Glossary

The model

Token: the unit the model reads and writes; a subword chosen by the tokenizer, not a word. Cost and context limits count tokens.

Byte-pair encoding (BPE): a tokenizer trained by repeatedly merging the most frequent adjacent pair; more merges mean a larger vocabulary and shorter sequences.

Next-token objective: the training loss, cross-entropy on the next token; the only thing a language model is trained to do (Principle 1).

Logits: the raw scores over the vocabulary before the softmax.

Temperature: a divisor on the logits before the softmax; low sharpens, high flattens; it never adds knowledge.

Embedding: the vector a token id becomes; the unembedding maps the final vector back to logits.

Attention: a weighted average of value vectors where weights come from query-key similarity; the causal mask hides the future; the sqrt(d) scaling keeps the softmax from saturating (Principle 4).

Residual stream: the vector carried through the layers that each block adds to; why deep Transformers train.

KV cache: the stored keys and values of earlier positions so each decode step computes one new row (Principle 5). Size per token = 2 x layers x KV heads x head size x bytes.

Grouped-query attention (GQA): fewer KV heads than query heads; shrinks the KV cache and multiplies the feasible batch.

Prefill and decode: processing the prompt all at once (compute-bound) versus producing tokens one at a time (bandwidth-bound at small batch).

Scaling law: loss as a power law in data, parameters and compute; Chinchilla gives the compute-optimal split (Principle 7).

Pretraining: next-token training on a large corpus; gives knowledge and a prior.

Supervised fine-tuning (SFT): training on instruction-response pairs; changes format and behaviour, not reliably facts (Principles 8, 10).

Direct preference optimisation (DPO): training on preferred-versus-rejected pairs without a separate reward model.

RLVR and GRPO: reinforcement learning with verifiable rewards; GRPO uses group-relative baselines; the KL penalty keeps the policy near the reference.

Reward hacking: the policy finding a loophole in the reward specification; lab 09's "please" spam.

LoRA: low-rank adapters trained while base weights stay frozen; cheap fine-tuning that fits one GPU.

Inference-time scaling: best-of-N, majority voting, self-refinement; buys quality with compute at inference, sometimes negatively.

Systems

Retrieval-augmented generation (RAG): putting retrieved documents into the context so the model can answer from them (Principle 9).

Hybrid search: lexical (BM25) plus vector retrieval, fused by rank.

Reciprocal rank fusion: combining ranked lists by summing 1/(k + rank); robust to score scales.

Reranker: a second model that reorders the top results; costs latency.

Golden set: a versioned set of real questions with expected sources; the yardstick for retrieval (Principle 13).

Context window: the finite attention budget; every token competes (Principle 9).

Context engineering: choosing the smallest set of high-signal tokens: system-prompt altitude, tool design, just-in-time retrieval, compaction, notes outside the context, sub-agents.

Compaction: summarising history to fit the context; can raise cost by breaking the cached prefix and cut recall.

Prompt caching: reusing the computation for a stable prefix; put dynamic content at the end.

Agent: a loop in which the model proposes, code validates, a tool runs, and the result returns as tokens; rules live in code.

Tool invariant: a condition checked in code before a tool executes; schema validation is not safety.

Model Context Protocol (MCP): a protocol for exposing tools to models; pin tool definitions, least-privilege tokens, never forward the client token upstream.

Lethal trifecta: private data, untrusted content and an outbound channel in one agent; remove one leg in code (Principle 16).

Prompt injection: instructions arriving as data (documents, tool results) that the model may follow; prompts are not a security boundary.

Roofline: the plot of achievable throughput against arithmetic intensity; says whether a kernel is bound by bandwidth or compute (Principle 18).

Speculative decoding: a small draft model proposes tokens the large model verifies in one pass; the benefit depends on the acceptance rate.

Paged attention: managing the KV cache in fixed-size blocks like virtual memory; the vLLM idea.

Measurement

Error analysis: reading 20 to 50 outputs by hand, labelling pass/fail, grouping notes into failure categories; precedes tooling (Principle 12).

Code grader: a deterministic check (parses, cites, abstains); used wherever a rule can catch the failure.

LLM-as-judge: a model grading outputs; a model with error rates that needs calibration (Principle 14).

True-positive and true-negative rate (TPR, TNR): how often the judge passes real passes and fails real failures, against human labels.

Groundedness: whether an answer is supported by a retrieved chunk it cites.

Abstention: saying "I don't know" when retrieval returns nothing relevant; the measured answer to "how did you measure hallucinations".

recall@k: the fraction of golden questions whose relevant chunk is in the top k results.

pass@k and pass^k: the probability that any of k trials succeeds versus that all do; capability versus reliability (Principle 15).

Noise band: the spread of a suite's score across repeated runs of the same model; a difference inside it is sampling (Principle 19).

Regression suite: cases from past failures run on every change; overfit to the incumbent model unless tagged and re-baselined.

Leakage: test information reaching training or tuning; group-aware and time-ordered splits prevent it.

p50 and p95: median and 95th-percentile latency; users feel p95.

Time to first token (TTFT) and time per output token (TPOT): the two latency numbers a serving benchmark must report.

Career

Adjacent role: a job that touches AI without an AI title; the most common route into ML in the evidence.

Public record: posts, repos and talks a stranger can cite; what produces referrals.

Depth probe: the interview round that asks "why" five levels deep on one thing you built.

Expert markers: own a problem end to end; known outside the company; arbitrarily deep under questioning; learn on a cadence.