glossary · 28 terms
The working glossary
The AI terms I use constantly in essays and at work, defined the way I actually mean them. Plain language, a practitioner’s angle, and an opinion where one is earned.
last reviewed · August 2026
the basics · 8 terms
model#
Software that learned its behavior from examples instead of being programmed rule by rule. Training tunes billions of internal numbers until the outputs look right. On this page, and in most current conversation, "model" means a large language model.
see: large language model · parameters
large language modelLLM#
A model trained on a huge body of text to predict the next token. That single skill, scaled far enough, turns out to cover drafting, summarizing, translating, and writing working code. Claude, GPT, Gemini, and Llama are all LLMs.
see: token · pre-training
classifier#
A model that assigns a label instead of generating text: spam or not spam, positive or negative, safe or unsafe. Classifiers are smaller, cheaper, and older than LLMs, and they still carry a large share of production machine learning. Not every problem needs a conversation.
see: model · large language model
token#
The unit a model reads and writes: a chunk of text, roughly three-quarters of an English word on average. Context windows, API pricing, and speed are all measured in tokens, which is why the word appears on every AI invoice.
see: context window · inference
parameters#
The adjustable numbers inside a model. Training sets them, and inference uses them. Parameter count is the standard size measure, so "a 70B model" has 70 billion parameters. More parameters buy capacity, but they also raise memory and cost, and past a point the training data matters more than the count.
see: weights · training data
weights#
The learned values of a model’s parameters, saved as a file. The distinction is subtle and mostly ignored: parameters are the slots, weights are what training left in them. The file is the model. Hand someone the weights and they can run it anywhere.
see: parameters · open-weight model
training data#
The text, code, and images a model learned from. Its knowledge, gaps, biases, and cutoff date all trace back to this corpus. A model trained through January knows nothing that happened in February, no matter how confident it sounds.
see: pre-training · hallucination
context window#
The amount of text a model can consider at once, measured in tokens: your prompt, the conversation so far, and any documents you attach. Anything outside the window does not exist to the model. Windows keep growing, but models still attend unevenly across a long one, so what you put in the context, and where, is engineering work now.
how a model gets made · 5 terms
pre-training#
The first and most expensive phase, where a model learns general language and world knowledge from a massive corpus. It takes months of GPU time and produces a base model: capable, but raw and unhelpful until further training shapes it.
see: training data · fine-tuning
fine-tuning#
Additional training on a narrower dataset to specialize a pre-trained model: a company’s support tone, a domain like law, a strict output format. It costs a small fraction of pre-training because the heavy lifting is already done.
see: pre-training · distillation
RLHFreinforcement learning from human feedback#
People rank a model’s answers, and those rankings train the model toward responses people prefer. RLHF is much of the difference between a raw base model and an assistant you would let near your work.
see: pre-training
distillation#
Training a small model to imitate a larger one’s outputs. You give up some capability in exchange for a model that is far cheaper and faster to run. A lot of the industry’s falling cost curve is distillation doing quiet work.
see: quantization · inference
quantization#
Storing weights at lower numeric precision so the model fits in less memory, at a small cost in quality. Quantization is why a serious model can run on a laptop instead of a rack of GPUs.
see: weights · open-weight model
open and closed · 3 terms
open-weight model#
A model whose weights are published for anyone to download and run. The training data and code usually stay private, which is why "open weight" is a narrower claim than "open source." Llama and Mistral made the pattern mainstream. If you can self-host it, it is probably this.
see: weights · open-source model
open-source model#
The stricter claim: weights, training code, and training data all released under an open license. It is rare, because the data is the part nobody wants to publish. Most models marketed as open source are actually open weight, and the difference matters when you are betting a product on one.
see: open-weight model
frontier model#
A model at the current edge of capability, trained at a cost only a handful of labs can carry. Frontier models are usually closed and reached through an API. The edge moves every few months, and last year’s frontier has a habit of showing up on your laptop.
see: open-weight model · quantization
working with a model · 7 terms
inference#
Running a trained model to produce output. Training is paid once; inference is paid on every request, forever. When people talk about the economics of AI, inference is usually the bill they mean.
see: token · distillation
prompt#
Everything you send a model to get a response: instructions, questions, examples, context. Output quality tracks input quality closely, which makes prompt writing a real skill and not an incantation. I treat prompts like code, reviews included.
see: system prompt · context window
system prompt#
Standing instructions a model receives before the conversation starts: its role, its rules, its tone. Users rarely see it, and it shapes every answer they get.
see: prompt
context engineering#
The discipline of deciding what goes into a model’s context window and when: instructions, documents, tool results, history. Prompt writing was the first draft of this job. Agents made it an engineering problem, with token budgets, retrieval, and real tradeoffs. I argue context is becoming a platform capability, not a per-team trick.
see: context window · prompt · agent
temperature#
A setting that controls how much randomness the model’s output allows. Low temperature gives consistent, predictable answers. High temperature gives variety and surprises. For anything that has to be correct, keep it low.
see: inference
hallucination#
The industry’s word for a model stating something false with complete confidence. I prefer "mistake." The model is not seeing things; it is predicting text, and the prediction is wrong. Whatever you call it, it is why verification is the actual job when you work with AI.
see: evals · training data
putting models to work · 5 terms
RAGretrieval-augmented generation#
Fetch the relevant documents first, put them in the prompt, then ask the question. RAG is the standard way to make a model answer from your data instead of from whatever it memorized in training. Most "chat with your docs" products are RAG underneath.
see: embedding · context window
agent#
A model in a loop with tools. It plans, takes an action, reads the result, and goes again until the job is done. The loop is the difference between autocomplete and a system that ships work, and it is where most of my writing lives now.
tool use#
Letting a model call real software: search the web, query a database, run code, edit files. The model writes a structured request, your system executes it, and the result goes back into the context. Tools turn text prediction into work, which is exactly why they deserve gates and guardrails.
see: agent
MCPModel Context Protocol#
An open standard, started by Anthropic in 2024, that gives models one common way to connect to tools and data sources. Build an MCP server for your system once and any compatible agent can use it, instead of one custom integration per pair. The usual analogy is USB-C for AI tools, and for once the analogy holds.
evals#
Automated tests for model behavior. You cannot hand-review every output, so you write checks for the cases that matter and run them on every change, the way unit tests guard code. Teams that ship LLM features without evals are shipping on vibes.
see: hallucination · agent
definitions favor plain language over precision theater · where a term is contested, the entry says what I mean by it