Artificial intelligence has developed its own vocabulary, and for anyone trying to follow the technology industry, it can sometimes feel as though a new term appears every week.
Large language models, hallucinations, agents, inference, tokens, multimodal AI, reasoning models, context windows and vibe coding have all entered everyday technology conversations. Some describe genuinely important technical ideas, while others are industry shorthand that has become popular as companies race to describe what their latest systems can do.
For people outside the technology industry, the language can make AI appear considerably more complicated than it needs to be.
Understanding these terms matters because artificial intelligence is no longer a subject confined to software companies and research laboratories. It is increasingly being introduced into banking, education, healthcare, government, agriculture and ordinary office work. As Rwanda continues developing its digital economy, more people will encounter AI products without necessarily understanding what is happening behind them.
So what do the words actually mean?
This Rwanda Tech News guide explains some of the most important AI terminology in plain English, including the slang increasingly being used by developers and technology companies.
Artificial Intelligence — AI
Artificial intelligence is the broad term covering computer systems designed to perform tasks normally associated with human intelligence.
Those tasks can include recognising images, understanding language, making predictions, solving problems and generating content.
AI itself is not new. Banks have used forms of artificial intelligence to detect suspicious transactions for years, while email providers have long used machine learning to identify spam.
What has changed recently is the capability and accessibility of AI systems.
When somebody uses ChatGPT to summarise a report, asks an AI system to create an image or uses software to predict whether a financial transaction may be fraudulent, these are different applications of artificial intelligence.
Think of AI as the overall field, rather than one particular product.
Machine Learning — ML
Machine learning is a branch of AI in which computers learn patterns from data rather than having every possible rule explicitly programmed by a developer.
Imagine trying to build software that identifies fraudulent mobile-money transactions.
A programmer could attempt to write thousands of rules describing what fraud looks like. Alternatively, a machine-learning system could be trained using large numbers of previous transactions and learn patterns associated with fraudulent behaviour.
The distinction is important.
Traditional programming is roughly:
Rules + Data → Answer
Machine learning is closer to:
Data + Previous Answers → Learned Rules
The computer is effectively learning patterns from examples.
Deep Learning
Deep learning is a more advanced form of machine learning built around artificial neural networks containing many layers.
These systems have become particularly effective at recognising speech, understanding images and processing language.
Much of the extraordinary progress seen in modern AI has been driven by deep learning combined with enormous datasets and increasingly powerful computing infrastructure.
You do not need to understand the mathematics behind neural networks to follow the AI industry, but it helps to remember the hierarchy:
Artificial Intelligence → Machine Learning → Deep Learning
Deep learning is therefore part of machine learning, which itself sits within the broader field of artificial intelligence.
Generative AI
Traditional AI often analyses or classifies information.
Generative AI creates something new.
That could be text, computer code, an image, music, audio or video.
ChatGPT generating an email is generative AI. An image generator creating a photograph that never existed is generative AI. A coding assistant producing Python code is also generative AI.
This ability to generate new content is one reason the current AI revolution feels different from previous waves of artificial intelligence.
Large Language Model — LLM
This is one of the most important terms to understand.
A large language model, usually shortened to LLM, is an AI model trained on enormous amounts of text and other data so that it can understand and generate language.
Modern systems can answer questions, translate languages, write computer code, summarise documents, analyse information and perform many other tasks.
One simplified way of understanding an LLM is as an extremely sophisticated prediction engine.
If you write:
“The capital of Rwanda is…”
the model predicts that Kigali is the most appropriate continuation.
The remarkable part is what happens when this prediction process is scaled across enormous models trained on huge quantities of information. Capabilities emerge that allow the systems to perform surprisingly complicated tasks.
Model
People frequently use the words AI, model and chatbot interchangeably, but they are not exactly the same.
The model is essentially the underlying AI engine.
A chatbot or application is the product through which you interact with that model.
Think about a car.
The application is the car you interact with, while the AI model is closer to the engine powering it.
Companies can also build many different applications around the same underlying model.
Training
Before an AI model can perform useful tasks, it has to be trained.
Training involves exposing the system to enormous amounts of data and allowing it to learn patterns and relationships within that information.
Training frontier AI models can require huge amounts of computing power, specialised chips, electricity and money.
When technology companies talk about spending billions of dollars on AI infrastructure, much of that investment supports the computing required to train and operate increasingly powerful models.
Training Data
Training data is the information used to teach an AI model.
Depending on the system, this might include text, images, computer code, audio, video or specialised datasets.
Training data has become one of the most controversial areas of the AI industry because publishers, artists, authors and other copyright holders have questioned whether their work should be used to train commercial AI systems without permission or compensation.
There is another important issue for Africa.
If African languages, cultures and local information are poorly represented in training datasets, AI systems may perform considerably better for users elsewhere than they do for African users.
Parameters
You may see an announcement saying an AI model has “70 billion parameters” or considerably more.
Parameters are internal numerical values the model learns during training.
They help determine how the system processes information and generates responses.
It is tempting to assume that more parameters automatically mean better AI, but that is not necessarily true. Model architecture, training data, training methods and optimisation can all influence performance.
Think of parameter count as one indication of a model’s scale rather than a simple intelligence score.
Prompt
A prompt is simply the instruction you give an AI system.
“Explain artificial intelligence” is a prompt.
“Write a professional email asking for a meeting” is another.
The quality of the prompt can significantly influence the quality of the response, particularly when the user provides useful context and clearly explains the desired outcome.
Prompt Engineering
Prompt engineering is the practice of designing instructions that help AI systems produce better results.
Early in the generative AI boom, it was sometimes presented almost as a new profession.
The idea is simpler than the name suggests.
Instead of writing:
“Write about Rwanda.”
you might specify the audience, subject, length, tone, information to include and what you want the finished output to achieve.
Better instructions generally produce better results.
As models improve, elaborate prompt tricks may become less important, but communicating clearly with AI will remain useful.
Tokens
AI models do not necessarily process language one complete word at a time.
They divide information into smaller units called tokens.
A token might represent a whole word, part of a word or punctuation.
Tokens matter because AI companies frequently calculate usage and API pricing according to how many tokens a system processes.
If a developer builds an application using an AI company’s model, the amount of text sent to and generated by the model can therefore affect the cost of operating the application.
Context Window
The context window describes how much information an AI model can consider at once.
Imagine giving an AI system a 300-page report and asking questions about it. Whether the model can process the entire document at once depends partly on the size of its context window.
Larger context windows allow models to work with longer conversations, larger documents and more information simultaneously.
A useful analogy is short-term memory.
The larger the context window, the more information the model can have “on the desk” while working on a problem.
Hallucination
This is one of the most important AI terms for ordinary users.
An AI hallucination happens when a system produces information that sounds convincing but is incorrect or invented.
An AI might confidently provide a statistic that does not exist, misrepresent what somebody said or even invent a source.
This happens because language models are designed to generate plausible outputs, not because they inherently possess a human understanding of truth.
It is why important AI-generated information should still be verified, particularly in journalism, medicine, finance, law and academic work.
Confidence is not the same as accuracy.
Multimodal AI
Early generative AI systems were largely text-based.
Modern systems are increasingly multimodal, meaning they can work across several types of information.
A multimodal AI might be able to read text, analyse a photograph, understand audio, watch video and generate images.
You could photograph a broken piece of equipment, for example, and ask an AI system what appears to be wrong.
The system is no longer simply processing words.
Reasoning Models
Reasoning models are AI systems designed to spend more computational effort working through complex problems before producing an answer.
They are particularly useful for areas such as mathematics, programming, scientific questions and complicated analysis.
The term can sometimes create the impression that an AI “thinks” exactly like a person. That remains a much more complicated question.
In practical terms, reasoning models are designed to perform better on problems requiring multiple steps rather than immediately generating a response.
Inference
Training is how an AI model learns.
Inference is what happens when you actually use it.
When you ask an AI assistant a question and the model calculates its response, inference is taking place.
This distinction matters economically.
A company might spend enormous amounts of money training a model once, but then potentially spend substantial amounts continually providing inference to millions of users.
AI Agent
Agents are one of the most important developments happening in AI.
A normal chatbot generally waits for you to ask something and then responds.
An AI agent can potentially be given an objective and perform multiple actions towards completing it.
For example, instead of saying:
“Find me three hotels in Kigali.”
you might eventually tell an agent:
“Organise my business trip to Kigali next month within this budget.”
The agent could potentially research flights, compare hotels, check your calendar, prepare an itinerary and carry out approved actions using connected services.
The important shift is from AI that answers questions to AI that can perform tasks.
Agentic AI
You will increasingly hear technology companies describe systems as agentic.
It simply means the AI has greater ability to plan and take actions towards a goal rather than responding to one instruction at a time.
The more autonomous these systems become, however, the more important safeguards, permissions and human oversight become.
Giving an AI the ability to suggest an email is very different from giving it permission to send emails, transfer information or interact with company systems independently.
Human in the Loop
“Human in the loop” means a person remains involved in an AI-controlled process, particularly at important decision points.
A medical AI might analyse an X-ray, for example, but a doctor makes the diagnosis.
An AI could identify a suspicious banking transaction, while a human investigator decides whether action should be taken.
As AI becomes more autonomous, deciding when humans must remain in the loop will become an increasingly important policy question.
Fine-Tuning
Companies do not always need to build an AI model from scratch.
Instead, they can take an existing model and fine-tune it using specialised examples so it performs better for a particular task.
A financial institution might want a model better suited to financial language, while another organisation might want AI adapted for customer-service conversations.
Fine-tuning essentially specialises an existing model.
RAG — Retrieval-Augmented Generation
RAG sounds complicated but the idea is extremely useful.
Retrieval-Augmented Generation allows an AI system to retrieve relevant information from an external source before answering a question.
Imagine a Rwandan company has 20,000 pages of internal documents.
Rather than expecting an AI model to somehow know everything contained in them, a RAG system can search those documents for information relevant to the user’s question and provide it to the model before the model answers.
This can make AI systems more useful for businesses and can reduce hallucinations.
API
An Application Programming Interface, or API, allows different software systems to communicate.
When developers say they are “using an AI API”, they generally mean their application sends information to an AI model operated by another company and receives a response.
This is how a small startup can add sophisticated AI capabilities without spending billions developing its own foundation model.
Open Source and Open-Weight AI
These terms are frequently confused.
Broadly speaking, open-source software allows people to inspect, modify and redistribute its underlying code under specified licences.
In AI, companies also release open-weight models, where the trained model weights are available for developers to download and run themselves, although the complete training data or development process may not necessarily be available.
This matters because organisations can potentially operate these models on their own infrastructure, customise them and reduce dependence on a single AI provider.
AI Slop
Now we enter AI slang.
AI slop is a dismissive term used to describe large amounts of low-quality, mass-produced AI content flooding websites and social platforms.
It might be meaningless AI-generated articles, bizarre images, repetitive videos or content created primarily to attract clicks.
The term reflects a growing concern that making content extremely cheap to produce does not necessarily make that content valuable.
Vibe Coding
Vibe coding describes building software largely by telling AI what you want rather than manually writing every line of code yourself.
Someone might tell an AI coding tool:
“Create a booking system with user accounts, payments and an admin dashboard.”
They then continue describing changes while the AI generates much of the underlying code.
It has made software development accessible to many people who are not traditional programmers.
However, there is an important limitation.
Being able to generate software is not the same as understanding whether that software is secure, scalable or reliable.
Vibe coding can be extremely useful for prototypes and increasingly for real products, but blindly trusting generated code can create serious problems.
Wrapper
Calling an AI product a wrapper can sometimes be an insult in technology circles.
It generally refers to an application that takes an existing AI model and places a relatively simple product or interface around it.
For example, if somebody builds a writing application that simply sends prompts to an existing model and returns the answers, critics might call it “just an AI wrapper”.
But the term can be misleading.
Many successful technology businesses are ultimately built on technology supplied by other companies. What matters is whether the product solves a valuable problem for customers.
Frontier Model
A frontier model generally refers to one of the most advanced AI models available at a particular time.
The phrase is commonly associated with systems pushing the limits of current AI capabilities.
There is no universally fixed threshold determining exactly when a model becomes “frontier”, so the term should not be treated as a precise technical classification.
Foundation Model
A foundation model is a large AI model trained broadly enough that it can serve as the basis for many different applications.
Rather than developing separate models for writing, customer service, coding and research, developers can take a powerful foundation model and adapt it to many tasks.
This is part of the reason the AI industry is developing so quickly.
AI Alignment
Alignment deals with a deceptively difficult question:
How do we make sure powerful AI systems behave according to human intentions and values?
If an AI is given an objective, researchers want it to achieve that objective without finding dangerous, deceptive or unintended shortcuts.
The more capable and autonomous AI becomes, the more important alignment research is expected to become.
Guardrails
Guardrails are restrictions placed around an AI system to reduce harmful or unwanted behaviour.
They might prevent a model from providing certain dangerous information, limit which company data it can access or require human approval before an agent performs a sensitive action.
Think of guardrails as boundaries around what an AI system is allowed to do.
Jailbreaking
A jailbreak is an attempt to manipulate an AI system into bypassing its safeguards.
Users may deliberately construct unusual prompts or techniques intended to persuade the model to produce something its safety systems would normally prevent.
AI companies continually try to strengthen their systems against these methods.
AGI — Artificial General Intelligence
Few AI terms generate more debate than AGI.
Artificial General Intelligence generally refers to a hypothetical AI capable of performing intellectual tasks across a very broad range of areas at or beyond human levels.
The difficulty is that there is no universally accepted definition.
Different researchers and companies use different benchmarks, which means claims about how close the world is to AGI should be treated carefully.
Superintelligence
If AGI broadly describes AI reaching or exceeding human-level general capability, superintelligence describes the theoretical stage beyond it.
It refers to AI dramatically more capable than humans across most intellectual tasks.
No generally recognised artificial superintelligence exists today.
The possibility nevertheless plays a major role in debates about AI safety because the consequences of creating such a system could be enormous.
AI Bubble
When investors talk about an AI bubble, they are questioning whether the enormous amounts of money flowing into AI companies, chips and data centres can ultimately generate sufficient financial returns.
The comparison is often made with previous investment booms, particularly the dot-com era.
It does not necessarily mean AI itself is useless. A technology can transform the world while investors simultaneously overvalue some of the companies building it.
Both things can be true.
AI Winter
An AI winter describes a period when excitement, investment and research interest in artificial intelligence collapse after expectations fail to materialise.
AI has experienced several such periods historically.
Today’s industry is experiencing the opposite: enormous investment and public attention.
Whether the current boom eventually cools substantially remains an open question.
Why This Language Matters
It would be easy to dismiss much of this vocabulary as technology-industry jargon, but understanding it is becoming increasingly important.
A Rwandan student hearing that AI will affect employment should understand the difference between generative AI and traditional automation. A business owner considering an AI product should know what an API, hallucination and RAG system are. A policymaker discussing AI regulation needs to understand the difference between a chatbot and an autonomous agent.
And ordinary users need to understand that an AI-generated answer sounding intelligent does not guarantee that it is correct.
This is perhaps the most important lesson behind all of the terminology.
Artificial intelligence is powerful, but it is not magic.
Once terms such as LLMs, tokens, agents, inference, hallucinations and context windows are understood, much of the mystery disappears. What remains is an extraordinarily capable and rapidly developing technology with genuine strengths, serious limitations and consequences that societies are only beginning to understand.
For Rwanda, AI literacy should therefore mean more than teaching people how to use a chatbot.
It should mean giving people enough understanding to question these systems, use them productively, recognise their weaknesses and eventually build with them.
Because as AI becomes part of everyday life, understanding its language will increasingly become part of understanding technology itself.
