AKP Anthropic Kernel Path

0/21 read · no cycles yet

How Anthropic hires kernel / performance people

Sources: anthropic.com/careers (fetched 2026-08-18), Greenhouse reqs 4720576008 (TPU Kernel Engineer) / 4926227008 / 4020350008 / 5224564008 / 5257650008 / 5097742008 (London Inference Staff), Tristan Hume "Designing AI-resistant technical evaluations" (2026-01-21), sampling postmortem (Sep 17, 2025), RSP v3.4 (effective Jul 8, 2026), Claude's new constitution (Jan 22, 2026), candidate reports 2025–2026 (treat as noisy). Full paper list: 01-prior-art/corpus.md.

The two gates

Gate A — Can you make a strange accelerator fast? The performance take-home. Originally a 4-hour (then 2-hour) optimize-a-kernel-on-a-simulated-TPU-like-machine. Now the live version is a constrained-ISA puzzle set because Claude beat the realistic one. The original is public:

https://github.com/anthropics/original_performance_takehome

You start at 147,734 cycles. The email-us bar is < 1,487. That is a ~99× speedup. People get there by SIMD, VLIW packing, scratch management, list scheduling, hash fusion, and problem-structure tricks that look like "I hit a memory-bandwidth wall" until they don't.

They designed it with these properties (Hume). Internalize them; they are also how the team works:

  • Representative of real work (or it was, until models forced them weirder)
  • High signal, wide score distribution, no single-insight gotcha
  • No required DL domain knowledge
  • Fun, fast iteration, Perfetto-style traces
  • AI assistance allowed on this take-home (unlike most Anthropic take-homes)

The TPU Kernel Engineer req (4720576008, $280k–$850k) is the same Gate A with a different ISA: sampling, low-precision, custom collectives, debug at assembly. The Sep 2025 postmortem is the public exam — bf16/XLA/approximate top-k on TPU, plus the fact they serve on TPU + Trainium + NVIDIA. A GPU-only story is incomplete.

Gate B — Can you work here? Live coding that looks like production (crawler, profiler traces, inference pipeline pieces), an inference-serving system design, a project deep-dive, and a culture/safety interview they warn may make you uncomfortable. Plus a written "Why Anthropic?" they score hard. Culture reading that is current: RSP v3.4 (effective Jul 8, 2026, page updated Aug 14, 2026, https://www.anthropic.com/responsible-scaling-policy) and Claude's new constitution (Jan 22, 2026, https://www.anthropic.com/news/claude-new-constitution) — not the 2023 principle list.

Official process (from careers page)

  • Google Meet, they accommodate timezones
  • Technical: Colab and CodeSignal. Looking things up is fine. Syntax fumbling is not.
  • They will ask what motivates you
  • No internships
  • Reapply after 12 months, or sooner if something material changed
  • They sponsor visas when they offer, not as a promise up front
  • 25% in-office default; some roles more
  • Recruiters only email from @anthropic.com

Application questions that matter

From the live Greenhouse forms:

  1. Why Anthropic? 200–400 words. They say they value this highly. Do not write "I want to work on AGI." Write a specific argument that uses their research, RSP v3.4, the 2026 constitution, their public benefit corp structure, and a concrete thing you would do in the first six months (name the platform: TPU / Trainium / NVIDIA). If you cannot write this without generic safety language, you are not ready to apply.
  1. Performance Engineer extra: "the most impressive low-level or performance thing you've done" (1 paragraph + link). This is why the artifact list exists. A from-scratch GPT does not count as low-level. A Triton FlashAttention that you profiled against a roofline, with a writeup of the bank conflicts you hit, does.
  1. AI partnership guidelines confirmation.

Reported loop (noisy, 2025–2026)

Typical shape, not a guarantee:

  1. Recruiter screen (mission, logistics, which team)
  2. CodeSignal or timed practical (90 min common). Implementation speed, multi-part problems that grow. In-memory DB style and systems problems more than leetcode puzzles.
  3. Live pair-programming (Colab / CoderPad). Practical, not "invert a binary tree."
  4. System design: inference serving, not URL shorteners. Continuous batching, KV cache, prefill vs decode, GPU memory, streaming, tail latency vs throughput, autoscaling signals that are not "GPU util."
  5. Second coding: messy real artifacts (stack-sampling profiler → traces is one reported question).
  6. Hiring manager / project deep-dive.
  7. Culture / safety. Interrogative. Have opinions you can defend, not slogans.

Research-engineer variants have added: open-ended alignment brainstorm with almost no feedback; Colab "finish this inference pipeline" with a GPU already in the notebook.

Performance-engineer variants center the take-home. Everything else is confirmation.

What they will grill (evidence-backed)

From Hume + reqs + candidate reports:

  • Roofline thinking: is this math-bound, memory-bound, or latency-bound? Prove it.
  • Scratch / shared memory as a resource you allocate, not a magic cache
  • SIMD + instruction packing + dependency hazards
  • Building your own debugging tools when the starter kit is hostile
  • Prefill vs decode, KV cache footprint (2 * n_layers * n_kv_heads * seq * d_head * bytes * batch)
  • Continuous batching / in-flight batching
  • Quantization that does not silently change outputs (correctness is performance)
  • NCCL / collectives / why an all-reduce is on the critical path
  • Why a kernel fusion helps (bytes moved, not "it's faster")
  • FlashAttention in 2026 means FA3 (TMA + WGMMA, GEMM/softmax overlap, FP8 block quant). FA1/2 are the algorithm you implement first; FA3 is the hardware conversation.
  • Sampling / XLA numerics (the postmortem). Exact vs approximate top-k. bf16 vs fp32 accumulators.
  • Three platforms: a change that is correct on H100 and wrong on TPU.
  • Safety: Constitutional AI as a procedure, the 2026 constitution priority order (broadly safe → ethical → Anthropic guidelines → helpful), RSP v3.4. What you would refuse to ship.
  • Course-shaped questions they assume you can take: CS149 Fall 2025 (H100 kernels + Trainium2, https://gfxcourses.stanford.edu/cs149/fall25/) and CMU 10-414/714 (https://dlsyscourse.org/).

What they will not care about

  • LeetCode streak
  • A fine-tune of Llama on a notebook
  • "I used LangChain"
  • Course certificates
  • Being able to recite the transformer equations without being able to implement or cost them

Culture they are selecting for (from careers + "How we're different")

  • Big science, few large bets, not a pile of small papers
  • Empirical: common with physics/biology
  • Pair programming is not optional flavor text. Both GPU and generic PE reqs mention it.
  • "Do the simple thing that works"
  • "Pick up slack even if it goes outside your job description"
  • Societal impact is a first-class interview topic
  • Communication is a hiring criterion because the research is communal

If you cannot enjoy pairing on a nasty kernel for three hours, this is the wrong company, not the wrong prep.

The email door

If your unlimited-time take-home is under 1,487 cycles (and honestly, if you can see a path under 1,363), email performance-recruiting@anthropic.com with the code and a resume. That is a real door they published. Do not email a 20k-cycle solution.