Anthropic system design interview

Round 6 of 8 · 26 questions reported

See the whole loop
Length
55 min
Format
Design discussion, often typed into a shared Google Doc rather than drawn; some rounds let you share any drawing tool
They grade
APIs, data model, request flow, handling the stated load, and at least one failure case
AI use
No AI
Heads-up
Leans toward AI infrastructure: batch inference APIs, keeping GPUs busy under batching limits, shipping large model files to many machines. Other prompts are a prompt playground (closer to product design), a one-to-one chat system and a data platform. Managers may review a design doc instead, and some loops use two design rounds on the team's own domain.

Questions Anthropic asks in this round 26

  • Batched LLM inference API over GPUsAI systemDesign an inference API that batches requests across GPUs efficiently. The discussion goes deep on a few axes: utilization, scaling under spikes, and separating prompt processing…
    Last asked Oct 2026
  • Design a distributed AI model downloaderAI systemGet a new multi-hundred-gigabyte model checkpoint onto thousands of GPU hosts as fast as possible without overloading the storage it comes from.
    Last asked Oct 2026
  • Design an AI prompt playgroundAI systemDesign a web workspace where developers write prompts, run them against several models and settings, compare outputs side by side, and save and share versions.
    Last asked Oct 2026
  • One-on-one chat system with presence and offline queueSystem designDesign direct messaging for single-device users: delivery, presence, offline-recipient queueing, session storage, and the choice of message bus. Group chat is explicitly out of…
    Last asked Oct 2026
  • Critique a design doc from the team's backlogSystem designThe interviewer hands over a real design doc and asks for a critique: surface risks, propose alternatives, and say what is missing before this could ship.
    Last asked Sep 2026
  • Matmul performance modeling on a fixed GPUML designEstimate wall-clock time for matrix multiplications of varying shapes on a given accelerator from arithmetic intensity and memory movement, then extend the model to a multi-layer…
    Last asked Sep 2026
  • ML configuration system with inheritance and validationML designDesign and partly implement an ML configuration system covering schemas, inheritance and overrides, validation, and reproducibility. python file=solution.py class Field: def…
    Last asked Apr 2026
  • Debug a GRPO training script and defend the RL mathML designBelow is a small group relative policy optimization (GRPO) trainer in plain Python. The policy is a table of logits per prompt, and the reward is the share of tokens that match…
    Last asked Mar 2026
  • Self-contained ML take-home with live reviewML designA few-hour modeling problem you solve alone on your own machine, followed by a live review: you present your results, code and trade-offs, and the interviewers ask questions…
    Last asked Feb 2026
  • Prompt design with evaluation scaffoldingML designIn a notebook, design prompts for a defined task and write the surrounding scaffolding that evaluates them. The exercise tests task framing and measurement more than clever…
    Last asked Jan 2026
  • Mechanistic interpretability experiment on double descentML designDesign and run a small-model experiment reproducing sample-aspect double descent, explain the phenomenon from theory, and propose a remedy. The deliverable is the experiment plus…
    Last asked Apr 2026
  • Design a GPU scheduling platformAI systemShare a large GPU cluster among research and production teams: queue training and inference jobs, enforce quotas and priorities, place jobs on topology-aware hardware, and preempt…
  • Design a RAG system over company docsAI systemEmployees should be able to ask questions about internal documents (Drive, Confluence, Slack, GitHub) and get grounded answers with citations.
  • Design a Slack-like chat systemSystem designWorkspaces, channels, threads and direct messages with real-time delivery, unread state, search and history, for large companies.
  • Design an AI coding agent sandboxAI systemAgents run untrusted code for many users. Design isolated, fast-starting sandboxes with file and network controls.
  • Design ChatGPTAI systemDesign a consumer chat product backed by a large language model: conversations, streamed replies, history across devices, at 100M weekly users.
  • Design Facebook MessengerSystem designOne-to-one and group messaging for a billion users: real-time delivery, offline sync, read receipts and message history on every device.
  • Design review: an eval gate for model launchesML designYou are reviewing the design of the evaluation gate that decides whether a new model checkpoint can launch. Critique it: what it can miss, how it can mislead, what you would…
  • Design review: checkpointing a large training runAI systemYou are reviewing the checkpointing design for a month-long training run. Critique it: what fails, what it costs, what you would change, and what the doc does not cover. You have…
  • Design review: rolling out a new model versionAI systemYou are reviewing a launch plan written by the serving team. Critique it: rank the risks, propose changes, and say what must be true before you would approve it. You have 45…
  • Design review: usage billing for an LLM APISystem designYou are reviewing a teammate's design doc before it goes to the wider team. Read it, then lead a critique: what is wrong, how bad each problem is, what you would change, and what…