Roadmap: Oct 2026 → Sep 2027 (~25 h/week ≈ 1,250 h)¶
Milestone-based: move on when the exit criteria are met, not when the calendar says so. Adjust in the monthly review. Every week also includes DSA 3 h (12-week plan), one journal entry, and, from stage 4, one customer conversation (founder track). Read each stage's topics in the library before its labs.
Note
No GPU? Every exit criterion below can be met with a CPU laptop plus free Colab or Kaggle T4 sessions for the scale-up runs (S1–S5). SETUP.md lists the minimum hardware for each experiment.
| stage | when (target) | hours | do | exit criteria |
|---|---|---|---|---|
| 0 Setup | week 1 | 10 | SETUP; measure_gpu.py; self-rate modules 0–5; read 00 |
GPU numbers in the journal; baseline ratings |
| 1 Foundations | Oct–Nov | 150 | 01–03; labs 01–03 | labs pass; derive the softmax-CE and LayerNorm backward on paper; napkin drills ≥ 80% |
| 2 Transformers | Nov–Dec | 150 | 04; labs 04–07; S1 pretraining run | your GPT trains on TinyStories; Triton kernel correct on GPU; blog post 1 |
| 3 Scale | Jan 2027 | 100 | 05; labs 08–09 and 18 (MoE); S2 scaling sweep | your own fitted scaling law; a 7B run planned on paper and defended; dense vs MoE at matched active parameters |
| 4 Post-training | Feb–Mar | 175 | 06; labs 10–12; S4/S5 | GRPO improves a verifiable task (with CIs); write-up 2; first OSS PR |
| 5 Inference | Mar–Apr | 125 | 07; labs 13–14; serving benchmark | measured tokens/s within 30% of your prediction; write-up 3 |
| Gate A | end Mar | — | research engineer vs founder evidence (founder README) | decide the emphasis for the next 3 months against thresholds written down in advance |
| 6 Eval & research | Apr–May | 100 | 08; lab 15; pick a research project | project proposal plus first results with error bars |
| 7 Interp & safety | May | 75 | 09; lab 16; values answers | SAE study; the values section of stories written |
| 8 Applied & founder | Jun | 75 | 10; lab 17; founder 01–03; upgrade FinSentinelAI evals | measured RAG; 10+ problem interviews; 3 scored ideas |
| Gate B | end Jun | — | offers pipeline vs paid pilots or LOIs, against pre-registered thresholds | commit: apply broadly, found, or both in sequence |
| 9 Launch | Jul–Sep | 250 | research write-up (workshop-ready); 2–3 merged PRs; applications (career) or paid pilots | onsites or signed pilots |
Gates: decide on evidence, not mood¶
Two weeks before each gate, write your thresholds into the journal and date them, using the decision sheet in the founder track. For example: "founder if 2 paid pilots or 5 LOIs with a price; research engineer if 3 onsites or 1 offer". The numbers are yours to choose; what matters is that you choose them before you see the evidence. At the gate, compare the evidence with the thresholds and move your hours towards the stronger side.
Research-engineer applications (from May)¶
Mocks start earlier: one every two weeks from March (month 6), per the mock protocol. From May: résumé v2 with a "Selected technical work" section → application waves: practice targets → core targets → frontier labs. Use interview-loops, the question bank, and the company guides for your two targets.
If you fall behind¶
Cut breadth, never depth: skip stretch goals and spec-lab extras before skipping a core lab or a write-up. Two weeks behind: extend the stage. Six weeks behind: drop stage 7 to reading-only and keep 4, 5 and 6.