The AI Glossary

A system of connected layers made of neurons that processes data and learns by adjusting the strength of connections (weights) between neurons.

The basic unit of a neural network. Takes in numbers, multiplies them by weights, adds them up, then decides whether to "fire" a signal using an activation function.

A number stored in a connection between neurons that controls how much influence one neuron has on another. Training = adjusting millions (or billions) of weights until outputs are accurate.

Using a model pre-trained on a large dataset as the starting point for a new task. Like starting a new job already knowing how to type — you adapt existing skills instead of starting from scratch.

Breaking text into smaller units (tokens) that a model can process. "unhappy" might become ["un", "happy"]. Most LLMs use subword tokenization — roughly 1 token ≈ ¾ of a word.

Converting tokens (or images, or audio) into vectors of numbers in a high-dimensional space where similar things sit close together. "King" and "Queen" are far from "Table" but close to each other.

A mechanism that lets every token look at every other token and decide what's most relevant. When processing "it" in a sentence, attention figures out what "it" refers to by weighing all the other words.

The architecture behind almost every modern AI model. Introduced in "Attention is All You Need" (2017). Stacks of attention layers that process entire sequences in parallel — far more efficient than older RNNs.

Large Language Model. A transformer trained on massive text datasets to predict the next token. GPT, Claude, Gemini, Llama are all LLMs. "Large" refers to the number of parameters (weights).

The maximum amount of text an LLM can "see" at once (its working memory). GPT-4 has ~128K tokens; Claude 3 up to 200K. Text outside the window is forgotten entirely.

A setting (0–2) that controls how random an LLM's outputs are. Temperature 0 = deterministic (always picks the most likely next token). Temperature 1+ = creative but sometimes chaotic.

When an LLM confidently states something false. Happens because LLMs predict plausible text, not verified facts. The model has no internal "truth checker" — it generates the most statistically likely continuation.

Taking a pre-trained model and continuing to train it on a smaller, task-specific dataset. You keep most of the pre-trained knowledge but specialize the model for your use case.

Reinforcement Learning from Human Feedback. Humans rank model outputs, those rankings train a reward model, and then the LLM is tuned to maximize that reward. How ChatGPT became helpful instead of just predicting text.

Low-Rank Adaptation. A parameter-efficient fine-tuning method that adds small trainable matrices to each layer instead of updating all billions of weights. 100× cheaper than full fine-tuning, nearly the same result.

Reducing the precision of a model's weights (e.g. 32-bit floats → 4-bit integers) to shrink memory usage and speed up inference. Makes it possible to run large models on a laptop GPU.

The practice of crafting inputs to AI models to get better outputs. Techniques include few-shot examples, chain-of-thought instructions, role assignment, and output format specification.

Retrieval-Augmented Generation. Instead of relying solely on training data, the model retrieves relevant documents at inference time and uses them as context. Reduces hallucination and adds up-to-date knowledge.

A database optimized for storing and searching embeddings (high-dimensional vectors). Used in RAG pipelines to find the most semantically similar documents to a query. Examples: Pinecone, Weaviate, pgvector.

An LLM given access to tools (search, code execution, APIs, file system) and a feedback loop so it can plan multi-step tasks, take actions, observe results, and adjust — rather than just responding once.

An image generation technique that adds random noise to training images step-by-step, then trains a model to reverse that process. At inference, start from pure noise and denoise toward a coherent image. Powers DALL-E, Midjourney, Stable Diffusion.

When an AI system produces systematically unfair or prejudiced results for certain groups. Usually caused by biased training data or flawed objective functions. Can affect hiring, lending, healthcare, and more.

The challenge of making AI systems do what humans actually intend, not just what was literally specified. A misaligned AI might achieve its stated goal in ways that violate human values or cause unintended harm.

Deliberately trying to break or find harmful outputs in an AI system before it's deployed. Teams attempt jailbreaks, adversarial prompts, and edge cases to surface safety failures. Standard practice at major AI labs.

Anthropic's approach where the AI critiques its own outputs against a set of written principles (the "constitution") before responding. Reduces the need for human labellers while still aligning behaviour.

AI that can process and generate multiple types of data simultaneously — text, images, audio, and video. GPT-4V, Gemini, and Claude 3 are multimodal. The transformer architecture scales to all of these.

A very large model trained on broad data that can be adapted (fine-tuned) for many downstream tasks. GPT-4, Claude, Gemini are all foundation models. The "general" model that specialists are built on top of.

Artificial General Intelligence — a hypothetical AI that can perform any cognitive task a human can, at or above human level. No consensus on whether or when it will exist. The timeline is genuinely contested among experts.

AI systems that act autonomously over long time horizons — using tools, browsing the web, writing code, spawning sub-agents — to complete complex multi-step tasks with minimal human oversight.