Must-read prior art (kernel / PE track)
Status: short MUST stack, 2026-08-18. Full annotated corpus with hours and URLs: 01-prior-art/corpus.md. This file is the interview-week subset, not a substitute.
Priority: MUST = you cannot interview without it. SHOULD = you will be a worse engineer without it. BG = background, do not hide here.
Anthropic, about the work itself
| Pri | Item | Why |
|---|---|---|
| MUST | Designing AI-resistant technical evaluations (Hume, 2026) | The hiring instrument. Read twice. Then do the repo. |
| MUST | https://github.com/anthropics/original_performance_takehome | The job. |
| MUST | TPU Kernel Engineer | Sampling, low-precision, custom collectives, debug at assembly. $280k–$850k. Same band as GPU PE. |
| MUST | Performance Engineer, GPU and Performance Engineer | The scoreboard. Representative projects are the curriculum. |
| MUST | Careers / how we hire | Apply-as-engineer, what they put on the resume, 25% office, no internships. |
| SHOULD | Performance Engineer, Inference Systems | Correctness-as-performance. Cross-layer tracing. |
| SHOULD | Staff+ Inference Runtime | Multi-accelerator reality (GPU/TPU/Trainium), Rust runtime. |
| SHOULD | London Inference Staff | Same org, £325k–£390k. |
| BG | AI partnership guidelines for candidates (linked from reqs) | When you may use Claude in the process. |
Anthropic research a PE must be literate in
You will not be on the interpretability team. You will sit next to them. "I just do kernels" is a fail in the culture interview.
| Pri | Item | Why |
|---|---|---|
| MUST | Bai et al., 2022. Constitutional AI. https://arxiv.org/abs/2212.08073 | The company's namesake method. Procedure, not slogan. |
| MUST | Bai et al., 2022 HH-RLHF. https://arxiv.org/abs/2204.05862 | Post-training stack they built the company on. |
| MUST | Christiano et al., 2017. Deep RL from Human Preferences. + Askell et al. 2021 HH | Lineage. |
| MUST (interview) | Claude's new constitution (Jan 22, 2026) | Living constitution. Priority order: broadly safe → ethical → guidelines → helpful. |
| MUST | RSP v3.4 effective Jul 8, 2026 (page updated Aug 14, 2026) | Current policy, not the 2023 blog memory. |
| MUST | A postmortem of three recent issues (Sep 17, 2025) | bf16 / XLA / approximate top-k. THE public kernel/numerics look. Three platforms. |
| MUST | Many-shot Jailbreaking (Anil, Apr 2024) | Long-context is a kernel + product problem. |
| MUST | Prompt caching (Aug 2024) | Prefix KV reuse. Implement mentally. |
| SHOULD | Kaplan et al. 2020 scaling laws; Hoffmann et al. 2022 Chinchilla | Why kernels unlock capability. |
| SHOULD | Elhage / Olsson transformer-circuits / induction heads (Anthropic, 2021–2022) | Residual stream as a working memory. Helps you instrument models. |
| MUST | Templeton et al. 2024 Scaling Monosemanticity | What interp people will ask you to make fast. |
| SHOULD | Constitutional Classifiers++ (~Jan 2026, arXiv:2601.04603) | Production safety has a compute budget. |
| SHOULD | August 2026 Risk Report (linked from the RSP page) | How they talk about catastrophic risk now. |
| BG | Natural Language Autoencoders (2026); Global workspace (2026) | Literacy, not your day job. |
Also read the "How we're different" paragraph on every req: they consider their lineage GPT-3, circuit interpretability, multimodal neurons, scaling laws, AI & Compute, Concrete Problems in AI Safety, Learning from Human Preferences. Those six names are a reading list.
The kernel canon (not Anthropic, table stakes)
| Pri | Item | Why |
|---|---|---|
| MUST | Williams, Waterman, Patterson. Roofline (CACM / tech report 2008–2009) | The diagram of your career. |
| MUST | Dao et al. 2022 FlashAttention; Dao 2023 FA2; FA3 (2024) is what "FlashAttention" means in 2026 | Attention as an IO problem. FA1/2 = algorithm you implement. FA3 = Hopper conversation (TMA/WGMMA/FP8). |
| MUST | CUDA C++ Programming Guide (current) + PMPP 4e MUST; PMPP 5e (Jun 2026) SHOULD | The craft. |
| MUST | Triton docs + Tillet et al. 2019 | Daily driver language. |
| SHOULD | CUTLASS / CuTe docs (NVIDIA) | How serious GEMM is structured. |
| SHOULD | Kwon et al. 2023 vLLM / PagedAttention | Decode serving. |
| SHOULD | Shoeybi et al. Megatron-LM; Rajbhandari ZeRO; PyTorch FSDP | Parallelism zoo. |
| SHOULD | Pope et al. 2022 (Google) efficiently scaling transformers (inference) | Prefill/decode, batching, multi-query. |
| SHOULD | DeepSeek MLA / MoE papers (2024–2025) | Architecture that changes kernels. |
| SHOULD | NVIDIA PTX / Hopper (H100) / Blackwell whitepapers: TMA, WGMMA, tensor memory | Hardware you are targeting. |
| BG | ThunderKittens, TileLang, Helion, Transformer Engine | 2025–2026 kernel DSLs. Read after you can write Triton without them. |
Foundations (do not skip)
| Pri | Item | Why |
|---|---|---|
| MUST | CSAPP, Bryant & O'Hallaron, 3e, labs | Systems taste. Non-negotiable. |
| MUST | A real linear algebra (Axler or Trefethen & Bau) | Attention is linear algebra with a softmax. |
| MUST | IEEE-754 + a numerics chapter (Higham or Trefethen Lectures on NLA) | Softmax, fp8, accumulation width. |
| SHOULD | Hennessy & Patterson, Computer Architecture (selected: memory, ILP, GPU chapters) | Vocabulary with hardware vendors. |
| SHOULD | Cover & Thomas, Elements of Information Theory (selected) | Cross-entropy, KL, bits. |
| SHOULD | GPU Mode lectures (YouTube, 2023–2026) | Current craft, not a textbook. |
| MUST (one) | CS149 Fall 2025 https://gfxcourses.stanford.edu/cs149/fall25/ (H100 kernels + Trainium2). Alt: CMU 15-418. | Parallel course that matches three-platform reality. Do the labs. |
| MUST | CMU 10-414/714 Deep Learning Systems https://dlsyscourse.org/ | Compilers + autodiff + systems for the stack. |
| BG | Dragon book or Cooper/Torczon, Engineering a Compiler (selected) | XLA / Triton / Neuron mental model. |
Repos to read like literature (clone, do not pip-install-and-pray)
- anthropics/original_performance_takehome
- Dao-AILab/flash-attention
- triton-lang/triton
- NVIDIA/cutlass
- vllm-project/vllm
- NVIDIA/TransformerEngine
- karpathy/nanoGPT (for the tiny trainer, then discard training-as-identity)
- pytorch/pytorch (Inductor / generated Triton — later)
- gpu-mode/lectures (if still the canonical org)
- A Megatron-LM or Megatron-Bridge tree current in 2026
Full annotated hours, URLs, jobs table, kernel canon, and honesty notes: 01-prior-art/corpus.md.