Anti-patterns
How strong engineers waste months here
✗Math-first paralysis
"I'll learn all the linear algebra first." You won't, and you don't need to. Pull math in when a model demands it.
✗Tutorial hell / course collecting
Finishing lectures feels like progress and isn't. If you can't implement it from a blank file, you don't know it. Watched ≠ can-build.
✗Skipping reproduction
Chasing "novel" before you can reproduce. You'll generate confident nonsense. Reproduce first.
✗Learning in private
No public repo, no notes, no community. Invisible, and no feedback loop. Fix this in week 1.
✗Perfecting the plan
Re-reading roadmaps, re-sequencing, optimizing your setup. This site included: read it once, start Phase 0 today, deviate freely.
✗Over-respecting the phases
They're a spiral, not a gate. Build the tiny transformer before you "finish" the math.
Start now
Open Phase 0 and do the week-1 vertical slice today. Not tomorrow, after more reading. Today.
The roadmap works only if the first artifact exists by the end of this week.
phase 0 — orientation →