SI Words in Plain English: A Short Glossary for This Blog

If you've read a few of my posts about SI, you've probably hit a word I used as if everyone already knew it. Not everyone does, and that's fair.

So here's a short glossary. Each word gets a plain definition and, where I've written more about it, a link. It isn't a complete dictionary. It's the words that keep showing up here.

One note first. SI stands for Super Intelligence. It's the term I use now for where this technology is heading. Older pages, and the published book, say AI. The start-here guide covers why you'll see both: https://johndclay.com/blog/new-to-the-shadows-series

The basics

Model. The trained system that takes text in and produces text out. When people say "the model," they mean the part that does the predicting, separate from the app or tool wrapped around it.

Prompt. What you send the model: the instructions, the question, and any documents you include. Everything the model knows about your task, it gets from what you send.

Token. The small pieces text is cut into before a model reads it. A token is often part of a word rather than a whole word. Models are usually priced and limited by tokens, which is why the word comes up whenever cost does.

Context window. How much the model can take in at once: the prompt, the conversation so far, and any files or tool results. When it fills up, something has to be dropped or summarized.

Temperature. A setting for how much randomness goes into the model's word choices. Lower means more predictable. Higher means more varied. Setting it to zero still doesn't promise the same answer every time, which surprises a lot of people.

Inference. Running a trained model to get an answer, as opposed to training it. Every time you send a prompt and get a reply, that's inference.

Hallucination. When a model states something false as if it were true. It isn't lying on purpose. It's producing text that sounds right without anything checking that it is right. That's why it pays to check facts, links, and numbers before you rely on them.

Agents

Agent. A model set up to take steps on its own: read files, use tools, make changes, then decide what to do next based on what happened. A chatbot answers. An agent acts.

Tool. Anything an agent can call to do something outside itself, like searching, reading a file, running code, or sending a message. An agent's tools decide what it can touch.

Session. One working stretch for an agent, from start to finish. When a session ends, what the agent was holding in its context window ends with it. That's why I suggest leaving it a handoff note: https://johndclay.com/blog/agent-handoff-note

Read access and write access. Read access lets an agent look. Write access lets it change, delete, send, or spend. I'd start every new agent on read only: https://johndclay.com/blog/agent-read-access-first

Stop rule. A plain condition that tells an agent to pause and ask a person before it goes on. For example: stop before deleting anything, or stop when the instructions could mean two different things.

Harness. The software around a model that turns it into a working tool. It picks the model, sets the defaults, manages the context, and decides when to call helper agents. The harness often shapes the bill more than the price sheet does.

Cache. Saving work on parts of a prompt that repeat. In an agent's loop, the same instructions and history get sent again on every turn. A cache can make those repeated parts cheaper, but only if it's actually being hit.

Testing

Eval. Short for evaluation. A set of test cases you run a model or agent against to see how well it does.

Held-out set. Test cases kept apart from anything used to build or tune the system, so the score reflects new work and not memorized work.

Frozen eval set. A held-out set you don't change, so this month's score and next month's score can be compared fairly: https://johndclay.com/blog/frozen-eval-set

Benchmark. A public, shared set of tests that lots of people use to compare models. Useful for a rough idea. Less useful for your own work, which no public test was built around.

Training and data

Training. Teaching a model by showing it large amounts of data and adjusting it until its predictions get better.

Fine-tuning. Extra training on a smaller, focused set of data, so a model that already works in general gets better at one kind of task.

Synthetic data. Data made by a model instead of by people. Useful as an addition. Risky as a replacement: https://johndclay.com/blog/training-on-model-outputs

Open weights. A model whose trained settings are published, so you can download it and run it yourself. Open doesn't mean free to run: https://johndclay.com/blog/self-hosting-vs-api

API. A way for your software to call a model that someone else runs, usually paying for each use. It's the alternative to running the model yourself.

Words I try not to use

Some words show up a lot in SI writing and explain very little. "Revolutionary." "Game-changing." "Magic." If I can't say what something does in a sentence, I'd rather not say it at all.

That's the rule for this glossary, too. If a definition here doesn't make sense, that's on me, and I'd like to know. The contact page is open: https://johndclay.com/contact

Where the longer pieces live

The short posts live here on the blog. The longer SI essays go out on X as Articles, and they're listed with a short summary on the Articles page: https://johndclay.com/articles

And if you came for the SI and you're curious about the fiction, The Shadows Series is hard dystopian science fiction about memory, identity, and the people who refuse to disappear. All five books are on one page: https://johndclay.com/books