Interview questions in AI & ML, page 26
Penalizing sycophancy in a reward model
Sycophancy is reward proxy gaming where agreeableness substitutes for correctness; counter it with truth-anchored labels, perturbed-premise pairs, and consistency checks.
repartition() versus coalesce() in Spark
Repartition does a full shuffle and can increase or balance partitions; coalesce avoids a full shuffle and only reduces them.
The role of anchor boxes in detectors
Anchors are reference boxes of set scales and ratios; the model regresses offsets, framing detection as classification plus refinement; downsides are tuning, imbalance.
How would you design drift detection for high-dimensional embeddings?
Tests distribution shift in latent spaces beyond per-feature stats. Strong answers use maximum mean discrepancy, k-NN two-sample tests, or domain-classifier AUC, plus windowing. Red flag: per-dimension KS tests or mean-difference thresholds.
The alignment tax and capability trade-offs
Alignment tax is capability lost from safety tuning, measured as benchmark or task-success deltas before and after; a product decision weighs over-refusal against harm risk.
Purpose of watermarks in Spark Structured Streaming
A watermark sets a threshold on event-time lateness, lets late data update windows up to that bound, and tells Spark when to finalize and drop old state.
Why averaging shots reduces image noise
Random noise averages out while signal stays, so SNR rises with the square root of frame count; limits include motion and fixed-pattern noise.
Explain a model registry's purpose and what to store per version
Tests if you treat the registry as a governance bridge between training and production, not just storage. Strong answers cite versioned artifacts, lineage, metrics, dependencies, and approval gates. Red flag: calling it a file dump or experiment tracker.
Scalable oversight of superhuman models
Humans cannot judge outputs beyond their expertise, so feedback degrades; techniques like AI debate or recursive reward modeling decompose judgment.
Communicate a forecast interval to an executive
Give the point estimate but frame the range as scenarios, use a fan chart, tie the interval to planning decisions and risk.
PTQ versus QAT for model quantization
PTQ quantizes a trained model with calibration data, fast but may lose accuracy; QAT simulates quantization in training so weights adapt.
How does a model registry differ from cloud storage like S3?
This tests model governance beyond raw storage. A strong answer contrasts storage with stage transitions, lineage, and ACLs, then lists metadata like metrics, dependencies, and schemas. A red flag is treating S3 folders with naming conventions as a registry.
Declarative vs imperative ML platform design
Declarative GitOps gives auditable, reproducible, reviewable desired-state config with strong governance but a steeper learning curve; imperative SDKs are flexible and fast for scientists but harder to…
Designing an LLM red-teaming framework
Taxonomy of harms, automated adversarial prompt generation via attacker models and mutation, a classifier to triage outputs, and severity-by-likelihood prioritization.
A/B test two fraud models in production
Randomize by entity, consider shadow mode first, collect precision/recall and business loss, decide with significance and guardrails.
Homography vs fundamental matrix degeneracy
Homography fits planar scenes or pure rotation; fitting a fundamental matrix there is degenerate because points lack depth variation.
Epsilon in differential privacy and its trade-offs
Epsilon is the privacy budget bounding how much one record can change outputs; smaller epsilon means stronger privacy but more noise and lower accuracy.
AWS Bedrock versus a direct provider API
Bedrock unifies many models with IAM, VPC, and cloud integration; a direct provider API gives earliest models, full feature parity, and simpler vendor terms.
Audit an ML pipeline for GDPR compliance
Inventory data and check minimization, verify processing matches stated purpose, build lineage to trace any prediction's inputs.
How is IoU computed and why prefer mIoU?
IoU is intersection over union of predicted and true pixels; mIoU averages per class; pixel accuracy is dominated by background.
We are hiring for this. Every open role lists the topics its interview covers, so you can prepare for the real thing rather than guessing.
See open roles