# AI Glossary: Every Reading Level Three lines per term, one for each stop on the site's reading-level dial. Level 1 is for the person running the shop, level 3 for the person who has to make it work on Monday, level 2 for the two of them reading together. Every line is true; only the words change. Industry words (Metrc, SOP, 280E) keep their real name at every level. | level | who it is for | |---|---| | 1 · Plain English | operators, owners, budtenders | | 2 · Bridge | an operator and their tech person, together; real terms in brackets | | 3 · Technical | developers and data people; full jargon, terse | ## How the model works ### LLM (large language model) 1. The AI 2. The language model (the AI itself, like Claude) 3. The LLM ### Model 1. The AI brand you pick 2. The model (the specific AI you are talking to, like Claude Sonnet or GPT) 3. The model (a specific checkpoint, e.g. claude-sonnet-5) ### Token 1. Little chunks of words 2. Tokens (the word-chunks the AI reads and bills in) 3. Tokens (tokenizer vocabulary units; billing and context are metered in them) ### Context window 1. How much it can hold in mind at once 2. Its working memory (which people call the context window) 3. The context window (full per-turn payload, ~200K–1M tokens, re-sent every turn) ### Prompt 1. What you type to it 2. Your prompt (the message you send) 3. The prompt (user turn, plus whatever system prompt precedes it) ### System prompt 1. The standing instructions the AI reads before every message 2. The system prompt (instructions loaded before your message, every time) 3. The system prompt ### Hallucination 1. Making something up and sounding sure 2. Confidently making things up (a hallucination) 3. Hallucinating (fluent, high-confidence output unsupported by the context) ### Training cutoff 1. A fixed date in the past 2. Its training cutoff (the date the text it learned from stops) 3. Its training cutoff (no knowledge past it without retrieval) ### Structured output 1. Making the AI answer in an exact fill-in form 2. Structured output (forcing the answer into exact fields instead of prose) 3. Structured output (schema-constrained decoding) ### Schema 1. A fixed fill-in-the-blanks shape for the answer 2. A schema (the exact fields and types the answer has to come back in) 3. A schema (JSON Schema for structured output; validate before trusting) ### Deterministic 1. Pure arithmetic, same answer every time 2. Deterministic (same input, same output, no judgment in the middle) 3. Deterministic (no model in the control path; identical on rerun) ### Eval 1. A quick test to see if the AI got it right 2. An eval (a repeatable test with known right answers you run the AI against) 3. An eval (fixed inputs, expected outputs, scored; rerun on every model change) ### Local model 1. An AI that runs on your own machine, no internet 2. A local model (an open model running on your own hardware, data never leaves) 3. A local model (e.g. Llama via Ollama; weights on-prem) ### Fine-tuning 1. Retraining an AI on your own examples 2. Fine-tuning (adjusting a model's weights on your own examples) 3. Fine-tuning (SFT/LoRA on a base checkpoint; rarely the first move) ### Model-agnostic 1. Works with any AI brand 2. Model-agnostic (works with whichever AI you point it at) 3. Model-agnostic (provider-neutral; swaps models behind a common completion interface) ### Open-source 1. Free to take, read and change 2. Open-source (the code is public and free to use and change) 3. Open-source (check the license) ## Skills, connectors and what loads when ### Skill 1. A written checklist it follows 2. A skill (a checklist file it follows when the job matches) 3. A skill (SKILL.md: description as trigger, body as procedure, bundled scripts and references) ### Trigger 1. The sentence that tells the AI when to use this checklist 2. The trigger (the description sentence that tells the AI when this skill applies) 3. The trigger (the `description` field, matched against the request) ### Progressive disclosure 1. The read-the-top-first rule (it only opens a checklist whose top line matches the job) 2. Progressive disclosure (the AI reads only each skill's name and description, and opens the full body when one matches) 3. Progressive disclosure (metadata always in context; body loaded on description match) ### MCP (Model Context Protocol) 1. A plug that lets it use another program 2. An MCP connector (the standard plug that gives it tools in another program) 3. An MCP server (Model Context Protocol; exposes tools, resources and prompts over JSON-RPC) ### Connector 1. A plug into one of your programs 2. A connector (a plug that gives the AI a set of tools inside one program) 3. A connector (an MCP server bound to one system; its tool definitions load every turn) ### Tool 1. Something the AI is allowed to do, like read a file 2. A tool (one action the AI can take, like read a file or search orders) 3. A tool (a typed function the model can call; schema + handler) ### Tool definition 1. The description of each action the plug offers 2. A tool definition (the name, description and input fields of one tool, sent to the AI on every message) 3. A tool definition (name, description, JSON Schema; in context every turn) ### Resource 1. A reference document the plug keeps on hand 2. A resource (a document behind a connector that loads only when something asks for it) 3. A resource (URI-addressed content; enters context on resources/read) ### Plugin 1. A bundle of checklists and plugs installed together 2. A plugin (a package of skills and connectors installed as one unit) 3. A plugin (packaged skills + MCP servers; per-turn cost depends on contents) ### Memory 1. Notes the AI keeps between chats 2. Memory (notes the AI saves and reloads so the next chat is not a blank slate) 3. Memory (persisted notes injected into context; loads every turn, audit it) ### Allowlist 1. A short list of the actions you have switched on, everything else off 2. An allowlist (the explicit list of tools the AI may call, with everything else off) 3. An allowlist (per-tool permission set; deny by default) ### API 1. A way for two programs to talk directly 2. An API (the door a program opens so other programs can ask it for data) 3. An API ### Webhook 1. A message one program sends another when something happens 2. A webhook (an HTTP call one system fires at another when an event happens) 3. A webhook (HTTP POST on event; verify the signature) ### JSON 1. A tidy, machine-readable format 2. JSON (the curly-brace text format programs use to pass data around) 3. JSON ### CSV 1. A plain spreadsheet file 2. A CSV (the plain-text spreadsheet export every system can produce) 3. A CSV (RFC 4180-ish; quoting and encoding will bite) ### Markdown 1. Plain text with a few symbols for headings and lists 2. Markdown (plain text where # means heading and - means bullet) 3. Markdown ### Frontmatter 1. The label block at the top of a file 2. Frontmatter (the block of labels between dashes at the top of a text file) 3. Frontmatter (YAML header; parsed, not prose) ### Repo 1. A shared folder of code and files with history 2. A repo (a version-controlled folder of code and files, usually on GitHub) 3. A repo ### Terminal 1. The black text window where you type commands 2. The terminal (the command-line window engineers type into) 3. The shell ## Agents ### Agent 1. An AI that keeps working on its own 2. An agent (an AI that loops on its own, using tools, until it is done) 3. An agent (LLM in a tool-use loop with a stopping condition) ### Harness 1. The machinery around the AI 2. The harness (the software around the model that feeds it, gives it tools and stops it) 3. The harness (orchestration layer: context assembly, tool dispatch, loop control, stop criteria) ### Agent host 1. The AI program you use (the chat app or the coding tool) 2. The host (the AI program, like Claude Code or a chat app, that loads your skills and connectors) 3. The host (agent runtime; loads skill metadata, dispatches tool calls) ### Orchestrator 1. The lead AI that hands out jobs 2. The orchestrator (the lead agent that splits work and hands pieces to helpers) 3. The orchestrator (parent context; spawns sub-agents, merges summaries) ### Sub-agent 1. A helper copy of the AI 2. A sub-agent (a helper AI spun up for one job with its own clean memory) 3. A sub-agent (child context spawned by an orchestrator; isolated window, returns a summary) ### Stopping condition 1. A clear rule for when it should stop 2. A stopping condition (the written rule for when it is done or must ask a person) 3. A stopping condition (explicit termination predicate; missing one means it runs until budget or error) ### Write gate 1. The point where it would send, save or delete something real 2. The write gate (the point where the AI would change a real record, so a person confirms first) 3. The write boundary (human confirmation before any mutating call) ### Idempotent 1. Safe to run twice without doubling anything 2. Idempotent (running it again changes nothing the first run did not) 3. Idempotent ### Automation 1. A set of steps that runs by itself the same way every time 2. An automation (fixed steps, same order every time, no judgment in the middle) 3. An automation (deterministic pipeline; no model in the control path) ### n8n 1. A free tool for chaining steps together 2. N8n (a self-hostable tool where you draw a flow of steps and it runs them) 3. N8n (self-hostable workflow runner; nodes, credentials, webhooks) ### Self-hosted 1. Running on your own computer or server 2. Self-hosted (running on a server you control, not the vendor's) 3. Self-hosted ### Sandbox 1. The rented machine it works on 2. The sandbox (a rented, throwaway machine the AI works in, not yours) 3. The sandbox VM (ephemeral, per-session; artifacts only) ### Blast radius 1. Mess if it goes wrong 2. Blast radius (the size of the mess if it goes wrong) 3. Blast radius ## Decision models (System One) ### Decision model (System One) 1. A decision model, an AI that answers questions with a pick, a score or a yes/no instead of writing text 2. A decision model, also called a System One model (it takes your data and typed questions and returns choices, scores and probabilities instead of writing text) 3. A System One decision model (typed questions over a state → calibrated choice/score/noul answers; no generation) ### System One 1. System One, the fast gut-check kind of AI 2. System One (TypeSafe's name for models that make fast, structured judgments instead of writing, after Kahneman's fast-thinking System 1) 3. System One (TypeSafe's decision-model class; Jev is the first) ### Jev 1. Jev, TypeSafe's decision model 2. Jev (TypeSafe's first System One decision model) 3. Jev (jev-latest → jev-1.13.0; POST /v1/systemone) ### State 1. The information you hand the AI to judge 2. The state (the text or record you hand the model to judge, like the ticket, the order and the policy together) 3. The state (string, JSON object or array of text; one state per request, all questions see it) ### Typed question 1. A fixed-answer question 2. A typed question (one whose possible answers you define up front, so the answer comes back as a value your software can use) 3. A typed question (Choice / Score / Noul primitive with instructions and criteria) ### Choice 1. A pick-one question 2. A Choice question (pick one option from a list you define, with a probability for each option) 3. A Choice (returns choice, probabilities per option, confidence) ### Score 1. A rate-it question 2. A Score question (rate the content against ordered levels you describe, returning a number and a probability per level) 3. A Score (ordered levels; returns expected score, probabilities per level, confidence) ### Noul (yes/no probability) 1. A yes/no question that comes back as a percentage 2. A Noul question (is this statement true, answered as a probability from 0 to 1) 3. A Noul (P(true) in 0–1; no separate confidence field) ### Probability 1. How likely the AI thinks each answer is, as a percentage 2. A probability (the model's estimate, 0 to 1, that each possible answer is the right one) 3. A probability (per option or level; the distribution shape is the uncertainty signal) ### Calibrated 1. Honest about how sure it is 2. Calibrated (when it says 80 percent, it is right about 80 percent of the time across many answers, not a promise about any single one) 3. Calibrated (probabilities match observed frequencies in aggregate; trained via RLCD) ### Confidence 1. How sure the AI is, as one number 2. Confidence (a single 0-to-1 number summarising how concentrated the probabilities are, so your code can decide whether to act) 3. Confidence (statistic over the probability distribution; on Choice and Score answers) ### Threshold 1. The cut-off you set for when the AI is sure enough to act 2. A threshold (the confidence cut-off your code uses to decide act, ask a person, or stop) 3. A threshold (per-action confidence gate; scale with the cost of being wrong) ### Classification 1. Sorting things into buckets 2. Classification (putting each item into one of a fixed set of categories) 3. Classification ### Entity alignment 1. Matching two descriptions of the same thing, like a texting number and a customer account 2. Entity alignment (deciding whether two records describe the same thing, like a sender and a customer account, with a middle answer for a person to settle) 3. Entity alignment (pairwise same / related / different judgment over two records in one state; middle level escalates) ### RLHF 1. Training an AI on what people like to read 2. RLHF, reinforcement learning from human feedback (training a model to produce answers people prefer, which made chatbots pleasant and sometimes sycophantic) 3. RLHF (preference optimisation; rewards agreeable text; mode dropping) ### RLVR 1. Training an AI on problems with checkable right answers, like math 2. RLVR, reinforcement learning with verifiable rewards (training on tasks with checkable answers, which produced the slow, expensive reasoning models) 3. RLVR (reinforcement learning with verifiable rewards; reasoning models, higher latency and cost) ### RLCD 1. Training an AI to be honest about how sure it is 2. RLCD, reinforcement learning for calibrated decisions (TypeSafe's training that rewards decisions whose probabilities match reality) 3. RLCD (calibration-optimised post-training; decisions + probabilities, no text) ### Playground 1. TypeSafe's free try-it page in the browser 2. The TypeSafe Playground (a browser page where you paste a record and questions and see the probabilities) 3. The TypeSafe Playground (console.typesafe.ai/playground; shareable state + questions links) ### SDK 1. The ready-made code library for talking to a service from your own program 2. An SDK (the ready-made code library for calling a service from your own program) 3. The SDK (typesafe-sdk on PyPI; TypeSafeClient().system_one(state, questions)) ## Patterns for building with decisions ### Speculative fan-out 1. Asking many small questions at once 2. Speculative fan-out (ask every question you might need in one call, since each is judged independently, then let your code pick what matters) 3. Speculative fan-out (N questions, one state, one request; parallel, isolated, near-flat latency) ### Composite scoring 1. Scoring a few small things and adding them up your way 2. Composite scoring (break a big judgment into small scores and combine them with weights in your own code) 3. Composite scoring (atomic Score questions; weights and formula live in code, not the prompt) ### Confidence-gated routing 1. Letting how sure the AI is decide whether it acts, asks, or stops 2. Confidence-gated routing (the answer says what, the confidence says whether to act, confirm with a person, or hand off) 3. Confidence-gated routing (three bands: act / confirm / escalate; thresholds per action) ### Intent routing 1. Working out what someone wants and sending it to the right place 2. Intent routing (classify an incoming message and route it to plain code, a bigger AI, or a person) 3. Intent routing (Choice over intents → deterministic handler / specialist LLM / human) ### Guardrail 1. A check that stops the AI from doing or saying something it should not 2. A guardrail (a cheap check on an AI's input, output or action before it goes through, often a yes/no question to a decision model) 3. A guardrail (semantic check on prompt / completion / tool call; Noul or Choice, gated by threshold) ### Tool call 1. A moment where the AI actually does something, like create a meeting or send an email 2. A tool call (the AI asking a program to do one thing, like create a meeting or search orders) 3. A tool call (structured invocation of a tool; name plus arguments, executed by the harness) ### Coding agent 1. An AI that writes software for your tech person 2. A coding agent (an AI such as Claude Code or Codex that reads and edits a codebase on request) 3. A coding agent (Claude Code, Codex, or similar agent harness with file and shell tools) ## Cannabis operations ### Metrc 1. Metrc 2. Metrc (the state track-and-trace system every plant and package is reported to) 3. Metrc (state track-and-trace; per-facility REST API, rate-limited) ### Metrc tag 1. The Metrc tag on every plant and package 2. The Metrc tag (the 24-character barcode ID on every plant and package) 3. The Metrc tag (24-char UID; join key for every reconciliation) ### Manifest 1. The manifest (the state paperwork for every shipment) 2. The manifest (the state's transfer record listing every package on the truck) 3. The Metrc transfer manifest (incoming transfers need an explicit accept before packages go active) ### ERP (inventory and order system) 1. Your inventory system (Distru, for most people reading this) 2. Your ERP (the system of record for inventory, orders and invoices, like Distru) 3. The ERP (system of record; Distru here) ### SOP (standard operating procedure) 1. An SOP, your written how-we-do-it checklist 2. An SOP (the standard operating procedure you would hand a new hire) 3. An SOP ### 280E 1. 280E, the tax rule that lets you deduct almost nothing but product cost 2. 280E (the federal tax rule that lets cannabis businesses deduct product cost and little else) 3. IRC §280E (only COGS deductible; classification drives the whole ledger) ### Cogs 1. Cost of goods sold, what you paid for the product you actually sold 2. Cost of goods sold (COGS): what went into the product you sold, and under 280E the only thing that survives 3. COGS (§471 inventoriable costs; the sole survivor under §280E, method-dependent) ### Chart of accounts 1. The list of buckets every dollar gets sorted into 2. The chart of accounts (the list of buckets every dollar gets sorted into) 3. The chart of accounts (account_code enumeration; closed set for classification) ## Working with your data ### CSV 1. A plain spreadsheet file 2. A CSV (the plain-text spreadsheet export every system can produce) 3. A CSV (RFC 4180-ish; quoting and encoding will bite) ### Normalization 1. Making different spellings of the same thing read the same 2. Normalization (making different spellings of the same thing read the same) 3. Normalization (lowercase, alias expansion, unit folding; canonical form before any comparison) ### Fuzzy match 1. A close-enough match, with a number for how close 2. A fuzzy match (close-enough, scored, not exact) 3. A fuzzy match (token overlap plus sequence ratio, thresholded; difflib here, rapidfuzz at scale) ### Chunking 1. Splitting a big file into pieces it can read one at a time 2. Chunking (splitting a file too big for the context window into pieces, each with its own header, asked about separately) 3. Chunking (partitioning input to fit the context window; split by group, repeat headers, never aggregate across chunks in the model) ### Regex 1. A word-pattern search that never gets tired 2. A regex (a word-pattern search that runs the same way every time) 3. A regex (case-insensitive lexical gate; no semantics) ### Bulk import 1. Uploading one spreadsheet to change many products at once 2. A bulk import (one spreadsheet upload that changes many products at once) 3. A bulk import (templated CSV; rejects on schema mismatch, may overwrite on duplicate keys) ### Pii 1. Anything that points at a real person, like a name, phone number or ID 2. Personal information (PII): names, phones, emails, IDs 3. PII (personally identifiable information: names, contact details, government and customer IDs) ### Redaction 1. Blacking out the personal bits before you share 2. Redaction (blacking out the personal bits before you paste) 3. Redaction (deterministic PII substitution before the payload leaves your machine) ### Data retention 1. How long the company keeps what you pasted 2. How long the vendor keeps what you pasted (its data-retention policy) 3. The retention policy (how long prompts and outputs persist server-side; varies by tier and contract) ### Dry run 1. A practice run that shows what it would do without doing it 2. A dry run (the script prints every request or change it would make and sends none of them) 3. A dry run (--dry-run; log the intended calls, mutate nothing) ### Spot check 1. Checking a handful of rows by hand instead of all of them 2. A spot-check (a random sample of rows checked by hand, with a written rule for when one wrong row means you check everything) 3. Spot-check sampling (random n of N; escalation rule on first failure; written down so it does not drift) ### False positive 1. A false alarm, where it flags something that is actually fine 2. A false positive (a flag on something that is actually fine) 3. A false positive (precision loss; every one costs a human a look) ### Few shot 1. An example of the finished thing, pasted so it can copy the style 2. A few-shot example (one or more finished samples in the prompt that show the shape and voice you want) 3. A few-shot example (in-context input/output pairs; format and tone transfer better than from instructions) ## Code, APIs and running things ### Api key 1. The badge a program shows to prove it is allowed in 2. An API key (a long secret string that proves to Distru which account is asking) 3. An API key (bearer token; carries the creating admin's permissions) ### Pagination 1. Getting a long list one page at a time 2. Pagination (the server hands you a long list one page at a time and tells you where the next page is) 3. Pagination (cursor-based; follow next_page until null) ### Rate limit 1. A cap on how often a program may ask before it is told to wait 2. A rate limit (a cap on requests per minute or day; over it, the server answers 429 and says how long to wait) 3. A rate limit (429 + Retry-After; sliding window) ### Cron 1. A timer that runs a task at the same time every day 2. A scheduled timer (which people call cron) 3. Cron (crontab entry; consider systemd timers or the ERP's scheduler where available) ### Environment variable 1. A setting kept on the computer itself, outside the chat 2. An environment variable (a setting the computer hands the program at start, never typed into the chat) 3. An environment variable (process env; the only place a credential should live for a local server) ### Stdio 1. A program running on the same computer, talking to the AI app directly 2. Stdio (the plug talks to a program on the same computer over its input and output, no network) 3. Stdio (newline-delimited JSON-RPC over the child's stdin/stdout; stderr is free for logs) ### Json rpc 1. The fixed question-and-answer format the plug speaks 2. JSON-RPC (the fixed request-and-response message format MCP uses under the hood) 3. JSON-RPC 2.0 (id, method, params; MCP's wire format on every transport) ### Credential 1. A password or key 2. A credential (a password, login or API key) 3. A credential (password, API key, OAuth token; anything that authenticates as you) ### License 1. The note that says whether you are allowed to use it 2. The license: the file that says what you may do with the code (MIT and Apache let you use it freely) 3. The license (SPDX id in a LICENSE file; no file means all rights reserved) ### Commit 1. A saved change with a date on it 2. A commit (one saved, dated change to the code) 3. A commit ### Fork 1. Your own copy that you are free to change 2. A fork (your own copy of the repo, which you change and can still pull fixes into) 3. A fork (tracking upstream; rebase or merge to pull fixes) ### Readme 1. The front page that explains what it is and how to set it up 2. The README (the front-page file with what it does and how to install it) 3. The README ### Project 1. A saved workspace in the chat app that remembers your standing instructions and files 2. A project (the saved workspace in the chat app that holds standing instructions plus files) 3. A project (persisted system prompt plus attached files, prepended to every turn) ## Cost and measurement ### Input token 1. Word-chunks you send in 2. Input tokens (the word-chunks you send in, billed at the reading rate) 3. Input tokens ### Output token 1. Word-chunks it writes back 2. Output tokens (the word-chunks it writes back, billed at about five times the reading rate) 3. Output tokens ### Prompt caching 1. A way to have it remember the part that never changes so you stop paying full price to re-send it 2. Prompt caching (marking the part of the message that never changes so the vendor re-uses it at a tenth of the price) 3. Prompt caching (cache_control breakpoints; reads 0.1x input, 5-min writes 1.25x) ### Batch api 1. The overnight queue: half price if the job can wait until tomorrow 2. The Batch API (a queue that runs your jobs within a day for half price) 3. The Batch API (async, 24h window, 50% off input and output) ### Benchmark 1. A fixed set of test questions with known answers, used to compare AIs 2. A benchmark (a fixed task set with expected answers and a scoring rule, so every model is graded the same way) 3. A benchmark (versioned task set + gold answers + scorer; public ones leak into training data) ### Rubric 1. The marking guide that says what counts as a right answer 2. A rubric (written criteria for grading an answer when exact match is not possible) 3. A rubric (graded criteria; model-graded with human audit of a sample) ### Held out set 1. Test questions you keep locked away so the AI never practises on them 2. A held-out set (examples kept out of training so the score on them is honest) 3. A held-out set (never in the training split; version it) ### Contamination 1. When the AI has already seen the test questions during training 2. Contamination (test items that leaked into the model's training data, so the score measures memory, not skill) 3. Benchmark contamination (test-set leakage into pretraining; inflates public scores) ### Expected calibration error 1. One number for how far the AI's sureness is from its real hit rate 2. Expected calibration error, ECE (bucket answers by claimed confidence, compare each bucket's accuracy to its confidence, average the gaps) 3. ECE (count-weighted mean |confidence − accuracy| over bins; Guo et al. 2017) ### Sampling 1. The small dice roll it makes when picking each word 2. Sampling (the small dice roll when it picks each token) 3. Sampling (stochastic decoding at temperature > 0; same prompt, different tokens) ## Training and the far edge ### Retrieval 1. Handing the AI the right pages from your files before it answers 2. Retrieval, or RAG (search your own documents for the relevant pages and put them in front of the model before it answers) 3. Retrieval (RAG: index, retrieve top-k, inject into context; cite the retriever) ### Weights 1. The AI's built-in memory from training, fixed once training ends 2. The weights (the billions of numbers inside the model that training set and fine-tuning nudges) 3. The weights (checkpoint parameters; what SFT/LoRA modifies) ### Genotype 1. The plant's DNA fingerprint 2. The genotype (which version of each DNA marker the plant carries) 3. The genotype (SNP calls per marker; array or low-pass sequencing) ### Phenotype 1. What the plant actually did: height, yield, smell, test results 2. The phenotype (the measured traits of a grown plant, from height to COA numbers) 3. The phenotype (measured trait values per individual per environment) ### Genomic prediction 1. Predicting how a plant will turn out from its DNA before you grow it 2. Genomic prediction (a model trained on plants with both DNA and measurements, used to predict traits for a seedling that only has DNA) 3. Genomic prediction (GBLUP / Bayesian or ML regressors from marker matrix to trait; accuracy bounded by training set size and heritability) ### Heritability 1. How much of a trait comes from the genes rather than the room 2. Heritability (the share of a trait's variation that genetics explains, the rest is environment and noise) 3. Heritability (h²; the ceiling on marker-based prediction accuracy for that trait) ## Other ### Machine learning 1. Software that learns a pattern from past examples instead of being told the rules 2. Machine learning (software that learns its rules from examples instead of having them written in) 3. Machine learning (parameters fitted to data by minimising a loss) ### Expert system 1. A program that follows a long list of if-this-then-that rules written by an expert 2. An expert system (hand-written if-then rules captured from a specialist) 3. An expert system (symbolic rule base plus inference engine) ### Neural network 1. A big web of tiny number-crunching deciders, each passing its result to the next 2. A neural network (layers of simple units, each weighing its inputs and passing a number on) 3. A neural network (stacked affine maps with non-linear activations) ### Perceptron 1. One tiny decider that weighs a few facts and says yes or no 2. A perceptron (a single artificial neuron: weigh the inputs, add them up, say yes above zero) 3. A perceptron (y = sign(w·x + b), trained with the perceptron rule) ### Feature 1. One fact about each example, like price or THC percent 2. A feature (one measured fact about each example that the model gets to look at) 3. A feature (an input dimension of x) ### Label 1. The right answer written next to each past example 2. A label (the known right answer attached to each training example) 3. A label (target t for a supervised example) ### Learning rate 1. How big a correction it makes each time it gets one wrong 2. The learning rate (how far each correction moves the weights) 3. The learning rate η (step size of each weight update) ### Gradient descent 1. A way of getting less wrong by taking small steps downhill on a map of mistakes 2. Gradient descent (repeatedly nudging every weight in whichever direction lowers the error) 3. Gradient descent (w ← w − η∇L(w)) ### Deep learning 1. Machine learning with very big webs of deciders, many layers deep 2. Deep learning (neural networks with many layers, trained on a lot of data) 3. Deep learning (many-layer networks trained end to end by backprop) ### Embedding 1. Its meaning written as a long list of numbers 2. An embedding (a long list of numbers standing for what a piece of text means) 3. An embedding (a dense vector; nearby vectors mean similar text) ### Cosine similarity 1. A closeness score 2. Cosine similarity (a score for how closely two lists of numbers point the same way) 3. Cosine similarity (a·b / ‖a‖‖b‖) ### Overfitting 1. Memorising the examples instead of learning the pattern 2. Overfitting (fitting the noise in the training examples, so new cases go worse) 3. Overfitting (low training error, high held-out error; excess variance) ### Attention 1. A way for each word to look back at the other words that explain it 2. Attention (each word weighs every other word in the text to work out what it refers to) 3. Attention (softmax(QKᵀ/√d)V over the context) ### Transformer 1. The design behind today's chat AIs, built around looking back at every word at once 2. A transformer (the model design behind modern LLMs, built on attention) 3. A transformer (stacked self-attention and feed-forward blocks) ### Temperature 1. A randomness dial 2. Temperature (a setting for how adventurous the AI is when picking each next word) 3. Temperature (divides logits before softmax; T→0 is greedy) ### Prompt injection 1. A message with hidden orders for the AI inside it 2. Prompt injection (text in an email, document or web page that tries to give the AI new instructions) 3. Prompt injection (untrusted input interpreted as instructions; direct or indirect)