AWS Bedrock versus a direct provider API
WHAT IT TESTS: build-versus-aggregate LLM sourcing. OUTLINE: Bedrock unifies many models with IAM, VPC, and cloud integration; a direct provider API gives earliest models, full feature parity, and simpler vendor terms.
Designing an LLM red-teaming framework
WHAT IT TESTS: systematic safety probing. OUTLINE: taxonomy of harms, automated adversarial prompt generation via attacker models and mutation, a classifier to triage outputs, and severity-by-likelihood prioritization.
Scalable oversight of superhuman models
WHAT IT TESTS: supervising models you cannot fully evaluate. OUTLINE: humans cannot judge outputs beyond their expertise, so feedback degrades; techniques like AI debate or recursive reward modeling decompose judgment.
The alignment tax and capability trade-offs
WHAT IT TESTS: cost of safety interventions. OUTLINE: 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.
Penalizing sycophancy in a reward model
WHAT IT TESTS: reward hacking and truthfulness. OUTLINE: sycophancy is reward proxy gaming where agreeableness substitutes for correctness; counter it with truth-anchored labels, perturbed-premise pairs, and consistency checks.
Constitutional AI versus standard RLHF
WHAT IT TESTS: alignment methods beyond human feedback. OUTLINE: a written principle set guides self-critique and revision, plus AI feedback (RLAIF) replaces human preference labels. RED FLAG: calling it just RLHF with extra steps or human-only labeling.
Preventing PII in LLM outputs: curation, fine-tuning, or guardrails
WHAT IT TESTS: Choosing the right layer for PII control. OUTLINE: Favor post-processing guardrails as the enforceable last line, backed by data curation; note each layer's tradeoffs and that defense-in-depth is best.
Red teaming a generative model
WHAT IT TESTS: Adversarial safety evaluation. OUTLINE: Deliberately probe for harmful outputs across categories, document jailbreaks, and automate with adversarial prompt generators plus classifier-based judging.
Evaluating faithfulness and compositionality in multimodal models
WHAT IT TESTS: Going beyond n-gram metrics. OUTLINE: Use targeted probes with hard negatives, attribute-relation binding tests, and structured grounding checks; note BLEU rewards surface overlap not correctness.
CLIP's contrastive objective and zero-shot classification
WHAT IT TESTS: Contrastive vision-language pretraining. OUTLINE: Train image and text encoders to align matched pairs and repel mismatched ones in a shared space; classify zero-shot by comparing an image to text prompts of class names.
Handling a 401 error in an LLM agent's tool call
WHAT IT TESTS: Robust agent error handling. OUTLINE: Catch the tool error, return a structured observation to the LLM, and distinguish recoverable retries from terminal failures needing re-plan or escalation.
Why chain-of-thought helps large models but not small ones
WHAT IT TESTS: Understanding emergent abilities and tier-aware prompting. OUTLINE: Small models lack reliable multi-step reasoning, so CoT just adds error-prone steps; adapt by using few-shot/fine-tuning or distillation for small tiers.
Self-consistency over chain-of-thought
WHAT IT TESTS: Sampling-based reasoning improvement and its cost. OUTLINE: Sample multiple CoT paths at nonzero temperature and majority-vote the final answer; cost scales with the number of samples.
Prompt engineering to curb extraction hallucinations
WHAT IT TESTS: Practical hallucination control plus honesty about limits. OUTLINE: Ground strictly in source, allow null for missing fields, enforce a schema, and use few-shot examples; acknowledge prompting cannot fully eliminate it.
Fault-tolerant checkpointing for thousand-GPU pre-training
WHAT IT TESTS: Resilient large-scale training. OUTLINE: Checkpoint weights, optimizer state, RNG, and data position together; use asynchronous sharded writes and automated detect-restart-resume.
Tokens and vocabulary-size tradeoffs
WHAT IT TESTS: Tokenization fundamentals. OUTLINE: A token is a subword unit; larger vocab shortens sequences but bloats the embedding matrix, smaller vocab generalizes but lengthens sequences.
Cross-attention versus self-attention in encoder-decoder Transformers
WHAT IT TESTS: Information flow in encoder-decoder models. OUTLINE: Cross-attention draws Queries from the decoder and Keys/Values from the encoder, letting the decoder condition on the source.
How Transformers encode token position
WHAT IT TESTS: Why and how positional information is injected. OUTLINE: Attention is permutation-invariant, so positional encodings (sinusoidal, learned, or rotary) are added or applied.
Self-attention and the Query, Key, Value matrices
WHAT IT TESTS: Core Transformer mechanics. OUTLINE: Queries score against keys via scaled dot product, softmax yields weights, and those weight the values into the output. RED FLAG: Confusing the three roles or omitting the scaling and softmax steps.
Why RAG persists despite million-token context windows
WHAT IT TESTS: RAG versus long-context tradeoffs. OUTLINE: Cost and latency scale with context, attention degrades in the middle, and RAG adds freshness, access control, and citations. RED FLAG: Assuming a huge window equals reliable use of all of it.