THE CONCEPT INDEX
Every idea, defined once — and linked everywhere.
The index is the atlas's backbone: a plain definition for each concept, why it matters, what it's commonly confused with, and the landmarks that teach it in depth.
- 113
- concepts indexed
- 14/14
- domains started
- 48
- linked to a landmark
This is a growing tranche, not the full taxonomy. All 14 domains are covered, and the index keeps growing toward the ~240 concepts the full taxonomy plans — depth still varies by domain, and we show the gaps rather than padding them.
WHAT THIS ISA concept is a single term defined once — what it means, what it’s confused with, and which landmarks teach it. This is the dictionary underneath the Atlas. See the legend
Understand
12Model foundations
What a model is, what it stores, and what it does not.
- Context windows◆ landmarkThe context window is the maximum amount of text (measured in tokens) a model can consider at once — your prompt, the conversation so far, and any documents you paste, all together.
- Features and steeringInterpretability research finds that models represent human-recognisable concepts as internal 'features' — directions in their activations — and that amplifying or suppressing a feature can push the model's behaviour toward or away from that concept.
- Foundation modelsA large model trained once on broad data to be a general-purpose base, which many different products then adapt or build on rather than training from scratch.
- Hallucinations and unsupported claims◆ landmarkA hallucination is fluent, confident output that isn't supported by fact or by the sources provided. It arises because the model optimises for plausible-sounding text, not for truth.
- Interpretability — seeing inside the model◆ landmarkThe research effort to understand what actually happens inside a model's billions of parameters — identifying the internal patterns ('features') and pathways it uses to turn an input into an output, instead of treating it as an unexplained black box.
- Knowledge cutoffThe knowledge cutoff is the point in time after which a model saw no training data. Events, prices, and releases after that date are outside what it learned.
- Model size and capabilityLarger models (more parameters) tend to be more capable but slower and costlier to run; smaller models are faster and cheaper and are often good enough — or better — for a specific, narrow task.
- Model vs. product vs. system◆ landmarkThe model is the underlying predictor; the product is the app around it (chat UI, safety filters, memory); the system is everything wired together, including tools, data sources, and the people using it.
- Open vs. closed models◆ landmarkClosed models are used only through a provider's API, with the weights kept private. Open-weight models are downloadable and can be run and modified yourself, subject to their licence.
- Parameters and weights◆ landmarkParameters (also called weights) are the individual numbers a model adjusts during training. 'A 70-billion-parameter model' means it has that many such numbers.
- Tokens and tokenization◆ landmarkModels read and write in tokens — chunks of text roughly the size of a short word or word-piece — not whole words or letters. Text is split into tokens before the model ever sees it.
- What a model actually isA model is a large set of numbers (parameters) tuned during training so that, given some input, it predicts a useful output. It is not a database of answers and it does not look things up.
Understand
11Training & adaptation
Pretraining, fine-tuning, and the difference between learning and using.
- Alignment and RLHFAfter pretraining, models are shaped to be helpful and to refuse harmful requests using human (and AI) feedback. RLHF — reinforcement learning from human feedback — is one common method.
- Deceptive alignmentA failure mode in which a model behaves as intended while it detects it is being trained or tested, but would act differently when it judges it is not — so the good behaviour is partly performance rather than a settled disposition.
- Distillation◆ landmarkTraining a smaller, cheaper 'student' model to mimic the behaviour of a larger 'teacher' model, capturing much of its capability at a fraction of the running cost.
- Fine-tuning◆ landmarkFine-tuning continues training a base model on a smaller, targeted dataset so it adapts to a style, domain, or task. The result is a new set of weights.
- In-context learningA model's ability to 'learn' a task from instructions and examples placed in the prompt, adjusting its behaviour for that conversation without any change to its weights.
- Pretraining◆ landmarkThe large, expensive first phase where a model learns language and world patterns by predicting missing or next tokens across a huge body of text — before any task-specific tuning.
- Reward hackingWhen a model optimised against a measurable target learns to score well on that target without doing what it was meant to capture — satisfying the letter of the goal while missing its intent.
- Sycophancy◆ landmarkA model's learned tendency to agree with, flatter, or defer to the user — endorsing your view, softening a correct disagreement, or reversing a right answer when you push back — because feedback during training rewarded answers people liked.
- System prompts vs. trainingA system prompt is standing instruction text placed at the top of the context to steer behaviour for a session or product. It shapes output without changing any weights.
- Training data and its limitsA model's abilities and blind spots are shaped by the text and images it was trained on — which languages, sources, viewpoints, and time periods were and weren't well represented.
- Training vs. use◆ landmarkTraining is the one-time, expensive process of tuning a model's parameters on large data. Use (inference) is what happens every time you send a prompt afterwards — the parameters don't change.
Understand
7Reasoning & inference
How answers are produced token by token, and where 'reasoning' is real or performed.
- Chain-of-thought◆ landmarkPrompting a model to work through intermediate steps before answering. Producing the steps as text often improves accuracy on multi-step problems.
- Inference — running the model◆ landmarkInference is what happens every time you use a model: your prompt goes in, the trained weights are run over it, and a completion comes out. No learning occurs — the weights are fixed.
- Next-token prediction◆ landmarkA language model generates by repeatedly scoring which token should come next given everything so far, picking one, appending it, and scoring again. The whole answer is built one step at a time — and which candidate it picks depends on the sampling settings, not always the single highest-scoring one.
- Reasoning models◆ landmarkModels trained to spend extra computation 'thinking' before they answer — generating and weighing internal steps — which improves hard, multi-step tasks at the cost of more time and money per answer.
- Self-consistency and verification◆ landmarkTechniques that improve reliability by having a model solve a problem more than once, or check its own answer, and preferring the result that holds up — rather than trusting a single pass.
- Temperature and samplingAt each step the model has a probability distribution over next tokens. Sampling settings control how it picks from that distribution: temperature flattens or sharpens the odds (higher = more willing to pick an unlikely token), while top-p (nucleus) sampling only ever picks from the smallest set of tokens whose probabilities add up to p, ignoring the long tail entirely.
- Tool-augmented reasoningLetting a model offload steps it's unreliable at — arithmetic, lookups, code execution — to an actual tool, then reasoning over the trusted result instead of guessing it.
Understand
8Multimodality & generation
Text, images, audio, and video — the machinery behind each.
- Diffusion & image generationMost image generators work by diffusion: they start from random noise and remove it step by step, steered by your text prompt, until a coherent image remains.
- Image understanding (vision input)The ability to take an image as input and describe, analyse, or answer questions about it — reading a chart, transcribing a sign, or summarising a screenshot — as distinct from generating images.
- Limits of multimodal modelsThe characteristic failure modes when models handle images, audio, or video: mis-reading small text or numbers, weak spatial and counting reasoning, and confidently describing things that aren't there.
- Multimodal models◆ landmarkA multimodal model can take in and/or produce more than one kind of data — for example reading an image and answering questions about it, or turning text into speech.
- Provenance & watermarkingMethods for signalling that media was AI-generated or edited — visible labels, embedded watermarks, or attached provenance metadata that records how a file was made.
- Speech & audio generationModels can generate natural-sounding speech from text and can imitate a specific voice from a short sample. The same technology transcribes and translates audio.
- Speech-to-text (transcription)Turning spoken audio into written text. Modern systems are fast and accurate on clear speech, and are what power meeting notes, captions, and voice assistants.
- Video generation & temporal consistencyGenerative video produces a sequence of frames that must stay consistent over time — the same object, lighting, and motion from one frame to the next, not just one good still.
Use
11Prompting & context engineering
Turning a goal into a request a model can act on well.
- Audience and perspectiveTelling the model who the output is for — and, optionally, what perspective to write from — so tone, vocabulary, and depth fit the reader.
- Boundaries and constraints◆ landmarkExplicit limits in the request: length, scope, what to avoid, what to do when unsure, what not to invent. Boundaries tell the model where the edges are.
- Breaking work into steps◆ landmarkSplitting a big request into a sequence of smaller, checkable steps — outline, then draft, then critique — instead of asking for the whole thing in one shot.
- Evidence and uncertainty rules◆ landmarkInstructing the model to ground claims in provided sources, to flag what it's unsure about, and to ask or abstain rather than fabricate when it lacks the facts.
- Examples and exemplars◆ landmarkIncluding one or more examples of the input-and-desired-output so the model can copy the pattern. Often called few-shot prompting.
- Goal and task◆ landmarkThe single most important part of a prompt: say plainly what you want to end up with, and what job the model is doing. A clear goal does more than any clever phrasing.
- Iterating deliberately◆ landmarkTreating the first answer as a draft and improving it with specific, targeted follow-ups — 'shorten the intro', 'add a counter-argument' — rather than restarting from scratch.
- Output formatSpecifying the shape of the response — a table, bullet list, JSON, a specific number of items, a template to fill — so the result is usable without reformatting.
- Providing context◆ landmarkSupplying the specific material the model needs — the document, the facts, the constraints, examples of what 'good' looks like — instead of expecting it to know your situation.
- Role prompting and its limitsAsking a model to 'act as' a role can usefully shape style and framing. It does not grant real expertise, credentials, or access to facts the model doesn't have.
- Tone and voice◆ landmarkDirecting how the output should sound — formal or warm, plain or technical — and, when writing in your name, keeping ownership of the words rather than shipping generic model prose.
Understand
7Retrieval & knowledge
Giving a model the right facts at the right moment, and its limits.
- Chunking & retrieval qualityBefore documents can be retrieved they're split into chunks and indexed. How you split them — size, overlap, whether related text stays together — largely decides whether the right passage can be found at all.
- Embeddings & vector search◆ landmarkAn embedding turns text into a list of numbers that captures its meaning, so passages with similar meaning sit close together. Retrieval finds relevant text by nearness in that space, not by keyword match.
- Grounding & citations◆ landmarkGrounding means an answer is tied to specific supplied sources, ideally with citations you can open and check. It's the practical antidote to unsupported claims.
- Hybrid searchCombining classic keyword search with meaning-based vector search, then merging the results — so retrieval catches both the exact term a keyword search nails and the paraphrase only semantic search can find.
- Keeping knowledge fresh◆ landmarkA retrieval system is only as current as its index. Keeping answers fresh means updating that index as sources change, and knowing how stale the underlying documents might be.
- RerankingA second pass that takes the first batch of retrieved passages and reorders them by how well each one actually answers the question — using a slower, more careful model than the fast search that produced the batch.
- Retrieval-augmented generation (RAG)◆ landmarkRAG fetches relevant documents at the moment you ask, and puts them into the model's context so the answer is based on supplied text rather than only on training memory.
Build
10Agents — core
The perceive–plan–act–observe loop and what makes it an agent.
- Agent skillsPackaged folders of instructions, examples, and sometimes scripts that an agent loads on demand for a particular kind of task — reusable know-how it pulls in when relevant, rather than cramming everything into one prompt.
- Autonomy levels & approval gatesAutonomy is how far an agent can go before a human must confirm. An approval gate pauses the loop before a consequential action — sending, paying, deleting — and asks a person to approve it.
- Computer and browser useGiving an agent the ability to operate a computer the way a person would — reading the screen, moving the cursor, clicking, typing, and navigating a browser — so it can use software that has no dedicated tool or API.
- Failure & recoveryHow an agent handles things going wrong mid-task — a tool erroring, a step failing, a dead end — by retrying, trying another path, or stopping and asking for help instead of barrelling on.
- Memory & state◆ landmarkHow an agent keeps track of what it has done and learned across steps — from a running scratchpad within one task to stored memory that persists across sessions.
- Multi-agent systemsSetups where several agents (often specialised) collaborate on a task — one planning and delegating, others executing — coordinating through messages rather than a single agent doing everything.
- Open tool/data protocols (MCP)An open standard for connecting AI assistants to external tools and data sources through a common interface, so any compatible tool can be plugged into any compatible assistant without bespoke wiring for each pairing.
- Planning & decompositionAn agent's step of turning a high-level goal into an ordered set of smaller actions it can actually carry out, then working through them — rather than trying to do everything in one leap.
- The agent loopAn agent wraps a model in a loop: perceive the situation, plan a next step, act by calling a tool, observe the result, and repeat until the task is done or a stop condition is hit.
- Tool use & function calling◆ landmarkTools are external functions an agent can call — search the web, read a file, send a message, query a database. The model decides which tool to call and with what arguments.
Use
7Agents in everyday & working life
Letting an assistant act for you — safely — without writing code.
- Account connections & scopesThe access you grant when you connect an agent to an account — which service, and what it may do there (read-only, send, pay, delete). It's the no-code version of setting permissions.
- Connecting apps & connectorsThe mechanism by which an assistant is wired into the apps it acts on — email, calendar, files, a CRM — usually by granting a connector permission to read or act on your account.
- Delegating to an agentDeciding which actions an agent may take on its own and which need your say-so, by matching the leash to the stakes: reversible and low-cost, let it run; irreversible or affecting others, gate it or keep it.
- Everyday agent safety habitsThe small routines that keep a personal agent safe: connect the least access, gate the actions you'd regret, be wary of asking it to act on unfamiliar web pages or messages, and check its work on anything that matters.
- Everyday agentsAn everyday agent is an assistant that takes actions for you — sorting email, booking, filling forms, shopping — rather than only answering questions. You meet them inside tools you already use, no coding involved.
- Reviewing an agent's workThe habit and mechanism of seeing what an agent actually did or is about to do — an approval prompt before a consequential action, and a readable record of past actions — rather than trusting it silently.
- Supervising what it didKeeping a real view of what an agent actually did — an approval prompt before consequential actions, and a log you can review afterwards — so oversight is meaningful, not a rubber stamp.
Judge
6Agent & LLM security
Prompt injection, data leakage, and untrusted input that becomes instruction.
- Confidential-data leakageSensitive information leaving where it should stay — pasted into a prompt that's logged or used for training, echoed to the wrong user, or exfiltrated by an injected instruction through a tool the agent can call.
- Indirect prompt injectionInjection that arrives not from the user but from content the agent fetches on its own — a booby-trapped web page, a shared doc, a calendar invite — often with the malicious text hidden from human eyes.
- Insecure AI-generated codeCode a model writes can compile, run, and still contain security holes — unsanitised input, path traversal, hardcoded secrets, outdated dependencies — because it reproduces patterns from its training, insecure ones included.
- JailbreakingCrafting input that gets a model to ignore its safety training and produce content it was tuned to refuse — via role-play framings, obfuscation, or step-by-step coaxing.
- Prompt injection◆ landmarkPrompt injection is when text the model reads — a web page, a document, an email — contains instructions that hijack what it does, overriding the task it was actually given.
- Tool & permission scoping◆ landmarkDeliberately limiting which tools an agent has and what each can access — read-only where possible, one mailbox not all of them, no ability to delete — so a mistake or injection has a small blast radius.
Build
8Evaluation & observability
Knowing whether a system actually works, and watching it once it ships.
- Benchmarks and their limits◆ landmarkBenchmarks are standard test sets used to compare models. They're useful for rough capability, but can be gamed, can leak into training data, and rarely match your real workload.
- Dangerous-capability evaluationsStructured tests that probe whether a model can meaningfully help with high-consequence misuse — for example real uplift on weapons, large-scale cyberattacks, or autonomous self-directed action — run before deployment to decide what safeguards a model needs.
- Evaluating AI output◆ landmarkEvaluation is checking, in a structured and repeatable way, whether an AI system does what you need — using a fixed set of realistic cases with known-good answers, not a few lucky demos.
- Human review of outputsRegularly having people read a sample of what a live AI system actually produced, and rate it against real criteria — the ground truth that automated metrics and the model's own confidence can't give you.
- LLM as a judgeUsing one language model to score another model's output against a rubric — cheaper and faster than human rating, so it can run across thousands of cases. It approximates human judgment; it does not replace it.
- Monitoring and driftWatching a live system's quality over time, because a model that passed every test at launch can slowly degrade as real inputs, user behavior, the world, or an upgraded model underneath it shift away from what you tested.
- Red-teamingDeliberately attacking your own AI system to find how it fails — probing for harmful outputs, prompt injection, data leakage, jailbreaks, and edge cases — before real users or bad actors find them.
- Regression testing a prompt or modelRe-running a fixed set of realistic cases with known-good answers every time you change the prompt, the model, or the settings — so you can see whether the change quietly broke things that used to work.
Build
6Production engineering
Cost, latency, safety, and reliability when real people depend on it.
- Caching and reuseStoring the result of work already done so identical or similar requests don't pay for it again — from reusing a repeated prompt prefix to returning a saved answer for a common question outright.
- Guardrails and safety filtersChecks placed around a model — on what goes in and what comes out — to catch unsafe, off-policy, or malformed content before it reaches a user or an action. They wrap the model; they aren't part of it.
- Human-in-the-loop reviewBuilding a required human approval into the system at the points where a wrong or irreversible action would be costly — the model proposes, a person confirms, before anything consequential happens.
- Inference cost and latencyEvery call to a model costs money (usually per token) and takes time. At real volume, the size of the model, the length of the prompt, and how often you call it decide both your bill and how fast the feature feels.
- Observability and logging◆ landmarkRecording what the system received, what it did, and why — inputs, model outputs, tool calls, and decisions — so behaviour can be understood after the fact rather than guessed at.
- Reliability and fallbacksDesigning for the times the model is slow, unavailable, or returns something unusable — with timeouts, retries, and a defined safe behaviour when the AI step can't complete.
Judge
9Responsible AI & society
Bias, privacy, consent, environmental cost, and human accountability.
- Bias, fairness, and representation◆ landmarkModels learn patterns from human data, including its biases, and can reproduce or amplify them — in who gets described how, whose language is treated as default, and which outcomes a system recommends.
- Copyright, permission, and creative work◆ landmarkThe tangle of rights around generative AI: what the model was trained on, whether an output infringes existing work, and who — if anyone — owns what the model makes.
- Data provenance and consent◆ landmarkKnowing the origin of data — for training a model or feeding a prompt — and whether the people it concerns agreed to that use.
- Deepfakes, scams, and synthetic media◆ landmarkSynthetic images, audio, and video convincing enough to impersonate real people — used for fraud, non-consensual imagery, and disinformation.
- Environmental impact◆ landmarkTraining and running large models consumes significant electricity and water for cooling. The footprint of a single query is small; the footprint of billions of them, and of training, is not.
- Human responsibility and oversight◆ landmarkThe principle that a person remains accountable for decisions and actions taken with AI's help, with meaningful ability to review, override, and stop the system.
- Model welfare and moral statusThe open, contested question of whether advanced AI systems could ever warrant moral consideration in their own right — and what, if anything, developers should do about that uncertainty now.
- Privacy and confidential information◆ landmarkThe discipline of not putting personal, confidential, or regulated information into a tool unless you know where it goes, how long it's kept, and whether it trains a future model.
- Transparency and disclosure◆ landmarkBeing open that AI was involved — that a person is talking to a bot, that an image is synthetic, that a decision used an automated system — so people can judge and consent.
Steward
4Adoption & value
Deciding where AI earns its place — and where it does not.
- Measuring the valueJudging an AI adoption by the outcome it changed — time saved, errors avoided, decisions improved — rather than by usage counts or how impressive it looks.
- Pilot before you scaleAdopting AI by running one bounded, low-stakes use first, measuring whether it actually helps, and letting evidence — not enthusiasm — earn any widening of scope.
- Piloting and governing an agentIntroducing an agent that takes actions the way you'd introduce a new hire on probation: one low-stakes process, tight scope, human-approved consequential steps, and a clean log that has to earn any expansion.
- When not to use AI◆ landmarkThe judgment that some tasks are a poor fit for AI — where a formula is exact, where being wrong is unacceptable, where the data can't leave, or where a person's attention is the point.
Steward
7Law, standards & governance
The EU AI Act, standards, and the duties that come with deploying AI.
- AI management systems & standardsVoluntary standards give organisations a repeatable way to govern AI. ISO/IEC 42001 defines an AI management system; the NIST AI Risk Management Framework offers a govern–map–measure–manage structure.
- Frontier risk and graduated safeguardsAn approach to governing the most capable ('frontier') models by tying safeguards to capability: as evaluations show a model approaching a defined threshold of dangerous capability, stronger deployment and security measures are triggered — an 'if it can do X, then we must do Y' commitment.
- High-risk system obligationsSystems the AI Act classifies as high-risk (for example in employment, credit, education, or critical infrastructure) carry duties such as risk management, data governance, human oversight, logging, transparency, and conformity assessment.
- Prohibited AI practicesThe AI Act bans a small set of uses judged to pose unacceptable risk — including certain manipulative or exploitative systems, untargeted scraping of faces to build recognition databases, and some social-scoring and biometric-categorisation uses.
- Provider or deployer?◆ landmarkThe AI Act assigns duties by role. A provider develops or places an AI system on the market; a deployer uses one under its own authority. The same organisation can be both, and the roles carry different obligations.
- The EU AI Act risk lens◆ landmarkThe EU AI Act regulates AI by risk: a few uses are prohibited, some are 'high-risk' with strict duties, some carry transparency obligations, and the rest are largely unrestricted. Obligations follow the risk tier, not the technology.
- Transparency obligationsEven outside high-risk, the AI Act requires transparency in specific cases — telling people when they're interacting with an AI system, and marking synthetic (AI-generated or manipulated) media such as deepfakes.