Top tier prep — Anthropic, OpenAI, Thinking Machines

Everything you need to know before each loop. Read this two days before the onsite.

Anthropic S

Loop overview (~20 days)

1. Recruiter screen (30 min) — résumé walk, motivation, "why Anthropic", role fit. Be specific about Claude / RSP / interpretability / alignment interest.
2. CodeSignal take-home — async, 90 min (sometimes 60 live). 4-level progressive spec. Pick from the "six" (see coding by company).
3. Hiring manager call (45 min) — past projects, tech depth, role match.
4. Onsite: ML round (Colab on prompting + LLM eng) + concurrency-flavored coding + ML config system design + behavioral. ~5 hours over 1–2 days.
5. Reference checks happen during the loop (up to 3).
6. Offer / debrief within ~2 weeks of onsite.

What they probe in behavioral

• Safety judgment: "Walk me through a time you slowed down a launch for safety/quality reasons."
• Calibrated confidence: don't oversell. Anthropic culture: "if you don't know, say you don't know."
• Why Anthropic: should reference Constitutional AI, Responsible Scaling Policy, model welfare, interpretability — not just "small frontier lab where I can have impact."
• Mission alignment: read Dario's Machines of Loving Grace essay. Reference if natural.

Technical content to know cold

• Constitutional AI (Bai 2022) — the paper. Both phases. Use cases.
• Anthropic Responsible Scaling Policy (RSP) — current ASL levels.
• Mechanistic interpretability — Transformer Circuits Thread, induction heads, sparse autoencoders, dictionary learning, monosemantic features.
• Claude 3 → 3.5 → 4 → 4.5 → 4.7 lineup (Opus / Sonnet / Haiku tiers); know which is the reasoning model and which the agentic.
• MCP — Model Context Protocol (their open standard).
• Agentic Claude Code / Computer use.

Comp data (Levels.fyi May 2026)

• Median SWE: $710k
• Senior: $445k–$575k
• Lead: $785k
• Staff: $920k
• T6 Tech Lead RS: $2.0–2.5m (the level above Staff IC)
• Vest: monthly after 1y cliff
• Valuation: $380B (Feb 2026 Series G)

Levers

• Mike Krieger (CPO, ex-Instagram founder) sources senior product IC directly via LinkedIn.
• Best entry: warm intro from MTS (Tom Brown / Sam McCandlish / Jared Kaplan area). Many ex-Meta in Pretraining.
• Public referral via Anthropic site is OK but slower.

OpenAI S

Loop overview (~2–8 weeks)

1. Recruiter screen — fast, 20–30 min.
2. CoderPad phone screen — 60–75 min, multi-part problem. Coding bar is binary: 2/4 ≠ pass even if everything else aces.
3. Hiring manager / team match calls.
4. 4–6 hr final loop (1–2 days): coding × 2, ML/research depth, behavioral (non-standard), system design or ML system design depending on team.
5. Reference checks (up to 3).
6. Offer.

What they probe in behavioral

• Agency & ambition: stories of high ownership, late-night ship-it moments.
• AGI views: opinions on timelines, alignment, race dynamics. Have a substantive answer.
• Research taste: "what would you build if you had unlimited compute?"
• Why OpenAI: be specific about which part of the mission — pretraining? post-training? alignment? products? Be ready to defend.

Technical content to know cold

• GPT-4o / o1 / o3 / o4 architecture insights (publicly disclosed).
• InstructGPT (Ouyang 2022) — the original RLHF paper.
• RLVR (RL with verifiable rewards) — used in o-series training.
• Spec for safety / alignment teams: Superalignment direction (post-Sutskever), preparedness framework.
• Agentic / Operator (computer use) — the responses API and Agents SDK.
• Inference systems at scale — disaggregated serving (Splitwise lineage), KV cache management, FP8 inference.

Comp data (RTF Apr 2026 + Levels.fyi May 2026 — highest band on the list)

• L5 Staff SWE: ~$875k
• L5 Staff RE/RS: ~$1.5m
• L6 Sr Staff SWE/RE: ~$1.2m–2.0m
• Vest: immediate quarterly (RSU model since Jan 2026 — switched from PPU)
• Valuation: $852B (Mar 2026 — $122B raised, world record). Coatue $2T 2030 estimate.
• Targeting ~8000 employees by EOY 2026.

Levers

• Many ex-Meta in Pretraining + Multimodal teams — easy network bridge.
• LinkedIn alumni filter on Meta → "Current company OpenAI" → message for warm intro.
• Recruiter cold-apply is OK but referral is much faster.
• Personalization / Memory team is your direct RecSys → LLM pivot.

Thinking Machines S

Loop overview (~19 days)

Less public data. Glassdoor n=14. Reportedly:

1. Recruiter screen.
2. Coding round — ex-OpenAI/Anthropic-tier bar. CoderPad-style multi-part.
3. Research / ML depth round — strong opinions on a paper, defend.
4. System design or ML system design.
5. Founder/leader chat — this matters most. Tiny team, culture is everything.

What they probe

• First-principles thinking: be ready to defend a contrarian view with rigor.
• Why TML: substantive — not "small team where I can have impact." Reference Mira's / John Schulman's / Lilian Weng's public writing.
• Reproducibility / careful science: TML is reportedly emphasizing rigorous methodology. Have stories about that.
• Comfort with ambiguity: tiny team, no infrastructure handed to you.

Technical content to know cold

• John Schulman's PPO/Trust Region work.
• Lilian Weng's blog posts (every one. Seriously, all of them).
• Mira Murati's public talks.
• Andrew Tulloch's work at Meta + OpenAI (ex-Meta — your direct bridge).
• OpenAI's pretraining lineage (since most of TML's RS team came from there).
• Their public products (Tinker if released; whatever else).

Comp (federal filings / press)

• Base $350–475k (per Apr 2026 RTF).
• Sign-on substantial.
• Equity at $50–60B Series A — world-record Series A valuation. $2B raised in seed alone.
• Total comp likely $800k–$2m+ to outbid Meta.
• Andrew Tulloch reportedly turned down $1.5B from Meta to stay/return → they pay seriously for ex-Meta hires.

Levers

• Ex-Meta GenAI alumni already at TML — warm intro is the path.
• Andrew Tulloch is your closest ex-Meta bridge.
• Tiny team means no recruiter buffer; every conversation matters.

Common pitfalls across all three

• Generic "why frontier lab": each has a different mission. Memorize the differences:
  • Anthropic: safety-first, RSP, Constitutional AI, interpretability.
  • OpenAI: build AGI safely + commercially, fast iteration, products + research.
  • TML: rigorous science of LLMs, careful reproducibility, post-training depth.
• Bashing Meta: never. They all have ex-Meta employees. They will hear about it.
• Overselling: especially at Anthropic. Calibrated confidence > confident overstatement.
• Generic STAR stories: each story should land for the specific theme they probe. Not the same stories for all three.
• Skipping the recruiter: the recruiter sets your level, your interview panel, your initial offer band. They are an advocate.
• Letting the loop happen to you: if you don't push for a specific team early, you may end up with a generic "you matched with X" assignment that's not your best fit.

Two-week study sprint before each onsite