5 rounds. each one harder than the last. 1. coding: implement transformer attention from scratch in numpy. no pytorch no jax. raw matrix math 2. ML theory: 45 min of rapid fire questions about RLHF, constitutional AI, interpretability. they went DEEP 3. system design: design a model evaluation pipeline that can handle 100k+ prompts/day with human review 4. research taste: discussed 3 recent papers. they wanted to know what i thought was promising vs overhyped 5. values alignment: long conversation about AI safety and why i want to work here specifically got rejected. feedback said my coding was strong but my ML theory depth wasnt at the level they needed. fair honestly

Anthropic research engineer - hardest interview of my life

52fb0f53

5 rounds.

f11f3c8d

implementing attention from scratch in numpy is actually insane. thats not an interview question thats a homework assignment. how long did they give you?

e2f877e5

anthropic has the highest interview bar in AI right now. even people from deepmind and google brain get rejected. dont feel bad

6c70acb6

the research taste round is what separates anthropic from other AI companies. they dont just want people who can code ML - they want people who have OPINIONS about where the field should go

2f45f5ff

curious what papers they asked about. was it recent interpretability stuff like SAE features or more alignment/RLHF focused?