A chart called “15 Must Know AI Terms” has been making the rounds, and it’s a decent cheat sheet. LLM, token, context window, hallucination, RAG, the alphabet soup that shows up every time somebody talks about AI.
What the chart can’t tell you is what any of it means when you’re writing a book.
AI is part of my working day. Claude reads my finished novels and hunts for plot holes the way my Plot Handbook teaches, and I paste passages into it when I want a grammar check.
So here are the fifteen, explained for writers.
Think of a language model as a new hire who has read every book in the library and remembers none of your conversation from yesterday. Brilliant, fast, eager to please, and with a very short memory. Keep that new hire in mind. Almost every term below is a description of how that new hire works.
| Term | What it means | Why a writer cares |
|---|---|---|
| LLM | Large language model, the engine behind ChatGPT, Claude and Gemini | It predicts likely words. It doesn’t look facts up unless it’s given a tool to do so |
| Machine learning | Software that learns patterns from examples | The model’s habits come from what it was fed |
| Token | A chunk of text, often part of a word | Limits and prices are counted in tokens |
| Context window | How much text the model can work with at once | Overfill it and early chapters fade |
| Hallucination | Confident output that’s false | Every quote, date and source needs checking |
| Training and inference | Building the model, then using it | Your chat doesn’t retrain the model |
| Weights | The billions of numbers the model learned | They hold patterns, not a copy of any book |
| Validation loss | A score for how well training is going | Only matters if you train models yourself |
| Fine-tuning | Extra training on a narrow set of examples | How a general model gets a specialty or a house style |
| RLHF | Reinforcement learning from human feedback | Why models sound agreeable and eager to please |
| Distillation | A small model trained to imitate a big one | Cheap, fast tiers that are weaker on hard work |
| RAG | Retrieval-augmented generation: the model pulls passages from documents before answering | Answers from your notes, not from memory |
| MCP | Model Context Protocol, a standard way to plug AI into apps and files | Lets an assistant read your calendar, email or folders |
| Chain of thought | Working a problem through in steps | Better on plot logic and timelines |
| AGI | Artificial general intelligence, AI that can do any task a person can | Still theoretical, and still mostly a marketing word |
What is a large language model, and how does it write?
A large language model is a program trained on an enormous amount of text to predict what comes next. Give it the start of a sentence and it calculates which word, or piece of a word, most likely follows. Do that thousands of times in a row and you get a paragraph.
That’s machine learning in action. Nobody wrote rules for grammar or plot into it. It picked up patterns from examples, billions of them, and those patterns live in its weights, the huge set of numbers that shape every answer. The weights don’t hold a copy of any book. They store statistics about how words tend to follow each other. That lets a model write a clean sentence about a book it has never read.
Model builders track something called validation loss while they train, a score that tells them whether the model is getting better at predicting text it hasn’t seen. You’ll never need it unless you start training models yourself. I’ve included it because it’s on the chart.
The difference between training and inference matters more. The training phase is the months of expensive work that builds the model. Inference is you typing a question and getting an answer. Your conversation doesn’t retrain the model, so the new hire won’t remember your book next week unless the tool saves notes for it.
What is a token in AI, and why does the count matter?
A token is the unit a model reads and writes. It can be a whole word or a piece of one, and punctuation counts too. For English, a common rule of thumb is that a token averages about three quarters of a word.
Writers care because everything is counted in tokens. The price of an API call, the length limit on an answer and the size of the context window all run on tokens.
That last one bites.
Why does AI forget the start of a long manuscript?
A context window is how much text a model can hold in view at once: your instructions, the chapters you pasted in, and its own replies, all counted in tokens. Modern windows are big enough to take a whole novel. Big isn’t the same as reliable.
The new hire can see the whole manuscript on the desk and still forget what happened in chapter two.
Your new hire can see the whole manuscript on the desk and still forget what happened in chapter two. As a conversation grows, early material gets less attention, and eventually the oldest parts fall out of the window entirely. The character’s blue eyes turn brown. The minor villain who died in chapter six walks into chapter twenty. I’ve written about this failure as AI drift, and it’s the failure nobody warns you about.
Keep a short story bible and hand it to the model at the start of every session. Characters, settings, rules, what’s happened so far.
Don’t trust the window to remember for you.
Why does AI make things up with so much confidence?
Hallucination is the model producing something false and saying it like a fact. A quote nobody said. A book that doesn’t exist. A date off by a decade.
It happens because the model predicts plausible text. It doesn’t look things up unless it’s connected to a search tool or your documents, and even then it can garble what it finds. Plausible and true overlap most of the time.
When they don’t, the model has no idea. I wrote a whole hallucination survival guide for anyone publishing under their own name, and the short version is this: check every fact, quote and citation before it goes out.
An AI can’t tell plausible from true. You have to check. – Richard LoweShare on X
Chain of thought helps a little. When a model works through a problem in steps, it does better on logic, timelines and math, and you can see where it went wrong. A request to reason through your plot’s timeline step by step catches more mistakes than a yes-or-no question about whether the timeline works.
How do AI models learn a specialty or a personality?
The base model comes out of training as a raw text predictor. Several more steps turn it into the assistant you talk to.
Fine-tuning trains a general model further on a narrow set of examples, like legal contracts, medical notes, or one company’s support tickets, so it gets better at that one thing.
RLHF, reinforcement learning from human feedback, is where people rate the model’s answers and the model is trained toward the ones people liked. The chart describes it as trial, reward and repeat, like training a dog. That’s close, but the reward comes from human judgment, and that’s where the model’s personality comes from. It’s also why so many models sound agreeable, flattering and eager to please. People rated those answers higher. Push back on a model that tells you your manuscript is brilliant.
Distillation trains a small model to imitate a large one. The result is faster and cheaper and nearly as good on everyday tasks. It’s usually weaker on the hard stuff, like keeping a complex plot straight. When an app offers a “fast” tier and a “smart” tier, distillation is often the difference.
How can a writer make AI work from their own notes and files?
So how do you get the model to stop guessing?
This is where the alphabet soup turns useful.
RAG, retrieval-augmented generation, means the tool searches your documents first, pulls the relevant passages, and gives them to the model along with your question. Your interview transcripts, your research notes, your series bible. The model answers from those passages instead of from memory, and hallucinations drop. They don’t vanish. A retrieval tool can grab the wrong passage, or the right passage from the wrong draft, and the model will answer from it just as confidently. I’ve seen AI distort a writer’s own data while quoting it, so check the answer against the source.
MCP, the Model Context Protocol, is an open standard Anthropic introduced in late 2024 for connecting AI assistants to other software. The chart’s comparison to USB-C is a good one. One plug, many devices. With it, an assistant can read your calendar, search your email or open files in a folder, if you allow it. Give it access to what it needs and nothing more.
Then there’s AGI, artificial general intelligence, a machine that can do any intellectual task a person can. Nobody has built one. Some people say it’s a few years away, and others say decades. I’ve written about a Yale economist’s AGI forecast and what it leaves out. For now, treat any product that claims it as a sales pitch.
My take
Learn four of these terms and you’ll beat most people using AI to write. Context window tells you when the model will start forgetting your book. Hallucination tells you to check everything. RAG and MCP tell you how to make it work from your material instead of its guesses. The rest is vocabulary for impressing people at conferences. Know it if you like, but those four decide whether your book comes out right.
Put the new hire to work
Every term on that chart describes the same new hire: well read, fast, agreeable, forgetful and sometimes wrong. You wouldn’t let a new hire publish under your name without checking the work, and you shouldn’t let an AI do it either.
Pick one project and try it the right way this week. Write a one-page story bible, load your research into a tool that supports RAG, and ask the model to reason through its answers step by step. You’ll find more on using these tools well in my AI and writing hub. For help setting it up on your own book, my AI services page explains what I do.
Train the new hire with your notes, check what it hands you, and it’ll earn its desk.
Once you know the vocabulary, this is the next step: how to use AI so it amplifies your thinking and makes you more capable, without handing the thinking over.
