AI's Hidden Vocabulary

The vocabulary behind every AI conversation, finally defined. Seven of the terms you keep hearing, explained.

Every AI conversation runs on the same vocabulary.

You can't read an article about AI today without running into the same vocabulary. LLM. Training. Tokens. The terms show up in headlines, product announcements, marketing copy, and yes, in this very newsletter.

Most people nod along. Few stop to ask what they actually mean.

That's understandable. The terms get used so often that there's an unspoken assumption everyone knows them already. But the result is a vocabulary gap that compounds quietly. You cannot understand AI tools, push back on overhyped claims, or make informed decisions about which one to use without understanding the words being used to describe them.

This week we slow down. Seven of the terms you keep hearing, defined. These are foundational pieces of AI, and we'll take a deeper dive into some of them in future issues. Today they get the introduction they should have had from the start.

🚀 MARTY SAYS

"On a pre-flight checklist, there's no room for words nobody defined. That standard is a good one to borrow."

LLM (Large Language Model)

This is the technology underneath every AI tool covered in the past three months. An LLM is a type of AI that has been trained on enormous amounts of text and learned to predict what words should come next in any given sequence. When you type something into ChatGPT, Claude, or Gemini, the model is using patterns it learned from that text to generate a response. "Large" refers to the size of the model itself, measured in parameters. "Language" means it works with text. "Model" is the system itself.

Training

The process of teaching the model how to do what it does. We covered the data side of training in Issue 14. The process itself is the months-long, multi-million-dollar operation that produces a working LLM. During training, the model reads through its dataset and adjusts its internal settings to get better at predicting what comes next. After training, the model is essentially frozen, and its behavior is fixed unless it gets retrained.

Tokens

The basic unit of text an LLM processes. A token is roughly a word or part of a word. "Hello" is one token. "Internationalization" might be three. AI companies charge by tokens (input and output), measure model capability by how many tokens a model can handle at once, and describe training data size in trillions of tokens. When you see a model advertised with a "1 million token context window," that translates to roughly 750,000 words of memory. Tokens are how AI counts everything.

Prompt

What you type into an AI tool. The prompt is the input. The response is the output. Prompt also refers to the broader skill of getting useful results out of an AI tool: clear instructions, relevant context, examples, structure. "Prompt engineering" is the discipline of writing prompts well. Two people typing into the same AI tool can get wildly different results based purely on how they prompted it. Most of the gap between "AI is useless" and "AI is incredibly useful" sits in the prompt.

Parameters

The model's internal settings. Think of parameters as dials that get adjusted during training. A modern frontier model has hundreds of billions to trillions of these dials. More parameters generally means more capability, but also more cost to train and run. When you see "Llama 3 405B" or "GPT-4 1.8T," those numbers refer to parameters. It's a useful but rough proxy for model size. A smaller well-trained model can outperform a larger poorly-trained one, so parameter count alone never tells the whole story.

Context Window

How much information the model can see at once during a conversation. Every conversation has a context window measured in tokens. Once you exceed it, the model starts forgetting earlier parts of the conversation. Claude has a 200,000 token context window, around 150,000 words. Gemini has a 1 million token window. ChatGPT varies by version. A bigger context window means the model can handle longer documents, longer conversations, or more attached files at once.

Open Source vs Proprietary

Two fundamentally different approaches to releasing AI. Proprietary means the company owns the model and you access it through their service. ChatGPT, Claude, and Gemini are all proprietary. You cannot download them, modify them, or run them yourself. Open source means the model's underlying weights are publicly available. Anyone can download it, run it on their own hardware, customize it, even build a business on top of it. Meta's Llama family is the highest-profile open source AI. The line between "open source" and "open weights" gets blurry, which is a conversation worth having on its own.

THE BIGGEST MISCONCEPTIONS

The words above get used a lot. They also get used wrong a lot. A few common mistakes worth knowing:

Training does not mean continuous learning. Most models are trained once and frozen. When you have a conversation with Claude or ChatGPT, the model is not learning from that conversation in real time.

More parameters does not always mean better. A smaller well-trained model can outperform a larger poorly-trained one. Parameter count is a rough proxy for capability, not a guarantee of it.

A bigger context window is not always an advantage. Bigger windows cost more, can hurt response quality at the edges, and most use cases do not actually need them.

Open source does not always mean fully open. Many "open" models restrict commercial use, hide training data, or come with strings attached. "Open weights" is closer to the reality for most of the open models you will encounter.

Why This Vocabulary Matters

Most of the noise around AI comes from people using these terms loosely. "Industry-leading parameters." "State-of-the-art training." "Massive context window." Once you understand what the words actually mean, the marketing claims start to look thinner. The questions you can ask get sharper. The decisions you make about which tool to use, and for what, get easier.

Safe Harbor: Three Things You Can Do This Week

Find one AI article or product announcement and underline every term you weren't sure about. Look up the ones not covered here. The vocabulary grows fastest when you fill the gaps you actually encounter, not the ones someone else picks for you.
Test the vocabulary in a conversation. Ask the AI tool you use most three questions using the terms from this issue. "What's your context window?" "Are you open source?" "What's your parameter count?" The answers themselves matter less than watching how the tool responds. You will learn a lot about what it does and doesn't know about itself.
Pick one term and go one level deeper. Choose the term from this issue that felt most unfamiliar. Spend 15 minutes reading about it. Not to become an expert, just to close the specific gap.

Next week: What an AI knows versus what it looks up. RAG, retrieval-augmented generation, is another important technology inside the AI products you use every day. And it explains a lot about why AI sometimes seems weirdly smart and sometimes seems weirdly off.