Glossario
Definizioni brevi della redazione. Non sono un saggio citato.
Queste definizioni sono in inglese.
- Agent
- Software that can take a sequence of steps toward a task, often by calling other tools. A person is still expected to check the result.
- Large language model
- A model that predicts the next token from the text it was given. It is not a database of checked facts.
- Token
- A chunk of text a model reads or writes. A token is not the same thing as a word.
- Context window
- The amount of text a model can take in at one time.
- Prompt
- The instruction and the source text given to a model for one run.
- Hallucination
- A fluent statement from a model that is not supported by the sources it was given.
- Retrieval-augmented generation
- A setup that fetches documents first and then asks a model to answer from those documents.
- Embedding
- A list of numbers that stands for a piece of text, so a system can compare which texts are close.
- Fine-tuning
- Further training of an existing model on a narrower set of examples.
- Inference
- Running a trained model to produce an output. It is not the same step as training.
- Training
- Fitting a model’s parameters on data. A training claim counts here only when the publisher’s own document says it.
- Transformer
- A neural-network design, widely used in language models, that weighs which parts of the input matter for the next token.
- Parameter
- A stored number inside a trained model. A parameter count is not a measure of truthfulness.
- Weights
- The stored parameters of a trained model.
- Open weights
- Model weights the publisher lets others download and run. The license still sets the rules.
- API
- A documented way for one program to request a result from another.
- Benchmark
- A fixed test used to compare models. A benchmark score is not a field report.
- Model card
- A document from a model’s publisher that describes intended use, limits, and evaluation.
- Alignment
- Work aimed at making a model’s behavior match a stated set of rules. It is not proof that the model understands.
- RLHF
- Reinforcement learning from human feedback: a training step that uses people’s preferences between outputs.
- Multimodal
- A model that accepts more than one kind of input, such as text and images.
- Diffusion model
- A family of image models that learns to turn noise into a picture.
- Temperature
- A setting that changes how varied a model’s next-token choices are. It is not a measure of accuracy.
- Synthetic data
- Data produced by a model, rather than collected from the event being described.
- Latency
- The time between a request and a result.
- SEO
- Work aimed at how a page appears in classic web search.
- GEO
- A label for work aimed at being cited by answer engines. The label is not itself a standard.
- Citation
- A named source a reader can open. A citation is not an endorsement.