- You're already qualified for real engineering roles. 13 years SWE + ML Eng at Meta puts you at L5/L6 equivalent. Anthropic actively hires this profile — half their technical staff had no prior ML background (Anthropic careers page). Aiming low here just costs you ~$300K/yr and 5x the equity grant.
- Random non-eng roles don't compound for you. A Customer Success or Recruiter role at Anthropic gives you the badge but caps your learning and equity. The whole reason Anthropic ranks where it does on your list is learning value and equity; the wrong role kills both.
- Internal transfers from non-eng → eng are harder than people think. External eng hires get the full bar. Internal transfers from non-tech orgs face additional friction. So "get in any door, transfer later" is more expensive than going straight at engineering.
- You already hold Anthropic SPV exposure. If you can't land an eng/research role, the marginal value of also taking a non-eng Anthropic role is low — you'd be doubling concentration risk in exchange for a credential. Better to take a real eng role at Microsoft AI Redmond or AWS AI and be a serious customer/partner of Anthropic from that side.
1. Seattle-compatible roles open today (sample)
From Anthropic's Greenhouse job board (fetched today). Many ML/RE roles list SF / NYC / Seattle as the location set — Seattle is genuinely supported for Research and Engineering org, not just business roles.
| Role | Team | Location | Fit for you |
|---|---|---|---|
| Research Engineer, Production Model Post-Training | Post-Training | SF / NYC / Seattle | Strong — applied side of model improvement, ML-systems heavy |
| Research Engineer / Research Scientist, Pre-training | Pre-training | Remote-Friendly (Travel); SF / Seattle / NYC | Strong — Seattle + remote-friendly |
| Research Engineer / Research Scientist, Tokens | Tokens (tokenizer / data) | NYC / Seattle / SF | Strong — tokenizer/data work is ML systems |
| Research Engineer, Reward Models Platform | Reward Models | Remote-Friendly (Travel); SF / Seattle / NYC | Strong — RL infrastructure |
| Research Engineer, Knowledge Team | Knowledge | Remote-Friendly (Travel); SF / Seattle / NYC | Strong — retrieval / memory adjacent to continual learning |
| ML Systems Engineer, RL Engineering | RL | SF / NYC / Seattle | Very strong — direct map to your ML eng background |
| ML Systems Engineer, Research Tools | Research Tools | SF / NYC / Seattle | Very strong — internal tooling, leverage Meta infra experience |
| Performance Engineer (GPU / Inference Systems) | Performance | SF / NYC / Seattle | Strong if you have/can build GPU/kernel chops |
| Engineering Manager (GPU / Inference / Routing) | Various ML eng | SF / NYC / Seattle | Possible if you've managed; otherwise IC route |
2. The interview process — what's verified
Cross-referenced from interviewing.io, Dataford, InterviewQuery, and Anthropic's own careers page. Confidence is medium-high on the structure; specific question samples vary by role.
| Stage | Format | What's tested | Failure rate (anecdotal) |
|---|---|---|---|
| 1. Recruiter screen | 30 min, video | Mission alignment, why Anthropic, basic background fit | Non-trivial — candidates do fail here |
| 2. Coding challenge | CodeSignal 90 min (or 60 min live) | 4 progressively harder tasks, Python, "first-principles" implementation | High filter (speed-gated) |
| 3a. HM call | 1 hr | Deep walkthrough of a past project, sometimes multi-language code review | Medium |
| 3b. Coding (live) | 1 hr, CodeSignal/Colab, Python | Concurrency, mutation, parsing, data structures from scratch | Medium |
| 3c. System Design | 1 hr, shared Google Doc | LLM-serving systems: batching, queueing, GPU utilization, rate limits | Medium |
| 3d. Role-specific coding | 1 hr | Domain depth (RL infra, performance, etc.) | Medium |
| 3e. Values interview | 1 hr, non-technical interviewer | Reflective questions, ethics hypotheticals, motivation | Highest reported |
3. Stage-by-stage — what to actually do
Stage 1: Recruiter screen (30 min)
Don't treat as a formality. They specifically test (a) whether you've thought about Anthropic's mission, (b) whether you can articulate why Anthropic vs OpenAI/DeepMind, (c) basic background fit.
Do say: a specific reason tied to safety mission, a specific Anthropic paper or product you've engaged with, a credible bridge from your Meta ML work.
Don't say: competing offers, salary expectations, or "I'll take any role" (this is read as low conviction).
Stage 2: Coding challenge — CodeSignal General Coding Framework
Standard format: 4 levels, 90 min total. Level 1 trivial, levels 2-4 build on the same problem with added complexity. The famous example is "implement a bank with transaction types" — multi-level: basic transactions → time-windowed queries → merge accounts → cashback rules. Look up the General Coding Framework practice tasks.
What grades well: partial credit for higher levels with clean code beats brute-forcing level 1 with messy logic. Plan structures so level-3 changes are local. Run tests; the harness gives feedback.
Prep: CodeSignal General Coding Framework sample (free practice exists). Two sessions of timed runs is enough; this is a known format.
Stage 3a: Hiring Manager call (1 hr)
You walk through one past project in depth — design, tradeoffs, what you'd change. They probe technical decisions.
Prep: pre-write a 3-minute summary + a 10-minute deep dive on one project (probably the Calendo AI assistant or a Meta ML systems project). Be ready for "why this and not X?" on every choice. Bring a story where you changed your mind based on data.
Stage 3b/d: Live coding (1-2 rounds, 1 hr each, Python)
Not LeetCode-style. Themes that recur: concurrency / multithreading, parsing, hash-map-heavy data structures, building things from scratch. You will not be asked to invert a binary tree.
Prep:
- Build a
RateLimiterin Python (sliding window, token bucket) from scratch, no libraries. - Build a
ThreadSafeLRUCachewiththreading.Lock; then withasyncio. - Build a small stream parser (state machine) for a log format.
- Build an in-memory bank ledger with the 4-level extension pattern.
- Practice talking through tradeoffs aloud while coding — they grade reasoning, not just final code.
Stage 3c: System Design (1 hr, Google Doc)
Questions are real Anthropic problems, not generic. Examples reported: Design an API for serving LLMs efficiently (batching, queueing, GPU utilization, KV cache), Design a Claude chat service, Design a system for multi-turn agentic conversation.
Prep:
- Read the vLLM paper (PagedAttention) + the SGLang design — both are the actual production patterns.
- Understand continuous batching, KV cache management, speculative decoding, prefix caching.
- Practice drawing: tokenizer → router → batcher → GPU pool → KV store → streaming response, with capacity math.
- Know one approach to rate limiting at API gateway level (Redis token buckets, leaky bucket per tenant).
- Be ready to discuss: what if a request needs 100K context, how does it affect batching?
Stage 3e: Values interview — the highest-failure round
Described by candidates as "closer to a therapy session than a job interview." Non-technical interviewer. They probe how you actually think about AI risk, not whether you say the right buzzwords. They reward thoughtful skepticism, not enthusiasm or alignment-signaling.
Common question shapes:
- Describe a time your values were tested at work. What did you do? How did you feel?
- Describe a belief you changed your mind on. What evidence shifted you?
- Hypothetical: a customer wants Claude to do X (mildly dual-use). How do you reason about it?
- Why Anthropic specifically? What's the distinction vs. OpenAI?
- What do you think Anthropic is wrong about?
Prep — required reading:
- Dario Amodei — "Machines of Loving Grace" (2024 essay)
- Anthropic — "Core Views on AI Safety: When, Why, What, and How"
- "Responsible Scaling Policy" (the RSP document)
- Dario on Dwarkesh Patel podcast (most recent 2 appearances)
- "Constitutional AI" paper (skim, not deep)
What to actually do: for each piece of reading, write down (a) one point you agree with and why, (b) one point you genuinely disagree with or are uncertain about, and (c) one point you think is missing. The values interview rewards depth and pushback, not recitation.
4. Compensation reality check
| Level (inferred) | Median total comp | Comment |
|---|---|---|
| Senior SWE (~L5/IC5) | ~$563K | levels.fyi median; cash + equity blended |
| Lead SWE (~L6/Staff) | ~$785K | levels.fyi median |
| SWE Manager | ~$398K | Possibly underreported sample |
| All-Anthropic median | ~$420K | Across all roles |
5. The referral question
Multiple sources confirm referrals can skip the coding challenge and meaningfully accelerate the recruiter→onsite step. This is the single highest-leverage thing you can do before applying.
Who to ask
- Anyone you know at Meta who left for Anthropic. Common path. LinkedIn search: "Meta" → "Anthropic" in employer history. Filter to engineering. Reach out with a specific role link, not "hi can you refer me."
- Researchers you've engaged with publicly — if you've commented on their work, contributed to OSS they care about, or shared substantive analysis. Cold but warmer than cold.
- Investors / Anthropic-adjacent people — your SPV had an issuer; if you know any investor on the cap table, an intro from them is gold.
What to send
Three things in one short message: (1) the specific role URL, (2) a 1-paragraph why-you-for-this-role-specifically, (3) a link to one concrete artifact (GitHub repo, paper, blog post, or your ravikant.dev). Not a résumé attachment in the first message.
6. The Fellows Program — alternative path
Anthropic Fellows: 4-month research fellowship, ~40% conversion to full-time Anthropic (~25-50% range across cohorts), $3,850/week stipend + $15K/mo compute, remote-friendly within US/UK/Canada, Berkeley + London shared workspaces.
Honest take for you: probably worse than going straight at an eng role. Reasons:
- You're already established (13 yrs SWE). Fellowship is more designed as a credibility-building program for people transitioning from non-ML backgrounds.
- Conversion isn't guaranteed — 40% is decent but means 60% don't get the offer.
- Stipend ($200K/yr equivalent) is lower than your current comp, and equity doesn't accrue during the fellowship.
- Time cost: 4 months in a research project that may or may not convert vs. ~6 weeks for the standard interview pipeline.
Use Fellows as plan B if you fail the standard pipeline twice, not plan A.
7. 8-12 week prep plan
- Read all 4 values-interview required pieces (above). Write your 3-point response per piece.
- Pick the 1-2 specific roles you'll target. Read the JD line by line.
- Update your résumé to lead with ML/LLM work; demote pure SWE history.
- Identify 5-10 Anthropic employees to reach out to for referral. Send 2-3 first messages.
- 2 timed CodeSignal General Coding Framework practice rounds.
- Build the rate limiter, threadsafe LRU, log parser, bank ledger from scratch in Python. Public GitHub repo with tests.
- Solve 10 concurrency problems in Python (threading, asyncio, multiprocessing).
- Read vLLM PagedAttention paper + SGLang design doc. Take handwritten notes.
- Build a toy continuous batcher for a tiny LLM in Python (even with a small open model, just to internalize the loop). Public repo.
- Write a blog post titled "How I'd design the Claude API serving layer." Post to your site + LinkedIn.
- Do 3 mock system-design interviews focused on LLM-serving scenarios.
- If targeting RL Engineering: implement GRPO on a toy environment, write it up.
- If targeting Tokens: implement a BPE tokenizer from scratch, benchmark vs tiktoken.
- If targeting Performance: write a fused-attention Triton kernel; benchmark vs PyTorch eager.
- Do 2-3 full mock onsites (HM walkthrough + 2 coding + system design + values), ideally with someone in the field.
- Apply via referral if you got one; otherwise direct.
- Onsite is typically a single day with breaks; ask for the time slot you function best in.
- If you get the offer: ask explicitly about Seattle work-from. Most Seattle-listed roles are genuinely Seattle.
- If you don't: ask the recruiter the one piece of feedback they'll share (they often won't share much, but ask). Reapply window is 12 months.
8. Mistakes to avoid (from candidate reports)
- Treating the recruiter screen as a formality. Real screen, real fail rate.
- Over-optimizing CodeSignal level 1 at the cost of levels 2-3. Partial higher-level credit beats perfect lower-level.
- Generic "I want to work on AGI safety" answers in the values interview. They've heard this 10,000 times. Specific pushback or specific disagreement is signal.
- Forgetting Anthropic is a B Corp. They explain this in the recruiter call; reflect it back somewhere.
- Ignoring concurrency. It comes up across multiple rounds; assume it'll appear.
- Showing up with no concrete artifact. A GitHub repo or blog post tied to LLMs / RL / inference is a force multiplier.
9. What I'd actually do if I were you
- Pick one role this week. Recommend: ML Systems Engineer, RL Engineering or Research Engineer, Production Model Post-Training (both Seattle-listed). Don't apply to 5 at once — Anthropic notices.
- Spend 1 week on referral outreach before any prep. Highest leverage step.
- Block 8 weeks of evening prep while still at Meta. Don't quit before signed offer.
- Publish 1 substantive artifact mid-prep — vLLM-style toy implementation + blog post. This doubles as referral material.
- If you fail the onsite: try one different role in 3-6 months (different team, same Anthropic), not the same one. Different interview panel.
- If you fail twice: pivot to Microsoft AI Redmond or Amazon AGI for 1-2 years, build the ML systems track record, reapply.
Sources
- interviewing.io — Anthropic interview questions
- Dataford — Anthropic SWE interview guide
- InterviewQuery — Anthropic interview guide
- Anthropic — Careers (values, hiring criteria)
- Anthropic — Greenhouse open roles
- levels.fyi — Anthropic comp data
- Anthropic Alignment — Fellows Program 2026
- Blind — Anthropic interview discussions
- refer.me — Anthropic referral marketplace (use cautiously, prefer warm)