Synthesizing recent data points across Anthropic, OpenAI, DeepMind, and a few well-funded AI startups. Patterns from the last 30-45 days.
Recruiter screen patterns: Most still gate on "why this company specifically": generic enthusiasm consistently fails this stage. Salary conversations happen earlier in the process at AI labs vs. traditional tech. Be prepared with a number in screen #1. Technical depth probes are now common at screen stage, not just onsite. ~15 min of "walk me through a project" where they go deep on one specific decision.
Technical interview shifts: Coding: increasingly tilted toward "modify this existing system" rather than "implement this from scratch." Be ready to read 200+ lines of code in interview environment. System design: heavy on tradeoffs around model serving, latency vs. cost, evals infrastructure. Less on "design Twitter." ML-specific questions are LESS algorithmic ("explain backprop") and MORE applied ("how would you debug a model that's silently degrading in prod").
Behavioral shifts: "Walk me through a time you disagreed with a strong opinion held by someone more senior than you" is showing up everywhere. Have a real story. "How do you decide what to NOT do": taste/judgment questions are heavy at staff+ levels.
The Anthropic-specific signal: writing samples (long-form, narrative) are being weighted more heavily in late-stage decisions. If you're interviewing there, polish your written communication.
Drop your data points in the comments: what did you get asked, what surprised you? The fresher the intel, the more useful it is for the next person prepping.