Month 6 — Serving papers, prompt cache, sampling postmortem ×2, RSP + constitution, Why Anthropic, take-home path, public artifact
Roadmap late P2 / early P3. This month is the first time the company-shaped reading is Deep 1, not leftover evenings. Kernels do not stop.
Month goals
- Serving papers: Orca, Sarathi-Serve, DistServe. Whiteboard each: what problem, what invariant, what kernel implication.
- Prompt caching implemented mentally (and a tiny simulator): prefix key, hit/miss, invalidation, TTFT vs decode, what KV pages you keep.
- Sampling postmortem twice (read, write, re-read): https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues
- RSP v3.4 + 2026 constitution. Specific, dated.
- Why Anthropic draft (200–400 words) that could not be pasted onto OpenAI.
- Take-home sport aiming at the < 1,487 path (you may not hit it; you must see the next 2×).
- One public artifact good enough to put at the top of a resume (even if you will not apply for a year).
Week 21 — Serving papers
| Day | Deep 1 | Deep 2 | Paper / book | Fri / Sat |
|---|---|---|---|---|
| Mon | Orca (OSDI '22). Iteration-level batching. Draw vs static batching. | Tiny continuous batcher (drill #3 in 05-interview/drills.md). | https://www.usenix.org/conference/osdi22/presentation/yu | |
| Tue | Sarathi-Serve (chunked prefill). Stall-free. | Extend the batcher with a token budget + chunked prefill. | https://arxiv.org/abs/2403.02310 | |
| Wed | Debug the batcher. | Take-home. | None. | |
| Thu | DistServe (prefill/decode disaggregation). When it wins. | Inference design #5: disagg vs colocated. | https://arxiv.org/abs/2401.09670 | |
| Fri | Take-home sport. Start the 1487-path list: remaining bottlenecks, ranked. | |||
| Sat | Long build: notes/serving-three-papers.md — one page each, kernel implications in bullets. | |||
| Sun | Off-ish. |
Week 22 — Prompt cache + postmortem pass 1
| Day | Deep 1 | Deep 2 | Paper / book | Fri / Sat |
|---|---|---|---|---|
| Mon | Prompt caching post (Aug 2024). https://www.anthropic.com/news/prompt-caching | Simulator: system+tools prefix frozen; user turns miss. Hit rate vs TTFT. | Context engineering essay (SHOULD): https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents | |
| Tue | Postmortem pass 1: timeline, three bugs, platforms. Hand notes only. | Write notes/postmortem-p1.md: what you would have monitored. | Postmortem. | |
| Wed | Debug cache simulator. | Take-home. | None. | |
| Thu | Many-shot jailbreaking (Anil Apr 2024) as a context-length / product paper, not an attack recipe. | Design #6: cache keying + tenancy + invalidation on weights publish. | https://www.anthropic.com/research/many-shot-jailbreaking | |
| Fri | Take-home. | |||
| Sat | Long build: cache simulator in the public repo. | |||
| Sun | Off-ish. |
Week 23 — Postmortem pass 2 + RSP + constitution
| Day | Deep 1 | Deep 2 | Paper / book | Fri / Sat |
|---|---|---|---|---|
| Mon | Postmortem pass 2: XLA top-k deep dive. bf16 vs fp32, xla_allow_excess_precision, approximate vs exact top-k. | Recite the three bugs to a camera, 12 min. | Postmortem again. | |
| Tue | RSP v3.4 current page + changelog (Jul 8, 2026 items). https://www.anthropic.com/responsible-scaling-policy | Write: what changed from "ASL as a model label" to thresholds + Risk Reports, in your words. | Aug 2026 Risk Report (SHOULD) if you can get the redacted PDF. | |
| Wed | Constitution (Jan 22, 2026). Priority order + hard constraints idea. | Culture drill: one real disagreement. | https://www.anthropic.com/news/claude-new-constitution | |
| Thu | Why Anthropic v1 (200–400 words). Specific. Include a platform (TPU/Trainium/NVIDIA) and a first-six-months artifact. | Take-home (shorter). | Core Views, 2h BG if never read. | |
| Fri | Take-home sport. Hunt 1487-path item #1. | |||
| Sat | Long build: Why Anthropic v2 after sleeping on v1. Throw v1 if generic. | |||
| Sun | Off-ish. |
Week 24 — Public artifact + close first 6 months
| Day | Deep 1 | Deep 2 | Paper / book | Fri / Sat |
|---|---|---|---|---|
| Mon | Pick one public artifact to make unmistakable: FA1/FA2 writeup or pager or GEMM worklog. Polish numbers. | Greenhouse paragraph (one para + URL). | None. | |
| Tue | Blank-page exam (6h energy, split): attention+FA, KV, TP, CAI, postmortem three bugs, RSP one change, constitution order. | Fix whatever failed. | None. | |
| Wed | Debug / polish the artifact. | Take-home. | None. | |
| Thu | Design #7 (full 45 min, as if onsite). | Timed 90 min incremental system. | None. | |
| Fri | Take-home. Write the honest number and the 1487-path (what is left). Email door is not this month unless you are actually under 1487. | |||
| Sat | Month 6 / half-year artifact. Tag month-6. | |||
| Sun | Off. Plan months 7–12 from months-07-18.md. |
Month 6 exit criteria
Artifact
The half-year repo should contain: take-home changelog, Triton FA, pager, TP notes, serving notes, postmortem notes, Why Anthropic, Greenhouse paragraph. If a stranger cannot find a number in 60 seconds, the artifact failed.
You are not ready to apply. Roadmap apply-gate is 00-the-job/roadmap.md. Burning the 12-month clock now is how people stay out.