Almost Timely News: 🗞️ Better Vocabulary for Better AI Results (2025-11-30)

Almost Timely News: 🗞️ Better Vocabulary for Better AI Results (2025-11-30) :: View in Browser

The Big Plug

🚨 Watch my latest keynote, How to Successfully Apply AI in Financial Aid, from MASFAA 2025.

Content Authenticity Statement

95% of this week’s newsletter was generated by me, the human. You will see 3 AI rendered images in the opening section by Google’s Gemini 3. Learn why this kind of disclosure is a good idea and might be required for anyone doing business in any capacity with the EU in the near future.

Watch This Newsletter On YouTube 📺

Click here for the video 📺 version of this newsletter on YouTube »

Click here for an MP3 audio 🎧 only version »

What’s On My Mind: Better Vocabulary for Better AI Results

This week, let’s talk vocabulary. Vocabulary is one of the skills that separates great users of AI from users struggling to get good results.

Here’s what I mean. Large language models, by their very nature as language models, run on language. The better quality and quantity of language going in, the better the results are coming out. This fundamental principle is what sets apart good from great.

Part 1: The Very Basics of Language Models

To understand AI and how vocabulary matters, we first have to understand how AI works at least in a simple way. This explanation is what I use to oversimplify AI. It is mathematically incorrect. It is NOT the linear algebra and calculus that makes AI models work. If you want to learn that, you can check out this interesting interactive tutorial that steps through the math of an AI model in great detail.

Imagine every word has a constellation of related words around it, like a word cloud. I say “cheese” and you might think “burger”. I say “Apple” and you might say “iPhone”. All language has these relationships, these associations. On top of that, all language has predictable patterns. “God save the Queen” makes sense to us in UK English. “Queen God the save” does not, even though it’s the exact same words.

Those patterns - word order, meaning, etc. - can be boiled down into statistics. Every word is related to another word, and when we start putting words together, those associations strengthen. “Cheese” and “Apple” put together have a closer association with “fondue” than they do “burger” because they co-occur more.

Imagine as you type a prompt, the word clouds around each individual word start to intersect. The more words you type, the more intersections there are, until it becomes clear exactly what it is you’re writing about. The intersections get very, very specific.

Conceptually, this is what’s going on under the hood at a rudimentary level - but it’s close enough to help make the rest of this newsletter make sense. AI models are all about how things relate to each other.

Part 2: Jargon

Today’s AI models can have massive inputs as prompts. ChatGPT and Claude can take an entire business book as a prompt, 75,000 - 90,000 words. Should you do this? No. Could you do this? Yes. I always get a chuckle reading so-called AI experts on LinkedIn urging people to be concise in their prompts for daily business use. That advice was relevant in 2022 when ChatGPT came out and could handle only 3,000 words maximum in a conversation, but times have changed.

As a sidebar, that advice IS still relevant if you are building AI agents and systems where you expect thousands of users to use the system. Conciseness matters at that scale because you’re paying for every word in your app’s prompts. But for the average user of ChatGPT, Gemini, or Claude who pays a flat rate per month? Conciseness isn’t nearly as important as context.

Gemini can handle even more. You could, if you so chose, put the entire works of William Shakespeare as a prompt, all 800,000 words. Again, should you do that? No. Could you? Yes.

So the maxim I’ve been urging people to use since 2023 is: the more relevant, specific words you use, the better your AI results will be. Jargon is your friend.

Let’s talk about why. Jargon - those terms that are specific to an industry, the inside baseball language - is incredibly effective with AI because it helps language models understand the domain and context very, very quickly. If I say “paying for college”, that’s fairly generic. It could mean a lot of things in a lot of places. If I say “FAFSA”, the Free Application for Federal Student Aid published and collected by the US Government, that one word immediately conveys that we’re talking about federal student aid for college within the United States. It is incredibly specific.

George Bernard Shaw once say, “Every profession is a conspiracy against the laity”. Every profession’s language, though, is the key to unlocking AI’s precision.

Remember that AI models fundamentally are about how things relate to each other. If you’re using generic language, there can be a TON of related concepts that aren’t very specific, and thus the output you get from AI is… well, not specific.

Take this absolutely awful prompt as an example:

Write a blog post about B2B marketing.

That’s terrible. Useless. And the slop it will produce is equally unhelpful.

Suppose you wrote:

Write a blog post about B2B marketing of pillow block bearings.

Just that one phrase, using jargon, immediately changes the model’s understanding of what we’re doing from B2B marketing (a giant topic) to the marketing of a specific kind of industrial ball bearing.

Part 3: Vocabulary

With all that in mind, that brings us to the topic of vocabulary. When it comes to AI, prompts are a lot like recipes. Recipes are important, to be sure. Cooking without a recipe generally yields less good results, especially if you are cooking something you’re unfamiliar with or in a cuisine you don’t know.

But the recipe is not the food. The map is not the territory. To actually cook, you need ingredients. To coax the best results out of AI, you need ingredients as well, and those ingredients are data.

The good news is that AI brings a lot of its own data to the party. Today’s language models are trained on enough data that if they were printed books, they’d be a bookshelf that wraps around the equator of the planet 8-12 times. But not all that data is correct or fresh, and just because something is high probability (which is what AI generates, probabilities) doesn’t mean it’s true.

Thus, the more ingredients you bring to the party, the better your results will be. Sometimes, the ingredients AI brings are enough, but if you don’t know what to ask for, you’re going to get suboptimal results.

Here’s an example. My friend Ruby was attempting to have AI generate a simple tower defense video game. She got most of the prompt right in terms of the game mechanics and theme, but she lacked the vocabulary around specific terms that would have helped AI know what to do or prevent it from reinventing the wheel.

Adding just this one sentence made AI behave very, very differently:

Use Typescript libraries and CDNJS libraries like phaser.io as the game engine.

What does this cryptic sentence mean? Typescript is the language that the AI was working in, a variant of JavaScript, a very popular language on the web. A Typescript library is like a plugin or addon; Typescript libraries exist for thousands of different purposes, and because they already exist, AI knows about them. More important, AI knows what they are and what they mean, which can save enormous amounts of time because it then doesn’t have to invent its own version of that functionality.

CDNJS is a free web-hosted library created by Cloudflare that lets users reference common, open source libraries totally for free. Instead of the user having to download and incorporate those libraries, they can write code that simply refers to that lib