Prior-art corpus (kernel / PE track)
Sourced 2026-08-18 from primary pages (Anthropic news/research/engineering, arXiv, Greenhouse, course sites). Flags:
- MUST — you cannot interview without it
- SHOULD — you will be a worse engineer without it
- BG — background; do not hide here
- MUST-DO — not a paper; you run it
Hours are deep hours (pen + compiler + profiler), not calendar time. A 12-week paper-read plan is not becoming the engineer; that is literacy. The real path is the 16–24 month plan in 00-the-job/roadmap.md. Sequenced curricula (do not re-list here): 02-foundations/curriculum.md, 03-kernels/curriculum.md, 04-llm-theory/curriculum.md.
Do not invent papers. If an item has no URL below, none was verified on 2026-08-18.
Anthropic alignment
| Pri | Hours | Item | Why |
|---|---|---|---|
| MUST | 8–12h | Bai et al., 2022. Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. https://arxiv.org/abs/2204.05862 · Anthropic: https://www.anthropic.com/research/training-a-helpful-and-harmless-assistant-with-reinforcement-learning-from-human-feedback | HH-RLHF. The post-training stack they built the company on. Procedure, data, HHH tensions — not a slogan. |
| MUST | 10–14h | Bai et al., 2022. Constitutional AI: Harmlessness from AI Feedback. https://arxiv.org/abs/2212.08073 | Namesake method. SL-CAI critique→revise, then RLAIF. Recite as a flowchart. |
| MUST (interview) | 4–6h | Claude's new constitution, Jan 22, 2026. https://www.anthropic.com/news/claude-new-constitution | Living constitution used in training. Helpful / guidelines / ethics / broadly-safe priority order. They will ask. |
| SHOULD | 6–8h | Constitutional Classifiers++, ~Jan 2026. https://www.anthropic.com/news/next-generation-constitutional-classifiers · arXiv:2601.04603 | Production safety has a compute budget. You may be asked to make a probe cheap. |
| SHOULD | 6–8h | Hubinger et al., Jan 2024. Sleeper Agents. https://arxiv.org/abs/2401.05566 | Deceptive alignment as an empirical object. Why "the model looks fine" is not a kernel-only answer. |
| SHOULD | 6–8h | Greenblatt et al., Dec 2024. Alignment faking. https://arxiv.org/abs/2412.14093 | Models that play along under oversight. Culture-round fuel. |
| BG | 2h | Core views on AI safety. https://www.anthropic.com/news/core-views-on-ai-safety | Company doctrine. Read once; do not camp here. |
Claude / system cards
Index of cards: https://www.anthropic.com/system-cards
| Pri | Hours | Item | Why |
|---|---|---|---|
| MUST | 6–8h | Claude 3 model card, Mar 2024. https://www.anthropic.com/news/claude-3-family · card index above | First public "how we evaluate a frontier model" they expect you to have opened. Capability + safety shape. |
| SHOULD | — | Claude Sonnet 4.5 system card, Sep 2025. https://www.anthropic.com/claude-sonnet-4-5-system-card · PDF: https://www-cdn.anthropic.com/963373e433e489a87a10c823c52a0a013e9172dd/Claude%20Sonnet%204.5%20System%20Card.pdf · news: https://www.anthropic.com/news/claude-sonnet-4-5 | Hybrid reasoning, computer use, first interp-in-the-card evals. What "aligned enough to ship" looks like in 2025. |
| SHOULD | — | Claude Opus 4.5 system card, Nov 2025. https://www.anthropic.com/claude-opus-4-5-system-card · PDF: https://www-cdn.anthropic.com/bf10f64990cfda0ba858290be7b8cc6317685f47/Claude%20Opus%204.5%20System%20Card.pdf · news: https://www.anthropic.com/news/claude-opus-4-5 | Frontier coding/agents + ASL-3 + sabotage/eval-awareness. The model the take-home was redesigned against. |
| SHOULD | — | Developing a computer use model, Oct 2024. https://www.anthropic.com/research/developing-computer-use | Tool/computer-use as a systems + safety problem. Adjacent to inference runtime work. |
Interpretability (literacy so you can sit next to them)
You will not be on the interp team. "I just do kernels" is a culture-round fail.
| Pri | Hours | Item | Why |
|---|---|---|---|
| SHOULD | 12–16h | Elhage et al., 2021. A Mathematical Framework for Transformer Circuits. https://transformer-circuits.pub/2021/framework/ | Residual stream, QK/OV. How they talk about the model you are making fast. |
| SHOULD | — | Olsson et al., 2022. In-context Learning and Induction Heads. https://transformer-circuits.pub/2022/in-context-learning-and-induction-heads · arXiv:2209.11895 https://arxiv.org/abs/2209.11895 | First circuit you must be able to draw. |
| SHOULD | — | Elhage et al., 2022. Toy Models of Superposition. https://transformer-circuits.pub/2022/toy_model | Why neurons are the wrong basis. |
| SHOULD | — | Bricken et al., 2023. Towards Monosemanticity. https://transformer-circuits.pub/2023/monosemantic-features | Dictionary learning / SAE idea. |
| MUST | 12–16h | Templeton et al., 2024. Scaling Monosemanticity. https://transformer-circuits.pub/2024/scaling-monosemanticity/index.html | SAEs on Sonnet. What interp people will ask you to make fast (hooks, dumps, sparse matmuls). |
| SHOULD | — | Circuit Tracing / On the Biology of a Large Language Model, 2025. https://transformer-circuits.pub/2025/attribution-graphs/biology.html · methods: https://transformer-circuits.pub/2025/attribution-graphs/methods.html | Attribution graphs at production scale. Systems-shaped interp. |
Inference / serving (Anthropic-authored)
| Pri | Hours | Item | Why |
|---|---|---|---|
| MUST | — | Anil et al., Apr 2024. Many-shot Jailbreaking. https://www.anthropic.com/research/many-shot-jailbreaking · arXiv:2404.01833 https://arxiv.org/abs/2404.01833 | Long-context is a product and a kernel problem. ICL attacks scale with window. |
| MUST | — | Prompt caching, Aug 2024. https://www.anthropic.com/news/prompt-caching | Prefix KV reuse as a serving primitive. You must be able to implement the idea mentally (cache key, invalidation, TTFT). |
| SHOULD | — | Effective context engineering for AI agents, 2025. https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents | Attention budget, compaction, note-taking, sub-agents. Why "just grow context" is not a serving plan. |
| MUST | — | A postmortem of three recent issues, Sep 17, 2025. https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues | THE public kernel/numerics look. Three overlapping bugs: context-window routing, TPU output corruption, approximate top-k XLA:TPU miscompilation (bf16 vs fp32 / xla_allow_excess_precision). They serve Claude on TPU + Trainium + NVIDIA. Sampling correctness is performance. Read twice. |
Safety / policy
| Pri | Hours | Item | Why |
|---|---|---|---|
| MUST | — | Responsible Scaling Policy, current v3.4, effective Jul 8, 2026, page updated Aug 14, 2026. https://www.anthropic.com/responsible-scaling-policy | They will ask what changed (capability thresholds, Risk Reports, unredacted-share rule). Read the current PDF, not a 2023 blog memory. |
| SHOULD | — | August 2026 Risk Report (redacted), posted Aug 14, 2026, coverage through Jul 15. Linked from the RSP page. | How they talk about catastrophic risk now. Culture-round specific. |
Stack / hiring instruments
| Pri | Hours | Item | Why |
|---|---|---|---|
| MUST | — | Expanding our use of Google Cloud TPUs and Services, Oct 23, 2025. https://www.anthropic.com/news/expanding-our-use-of-google-cloud-tpus-and-services | Three platforms, stated: Google TPUs, Amazon Trainium, NVIDIA GPUs. Up to ~1M TPUs, >1 GW in 2026. The job is accelerator-agnostic even when the req says CUDA. |
| SHOULD | — | Google / Broadcom, 2026 | TPU supply and custom-silicon reporting around Google↔Broadcom (TPU design partner) and Anthropic's multi-chip / in-house silicon hiring. No single paper; track official Anthropic + Google Cloud notes, not Twitter. |
| MUST | — | Hume, Jan 21, 2026. Designing AI-resistant technical evaluations. https://www.anthropic.com/engineering/AI-resistant-technical-evaluations | Primary interview signal. Why the take-home looks like a weird ISA. |
| MUST-DO | — | https://github.com/anthropics/original_performance_takehome | The published job interview. Baseline 147,734 cycles. Email door < 1,487. |
Jobs (open as of mid-2026)
Apply as an engineer. Comp from Greenhouse unless noted. Levels.fyi snapshot ~Aug 16, 2026 is secondary (Senior SWE median ~$591k TC, Lead ~$773k, Staff ~$1.25M — noisy).
| Role | ID / link | Comp (posted) | What the req actually says |
|---|---|---|---|
| TPU Kernel Engineer | https://job-boards.greenhouse.io/anthropic/jobs/4720576008 | $280k–$850k | Sampling, low-precision kernels, custom collectives, debug at assembly. The TPU-shaped twin of GPU PE. |
| Performance Engineer, GPU | https://job-boards.greenhouse.io/anthropic/jobs/4926227008 | $280k–$850k | CUDA/Triton/CUTLASS, FlashAttention-class, tensor cores, FP8, NCCL |
| Performance Engineer | https://job-boards.greenhouse.io/anthropic/jobs/4020350008 | $280k–$850k | Sampling, low-precision, load-balancing, fault-tolerant distributed, kernel-level network debug |
| Performance Engineer, Inference Systems | https://job-boards.greenhouse.io/anthropic/jobs/5224564008 | posted | Cross-layer kernels↔servers↔routing. Correctness is a performance bug. |
| Staff+ SWE, Inference Runtime | https://job-boards.greenhouse.io/anthropic/jobs/5257650008 | $405k–$485k (third-party listing) | Accelerator-agnostic runtime (GPU + TPU + Trainium). Rust + Python. |
| London Inference Staff | https://job-boards.greenhouse.io/anthropic/jobs/5097742008 | £325k–£390k | Same org, London. |
| Engineering Manager, GPU | Greenhouse / careers | — | Same org, not the IC path |
| Pre-training Distributed Systems | careers | — | Megatron-shaped: TP/PP/DP/SP/CP, fault tolerance |
| ML Networking | careers | — | NCCL / collectives / topology |
| Performance, RL | careers | — | RL sampling, RM forwards — inference+train |
| Hardware / Silicon | careers | — | Co-design with vendors; 2026 in-house silicon hiring |
| Fellows (ML Systems & Performance) | careers / fellows | stipend + compute | Closest external on-ramp |
Interview signal: Hume take-home is primary. Blogs (IGotAnOffer, Aceoffer, Glassdoor) are secondary and noisy. Comp: Greenhouse first, Levels.fyi ~Aug 16, 2026 second.
Kernel canon (not Anthropic; table stakes)
URLs are the canonical papers/pages. Hours live in 03-kernels/curriculum.md. FA3 is what "FlashAttention" means in 2026 (warp-specialized TMA + WGMMA, GEMM/softmax overlap, FP8 block quant). FA1/2 are the algorithm you implement first.
| Item | URL / id | Why |
|---|---|---|
| FlashAttention (FA1), Dao et al. 2022 | https://arxiv.org/abs/2205.14135 | IO-aware attention. Online softmax. You implement this. |
| FlashAttention-2, Dao 2023 | https://arxiv.org/abs/2307.08691 | Parallelism over Q; the Triton target. |
| FlashAttention-3, 2024 | https://arxiv.org/abs/2407.08608 · https://tridao.me/publications/flash3/flash3.pdf | 2026 meaning of FA. Hopper. |
| FlashDecoding | https://crfm.stanford.edu/2023/10/12/flashdecoding.html · https://pytorch.org/blog/flash-decoding/ | Decode-shaped attention. Different bottleneck than prefill. |
| ThunderKittens | https://arxiv.org/abs/2410.20399 · https://github.com/HazyResearch/thunderkittens | Tile DSL; FA3-class research kernels. |
| FlashMLA | https://github.com/deepseek-ai/FlashMLA | MLA decode kernel. Architecture that changes the kernel. |
| Orca (OSDI '22) | https://www.usenix.org/conference/osdi22/presentation/yu · PDF: https://www.usenix.org/system/files/osdi22-yu.pdf | Iteration-level / continuous batching. |
| PagedAttention / vLLM, Kwon 2023 | https://arxiv.org/abs/2309.06180 | KV as virtual memory. |
| Sarathi-Serve | https://arxiv.org/abs/2403.02310 | Chunked prefill + stall-free batching. |
| DistServe | https://arxiv.org/abs/2401.09670 | Prefill/decode disaggregation. |
| Megatron-LM 2019 | https://arxiv.org/abs/1909.08053 | Tensor parallel. |
| Megatron-LM 2021 | https://arxiv.org/abs/2104.04473 | 3D parallel on GPU clusters. |
| ZeRO, Rajbhandari 2019 | https://arxiv.org/abs/1910.02054 | Shard opt/grad/param. |
| FSDP | https://arxiv.org/abs/2304.11277 · PyTorch FSDP docs | ZeRO-3 as a PyTorch primitive. |
| Korthikanti sequence parallelism, 2022 | https://arxiv.org/abs/2205.05198 | SP + less rematerialization. Always-on with TP. |
| Ring Attention, Liu et al. 2023 | https://arxiv.org/abs/2310.01889 | Sequence/context parallel attention. |
| NVIDIA context parallelism (CP) | Megatron-Core / Transformer Engine parallelism guides | Production CP (ring / Ulysses / Llama-3-style). |
| Speculative decoding, Leviathan 2023 | https://arxiv.org/abs/2211.17192 | Exact acceptance. Verification is prefill-shaped. |
| Medusa | https://arxiv.org/abs/2401.10774 · https://github.com/FasterDecoding/Medusa | Multi-head draft. |
| EAGLE-3 | https://github.com/SafeAILab/EAGLE (EAGLE-1 https://arxiv.org/abs/2401.15077; EAGLE-3 NeurIPS 2025) | Tree attention / verify GEMM. |
| MQA, Shazeer 2019 | https://arxiv.org/abs/1911.02150 | One KV head. Decode bytes. |
| GQA, Ainslie 2023 | https://arxiv.org/abs/2305.13245 | Grouped KV. Default 2024–2026. |
| Mixtral | https://arxiv.org/abs/2401.04088 | Sparse MoE that shipped. |
| DeepSeekMoE | https://arxiv.org/abs/2401.06066 | Fine-grained experts + shared. |
| DeepSeek-V2 MLA | https://arxiv.org/abs/2405.04434 | Compressed KV. Kernel-relevant architecture. |
| DeepSeek-V3 | https://arxiv.org/abs/2412.19437 | Systems appendix is the read. |
| DeepSeek hardware reflections | arXiv:2505.09343 https://arxiv.org/abs/2505.09343 | How they think about the metal. |
| FP8, Micikevicius et al. 2022 | https://arxiv.org/abs/2209.05433 | E4M3/E5M2 contract. |
| OCP MX formats | Open Compute Project MX spec (MXFP8/MXFP4/…) | Block-scale formats the SKUs speak. |
| MXFP8 recipes | arXiv:2506.08027 https://arxiv.org/abs/2506.08027 | How you actually train/serve in MXFP8. |
| Triton, Tillet et al. 2019 | https://openai.com/index/triton/ · https://github.com/triton-lang/triton | Daily driver language. |
| CUTLASS / CuTe | https://github.com/NVIDIA/cutlass | How serious GEMM is structured. |
| CUDA Graphs | CUDA / PyTorch CUDA-graphs notes | Decode launch overhead. |
| Hopper / Blackwell | https://developer.nvidia.com/blog/nvidia-hopper-architecture-in-depth/ · https://resources.nvidia.com/en-us-blackwell-architecture | TMA, WGMMA, TMEM/tcgen05. Hardware you are targeting. |
| Anatomy of a Triton Attention Kernel | arXiv:2511.11581 https://arxiv.org/abs/2511.11581 | 2025 close read of a real Triton FA-class kernel. |
Courses / textbooks
| Pri | Item | Why |
|---|---|---|
| MUST | PMPP 4e (Hwu / Kirk / El Hajj, 2022) | The CUDA textbook. Start here if 5e is not in hand. |
| SHOULD | PMPP 5e, Jun 2026 (catalog also lists Feb 2026; take whichever printing you can buy). LLM + advanced GEMM chapters. | 2026 edition of the craft book. |
| MUST | CSAPP, Bryant & O'Hallaron, 3e + labs | Systems taste. Non-negotiable. See 02-foundations/curriculum.md module 7a. |
| SHOULD | Hennessy & Patterson, Computer Architecture: A Quantitative Approach (selected: memory, ILP, GPU/DSA) | Vocabulary with hardware vendors. |
| MUST (one parallel course) | Stanford CS149 Fall 2025 https://gfxcourses.stanford.edu/cs149/fall25/ — H100 kernels and Trainium2. Do the programming assignments. | The parallel course that matches Anthropic's three-platform reality. |
| MUST (systems-for-ML) | CMU 10-414 / 10-714 Deep Learning Systems https://dlsyscourse.org/ | Compilers + autodiff + systems for the stack you will sit in. |
| SHOULD | CS231n (vision / backprop fluency); CS224n https://web.stanford.edu/class/cs224n/ | NLP/transformer course, not a substitute for building. |
| MUST (papers, not courses) | Vaswani et al. 2017 Attention Is All You Need https://arxiv.org/abs/1706.03762 · Brown et al. 2020 GPT-3 https://arxiv.org/abs/2005.14165 · Kaplan et al. 2020 https://arxiv.org/abs/2001.08361 · Hoffmann et al. 2022 Chinchilla https://arxiv.org/abs/2203.15556 · Ouyang et al. 2022 InstructGPT https://arxiv.org/abs/2203.02155 | The lineage named on every Anthropic req ("How we're different"). |
Do one of {CS149, CMU 15-418, MIT 6.S894} as a full course — CS149 Fall 2025 is preferred here because of H100 + Trainium2. Then GPU MODE + CS336 A1/A2 (02-foundations/curriculum.md).
What they will grill
Evidence-backed from Hume + reqs + the postmortem + candidate reports (noisy):
- Roofline. Math-bound, memory-bound, or latency-bound? Prove it with bytes and FLOPs.
- Scratch / SMEM as a resource you allocate, not a magic cache. Hume's machine is the teaching version.
- SIMD + VLIW packing + hazards. The take-home is this.
- Sampling numerics. bf16 vs fp32, approximate vs exact top-k, XLA excess precision. The Sep 2025 postmortem is the exam question they already published.
- Prefill vs decode. Different kernels, different batching, different rooflines.
- KV cache bytes.
2 × n_layers × n_kv_heads × seq × d_head × bytes × batch. GQA/MLA/paging. - Continuous / chunked batching (Orca, Sarathi). A request in prefill must not stall decode.
- Quantization that does not silently change outputs. Correctness is performance (Inference Systems req + postmortem).
- Collectives. Why an all-reduce is on the critical path; RS+AG vs AR; overlap.
- Why a fusion helps (bytes moved, not "it's faster").
- FA as of 2026 = FA3 ideas, even if your artifact is FA1/FA2-shaped.
- Constitutional AI as a procedure + 2026 constitution priority order + RSP v3.4 as a living policy. What you would refuse to ship.
- Three platforms. A kernel change that is correct on H100 and wrong on TPU (XLA) is a production incident.
They will not care about: LeetCode streaks, a Llama fine-tune, LangChain, certificates, reciting transformer equations you cannot cost.
Honesty notes
- Literacy ≠ the job. Reading this corpus in 12 weeks makes you able to talk. The 16–24 month plan in
00-the-job/roadmap.md(artifacts, take-home sport, public writeups) is what makes you hireable. - Hume is primary; interview blogs are secondary. Comp from Greenhouse > Levels.fyi (~Aug 16, 2026 snapshot).
- System cards after Claude 3 are SHOULD unless you are interviewing on inference-systems / safety-adjacent staff loops — then Sonnet 4.5 + Opus 4.5 become MUST.
- Hours are estimates for a strong SWE who writes while reading. Skimming is not those hours.
- If a URL 404s, use the arXiv id or the system-cards index. Do not substitute a Twitter thread.
- Seed list and annotations:
01-prior-art/must-read.mdis the short MUST stack; this file is the full verified corpus.