Flashcard

Why do vector databases route new writes into a small separate segment instead of the main index?

Describe the dual write pattern teams use during an embedding model swap.

Define QK-norm and the training failure it prevents

Describe the attention sink: why token 0 hoards attention mass

Name the three fields every pairwise preference record needs

Why does attaching a screenshot make your assistant noticeably slower to start answering?

SigLIP versus CLIP: how a sigmoid loss changes vision encoder pretraining

Chunk overlap: what it protects against and what too much of it costs

Define sparse upcycling and what it reuses from a dense checkpoint.

Why do DeepSeek style stacks slice the FFN into many small experts instead of a few large ones?

Contrast RLAIF and RLHF as paths to safer model behaviour

Explain what Microsoft PyRIT contributes to LLM red teaming

Describe garak's role in the red team toolchain

Identify what Lakera Gandalf is and its training data value

Distinguish deception from sandbagging in frontier model evals

Explain alignment faking and what its existence implies for RLHF

Explain how SR 11-7 applies to LLM use in a bank

Tell apart the NIST AI RMF from its Generative AI profile

Describe the four NIST AI RMF functions and how the GenAI Profile extends them

Identify what ISO/IEC 42001 certifies and why auditors increasingly cite it

How LLM10 model theft happens against a hosted API

Describe LLM09 overreliance and one concrete UX mitigation

Walk through the three production paths through which LLM06 (sensitive disclosure) typically leaks

Walk through three distinct ways LLM04 (model DoS) shows up in production

Contrast LLM03 training data poisoning with LLM01 prompt injection