Phase 6 — Research & Mastery (Month 13)¶
“Knowing ML” and “doing ML at the frontier” are separated by one thing: the ability to form a technical opinion no one told you to have.
You have spent 12 months building the substrate. You understand the mechanics of transformers, the theory of optimization, the engineering of production systems, and the discipline of experimentation. Month 13 is not more of the same. It is a gear shift. The goal is no longer to understand what researchers have done — it is to evaluate it, extend it, and eventually produce it. This phase transforms you from a consumer of knowledge into a producer of technical opinion.
The distinction is concrete: a consumer reads a paper and absorbs its claims. A producer reads a paper, identifies what the ablations don’t test, runs a reproduction, finds where the method breaks, and writes down their finding. One of these people gets hired by AI labs. One gets passed over for “lacking research depth.” This phase is about becoming the first person.
What This Phase Transforms¶
Before Month 13 |
After Month 13 |
|---|---|
Reads papers for understanding |
Reads papers to interrogate claims |
Follows the field |
Has a calibrated opinion on contested questions |
Uses open-source tools |
Has contributed to at least one |
Can implement known methods |
Can reproduce a 2024-2025 paper from scratch |
Describes experience in job applications |
Has public technical artifacts (posts, code, talks) |
“I’ve worked with LLMs” |
“Here is my reproduction of MLA, here is where it diverges from the paper, here is my hypothesis about why” |
The Cycle: Paper → Reproduction → Extension → Contribution¶
This is not a linear path. Most papers go no further than annotated notes. That is correct behavior. The cycle is activated selectively — when a paper is foundational, contested, or directly relevant to your production work.
What Most People Get Wrong¶
Trying to read every paper instead of reading deeply. In 2025, arXiv publishes 100-200 ML papers per day — one paper every ~8 minutes. The person who skims 10 papers a week retains nothing. The person who reads 2 papers per month with full reproduction and written critique builds a compounding advantage. Breadth is a trap. Depth is the asset.
A second failure mode: treating paper reading as a solo activity. The engineers who advance fastest run or join reading groups, argue about papers publicly, and publish their disagreements. Disagreement is data. “I reproduced this paper and got different results on this dataset” is more valuable than “I read 500 papers.”
Exit Criteria — Month 13 (Final Exit Criteria of the Entire Roadmap)¶
You have completed this roadmap when you can demonstrate all of the following without preparation:
Technical Depth¶
Explain the core mechanism of at least 3 papers from 2024-2025 from memory, including their key innovation, their strongest evidence, and their most significant limitation
Reproduce the core algorithm of at least 1 recent paper from scratch and post it publicly
Identify, in real time during a conversation, when a technical claim lacks adequate ablation evidence
Contribution¶
Have at least 1 merged PR or accepted issue/documentation contribution in a major ML open-source repo (HF Transformers, PEFT, TRL, litgpt, or equivalent)
Have at least 2 published technical artifacts: blog posts, notebooks, or talks that demonstrate original analysis (not tutorials)
Opinion Formation¶
Hold and defend a calibrated technical opinion on at least 2 contested ML questions (e.g., “SSMs vs Transformers,” “RL for reasoning vs SFT,” “MoE scaling vs dense scaling”)
Be able to articulate why you hold that opinion with specific paper evidence and reproduction results
Interview Readiness¶
Complete 3+ mock ML system design interviews with documented feedback
Implement attention mechanism, backpropagation, and at least one training loop debugging exercise from scratch under time pressure
Have a portfolio walkthrough narrative that takes exactly 8–12 minutes and survives 20 minutes of follow-up questions
Identity¶
Have a clear 2-sentence answer to: “What is your specific area of depth in ML?” — with artifacts to back it
Have demonstrated the 5 dimensions of applied PhD-equivalent work (see
07_the_applied_phd_identity.md)
The standard is not “did you finish the curriculum.” The standard is: can you survive a 2-hour technical interview at an AI lab, then have a productive conversation with a research scientist about a paper they published last month? If yes — you have arrived.
Phase Structure¶
File |
Focus |
|---|---|
|
The 3-pass method, 5 critical questions, workflow |
|
13 must-read papers with practitioner-focused analysis |
|
Contribution ladder, 4-week sprint plan |
|
Forming, defending, and publishing technical POV |
|
Sustainable system: daily/weekly/monthly habits |
|
FAANG structure, ML system design, compensation data |
|
5 PhD dimensions, honest gaps, honest edges |
|
3 capstone projects with exact acceptance criteria |
Time Allocation (Month 13: ~50–60 hours total)¶
Activity |
Hours |
Rationale |
|---|---|---|
Deep paper reading (2 papers/week × 4 weeks) |
16h |
2h pre-read + 2h discussion/notes per paper |
Paper reproduction (1 core paper) |
12h |
Most important single investment |
Open-source contribution sprint |
10h |
1 merged contribution is the target |
Blog post / technical writing |
8h |
Permanent public artifact |
Interview preparation |
8h |
System design + coding under pressure |
Reading group / community |
4h |
Compounding return via peer discussion |
Begin with 01_how_to_read_papers.md — the skill that unlocks everything else in this phase.
Return to Phase 5 README · Next: 01_how_to_read_papers.md