UNAIVERSE / Learn AI
Somebody said “we should look at RAG” and the meeting moved on. This is the place to find out what they meant, in about ninety seconds, with an example you can picture.
Spelling does not have to be right. Acronyms work. So does describing it: “memory limit” finds the context window.
Sixteen sectors, from the words everything else is built from to the vocabulary of the rooms where AI rules get negotiated. Each one names a good place to start.
21 terms
The words people use before they use any of the others. What counts as artificial intelligence, what a model is, and why nobody agrees on the boundary.
Start with Artificial intelligence.
Next door: How Machines Learn, Large Language Models, AI Governance and UN Terminology
18 terms
Learning from examples instead of following written rules: what goes in, what changes inside, and what comes out.
Start with Training.
Next door: AI Foundations, Neural Networks, Model Quality and Evaluation
12 terms
The layered arrangement of simple numerical steps that most modern AI is built from.
Start with Neural network.
Next door: How Machines Learn, How Models Improve, Transformers and Attention
16 terms
The machinery of getting less wrong: measuring error, working out which settings caused it, and nudging them.
Start with Gradient descent.
Next door: Neural Networks, Model Quality and Evaluation, How Machines Learn
21 terms
How anyone can tell whether a system actually works, and the many ways a good-looking score can mislead.
Start with Accuracy.
Next door: How Machines Learn, Responsible AI, Human Rights and Safety, Data and Infrastructure
26 terms
The systems behind chat assistants: how they read text, how they produce it, and what they cannot do.
Start with Large language model.
Next door: Transformers and Attention, Search, Embeddings and RAG, AI Agents and Automation
11 terms
The architecture that made modern language models possible, explained without the mathematics.
Start with Transformer.
Next door: Large Language Models, Neural Networks, Search, Embeddings and RAG
14 terms
Finding the right passage by meaning rather than by matching words, and handing it to a model so the answer has a source.
Start with Embedding.
Next door: Large Language Models, Transformers and Attention, Data and Infrastructure
17 terms
The stages between a raw model and a usable one, and the several different things people mean by "training it on our data".
Start with Fine-tuning.
Next door: How Machines Learn, Large Language Models, Responsible AI, Human Rights and Safety
20 terms
Systems that do not just answer but act: calling tools, taking steps, and needing supervision.
Start with AI agent.
Next door: Large Language Models, AI Security, Responsible AI, Human Rights and Safety
13 terms
Producing images, audio, video and code, and the authenticity problems that follow.
Start with Image generation.
Next door: Large Language Models, Computer Vision and Language Technology, Responsible AI, Human Rights and Safety
Reading pictures and reading prose: the task-shaped words that predate the chatbot era and still describe most deployed systems.
Start with Computer vision.
Next door: Generative Media, How Machines Learn, Data and Infrastructure
25 terms
What models are built from and what they run on, including the parts that cost money and the parts that go stale.
Start with Data.
Next door: Model Quality and Evaluation, Search, Embeddings and RAG, AI Governance and UN Terminology
Who is affected, who is accountable, and what has to stay a human decision.
Start with Human oversight.
Next door: AI Governance and UN Terminology, Model Quality and Evaluation, AI Security
14 terms
The ways AI systems are attacked, and why an attack on a language model rarely looks like hacking.
Start with Prompt injection.
Next door: AI Agents and Automation, Responsible AI, Human Rights and Safety, Data and Infrastructure
32 terms
The vocabulary of the rooms where AI rules get negotiated, and the UN machinery the timeline records.
Start with AI governance.
Next door: Responsible AI, Human Rights and Safety, AI Foundations, Data and Infrastructure
Six to ten terms in a deliberate order, for a specific situation. Progress is kept in this browser, so you can stop halfway and come back.
The nine words everything else is built from. Start here if you have never started anywhere.
What is actually happening when a chat assistant answers you, and why fluent is not the same as correct.
The vocabulary that shows up in negotiated text, and what each word commits anyone to.
Enough to chair the meeting, ask the right question, and know what must not be pasted into a chat box.
How a prediction is made, how it is judged, and how it quietly goes wrong once it is in the field.
Where the harm actually lands: in the data, in the threshold, and in who never gets to appeal.
The most common question about AI vocabulary is not what a word means but how it differs from the word next to it. Each of these answers that in one line, then in detail.
AI vs machine learning vs deep learning
Are these three words for the same thing?
Generative AI vs predictive AI
Does the system produce something new, or score something that exists?
Training vs inference
Is the model changing, or just being used?
Parameter vs token
Both are counted in the billions. What is the difference?
Context window vs memory
If it has a million-token context window, does it remember me?
Prompt engineering vs fine-tuning
Both are called "customising the model". Only one changes it.
Some ideas do not survive being described. Turn the learning rate up and watch the walker overshoot; move the threshold and watch precision and recall trade against each other. Every one has a keyboard path, a prose alternative and a reduced-motion mode.
Attention
How attention shifts importance across a sentence
Backpropagation
How error moves backwards to guide weight updates
Context window
What fits inside the context window
Embedding
How embeddings place tokens in a space
False positives and false negatives
How changing the threshold changes false positives and false negatives
Gradient descent
How repeated small updates lower loss
UNAIVERSE is two things that belong together. This half explains the vocabulary. The other half is the timeline: every milestone the United Nations has produced on AI since 2017, each one sourced to the original document. Terms link into it, and it links back out.
The UN AI timeline
76 milestones from 2017 to 2027, with receipts, filters and a take for eight kinds of reader.
The Agenda
What is still to come, who owes what by when, and a calendar file.
Just handed the file?
Ten minutes on what exists and which meetings are worth staffing.
How this was written
The content pipeline, the review gate, and what is still awaiting a human.