Months 7–18 — phases (P2–P5)

Coarser than months 2–6. Matches 00-the-job/roadmap.md Track A. Daily clock remains operating-system.md. Point at curricula; do not restart resource lists.

Delta vs 05-interview/full-prep.md: that file's "M7 take-home <1487 / M12–14 apply" calendar is a different first-30-days (Karpathy-first) and a more aggressive apply window. This path prefers first-30-days.md + this file + roadmap.md. Apply when the apply-gate below is true, not when a blog says "month 12."

Hours below are deep hours. 12h×6 ≈ 9 deep h/day × 24 days/month ≈ ~200 deep h/month if you actually hold the OS. Life will cut that. The roadmap already says 16–24 months with a job.

Months 1–2   P0 foundations     (done if month-02 exit held)
Months 3–4   P1 first kernels   (done if month-04 exit held)
Months 5–6   P2 started         (FA3 literacy, serving papers)
Months 7–10  P2 finish          LLM kernels to a portfolio bar
Months 11–14 P3 systems         serving + multi-GPU + design fluency
Months 15–16 P4 post-training + culture
Months 17–18 P5 interview gym + apply-gate

If a month has no public artifact, the month did not happen.

Months 7–10 — P2 finish (LLM kernels)

Roadmap P2 exit: FA2-style fwd+bwd you understand; decode vs prefill kernels; paged KV; FP8 cast path; writeup with % of peak on a real GPU; take-home < 2,000 (you should already have this; now push).

Point: 03-kernels/curriculum.md Layer 3 level-up bar. Do not collect FA4 blog posts.

Month 7 — Backward and a second kernel family

Deep: FA-style backward (recompute tiles, store m/ℓ) or a written + tested tiny backward plus a fused CE / SwiGLU (03-kernels L3 map).
Paper: FA3 reread; Anatomy of a Triton Attention Kernel (arXiv:2511.11581).
Friday: take-home. First serious run at < 1,800.
Artifact: backward (or fused CE) + numerical table + % of a library baseline.

Month 8 — Decode is the job

Deep: paged decode attention that is not a toy; prefix cache hits; FlashMLA read if MLA interests you.
Paper: FlashDecoding; vLLM kernel design docs; Medusa / EAGLE-3 as verify is prefill-shaped.
Design: 45 min weekly (keep the month 5 habit).
Artifact: decode tokens/s vs batch vs page size. Roofline.

Month 9 — CUTLASS / CuTe or ThunderKittens (pick one)

Deep: 03-kernels Layer 2 graded exercises 3–5. Type CuTe sgemm_1 or a TK tutorial kernel. Triton-only people fail pairing on GEMM structure.
Paper: CUTLASS 3.x design + one Colfax Hopper GEMM.
Friday: take-home < 1,600 stretch.
Artifact: GEMM TFLOPS table vs cuBLAS/rocBLAS-equivalent on your GPU. Name the GPU in line 1.

Month 10 — FP8 path + portfolio freeze

Deep: FP8 or MX cast path with a correctness eval (output regression is a bug). TE recipes + from-scratch scales if the SKU allows.
Paper: MXFP8 recipes 2506.08027; DeepSeek-V3 numerics bits you skipped.
P2 exit review (all must be true):
FA2-style fwd you can defend; backward honest
Decode ≠ prefill in code
Pager that is not vapor
% of peak written for one kernel
Take-home < 2,000 (better: a 1487-path with experiments left)
Artifact: PORTFOLIO.md — three links (GEMM, attention, pager/decode). This is the resume.

Months 11–14 — P3 systems

Roadmap P3 exit: implement or surgically modify TP all-gather/reduce-scatter overlap; a toy vLLM-like pager (you have v0); a continuous batcher; 45 min inference serving whiteboard without collapsing.

Point: 03-kernels Layer 5; 04-llm-theory §D–F; month-06 serving notes.

Month 11 — Continuous batching that runs

Deep: scheduler: prefill chunks + decode tokens, pause/resume, KV reservation. Sarathi invariant in code.
Paper: Orca / Sarathi / DistServe reread against your scheduler.
Artifact: a service that accepts two fake clients and does not stall decode. Trace.

Month 12 — Multi-GPU or an honest single-GPU + calculated multi-GPU

Deep: TP=2 Linear + overlap study or, if one GPU, a correct simulator plus a rented 2-GPU weekend (operating-system.md compute).
Paper: Megatron 2021; NCCL "demystifying" if you want; Ring vs Ulysses volumes.
Artifact: numerical match to 1-GPU + a comm table. MFU if you actually train.

Month 13 — Disagg + runtime thinking

Deep: DistServe idea in a design doc and a tiny split (CPU prefill queue / GPU decode loop is enough). Staff+ Inference Runtime flavor: accelerator-agnostic API sketch (GPU/TPU/Trainium).
Paper: prompt caching; context engineering; postmortem (third read — sampling on the wrong platform).
Artifact: 8–15 page inference design: 7B and 70B-class, paging, batching, streaming, tail, one autoscaling signal, one failure, three platforms.

Month 14 — P3 exit + systems interview gym starts

Deep: close holes from 11–13. One OSS-shaped contribution or a standalone tool someone else can run (not a typo PR).
Drills: 05-interview/drills.md implementation 3×/week.
P3 exit review:
Batcher runs
TP (real or honestly simulated) costed and, if 2 GPU, overlapped
45 min design you can do cold
Design doc exists
Artifact: design doc + batcher demo + drill log.

Months 15–16 — P4 post-training + culture

Roadmap P4: CAI, RLHF, DPO, RLAIF as procedures; RSP read; interp enough to pair; Why-Anthropic that is specific.

Point: 04-llm-theory/curriculum.md §E + §G; 01-prior-art/corpus.md alignment + safety; 05-interview/full-prep.md Round G failure modes (use as a checklist, not a first-30-days replacement).

Month 15 — Procedures, not slogans

Deep: derive DPO; implement the loss (~30 lines) on a tiny pair set or a careful notebook against a known impl. Recite CAI two stages weekly.
Read: Bai HH-RLHF; Bai CAI; InstructGPT; Rafailov DPO; Gao overopt; 2026 constitution (again).
Interp: Elhage framework + induction heads + superposition toy (04-llm-theory §G labs G1–G2). Scaling Monosemanticity is the MUST; you need the idea, not an SAE on Sonnet.
Artifact: 2–4k word post-training writeup a researcher would not be embarrassed to read. One disagreement with CAI or RSP v3.4.

Month 16 — Culture that survives a second "why"

Deep: RSP v3.4 + Aug 2026 Risk Report; Sleeper Agents; Alignment faking; Core Views (BG, 2h). Write positions, argue the opposite, update.
Why Anthropic v3 (200–400 words) after the writeup. Include three-platform + postmortem + a first-six-months kernel.
Values mocks: 4× recorded, one with a human who will argue.
P4 exit review:
CAI as procedure cold
Constitution priority order cold
RSP v3.4 one concrete change
DPO derived
Why Anthropic is not portable to another lab
One interp object (induction head or SAE loss)
Artifact: essay + writeup + mock notes.

Months 17–18 — P5 interview gym + apply-gate

Roadmap P5: take-home as low as you can honestly get (aim < 1,487 for the email door); daily timed implementation; 10 recorded system-design drills; apply.

Point: 05-interview/drills.md, 05-interview/full-prep.md loop map (rounds A–H), Hume, original take-home.

Month 17 — Gym

Weekly	What
3×	90 min incremental systems (CodeSignal energy). No LLM in the timed window.
3×	25 min whiteboard derivations (drills list). Camera.
1–2×	45 min inference design.
Fri	Take-home from scratch in 2 hours, then unlimited-time hunt.
Sat	Artifact polish or a live mock (crawler / profiler / messy repo).

Target by end of month 17: take-home number you will put next to your name; 6/10 designs recorded; OA folder with 20+ timed builds.

Month 18 — Apply-gate month

Run the gate. Apply only if all are true (roadmap "ready to apply"):

Take-home is a number you will sign. If not under ~2,000 you are not a PE candidate. If under 1,487, email performance-recruiting@anthropic.com with code + resume.
Public kernel artifact with numbers and a postmortem.
Derive attention, cost KV, explain FA IO on a blank board.
Design inference: prefill/decode, paging, batching, streaming, tail, one autoscaling signal ≠ GPU util.
90 min multi-part systems problem without panic.
CAI as a training procedure + something non-generic about this lab (constitution / RSP v3.4 / three platforms / postmortem).
One-paragraph "most impressive low-level thing" + link.

If two or more fail, do not apply. The 12-month reapply clock is real. A new kernel repo with a 3× cycle win is material; a new certificate is not.

Which req: GPU PE (4926227008) if your artifact is CUDA/Triton; TPU Kernel (4720576008) if you can talk XLA/sampling/assembly and the postmortem is in your bones; Inference Systems (5224564008) if the month 13 design doc is the strongest piece. Apply as an engineer.

After you apply

Keep Friday take-home sport and the weekly artifact. The loop can take a month (team match longer). Do not go idle. Parallel processes are allowed; Anthropic is not a vow of poverty.

Monthly artifact cheat-sheet (7–18)

Month	Must exist
7	Backward or fused CE + numbers
8	Decode / paged tokens/s + roofline
9	GEMM vs library table
10	PORTFOLIO.md (3 links) — P2 exit
11	Running continuous batcher
12	TP (real or honest sim) + comm table
13	8–15 page inference design (3 platforms)
14	Design doc + drill log — P3 exit
15	2–4k post-training writeup
16	Why Anthropic v3 + values mocks — P4 exit
17	OA folder + 6 designs + take-home 2h redo
18	Apply packet or a written "not yet, here's the hole"

If you have a full-time job

Cut to 4–5 deep h/day and double the calendar (roadmap). Months 7–18 become ~18–24 calendar months. Do not pretend otherwise. Sunday stays off.