By Simon Royle and Kevin A. ‘KAM’ McGrail
There are so many terms flying around Artificial Intelligence right now that it can be difficult to separate what is useful from technobabble.
From our perspective, you do not need to become an AI engineer to use these AI systems effectively. However, understanding a few key terms makes the whole subject considerably easier to follow.
Here’s our top 10!
1. Multi-modality
The formal definitions of multi-modality can make something relatively simple sound complicated.
Think back to school. You heard something, read it, wrote it down, watched it being demonstrated and then tried it yourself. That was multi-modal learning.
In AI, multi-modality means the system can work with more than one type of information, including text, images, recorded video, live video, recorded audio, live video, documents, and more.
Instead of simply typing a question, you can photograph something, speak to the AI, upload a document or combine several types of input in the same task with a multi-modal solution.
2. Prompt-Fu, STAR and PTCF
A prompt is simply the instruction you give an AI system, but the quality of that instruction has a major effect on the quality of the response.
We call the skill of doing this well, Prompt-Fu.
Two useful frameworks are STAR, which stands for Situation, Task, Action and Result, and PTCF, which stands for Persona, Task, Context and Format.
The point is not which acronym you use. The important thing is to give the AI enough information to understand what you want, why you want it and what the output should look like.
“Write me a marketing plan” leaves a great deal for the AI to guess. A prompt that explains the company, market, objective and required format gives the system far more to work with.
So, “I am a marketing account executive. I am working on a marketing plan for October for Halloween in the United States. Please make a plan for how to give away candy corns and make people like them. Please put the plan into a Spreadsheet” would cover Persona, Task, Context, and Format.
3. Ethical AI
KAM’s definition of Ethical AI is deliberately simple:
“Ethical AI is the ability to recreate the decisions of the computer.”
If an AI system makes an important decision, somebody should be able to understand how that decision was reached well enough to examine it and challenge it.
If that is impossible, it becomes very difficult to identify bias, errors, or inappropriate assumptions.
Ethical AI therefore includes transparency, accountability, privacy, bias, and human oversight, particularly where AI is influencing decisions rather than simply helping someone write a document.
4. Generative Pre-Trained Transformer, GPT
GPT stands for Generative Pre-Trained Transformer.
Generative means the model can create new content. Pre-trained means it has already learned from very large amounts of information before you start using it. Transformer refers to the neural-network architecture behind modern language models.
You do not need to understand mathematics. What matters is that GPT describes a type of generative AI model, not simply the name of a chatbot. And you’ve been using transformers for a long time since Google invented them and started using them for https://translate.google.com/ over a decade ago!
5. NLP, Natural Language Processing
Natural Language Processing is the area of AI concerned with allowing computers to work with human language.
Historically, humans had to learn how to communicate with computers through commands, programming languages and rigid interfaces.
NLP increasingly reverses that relationship by allowing us to speak and write naturally while the machine interprets what we mean.
Translation, summarisation, sentiment analysis, document understanding and conversational AI are all applications of NLP.
The biggest point to remember is that most human languages are very easy to misinterpret. For example: “David and Mike rode their bikes. His bike was blue.” Whose bike was blue, specifically? A good recommendation is to pretend you are writing a deal with the Devil when you are working with NLP and writing your prompts.
6. LLM, Large Language Model
A Large Language Model, or LLM, is the technology behind many of the AI systems people now use every day.
An LLM has learned extremely complex relationships between words, ideas, and concepts from very large amounts of data.
When you ask it a question, it does not simply retrieve a stored answer in the way a search engine retrieves a webpage. It generates a response based on patterns learned during training and the context available at that moment.
That flexibility is why LLMs are so powerful, but it also explains why they can sometimes produce incorrect answers with confidence. However, many times, that confidence is based on statistics and probability but built on inaccurate or incomplete data.
Did you know that the Grandfather of AI is Thomas Bayes? His paper, “Doctrines on the Principles of Chance,” was published, posthumously in 1763?
7. RAG, Retrieval-Augmented Generation
RAG stands for Retrieval-Augmented Generation.
The concept is straightforward. Instead of relying only on what the model already knows, the system first retrieves relevant information from another source and then uses that information to help generate its answer.
That source might be company documents, research papers, databases, product manuals, or websites.
A useful way to think about RAG is like using information from the internet combined with your own private documents to give you and only you better answers from the LLM.
For businesses, RAG is particularly important because it allows AI work to be augmented with items such as proprietary corporate information without retraining the underlying model.
8. Tokens, Tokenization, and Tokenomics
AI systems do not process text exactly as humans do. They break it into smaller units called tokens.
A token may be a whole word, part of a word, punctuation or a character.
Tokens matter because AI systems have limits on how much information they can process at once, and many commercial AI services also calculate usage and cost partly according to the number of tokens being processed.
If tokens are the pieces, tokenization is the process of breaking text into those pieces.
A sentence containing ten words does not necessarily become ten tokens because some words may be divided into several parts.
This helps explain why different languages, unusual names and long documents can behave differently inside AI systems.
And going further, one unit to measure the complexity of work in AI systems is to measure it in Tokens. When this is used for billing, we refer to it as Tokenomics.
In simple terms, tokens are the pieces; tokenization is how the AI creates them; and tokenomics is when these measures are used for billing purposes.
9. SFT, Supervised Fine-Tuning and Context Windows
For AI, in uses such as with pre-trained models (GPT) and retrieval-augmented generation (RAG), supervised fine-tuning (SFT) is the critical process of taking a pre-trained generalist model and training it further on a specific dataset to adapt it for a particular task. SFT bridges the gap between general pre-training and specialized application use cases.
In some use cases, such as with Gemini AI for Workspace, since no customer data is used to train or fine-tune underlying models, massive Context Windows are used instead of SFT. This technique with Context Windows allows models to process vast amounts of text simultaneously, allowing them to ingest extensive documents, data, or thread histories directly into the prompt context rather than needing custom fine-tuning.
10. The “Mis”use of the Term Open: Open Models, Open-Weight Models, and Open-Source Models
What the term “Open” means is unknown and is rooted in misusing the long history of open source software. For open source software, open generally means:
Public Access: The code is open for anyone to read and study.
Modification Rights: Users can fix bugs, change features, or build new programs from the original code.
Free Redistribution: People can share the software or modified versions with others
So when you see the word open to describe AI models, do they just mean you are “open” to change a few parameters? Or are you open to using different companies’ models? Or are you just open to giving them your money?
Unfortunately, we see misuse of the word, “open”, intended to steal the gravitas and meaning of the word that open source software has earned over the past quarter century. An open source model would include everything needed to change, improve, and rebuild the model.
So beware the term Open with AI Models.
Why These AI Terms Matter
You can use Artificial Intelligence effectively without understanding how to build an LLM or train a transformer, but knowing these terms makes the technology much easier to understand.
Once you understand multi-modality, you stop thinking of AI as simply a chatbot.
Once you understand prompting, you can give AI better instructions.
Once you understand RAG, you understand how AI can be improved such as using your own proprietary data.
Once you understand tokens, you begin to understand some of the limits as well as the way it works.
And once you understand Ethical AI, you understand that responsibility for a decision does not disappear simply because a computer made it.