ML Take-Home (4–5 hours)

Research-track alternative to the live ML screen. A self-contained 4-hour ML problem on the candidate's own machine or Colab, immediately followed by a 40-minute live review where you present results, walk through code or slides, and field questions. Open-book on docs; no AI tools. Requirements From the recruiter blurb: "For this 4-hour take-home assignment, you can use your local environment (could be online like Colab or your own IDE). It's mostly coding work, but pen and paper could be useful. The problem will involve doing some exploratory research, obtaining results and showing them numerically. You should be comfortable training and evaluating basic models, and some knowledge of machine learning and visualization libraries (pandas, numpy, matplotlib, ...) will be helpful. Treat this like an open-book examination: feel free to look up documentation on code but please don't look up the whole problem, and ensure all work is your own." 4 hours of independent work + 40-minute live review. Bring a 10-minute presentation — slides or a code-notebook walkthrough are both acceptable, as long as results are clearly presented. The materials are sent at the start of the 4-hour window; you cannot prep on the actual problem. Notes Open-book on documentation; AI tools are not permitted. The 40-minute review is technical Q&A, not just a recap — interviewers probe how you chose the experiment, what you'd do with more time, and how you reason about uncertainty in the result. Available descriptions cover the structure rather than the specific dataset/task; the assignment varies across candidates and Anthropic asks candidates not to share specifics. Preparation Run a timed dry-run on an open ML dataset: 4 hours from blank notebook to slides covering hypothesis, methodology, results, and limitations. Build a reusable Colab/Jupyter scaffold (loading helpers, plotting helpers, results-table helpers) so the first hour isn't infrastructure. Practice the live review: 10 minutes of presentation, then defense — clear narrative beats raw breadth of analysis.