We share an AI term of the week in our Tuesday newsletter. Here are terms we have included so far.

Context window

The amount of text an AI model can “see” and work with at once, which includes your current message, plus everything else in the conversation so far (or in a document you’ve uploaded). Once a conversation goes past that limit, the model starts losing track of the earlier parts, the same way you’d lose the thread of a meeting if someone summarized the first hour in ten seconds.

Context windows are measured in tokens, not words (see: our earlier term of the week), and they’ve grown enormously – early models topped out around a few thousand tokens, while today’s frontier models handle hundreds of thousands, some over a million. That’s the difference between “paste in a paragraph” and “paste in an entire book and ask questions about it.” Bigger isn’t automatically better, though – models can still miss or misweight details buried in the middle of a huge context, so more room to work with doesn’t guarantee it’s used well.

Hallucination

When a model states something that sounds right but isn’t — a fake citation, a court case that never happened, a feature that doesn’t exist in the software, a quote nobody said. The term is a bad one, because it implies the model is seeing things. It isn’t. It’s doing exactly what it was built to do: predict text that fits the pattern. A plausible-looking source with a plausible-looking author and a plausible-looking page number is the pattern for a citation. Whether it’s true is a separate question the model was never asked.

That’s why hallucinations don’t come with a warning label. The model isn’t less confident when it’s wrong. It has no internal sense of “I know this” versus “I’m filling in the blank,” because both are the same operation. Newer models hallucinate less, and tools like web search and retrieval cut it down further by giving the model real material to work from instead of asking it to reconstruct from memory. But it doesn’t go to zero. The practical rule hasn’t changed: if a specific fact matters (a name, a number, a citation, a date) verify it before you use it.

Large Language Model (LLM)

The thing you’re actually talking to when you use ChatGPT, Claude, or Gemini. It’s a transformer trained on enormous amounts of text — books, websites, code, transcripts — with one deceptively simple job: predict the next chunk of text. That’s it. Type “the capital of France is” and it predicts “Paris,” not because it looked it up, but because that pattern showed up constantly in what it read. The “large” part is doing a lot of work here: these models have hundreds of billions of parameters, the internal dials that get tuned during training, and they’ve read more text than a person could get through in a thousand lifetimes.

The surprising part is what falls out of that. Nobody programmed an LLM to write a haiku, debug Python, or explain a contract clause. Those abilities emerged from prediction at scale. Which is also why LLMs fail in the weird ways they do. A model that’s optimizing for plausible-sounding text will happily produce a plausible-sounding citation that doesn’t exist. It isn’t lying to you and it isn’t consulting a database. It’s finishing the pattern. Understanding that one thing explains most of what LLMs are great at and most of what they get wrong.

Multimodal

An AI model’s ability to work with more than just text, including images, audio, video, even code, all within the same conversation. You can hand a multimodal model a photo of your whiteboard and ask it to turn the notes into an outline, or describe a chart and have it explain the trend, or have a spoken conversation with it out loud.

While most people’s mental model of AI is still “type a question, get a text answer back,” that’s really just one mode. Under the hood, everything (text, images, audio) gets converted into the same kind of token-based representation the model reads, which is what lets one model reason across all of them at once instead of needing a separate tool for each. This multimodality is part the explanation why AI is showing up in more places than chatbots lately, such as in camera apps, voice assistants, and meeting transcription tools.

Plugin

An add-on that gives an AI tool new abilities it doesn‘t have out of the box, like connecting to your calendar, searching a company database, or following a specific workflow for tasks like contract review or sales reporting. On its own, a large language model can only work with what you type in and what it learned in training. A plugin extends its reach to your apps, files, and processes.

The term goes back to 2023, when ChatGPT launched its first plugins for things like booking travel and ordering groceries. Today the same idea shows up under different names (connectors, extensions, apps), and the packaging keeps changing, but the idea is the same. For organizations, using plugins might mean the difference between an AI that gives generic answers and one that works with your actual data and tools. It’s also worth a security check before you install one: a plugin can only do what you‘ve given it access to, so it‘s worth knowing exactly what that is.

Recursive self-improvement

When an AI system improves its own capabilities (code, training process, prompts) and then uses those improvements to get even better at improving itself. It’s the feedback loop behind a lot of AI doomer talk: get a model good enough to upgrade itself, and each version makes the next one smarter, faster.

Right now this mostly shows up as models helping write their own training code or agents that iterate on their own prompts – not full autonomy. But it’s the direction frontier labs are pointed.

Skill

A packaged set of instructions that teaches an AI how to do a specific kind of task your way, like formatting a weekly report, reviewing a contract against your company‘s checklist, or writing in your brand voice. Instead of retyping the same long instructions every time, you write them down once, and the AI pulls them up whenever the task comes around.

You can think of a skill as a training manual for a new hire. The employee is already smart, but they don‘t know how your organization does things. A skill supplies that missing know-how. Here’s where the analogy breaks down: A skill is loaded only when it’s relevant, so it doesn’t clutter the model’s context window the rest of the conversation. Skills can be as simple as a single text document of written guidelines or as involved as a set of instructions plus templates and scripts. For organizations, they‘re a way to capture how your best people do a task and make it repeatable for everyone else. Just like with plugins, it‘s worth reviewing what a skill instructs the AI to do before you turn it on.

Tokens

Tokens are the pieces an AI model actually reads. It doesn’t read letters or words, but rather chunks somewhere in between, usually about four characters each. Common words like “the” and “and” get their own token. Longer or rarer words get chopped up, so “unbelievable” might come through as four separate pieces. That works out to roughly 750 words per 1,000 tokens.

Tokens are how the tools price you, how memory limits get measured (a 200K context window is 200,000 tokens, not words), and why a model can write you a sonnet but can’t count the r’s in “strawberry.” It never sees the letters, just the chunks. Type your name (or the word “strawberry”) into OpenAI’s tokenizer to see how tokens work.

Transformer

The “T” in GPT. It’s the neural network architecture behind basically every major AI language model today (GPT, Claude, Gemini, you name it). Introduced by Google researchers in 2017, the key trick to transformers is “attention”: instead of reading a sentence word by word in order, the model looks at all the words at once and weighs how much each one matters to every other one. That’s how knows “it” refers to “the dog” three words earlier, or that “bank” means something different in “river bank” versus “bank account.”

Before transformers came on the scene, AI language models processed text sequentially and struggled with long-term context. By contrast, transformers process everything in parallel, which is both why they understand context so well and why they scale so efficiently on modern hardware that is designed specifically for parallel computation. That scalability is a big part of why the last few years of AI progress happened as fast as they did: throw more data and computing power at the same architecture, and it keeps getting better at processing language.