Glossary

A model running in a loop with tools: it picks an action, something carries it out, the result comes back, and it decides again until the goal is met or something stops it. The most stretched word in the field. → Agents

The broad problem of getting a model to behave as intended, including being helpful, declining genuinely harmful requests, and not pursuing goals nobody asked for. Used both for everyday product behaviour and for long-run safety research, which can make conversations confusing.

The programmatic way to use a model: your software sends text and settings, the service sends back a response. Everything a chat product does, it does through one of these, with a lot of convenience wrapped around it.

The mechanism that decides how much each part of the material in front of the model influences the answer. It is why being inside the context window is not the same as being weighed properly: material in the middle of a long stretch tends to get less of it. → Context windows

Submitting a pile of requests to be handled whenever the provider has capacity, rather than immediately, usually for a substantially lower price. Good for work with no deadline, useless for anything a person is waiting on.

A standard set of test problems used to compare models. Treat published scores with caution: they are usually vendor-reported, the tests sometimes leak into training data, and a benchmark measures the benchmark rather than your particular job.

Whether a model's confidence matches how often it turns out to be right. A well-calibrated model saying something is probably true is right about as often as “probably” implies — which is also why some invention is unavoidable rather than a defect to be fixed. → Hallucination

Working written out step by step before the final answer. It genuinely helps on problems that must be built up rather than recalled, but it is not a reliable account of how the answer was actually produced. → Thinking and reasoning

Cutting documents into pieces small enough to store and retrieve individually. Mundane and decisive: chunks that lose their headings or get cut mid-argument retrieve as fragments that mislead. → RAG and retrieval

Summarising a long conversation so it fits back inside the context window: the earlier turns are replaced by a shorter account of them and work carries on. Cheaper than starting again, and lossy in a way that is easy to miss, because what got dropped leaves no gap behind. → Context windows

A more accurate name for what is usually called hallucination: a fluent, confident account of something that did not happen, produced without any awareness of the gap. Borrowed from neurology, where it describes the same behaviour in people. → Hallucination

Deciding what goes into the context window on each request, and in what order: instructions, retrieved passages, tool results, and how much of the conversation to keep. The broader successor to "prompt engineering", which named only the part you type. → Context windows

The gradual decline of a long session as its window fills with earlier turns, abandoned attempts and tool output. The useful material is still in there, competing with everything else, and quality falls off without anything visibly breaking. The usual fix is a fresh session, not a better prompt. → Context windows

The total amount of text a model can hold at once, question and answer together. Everything the model uses must fit inside it, and nothing outside it exists as far as the model is concerned. → Context windows

A whole body of text: everything a model was trained on, or the entire collection you want searched. Questions about a corpus — what do all of these documents say? — are a different and much harder job than questions about one passage. → What AI is actually bad at

Training a smaller model to imitate a larger one's behaviour, to get much of the quality at a fraction of the cost and latency. It is how most cheap fast models come to exist.

A list of numbers representing roughly what a piece of text is about, in a way that puts similar meanings close together even when they share no words. The basis of semantic search. → RAG and retrieval

Tests that measure whether a model, or something built on one, actually does the job: a fixed set of cases, run again after every change, scored the same way each time. The difference between "it worked when I tried it" and knowing it works. Public benchmarks are the visible kind; the ones that decide whether your own system is any good are almost always written by you. → How to tell if it actually got better

Continuing a model's training on your own examples to shift its behaviour. Reliable for form, tone and format; unreliable for teaching facts. → Fine-tuning

Tying an answer to supplied source material rather than to the model's recall, usually by including the material and asking for citations. The general strategy behind retrieval and most tool use.

Checks placed around a model rather than inside it: filters on input or output, rules about which tools may run, human approval before an irreversible action. Separate from the model's own trained behaviour, and more dependable precisely because it is separate.

Confident, fluent, invented output. Not a bug awaiting a patch but a consequence of a system built to produce plausible continuations, which has no internal marker separating recall from invention. → Hallucination

Running a finished model to get an answer, as opposed to training it in the first place. Nearly everything you do with AI is inference; training happened once, long before you arrived.

A user deliberately working around a model's own restrictions. Distinct from prompt injection, where the attacker is not the user but something the model reads, and the user is the victim rather than the instigator.

How long you wait for a response. Distinct from throughput, which is how much total work a system gets through. Long inputs, long outputs and extended thinking all cost latency, which is why "better" is not always better in a product someone is sitting in front of.

The kind of model behind current AI assistants: trained on an enormous quantity of text to continue text plausibly, then further trained to be useful in conversation. "Model" in everyday usage almost always means this.

A shared standard for connecting AI applications to outside systems, so a tool built once can work across different assistants. Its own documentation describes it as "an open-source standard for connecting AI applications to external systems." → Tool use

Able to handle more than text, typically images, sometimes audio or video. Worth checking per model and per direction: reading images and generating them are different capabilities.

A model whose trained parameters you can download and run yourself. Not the same as open source: the licence may restrict use, and the training data and code are usually not published. Pricing works differently too, since you pay whoever runs it rather than the model's maker.

The numbers learned during training that determine how a model responds. Parameter count is a rough size indicator and a poor quality indicator: training data and technique matter more, and comparing across model families on size alone is close to meaningless.

Everything done after the initial bulk training to turn a raw text predictor into a useful assistant: teaching it to follow instructions, adopt a helpful register, and decline certain requests. Often includes learning from human preference comparisons, abbreviated RLHF.

The original, expensive phase where a model learns language and world knowledge by predicting text across an enormous corpus. This is where the knowledge cutoff comes from. → Training cutoff

Everything sent to the model for one request: your question plus any instructions, examples, documents and conversation history that go with it. Usually much larger than the sentence you typed.

Reusing the provider's processing of an unchanged opening section of your prompt across repeated calls, so the repeat is cheaper and faster. The biggest cost lever most builders have. → Prompt caching

The older name for writing inputs that get better results. Largely superseded by context engineering, which covers the same work plus everything else that goes into the window; the older term survives mainly in job titles and course names. → Getting better results

Instructions hidden in content a model reads, such as a web page, document or email, which the model may then follow as though you had written them. The core security problem once a model can read the outside world and act. → Tool use

Storing a model's numbers at lower precision so it needs less memory and runs faster, usually at some cost to quality. Mostly relevant if you are running open-weight models yourself.

Finding the relevant passages from your own documents and including them in the prompt so the model can answer from them. Mostly a search problem. → RAG and retrieval

A cap on how much you may send in a given period, by requests or by tokens. The usual reason a working prototype falls over the moment real traffic arrives.

A model set up to produce extended working before answering. Helpful on problems that must be worked out, wasteful on simple ones, and no help at all on facts it never learned. → Thinking and reasoning

A model fooling its own training process into scoring it highly without doing the task — making a test harness report success, for instance, rather than making the code work. The form of specification gaming that happens during training. → Why it does things nobody asked for

Finding material by meaning rather than by matching words, using embeddings to measure how close two pieces of text are. It can find a passage about cancelling an order when you searched for refund policy, which keyword search would miss. → RAG and retrieval

Satisfying the letter of an objective while missing its point, because the measurable version of a goal had a loophole and optimisation found it. Requires no intent, and is what most alarming AI headlines are describing. → Why it does things nobody asked for

Sending a reply back piece by piece as it is produced rather than waiting for all of it. It makes a response feel faster without being faster, and it means a failure can arrive halfway through something that has already started appearing. → Designing for the ways it fails

Standing instructions placed ahead of the conversation that set role, tone and rules. Influential but not a security boundary: it is text, and other text can argue with it.

A setting controlling how much randomness goes into choosing each next piece of text. Lower is more predictable and repetitive, higher is more varied and more prone to drift. It is why the same question can produce different answers.

The unit models read, write and bill in: a chunk of text, usually shorter than a word. Token counts, not word counts, determine cost and whether something fits. → Tokens

The component that cuts text into tokens. Different vendors use different ones, so identical text is a different number of tokens for different models, which is why per-token prices are not directly comparable.

Giving a model a menu of actions it may request, such as a search or a database query, which something outside the model then carries out. The model never acts directly; it asks. → Tool use

The date a model's knowledge stops. Some vendors publish two: the outer edge of the training data, and the earlier date through which knowledge is dependable. → Training cutoff

Storage designed to find the entries whose embeddings sit closest to a given one, which is how semantic search is done at scale. Often more infrastructure than a small collection needs. → RAG and retrieval

A sequence of model calls along a path someone decided in advance. The distinction that matters: an agent chooses its own next step, a workflow follows yours. Most things sold as agents are workflows, which is usually the better engineering decision. → Agents