The AI Field Guideby Stephen Hunt's SFcrowsnest
The AI Field Guide / L
8 terms, explained without the techno-murk.
The answer or category attached to a training example.
A label is like the answer written on the back of a flashcard. It tells supervised learning what the example represents, such as 'spam', 'cat' or a particular price. Incorrect or inconsistent labels teach the wrong lesson.
A photograph is paired with the label 'pedestrian' for training a vision model.
A model trained on vast amounts of language to work with text and, often, other media.
An LLM learns statistical patterns that help it predict and generate sequences of tokens. It can summarise, translate, answer and write, but fluent language should not be mistaken for guaranteed understanding or truth.
The model behind a general-purpose writing chatbot is usually an LLM.
The delay between making a request and receiving a response.
Latency depends on model size, workload, connection speed and how much reasoning or output is required. Lower latency makes a tool feel more responsive.
A voice assistant needs low latency so the conversation does not feel awkwardly slow.
A compressed mathematical map of patterns an AI model has learned.
Latent means hidden. On this map, items with similar learned qualities often sit near one another, even though its directions may not have simple human labels. Moving through the space can change features of generated content in gradual ways.
An image model may place pictures with similar poses or styles near each other in latent space.
A training setting that controls how large each model update should be.
Imagine walking downhill in fog. Steps that are too large may overshoot the valley; steps that are too small may take forever. The learning rate controls the size of those steps while gradient descent tries to reduce error.
A developer lowers the learning rate when training jumps around instead of settling toward a good result.
A type of neural network built to keep useful information across a sequence.
An LSTM has gates that act like decisions about what to remember, forget and pass on. It was widely used for language, speech and time-series data before transformers became dominant in many of those tasks.
An LSTM uses earlier readings from a sensor to help predict its next reading.
A resource-saving way to adapt a model without retraining all of its weights.
LoRA stands for low-rank adaptation. It learns a relatively small set of additions to a model, making customisation cheaper and the resulting adapters easier to store and swap.
An image model uses a small LoRA file to learn a particular visual style.
A mathematical score that tells training how good or bad the model's result was.
It acts like a scoreboard. Training adjusts the model to reduce the loss, or more generally to improve an objective. The two phrases often overlap, although an objective can also include extra goals or penalties beyond prediction error.
A large loss tells an image classifier that its prediction was far from the correct label.