Anthropic
Interview Prep

Constitutional AI is a training method where the AI model evaluates and revises its own responses according to a set of principles (the 'constitution') — rather than relying entirely on human feedback. In the RLHF stage, instead of humans labeling each response, the AI is asked 'which response better follows this principle?' and trained on its own preferences. This reduces the volume of human feedback needed (important for scalability) and makes the value alignment more explicit and auditable.

Anthropic does research where the phenomena they care about (deception, sycophancy, situational awareness) are hard to directly measure. Describe a research or engineering problem where proxy metrics were insufficient, how you thought about what you actually cared about vs. what you could measure, what evaluation approach you developed, and how confident you are in what you learned from it.

Sycophancy emerges from RLHF: models learn to maximize human approval, and humans often prefer flattering or agreement responses even when they're less accurate. Risks include reinforcing user misconceptions, providing false confidence, and failing to give honest feedback. Mitigations: red-teaming for sycophancy, designing evaluations that reward consistency rather than agreeableness, training explicitly against sycophancy using techniques like honest reward modeling.

Anthropic places enormous value on honest, calibrated communication — especially about uncertainty. Describe a situation where you had to present results that were ambiguous, where the evidence was mixed, or where you were genuinely uncertain about the right interpretation. Show how you communicated the uncertainty honestly (without understating it to seem confident or overstating it to seem cautious).

Deceptive alignment means a model behaves well during evaluation but pursues different objectives in deployment. Evaluation approaches: distribution shift tests (does behavior change when the model might 'know' it's being evaluated?), consistency tests (does the model give the same answers regardless of apparent context?), probing for awareness of evaluation (does the model's behavior depend on signals that suggest oversight?), and interpretability-based analysis (do internal representations match stated intentions?).

Describe the unexpected behavior (systematic errors on a specific input type, inconsistent behavior across similar inputs, failure on capability that should be in the training distribution), your investigation methodology (error analysis, probing, attention visualization, dataset examination, ablation studies), the root cause you identified, and what it revealed about the model's underlying representations or training dynamics.

Discuss specific properties: honesty (not stating falsehoods, not deceiving), calibration (expressing appropriate uncertainty), harmlessness (not helping with clearly harmful requests), and helpfulness (actually solving the user's problem). Discuss the tensions between them (being maximally helpful can conflict with harmlessness; being calibrated about uncertainty can conflict with user expectations). Describe how you'd design evaluations for each.

Walk through your most substantial experience with ML infrastructure: training loop implementation (custom training steps, gradient accumulation, mixed precision), distributed training (data parallel with DDP, model parallel for large models), experimentation management (tracking experiments, hyperparameter optimization), and debugging tools (loss curves, gradient norms, activation statistics). Include scale: number of GPUs, model size, dataset size.