What’s a Tuple? 10 AI Words You’ve Gotta Know, If You Want AI to Work. (Don’t worry, tuple isn’t one of them.)

Everyday AI tools are all built as much on language as software, and much of that language seems strange and unsettling. The AIs that run these products are large language models, whether ChatGPT, Claude, Gemini, Meta’s Muse, and the integrations into Outlook, Spotify, and every shopping app. They understand the world and even “think” in language (sort of). And, they speak in a way that is just weird. I’m not just talking about the annoying AI-isms that leave you scratching your head.

Here’s a recent example of a headscratcher from Claude:

"Send it Monday. Not because Friday is bad. Not because Monday is good. But because you deserve a weekend that isn't shadowed by an open loop. That's not procrastination — that's stewardship

Yeah, that’s right! I do deserve a weekend free of open loops. Right!? Um, What?

It’s not just the AIs that use confusing language. Humans talking about AI seem to use gibberish too. AI builders and researchers use a fused language that comes from software engineering (“sandbox,”¹ “harness”²), math (“vectors,”³ “tuples”⁴), neuroscience (“neuron,”⁵ “ablation”⁶), linguistics (“token”⁷), and philosophy (“alignment”⁸), among others. That’s why you can find sentences in the research like this:

"the system state at iteration k is represented by the tuple (S_k, W_k), where S_k = {s_1, ..., s_M} denotes the active procedural skill set and W_k denotes the persistent knowledge base (Wiki). While candidate skill updates are subject to validation gating and rollback upon score degradation, the knowledge base W_k persists and compounds across iterations."⁹

Don’t worry, there is some good news in this post. To use AI safely, you don’t need to speak AI engineer. Nor do you have to accept that bad AI writing. (My post about fixing bad writing, here.) You can have both safety and good outputs.

To do that, you do need some basic AI vocabulary. The following 10 words will help you decipher how AI works and how to make it work for you. If you know these ten, you can use consumer-grade AI tools with more control, ask better questions of the vendors and consultants selling to you, and understand most of what you read about AI.

My goal is that by the end of this post, you will be better equipped to get better work out of your AI, whether it’s Claude, ChatGPT, Copilot, Gemini, or Monday.com agents. I’m going to start from the very beginning of how AI learns, then go to how it learns about you and how you control it.

Training. Training is what creates a large language model. It takes enormous amounts of computing power and is incredibly expensive for state-of-the-art models. Essentially, AI engineers create an algorithm to read an enormous amount of data, usually most of the internet, to learn the shape of information (and some specific facts too). Training is how the machine gets its general abilities. For scale: Meta trained one recent model on more than 15 trillion tokens of text.¹⁰ A devoted human reader gets through a few hundred million words in a lifetime.¹¹ Once training is done, the large language model is set. It doesn’t add new knowledge after training ends. It only answers questions.

The model (LLM). The model is what training produces. Think of it as the brain, or the engine. Each one has a name and a version: Opus 5.5, GPT 6, Gemini 3. LLM, short for large language model, is the generic term. Companies replace their models often, which explains something you may have noticed: if Claude or ChatGPT suddenly starts behaving differently, the model underneath may have changed. And there are far more of these brains than the famous few. One public catalog, openrouter.ai, listed 338 different models the week I checked.¹²

Prompt. The prompt is what you type into the box: the question and the information you send up to the model. “Review this draft for typos.” “Give me ten suggestions for a grant application to this foundation.” Prompts drive the whole business. ChatGPT alone handles more than 2.5 billion of them a day.¹³

Context. Context is everything the model can see right now: your prompt, the documents you attach, and the conversation so far. Context is not infinite, but the “context window” (the context you send up to the model with each prompt) can hold enormous amounts of text. The largest tools can hold about 750,000 words in view at once, roughly the Bible, or the first five Harry Potter novels.¹⁴

Context is the lever you control most directly. You provide the model your documents and your best examples as context. Warning: just because you can send up the entire Bible doesn’t mean that you should. As the context window fills, the model is parsing more and more information to figure out an answer, and your conversation with it can get weird and unpredictable. At that point, you should ask for a summary of what matters and start fresh.

Memory. Memory is what carries over from one session to the next. Memory is what makes Claude and ChatGPT seem like they know you, or that they are learning. In fact, the LLM is not learning about you. But the software between the LLM and you (the harness, below) is storing personalized memories.

Claude, ChatGPT, Gemini, etc. will save some memories automatically (you can read and edit them in the settings). You can also build memory on purpose, through saved instructions and projects: folders of documents and standing directions the tool reads every time you return. Training, context, and memory are the three ways an AI knows things, and the difference between them explains most of AI’s baffling behavior. I wrote a whole post on that, here.

Harness. The harness is the software built around the model (between you and the LLM): the app you type into, the tools the model can reach, the limits placed on it, and the record of what it did. ChatGPT is a harness built around OpenAI’s models; the Claude and Gemini apps work the same way. The same model in two different harnesses is, for practical purposes, two different machines. Claude Chat is a chat window; Claude Cowork can read your files and take action. Same brain, different harness, very different machine.

Agent. An agent is a model in a harness, working toward a goal. It can read files, use tools, work in steps, and start again until it decides the job is done or a person stops it. Agents are improving fast: by one respected measure, the length of tasks agents can finish on their own has been doubling roughly every seven months, and lately faster.¹⁵ But an agent does not reliably know whether it has actually succeeded. Someone still has to check the work.

Hallucination. A hallucination is an answer that the LLM creates without having any evidence to support it. (Note: hallucinations can at times be correct by accident, so to speak. And sometimes, you might want something new and creative, i.e., a hallucination.) Models have learned the shape of a good answer during training, and they are built to answer rather than to admit they do not know.

The scale of the problem depends on the stakes. Way back in 2024, Stanford researchers found general-purpose chatbots hallucinated on 58 to 88 percent of legal questions,¹⁶ and even purpose-built legal AI tools got it wrong 17 to 33 percent of the time.¹⁷ (Things have improved a lot since then.) Three habits help: give the model good source documents, ask for real sources you can check, and tell it plainly that “I’m not sure” is a better answer than a made-up one.

Governance. Governance is the human oversight around the machine. It decides who may use AI, for which work, with what data, and who checks the output before it matters. Where, in other words, is the “human in the loop”? Basically, governance is just ordinary risk management. In IBM’s latest breach study, more than 20 percent of organizations reported a security incident involving an AI model or application.¹⁸

AI policy. An AI policy is your organization’s governance rules, written down. A borrowed policy, marked up to fit your organization, is a fine place to start. If you’d like a starter policy, ask me.

These ten words have helped give me confidence to ask questions about AI from experts and to try out new tools. Let me know if you find them helpful or have any questions about them.

Btw-here’s how Claude’s Opus 5 suggested I conclude this blog:

"Ten concepts, one tuple: (model, harness, context, memory, governance) — get those five right and you're not just using AI, you're running it."

Notes

  1. Sandbox (software engineering): a walled-off space where untrusted software can run without touching anything else on the computer. See the NIST glossary definition.
  2. Harness (software engineering): originally a “test harness,” the scaffolding programmers build around code to run it and check it. See Wikipedia: Test harness. For the AI sense, see Anthropic’s Effective harnesses for long-running agents.
  3. Vector (math): a list of numbers treated as a single object. Language models represent words as very long lists of numbers like this. See Wikipedia: Vector.
  4. Tuple (math): a finite, ordered list, such as (3, 7, 2). See Wikipedia: Tuple.
  5. Neuron (neuroscience): a nerve cell. In 1943, researchers borrowed the word for the simplest unit of a neural network, a small math function modeled on a brain cell. See Wikipedia: Artificial neuron.
  6. Ablation (neuroscience): removing part of the brain to learn what it did. AI researchers “ablate” one piece of a system and measure what breaks. See Wikipedia: Ablation (artificial intelligence).
  7. Token (linguistics): one occurrence of a word in a text. “The cat saw the dog” has five tokens but only four different words. For AI, a token is a word or a piece of a word, and a thousand tokens is about 750 words. See Wikipedia: Type–token distinction.
  8. Alignment (philosophy): the question of how to make an AI system pursue the goals and values people actually intend. For the philosophy, see Iason Gabriel, “Artificial Intelligence, Values, and Alignment,” Minds and Machines (2020).
  9. Liyan Tang and others, “WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution,” Google Research, August 2026, page 3.
  10. Meta, “Introducing Meta Llama 3,” April 2024. Llama 3 was “pretrained on over 15T tokens.”
  11. Based on adult reading rates of about 175 to 300 words per minute. Marc Brysbaert, “How many words do we read per minute? A review and meta-analysis of reading rate,” Journal of Memory and Language 109 (2019).
  12. OpenRouter model catalog, checked August 4, 2026. The count changes often.
  13. OpenAI’s figure, reported by Axios in July 2025. See TechCrunch, “ChatGPT users send 2.5 billion prompts a day,” July 21, 2025.
  14. The largest models hold 1 million tokens (Claude documentation on context windows), and a token is about three-quarters of a word (OpenAI, “What are tokens and how to count them?”). Word counts are common estimates: the King James Bible is about 783,000 words (Word Counter), and the first five Harry Potter books total about 717,000 (Word Counter).
  15. METR, “Measuring AI Ability to Complete Long Software Tasks,” March 2025, and “Time Horizon 1.1,” January 2026. The update puts the doubling time since 2024 at 89 days.
  16. Matthew Dahl and others, “Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models,” Journal of Legal Analysis 16 (2024). The rates ran from 58% for ChatGPT 4 to 88% for Llama 2.
  17. Varun Magesh and others, “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools,” Stanford RegLab and HAI, May 2024. The tested tools were Lexis+ AI and Westlaw’s AI research products.
  18. IBM, “IBM Study: One in Four Malicious Breaches are AI-Enabled, Costing Companies $6 Million on Average,” July 29, 2026, on the Cost of a Data Breach Report 2026. The figure was 21%, up from 13% the year before.

R. Michael Waller leads Georgia Appleseed, a nonprofit law center for Georgia’s kids, and writes here about what happens when civil legal services and AI collide.

If this was useful, you can buy me a coffee: https://buymeacoffee.com/MichaelWaller