Advanced interview questions in AI & ML, page 5
Setting up an LLM-as-a-judge evaluation
Define rubric, prefer pairwise comparison, randomize order, calibrate against humans; control position, verbosity, and self-preference bias.
Attention in diffusion U-Nets for text conditioning
Self-attention mixes spatial features at low-res blocks; cross-attention has image queries attend to text-token keys/values; placed inside transformer blocks.
Design an automated system to diagnose model performance drop root causes
Tests causal attribution between pipeline bugs and drift. Strong answers sequence schema/null audits, feature drift via PSI/KS, then concept drift via holdout decay. Red flag: skipping pipeline checks to retrain immediately.
How do you monitor thousands of per-customer models as a fleet?
Tests fleet-level statistical aggregation versus per-instance alerting. Strong answers propose tiered telemetry, cohort baselining for drift, and hierarchical alerting to prevent fatigue.
Design concept drift detection with automated retraining safeguards
Baseline monitors raise CloudWatch alarms; EventBridge triggers retraining with cooldowns; model registry gates promotion.
How would you estimate causal impact using a quasi-experimental method?
DiD with Canada versus Australia; assert parallel trends; validate with pre-period plots and placebo tests.
Classifier-free guidance in diffusion models
Train one model jointly on conditional and dropped-condition inputs; at inference extrapolate from unconditional toward conditional prediction via a guidance scale, sharpening prompt adherence…
How would you architect dynamic training resource provisioning?
Routes jobs to right-sized instances, provisions managed jobs using spot, and auto-tears down on completion.
Explain Regression Discontinuity Design and propose a real-world scenario
Compare units just above and below a threshold for local effects; propose scenario with forcing variable.
Automated rollback for a failed model deploy
Detect failure via health and metric checks, automatically route traffic back to the last known-good version (blue-green or canary), keep the registry entry but unpromote, and alert.
DDPM versus DDIM sampling trade-offs
DDPM is a stochastic Markov chain needing many steps; DDIM is a non-Markovian, deterministic sampler that skips steps for far faster inference and reproducible, invertible latents, trading a…
Tensor versus pipeline parallelism for large models
Tensor parallelism splits individual layers across GPUs needing fast interconnect; pipeline parallelism splits layers into stages across GPUs.
Designing a high-resolution photorealistic face generator
Weigh StyleGAN's fast, controllable style-based synthesis against diffusion's diversity and stable training; handle scale via progressive or multi-resolution synthesis; protect diversity to avoid mode…
Core insight behind GPTQ and AWQ
Not all weights matter equally; GPTQ minimizes layer output error using second-order info, AWQ protects salient weight channels tied to large activations.
NeRF limitations and advances for robotics
Original NeRF is slow to train and render, per-scene, static, and needs many calibrated views; address speed with explicit grids or Gaussian splatting, dynamics with time-conditioned fields, and scale with…
Differential privacy vs utility in LLM fine-tuning
Clipping plus calibrated noise per step, smaller epsilon means stronger privacy but degraded accuracy, tracking the privacy budget across epochs.
Formulating a multi-step robot manipulation task
Perception detects and localizes the mug, action space spans navigation and manipulation, and a reward shaped over subgoals (reach, grasp, transport, place) with sparse final success guides learning.
Design a multi-tenant ML platform with isolation, security, and cost attribution
Tests mapping tenancy to compute, network, and identity primitives. Strong answers compare hard vs soft isolation, use namespaces or node pools with network policies and IAM, and enforce chargeback via resource quotas and labels.
Exploration versus exploitation: epsilon-greedy and UCB
Exploit current best to earn reward, explore to discover better options; epsilon-greedy explores randomly with probability epsilon; UCB explores by an uncertainty bonus…
Propose an architectural solution for contended GPU training resources
Tests multi-tenant GPU scheduling design at scale. Great answers tier jobs by checkpointability, apply quota-based preemption, mix spot and on-demand instances, and use MIG or time-slicing to bin-pack. Red flag: buying GPUs without scheduling logic.
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