Anthropic interview prep: the real loop

Band mapping: Anthropic's internal ladder is not publicly documented; this KB's L5 means a senior engineer with strong individual scope. The recruiter screen is more substantive than the big-tech norm — motivation and mission interest are genuinely screened, not just logistics. Outside this taxonomy: a 60-90 minute coding screen (CodeSignal take-home or live) and onsite coding rounds, publicly reported as practical and build-from-scratch rather than verbatim puzzle-style; looking things up is allowed per the official careers FAQ, Python is the reported default, and AI assistance is prohibited in live interviews per Anthropic's published candidate guidance.

Design close to Anthropic's real surface: serving large-model inference (batching, queueing, GPU utilization), API and service design for a product growing faster than its capacity plan. 'Do the simple thing that works' is a published company value — gratuitous complexity reads as a miss, and interviewers reward a small design that ships over an impressive one that doesn't. First-person accounts of the technical rounds: pacing is abrupt by design — rounds can open with effectively no introductions, names confirmed and straight into the problem — and difficulty arrives as mid-round constraint escalation (a runtime cap, a memory ceiling, a new requirement stacked on the part you just finished) rather than as a harder initial prompt.

The values/culture conversation — publicly reported as the round candidates most often fail. Expect follow-ups pushed three or four levels deep on real ethical and interpersonal situations: pressure to ship something you doubted, beliefs you hold that others don't, conviction under disagreement. Leadership has publicly framed this as understanding how you engage with hard topics respectfully; rehearsed mission enthusiasm lands worse than honest, reasoned pushback.

Deep dive on past projects, motivation, and fit with the group. Public accounts describe consensus-based decisions with the manager breaking ties, interviewers drawn from one engineering group, and reference checks taken unusually seriously.

Band mapping: this KB's L6 approximates a staff-equivalent engineer — owns a major system or platform and influences adjacent teams. Anthropic publishes no per-band loop detail; public accounts of senior loops describe a two-day format, with the second day centered on career discussion and a technical project deep dive (and cancelled if day one fails). Coding screen and rounds remain outside this taxonomy: practical, build-from-scratch, lookups allowed, no AI assistance. Anthropic's engineering blog describes its take-home philosophy directly: problems representative of real work, a wide scoring distribution that avoids single-insight puzzles, no specific domain knowledge required, AI assistance explicitly permitted to mirror the job — and the assessment is repeatedly redesigned as each Claude generation beats the prior version, shifting toward out-of-distribution constraint puzzles where human reasoning still wins.

Platform-scale problems near Anthropic's reality: capacity and rate-limit design when demand structurally exceeds GPU supply, multi-tenant fairness, reliability for an API enterprises build on while the underlying models change. Simplicity discipline still applies — the published value cuts against architecture astronautics even at staff scope.

Values round, staff-calibrated: the publicly reported deep-probing style applied to bigger surfaces — decisions that traded commercial pressure against stated principles, conviction maintained against senior pushback, how they shape a team's honesty norms rather than just their own.

Technical leadership in a high-trust, low-process culture: setting direction through clarity and credibility rather than mandate, growing engineers in a flat organization, keeping a fast-scaling group coherent without importing heavyweight process.

Career walkthrough plus technical project deep dive — public accounts describe this as its own loop day for senior candidates: one significant project taken down as many levels as the interviewer cares to go, with personal contribution separated from team output. Consensus decision afterward, manager reported as tiebreaker.

Band mapping: this KB's L7 approximates a principal-equivalent engineer shaping infrastructure or engineering direction company-wide. Public information about Anthropic loops at this band is minimal — processes are bespoke, often network-driven, and the structure below is an extrapolation of the reported senior format. Confidence here is lower than the file's overall rating.

Company-scale working session: compute strategy under structural scarcity, inference economics across a model family, build/buy/partner calls for infrastructure the whole company stands on. Expect the conversation to run on Anthropic's actual problems rather than abstractions.

Org-level leadership for a company that has publicly committed to binding safety gates (the Responsible Scaling Policy is board-approved and public): operating under commitments that can pause the roadmap, preserving a high-trust culture through repeated doubling, and keeping engineering honest at executive altitude.

Values round at full seriousness: conviction with real stakes, candor upward when the truth is unwelcome, and critically examined views on AI risk. Public reporting suggests this round fails more candidates than any technical one — at principal scope, performed alignment is treated as a stronger negative signal than honest disagreement.

Thesis and charter conversation: what the candidate believes the company's infrastructure must become, mutual calibration at near-executive altitude. Bespoke at this band; reference checks publicly reported as a serious, weighted step rather than a formality.

Band mapping: this KB's L5 approximates a senior research engineer. Anthropic hires research engineers and scientists alongside product SWEs, and its careers page publicly notes about half its technical staff had no prior ML experience and about half have PhDs — capability over credentials is explicit. Outside this taxonomy: practical ML-flavored coding screens (public accounts describe building model components from scratch and debugging broken training code, on collaborative platforms, lookups allowed, no AI assistance per official guidance). Reported overall shape: a 30-min recruiter call that genuinely screens mission alignment (generic 'interested in AI' answers fail), the 60-90 min CodeSignal or live coding screen, then an onsite of five one-hour sessions — hiring manager, coding, system design, a role-specific second coding round, and a values discussion; team matching happens after the onsite, with consensus-based decisions and the hiring manager breaking ties.

ML systems with hands on the metal: fine-tuning and feedback-based training pipelines, eval harnesses, inference optimization, experiment infrastructure. The published simplicity value applies to research code too — the smallest rig that answers the question beats the platform that doesn't exist yet.

Values round with research framing: how the candidate reasons about capability work and its risks, honesty about what their results do and don't show, conviction under pressure. Publicly reported as the highest-failure round; dismissiveness toward safety is reported as a rejection pattern even for technically brilliant candidates. Practitioner hiring-manager guides add data-literacy and evaluation probes here: a misleading summary statistic the candidate caught, unexpected or biased model outputs and the guardrails around them, and what they do when performance breaches a threshold.

Project deep dive and fit: research taste versus engineering strength for the specific team, what they'd work on first. Consensus decision afterward; references weighted.

Band mapping: this KB's L6 approximates a staff-equivalent research engineer — owns an eval ecosystem, training platform, or research-infrastructure area. No public per-band loop detail exists; senior loops are publicly reported as two days with a career discussion and technical project deep dive on day two. Practical ML coding screens remain outside this taxonomy.

Design at the lab's real scale: evaluation ecosystems including the capability evaluations the company's published Responsible Scaling Policy depends on, data engines, training platforms that survive research paradigm shifts. Often runs close to live problems.

Values round, staff-calibrated: brokering between research ambition, product pressure, and safety commitments; honesty about what their measurements can and cannot support; conviction maintained against senior researchers.

Direction-setting where research doesn't know what it needs next quarter: choosing ahead-of-need platform bets, growing brilliant-but-raw researchers into system owners, creating just enough structure in a deliberately low-process culture.

Charter conversation plus deep project autopsy: which platform or measurement surface they'd own, their thesis on it, with the publicly reported deep-dive style — personal contribution separated from team output, several levels down. The hiring-manager hour goes deep on implementation details and trade-offs and may include code-review-style discussion across languages; senior candidates should expect their architectural taste, not just their results, to be probed.

Band mapping: this KB's L7 approximates a principal-equivalent research engineering leader — shapes how the lab trains, measures, or deploys at company level. Hiring at this band is rare, bespoke, and often network-driven; public process information is minimal and this structure is an extrapolation. Confidence at this band is lower than the file's overall rating.

Working session on frontier problems: training-systems strategy across paradigm uncertainty, evaluation at the edge of what's measurable, deployment architecture for capabilities without precedent — likely anchored in the team's live questions.

Lab-level leadership: multi-year bets under deep uncertainty, operating where published safety commitments (capability thresholds, pause conditions) intersect the research roadmap, building organizations that do unprecedented work reliably.

Values round at principal seriousness: calibrated honesty about what cannot yet be measured, decisions they would refuse, candor with executives, trust from both safety-focused and capability-focused colleagues. Performed alignment at this band is a stronger negative than honest disagreement.

Thesis conversation at near-executive altitude: what the lab's technical strategy should be in their area, mutual evaluation, references weighted heavily. Highly bespoke.